Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

feat(x-content): write the forge and quality-gate articles

nirholas457a401500a3
diff --git a/data/x-content/articles/agents-board-article.md b/data/x-content/articles/agents-board-article.md
new file mode 100644
index 000000000..1b9f46a3c
--- /dev/null
+++ b/data/x-content/articles/agents-board-article.md
@@ -0,0 +1,51 @@
+The public agent board on three.ws lists more than 2,500 agents and puts them in one order: the most recent action each agent logged. Sign-up date, popularity and chat counts do not decide who leads. Beside it sits the Agent Spotlight, where a person writes up what an agent is for, and a trending clock that counts days decides which write-ups stay near the top.
+
+Both are public, both need no account, and both are built so that a page full of default names cannot crowd out the agents that are doing something.
+
+## What counts as a real action
+
+Each agent keeps an append-only action log. A row lands in it when the agent does something in its owner's session: its avatar finishes loading, it speaks, a skill finishes. The write path checks ownership, so viewing someone else's agent adds nothing to its log. The browser skips the call for an agent you do not own, and the server would refuse it with a 403.
+
+The board sorts on the newest row in that log, then on creation date. An agent with no logged action sorts after the agents that have one. Chats are shown on each row, but they do not move an agent up. More than 600 of the agents on the board have logged at least one action.
+
+## Keeping placeholders off the board
+
+New agents start with a default name, and many keep it. The board treats six names as placeholders (My First Agent, Agent, Avatar, My Avatar, Untitled Agent and New Agent) along with the names a selfie import generates. An agent still wearing one of them is left off the board when it also has no logged action, no chat and no on-chain identity. It stays in the [full agent directory](https://three.ws/agents), where search and browsing still reach it. It just does not take a row on a board meant to show activity.
+
+The filter judges signal, not naming. A My First Agent that has done something keeps its row, because it did something, and several of them sit near the top of the board right now. Renaming an agent is its owner's call, and the board does not make it for them.
+
+## Reading the board
+
+The board is the fleet panel of the [Agent Monitor](https://three.ws/monitor). Each row carries the agent's action count, its chat count and a recency stamp that turns green under 15 minutes. Typing in the search box narrows the fleet on the server, by name and description. Picking a row loads that agent's own 3D avatar, live, in the panel beside it. The fleet panel polls every 60 seconds and the list behind it is cached at the edge for 30 seconds, so a board left open on a second screen stays close to live without hammering the database.
+
+![The Agent Monitor's fleet panel searched for "writing": Whatchget and Simply Sage lead with actions logged 45 and 50 days ago, followed by Orion agents, and agents with no logged action sit at the bottom marked never. Simply Sage is selected and stands in 3D in the Spotlight panel beside the list](/x-media/agents-board-article/fleet-search.png)
+
+In the capture above, a search for "writing" shows the ordering at work. The two agents with the most recent actions lead, the Orion rows follow by date, and the rows that never logged an action fall to the bottom even when they have a chat.
+
+## The Spotlight: a pitch on a live agent
+
+The board answers who is doing something. The [Agent Spotlight](https://three.ws/spotlight) answers what an agent is for. An entry is a pitch attached to a live agent, never a copy of it. It stores a headline, a one-liner, an optional write-up of up to 4,000 characters, a category, up to six tags and an optional demo link. The agent's name, avatar, skills and conversation count are read from the agent itself on each request, so renaming the agent or swapping its avatar updates the card. Each agent gets one entry.
+
+Some entries were written by us about someone else's public agent. Those carry a Curated badge, so a visitor is never told a builder said something they did not write, and the builder's own write-up replaces ours the moment they submit one.
+
+![A Spotlight entry page: the agent Glyph #21 standing in 3D on the left, and on the right the Curated badge, the headline "The translator that refuses to flatten your voice", its one-liner, the builder credit and a button to talk to the agent](/x-media/agents-board-article/spotlight-entry.png)
+
+## A trending clock that counts days
+
+Trending orders the Spotlight by one expression:
+
+```
+score = (upvotes + 1) / (age_in_days + 1) ^ 1.2
+```
+
+The + 1 on top means an entry published minutes ago with no votes still scores 1.0 and lands on the first page, so new work is seen at all. The day clock is the deliberate part. The hour-scaled curve most ranked feeds borrow from Hacker News assumes a lot of votes. At this volume it would put a day-old entry at about 2% of a fresh one, so any new empty entry would outrank a week-old entry people liked, and trending would quietly become newest. Counted in days, a day-old entry still carries a bit under half the weight of a fresh one, a week-old entry with twenty upvotes stays above a fresh empty one, and a month-old entry finally rotates off the front page.
+
+The database orders the page and the server reports each entry's score, so the expression exists twice, and a test fails if the two ever disagree. Entries seeded in one batch tie on everything above, so ties break on the agent's own activity and then on its id, and the same query returns the same page each time. Each account gets one upvote per entry, and votes are counted on read rather than kept as a running total, so a lost write cannot leave a count wrong.
+
+## What it does not do yet
+
+Ranking by logged action measures use, not quality. An agent its owner opens often will sit above a better one that is rarely run, and a chat with a visitor does not count toward rank at all.
+
+The Spotlight is small. It holds 15 entries, most of them curated by us, and no entry has an upvote yet, so today its trending order comes down to age. The clock is built to weigh votes; it has not had any to weigh.
+
+Both surfaces are open now without an account: the board on the [Agent Monitor](https://three.ws/monitor), and the write-ups on the [Agent Spotlight](https://three.ws/spotlight).
diff --git a/data/x-content/articles/forge-article.md b/data/x-content/articles/forge-article.md
new file mode 100644
index 000000000..83f13f3dc
--- /dev/null
+++ b/data/x-content/articles/forge-article.md
@@ -0,0 +1,61 @@
+Type a sentence into the Forge and two chains run one after the other: the first paints a reference picture of the object, the second reconstructs that picture into a textured GLB. Each chain has backup rungs, so when an engine fails partway through a job, the job moves to the next engine under the same job id and the progress bar keeps going. The first rungs are our own NVIDIA GPU workers on Google Cloud Run, and a Draft or Standard model costs nothing and needs no account.
+
+## A picture comes before the shape
+
+The reconstruction models that build the mesh read an image, not a sentence. So a text prompt is first rewritten and painted as one centered subject on a plain seamless background, under soft even light, with real-world materials. The realism of the final mesh is set almost entirely by the realism of that one reference image: a clean, photoreal picture reconstructs into a model that looks real, and a cluttered or cartoonish one does not.
+
+The picture has its own ladder. A Gemini image model on Vertex AI leads, because it draws the most photoreal reference. If it fails, the request falls through to FLUX on NVIDIA's hosted inference, then to further providers, ending in a keyless rung, so no single provider's failure is ever the terminal error for the picture.
+
+Photos skip this step: your photo is the reference. A sketch skips it too, and goes to a model built to read drawings.
+
+![The engine picker on the Forge: TRELLIS selected and marked free, with NVIDIA, both Hunyuan3D lanes, and engines that take your own key beside it, a legend for busy and unavailable lanes, and the line "Standard · Free · usually ~60s · renders a preview image, then builds 3D"](/x-media/forge-article/engines.png)
+
+## Which engine builds the shape
+
+The router picks a lane per quality tier. Draft and Standard go to TRELLIS on our own worker. High goes to Hunyuan3D 2.1 on our own worker, which paints a full PBR material set (colour, roughness and normal maps) rather than one baked colour map. At the High tier, a prompt that clearly reads as a hard-surface object (a robot, a vehicle, a piece of furniture) swaps TRELLIS ahead of Hunyuan3D, because single-hop reconstruction is crisper on mechanical shapes.
+
+Behind our own workers sit free external lanes: Hugging Face Spaces for photos, and NVIDIA's hosted TRELLIS for text, which keeps a text prompt answerable even when every one of our GPU workers is cold. Engines that need your own API key run only when you pick them, and the engine picker lets you name any lane yourself instead of taking the router's choice.
+
+Before a job starts, the router reads lane health. It takes the first lane in its order that is healthy or simply unknown; if none is clean, it takes the first that is not confirmed down. The tiers set the polygon budget: about 12,000 at Draft, about 30,000 at Standard, and up to 200,000 at High.
+
+## When a lane fails in the middle of a job
+
+Failure at submit time is the easy case: the request walks the chain until a lane accepts it. The harder case is a worker that accepts a job and then fails it minutes later, when the request that started it is long gone.
+
+That failure surfaces on the next status poll. The poll handler marks the failed lane unhealthy for 90 seconds, so new jobs steer around it at once. Then it recovers the original prompt and the stored reference picture from the job's record, resubmits them to the next lane, and binds the old job id to the new one. The page keeps polling the same id and sees the job still running, now on a different engine.
+
+The order for that rescue is our TRELLIS worker, then our Hunyuan3D worker, then a hosted TRELLIS lane where one is configured, then TripoSG. A job gets its first lane plus up to three backups. TripoSG closes the chain because it returns an untextured mesh, so it runs only once every textured lane has failed or is marked down. It earned its place during a worker outage from 18 to 28 September that took Hunyuan3D and TRELLIS down together: without it, roughly 400 photo jobs a day ended in a hard failure while the TripoSG worker sat healthy.
+
+When nothing is left to try, the failure still is not a dead end. The response names the lanes that can serve a fresh retry, so the page can offer a one-click switch of engine.
+
+## The GPU workers on Cloud Run
+
+Each self-hosted lane is its own Cloud Run service: a small Python worker wrapping one model, with an NVIDIA GPU attached. TRELLIS runs on an L4. Hunyuan3D 2.1 runs on an RTX PRO 6000 Blackwell, held warm with one instance, because a cold start means loading 14 GiB of weights before any work begins. Every worker speaks the same contract, which is what lets the router treat them as interchangeable rungs. The first two lines below are that contract; the health answer is what the Hunyuan3D worker returned when we asked it while writing this:
+
+```
+POST /infer 202 {"task_id": "...", "status": "queued"}
+GET /tasks/:id {"status": "done", "result_gcs_url": "...", "elapsed_ms": ...}
+
+GET /health
+{"ok": true, "model": "hunyuan3d-2.1", "gpu_available": true,
+ "gpu_name": "NVIDIA RTX PRO 6000 Blackwell Server Edition",
+ "pipeline_loaded": true, "ready": true, "load_error": null}
+```
+
+Two fields in that health answer carry the design. `ok` means the process is up; `ready` flips only when the model has actually loaded. A worker that opens its port before the model loads would otherwise look healthy while it is dying. The health check reads the second field, so a worker that is up but still loading is reported degraded, not healthy. Task state lives in object storage rather than in the instance, so a job submitted to one autoscaled instance and polled on another still resolves.
+
+The Blackwell move taught us two things worth passing on. Cloud Run backs /tmp with RAM, so weights staged there and weights loaded into the process are charged to the same memory budget: 18 GiB staged plus 14 GiB loaded is a 32 GiB peak against the L4 tier's 32 GiB ceiling, and the container was killed on every cold start. The RTX PRO 6000 tier's floor of 80 GiB clears it. And Blackwell needs its GPU code built with NVIDIA's 12.8 toolkit. The worker image compiles its extensions for both GPU generations, so which card runs it is a deploy flag rather than a rebuild.
+
+![A draft from the Hunyuan3D lane on its public model page: a noir detective in a trench coat and fedora, with the model information panel reading Engine hunyuan3d, 4 textures, 1 mesh, 30k triangles and glTF valid](/x-media/forge-article/hunyuan-detective.png)
+
+## What it does not do yet
+
+The last rung trades colour for completion. A job rescued by TripoSG comes back as untextured geometry, which is a shape you can keep, not the model you asked for.
+
+Not every lane can be a rescue. Hugging Face Spaces block for the whole generation and NVIDIA's hosted lane takes text only, so neither can pick up a job that failed on a poll; both appear only in the retry suggestions. A rescue also rebuilds from the one stored reference picture, so a job that started from several photos continues from one of them, and says so.
+
+Lane health is a probe, not a promise. Each check waits 4 seconds, and a worker that is still starting can read as unreachable to it while being perfectly able to take the job a minute later.
+
+Small detail is where reconstruction still struggles. The detective above holds together as a figure, and his face does not survive the trip from picture to mesh.
+
+Try it on the [Forge](https://three.ws/forge). The lane status the engine picker reads is public at [/api/forge?health=1](https://three.ws/api/forge?health=1), and the GPU field report behind the Blackwell move is [on the blog](https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell).
diff --git a/data/x-content/articles/quality-gate-article.md b/data/x-content/articles/quality-gate-article.md
new file mode 100644
index 000000000..b14db6784
--- /dev/null
+++ b/data/x-content/articles/quality-gate-article.md
@@ -0,0 +1,71 @@
+A model that the platform generates for its own catalog has to clear a structural gate before anyone sees it, and the Model Inspector runs that same gate, the same module with the same thresholds, against any .glb you give it. Its companion, Model Diff, compares two versions of a model and names what went missing: the joints, the clips, the meshes. Between them they answer the two questions a 3D pipeline keeps asking: whether a file is good enough to publish, and whether a change broke something that depends on it.
+
+## One module, two callers
+
+The gate lives in a single file, `seed-mesh-gate.js`. The seeding job that fills the catalog imports it, and so does the inspector page. There is no second copy of the rules for the browser, and no API that might drift from them. A threshold change lands in both places at once because there is only one place.
+
+That only works because the gate reads nothing but the glTF JSON chunk. It needs no renderer and no network, so it runs entirely in the browser: nothing is uploaded, and the verdict appears without a round trip. A build check walks the real bundler graph from the page and fails the moment a Node built-in becomes reachable, which is what keeps the shared module loadable in both places.
+
+Every verdict carries a gate version, because an accept rate is only comparable within one version of the rules.
+
+## What the bar measures
+
+The structural gate is five checks, and the inspector shows each one with what it measured, the bound it was held to, and why the bound exists:
+
+- Geometry density: between 1,500 and 1,500,000 vertices.
+- File weight: between 20 KB and 80 MB.
+- Surface texture: at least one.
+- Bounds: the bounding box must have some extent, because a collapsed mesh passes every triangle count and still draws as a speck.
+- Assembly: a character may arrive as at most 8 meshes. A prop is exempt, since a prop can legitimately be a loose collection of parts.
+
+A failed check comes with the fix, written next to the rule it explains, so the guidance cannot drift from the rule.
+
+The texture rule is deliberately stricter than the interactive Forge. Someone who generates an untextured mesh chose to. A visitor browsing the catalog chose nothing, and an entry there is the platform vouching for itself.
+
+![The Fox sample on the Model Inspector: it clears the catalog bar with 1,728 vertices against a floor of 1,500, one texture and one mesh, and is reported rigged with 24 joints](/x-media/quality-gate-article/inspect-fox.png)
+
+Fox is the sample labelled "Passes, barely" for a reason. It clears the density floor of 1,500 with 1,728 vertices, and a passing check still prints its numbers, which tells a creator how much headroom they have instead of a bare yes.
+
+The gate is the first of two stages. On the catalog path a vision model then renders the mesh and scores whether it looks like what was asked for. That stage needs a renderer and a model call, so the page runs only the structural half and says so.
+
+## Diffing two models by what they mean
+
+Byte comparison cannot tell a lossless recompression from a rig that quietly lost joints. Model Diff reads both files and builds a structural description of each: scenes, nodes, meshes, materials, textures, skeletons and animation clips. Then it pairs objects across the two the way git pairs renamed files: by name first, then by a hash of the content, then by similarity. Anything still unpaired is a real addition or removal.
+
+Vertex data is compared at a quantum of 1e-4 units, so the float drift of a re-export does not register while a genuine edit does.
+
+Every change set carries one severity: none, cosmetic, minor, major or breaking. The ladder is ordered by what a consumer of the model notices, not by how many bytes moved. A texture swap is minor because everything still loads and plays. Breaking means something a consumer references by name is gone: a clip, a joint, a mesh. That is the failure that renders fine and never moves, because a clip addresses joints by name and a missing joint is a silent no-op rather than an error anyone sees.
+
+![Model Diff comparing the X Bot sample with Michelle: one skeleton lost the joints mixamorig:LeftEye and mixamorig:RightEye, and five clips are gone (agree, idle, run, sneak_pose and walk), each with its duration](/x-media/quality-gate-article/diff-detail.png)
+
+The example pair on the page puts the X Bot sample before Michelle. The skeleton section names the two eye joints that are gone and says clips addressing them will not play. The animation section lists the five clips that no longer exist, with their lengths. That is a change set a person can act on.
+
+## The same engine in a build
+
+The page, a command line tool and a free HTTP endpoint all run the same published package, so a verdict in CI cannot disagree with what the page shows.
+
+```bash
+npm install --save-dev @three-ws/glb-diff meshoptimizer
+npx glb-diff base.glb candidate.glb --fail-on breaking
+
+# or, with no install, a report sized for a pull request comment
+curl "https://three.ws/api/3d/diff?a=<url-a>&b=<url-b>&format=markdown"
+```
+
+The exit codes are the contract a pipeline depends on. Zero means the diff ran and stayed below the threshold you set with `--fail-on`, one means the severity reached it, and two means the tool could not run at all: a bad argument, an unreadable file. A tool that returned one for both "your model regressed" and "I could not open the file" would be useless in a pipeline.
+
+The endpoint takes two public URLs and caps each side at 32 MB.
+
+## What it does not do yet
+
+The inspector judges structure, not taste. A model can clear every check and still look wrong; that judgment belongs to the vision stage, which does not run in the browser.
+
+Rename detection is a similarity score, and on animation clips it can be fooled. Two clips that drive the same joints on the same rig score as close even when one is a head shake and the other is a dance, so the X Bot pair above also reports headShake as renamed to SambaDance. Treat a rename on a clip as a prompt to look, not a verdict.
+
+The summary can also overstate a removal. When one of several skeletons disappears, the summary line says the model is no longer rigged even if another skeleton remains, as it does in this pair: Michelle still carries a full rig. The skeleton section underneath is the accurate record.
+
+The similarity pass is skipped when two models would need more than 40,000 comparisons, and the result says so rather than quietly turning renames into additions and removals.
+
+Most avatars on the platform ship with compressed geometry. The browser page decodes it, but the command line tool needs the optional `meshoptimizer` package installed next to it, which is why the install line above names both.
+
+Both tools are free and run in your browser: the [Model Inspector](https://three.ws/inspect) and [Model Diff](https://three.ws/diff). The command line tool is [on npm](https://www.npmjs.com/package/@three-ws/glb-diff).
diff --git a/data/x-content/articles/voice-article.md b/data/x-content/articles/voice-article.md
new file mode 100644
index 000000000..4b5b32ebb
--- /dev/null
+++ b/data/x-content/articles/voice-article.md
@@ -0,0 +1,61 @@
+The Voice Lab on three.ws puts more than 300 voices in one grid, and you can hear any of them before you choose, without signing in. Drive puts an agent on a car screen, where speaking is the main way in, and the phrases a driver uses to control it ("repeat that", "stop talking", "louder", "night mode") are answered by the page itself before anything is sent to a server.
+
+The two pages solve opposite halves of the same problem: how an agent should sound, and what it should not wait on a network to do.
+
+## One grid across three lanes
+
+The grid merges three synthesis lanes that work signed out: Microsoft Edge, Google Gemini and NVIDIA Magpie. The Edge list is fetched live from Microsoft and cached for six hours, so it is not a hardcoded snapshot, and it covers more than 100 locales. Gemini adds 30 prebuilt voices that take a written direction, such as "warm and unhurried". Magpie adds 11 personas built for the real-time avatar lane. Two more lanes, OpenAI and ElevenLabs, appear in the same grid and need an account.
+
+Search matches a voice's name, id, locale and labels, and the language filter is built from the locales actually in the list, so it offers no language the grid cannot play. Each card names the lane it runs on.
+
+![The Voice Lab grid searched for "warm": five voice cards, Sulafat on Google Gemini, Andrew and AndrewMultilingual and Xiaoxiao on Microsoft Edge, and Ash on NVIDIA Magpie, each with its labels and a Preview and a Use this voice button](/x-media/voice-article/voice-grid.png)
+
+## Hear it before you choose it
+
+Preview plays a voice's own sample clip when it has one, and otherwise renders one short line through the synthesis endpoint. Use this voice moves the pick into the playground, where you type up to 1,000 characters, set the speed and press Speak.
+
+Each clip is cached against a hash of the whole request: lane, voice, text, model, direction, speed and settings. Ask for the same line twice and the second answer comes back from storage with no new synthesis. The playground says which happened. In the capture below the line had already been spoken once, so Speak returned it from the cache.
+
+![The Text-to-Speech Playground with the Microsoft Edge voice Andrew selected, the line "Take the next exit. There is a charger two minutes past the bridge." typed in, and the status beside Speak reading 23 KB, cached, with the clip playing](/x-media/voice-article/playground.png)
+
+## Drive: the agent on a car screen
+
+[Drive](https://three.ws/drive) starts in Driving mode with the keyboard hidden and four controls: Hands free, Repeat, Stop and Type. Hold Talk and speak. The agent is told to reply in one or two short sentences that sound right out loud, and never to ask the driver to read anything, tap through a list or look at the screen. The reply is clamped to two lines on screen and spoken in full. Night is the default palette, because a bright panel at speed is glare.
+
+If the browser reports a real ground speed, moving above walking pace locks the keyboard again, and the Parked switch will not unlock it while the wheels are turning. Behind the screen is the platform's existing talk loop: microphone into speech recognition on NVIDIA Riva, with the browser recognizer as the fallback lane, transcript into the chat model, the reply into speech, and the audio into lip sync on a rigged avatar.
+
+![Drive parked at night with the avatar Michelle. Asked how long to drive before taking a break, the agent's answer is clamped to two lines on screen while it is spoken, with Hands free, Repeat, Stop, Type and Talk along the bottom](/x-media/voice-article/drive-reply.png)
+
+## The phrases the page answers itself
+
+Some things a driver says are control, not conversation. Sending "louder" to a language model costs a round trip and can come back wrong, so Drive keeps a small phrase table and checks each final transcript against it before the network is touched. There are eight commands, each with its own set of phrasings: repeat, stop, louder, quieter, night, day, parked and driving.
+
+```js
+const PHRASES = [
+ ['repeat', ['repeat', 'repeat that', 'say that again', 'again', 'what did you say', 'come again', 'one more time']],
+ ['hush', ['stop', 'stop talking', 'be quiet', 'quiet', 'hush', 'never mind', 'nevermind', 'cancel', 'shut up']],
+ ['louder', ['louder', 'speak up', 'turn it up', 'volume up', 'i can not hear you', "i can't hear you"]],
+ ['night', ['night mode', 'dark mode', 'go dark', 'dim the screen']],
+ // quieter, day, parked and driving follow the same shape
+];
+
+export function matchDriveCommand(transcript) {
+ const phrase = normalize(transcript);
+ if (!phrase || phrase.length > 40) return null;
+ return LOOKUP.get(phrase) || null;
+}
+```
+
+The match is against the whole utterance, not a word inside it. Before comparing, the page lowercases the transcript, strips punctuation, drops "hey", "ok", "please" or "can you" from the start and "please", "now" or "for me" from the end, and gives up on anything longer than 40 characters. That is how "stop" ends a reply while "stop at the next charger" still reaches the agent. "Louder" and "quieter" move the volume by 20%. A handled command never reaches the chat model.
+
+Typed text takes the same path as speech. Parked, typing "Day mode, please" switches the palette to day on the spot, with no request to a server.
+
+![Drive in the day palette after "Day mode, please" was typed while parked: the heard line shows the command above the previous answer, and the palette button now reads Day](/x-media/voice-article/drive-day-mode.png)
+
+## What it does not do yet
+
+Gemini voices can be listed and still fail. As of this writing the provider is refusing this deployment's credentials, so a Gemini preview answers with an error that names the lanes still working. A lane the provider refuses is then marked unavailable for a few minutes and its voices leave the picker, which is why the grid's count moves between 327 and 357. Edge and Magpie are unaffected.
+
+A voice picked in the Voice Lab reaches an agent once it is assigned in the agent editor. Drive answers on a head unit browser or a phone in a cradle today. The native CarPlay version is written and waits on Apple's entitlement grant, and there is no wake word, by design, because Apple's voice category does not allow one. A spoken answer waits on the chat model, and under heavy load that wait has run long enough to time out.
+
+The Voice Lab is at [three.ws/voice](https://three.ws/voice) and Drive at [three.ws/drive](https://three.ws/drive). Both work without an account.
diff --git a/data/x-content/proofs/agents-board-article.json b/data/x-content/proofs/agents-board-article.json
new file mode 100644
index 000000000..8119252bd
--- /dev/null
+++ b/data/x-content/proofs/agents-board-article.json
@@ -0,0 +1,210 @@
+{
+ "id": "agents-board-article",
+ "scenarioHash": "fe3688911035e1d9df2a219a91478f35c057596d1814c010d7ab77a0a085b2ff",
+ "ranAt": "2026-10-01T19:31:19.725Z",
+ "target": {
+ "commit": "68c8ec4db",
+ "revision": "three-ws-api-00475-lq6"
+ },
+ "passed": true,
+ "steps": [
+ {
+ "index": 1,
+ "kind": "goto",
+ "target": "https://three.ws/monitor",
+ "detail": "loaded https://three.ws/monitor",
+ "ok": true,
+ "ms": 24075,
+ "frames": 0
+ },
+ {
+ "index": 2,
+ "kind": "read",
+ "target": "fleet",
+ "detail": "fleet = \"2,557\"",
+ "ok": true,
+ "ms": 96,
+ "frames": 0
+ },
+ {
+ "index": 3,
+ "kind": "read",
+ "target": "active",
+ "detail": "active = \"648\"",
+ "ok": true,
+ "ms": 11,
+ "frames": 0
+ },
+ {
+ "index": 4,
+ "kind": "type",
+ "target": "writing",
+ "detail": "typed 7 characters; GET /api/agents/public answered 200 in 10.3 s",
+ "ok": true,
+ "ms": 24868,
+ "frames": 30
+ },
+ {
+ "index": 5,
+ "kind": "hold",
+ "target": "1800",
+ "detail": "1800 ms",
+ "ok": true,
+ "ms": 37544,
+ "frames": 54
+ },
+ {
+ "index": 6,
+ "kind": "click",
+ "target": "Simply Sage",
+ "detail": "clicked; GET /api/agents/42534db3-f8f8-48ae-a4cb-ad8b9b42b2d7 answered 200 in 0.3 s",
+ "ok": true,
+ "ms": 17288,
+ "frames": 21
+ },
+ {
+ "index": 7,
+ "kind": "expect",
+ "target": "creative writing guide",
+ "detail": "appeared after 0.2 s",
+ "ok": true,
+ "ms": 187,
+ "frames": 0
+ },
+ {
+ "index": 8,
+ "kind": "wait",
+ "target": "6000",
+ "detail": "6000 ms, not filmed",
+ "ok": true,
+ "ms": 6185,
+ "frames": 0
+ },
+ {
+ "index": 9,
+ "kind": "hold",
+ "target": "2200",
+ "detail": "2200 ms",
+ "ok": true,
+ "ms": 32968,
+ "frames": 66
+ },
+ {
+ "index": 10,
+ "kind": "goto",
+ "target": "https://three.ws/spotlight",
+ "detail": "loaded https://three.ws/spotlight",
+ "ok": true,
+ "ms": 9192,
+ "frames": 0
+ },
+ {
+ "index": 11,
+ "kind": "read",
+ "target": "showcased",
+ "detail": "showcased = \"15\"",
+ "ok": true,
+ "ms": 3,
+ "frames": 0
+ },
+ {
+ "index": 12,
+ "kind": "read",
+ "target": "upvotes",
+ "detail": "upvotes = \"0\"",
+ "ok": true,
+ "ms": 2,
+ "frames": 0
+ },
+ {
+ "index": 13,
+ "kind": "goto",
+ "target": "https://three.ws/spotlight/8fecd65a-4ab9-484a-9714-691d4040e9e5",
+ "detail": "loaded https://three.ws/spotlight/8fecd65a-4ab9-484a-9714-691d4040e9e5",
+ "ok": true,
+ "ms": 11328,
+ "frames": 0
+ },
+ {
+ "index": 14,
+ "kind": "expect",
+ "target": "Machine translation solved meaning years ago",
+ "detail": "appeared after 0.2 s",
+ "ok": true,
+ "ms": 247,
+ "frames": 0
+ },
+ {
+ "index": 15,
+ "kind": "wait",
+ "target": "4000",
+ "detail": "4000 ms, not filmed",
+ "ok": true,
+ "ms": 4303,
+ "frames": 0
+ },
+ {
+ "index": 16,
+ "kind": "hold",
+ "target": "2600",
+ "detail": "2600 ms",
+ "ok": true,
+ "ms": 36754,
+ "frames": 78
+ }
+ ],
+ "facts": {
+ "fleet": "2,557",
+ "active": "648",
+ "showcased": "15",
+ "upvotes": "0"
+ },
+ "saw": [
+ "creative writing guide",
+ "Machine translation solved meaning years ago"
+ ],
+ "responses": [
+ {
+ "method": "GET",
+ "path": "/api/agents/public",
+ "status": 200,
+ "seconds": 10.3
+ },
+ {
+ "method": "GET",
+ "path": "/api/agents/42534db3-f8f8-48ae-a4cb-ad8b9b42b2d7",
+ "status": 200,
+ "seconds": 0.3
+ }
+ ],
+ "cuts": [
+ {
+ "afterStep": 8,
+ "seconds": 6
+ },
+ {
+ "afterStep": 10,
+ "seconds": 9.1
+ },
+ {
+ "afterStep": 13,
+ "seconds": 11.1
+ },
+ {
+ "afterStep": 15,
+ "seconds": 4
+ }
+ ],
+ "video": {
+ "path": "public/x-media/agents-board-article/reel.mp4",
+ "sha256": "8a6032a715df2d546e71cc7ea3e1277c13457606c29222cf1b283e848b5a931d",
+ "frames": 249,
+ "motion": 1,
+ "durationSec": 8.3,
+ "width": 1280,
+ "height": 720,
+ "fps": 30,
+ "videoCodec": "h264",
+ "pixFmt": "yuv420p"
+ }
+}
diff --git a/data/x-content/proofs/google-cloud-production.json b/data/x-content/proofs/google-cloud-production.json
new file mode 100644
index 000000000..81ad50269
--- /dev/null
+++ b/data/x-content/proofs/google-cloud-production.json
@@ -0,0 +1,80 @@
+{
+ "id": "google-cloud-production",
+ "scenarioHash": "c790ce18a18bf6ee325fb72f3b8ced727d8cd381332e8f8e6586b43d1d09fbf3",
+ "ranAt": "2026-10-01T19:30:10.742Z",
+ "target": {
+ "commit": "68c8ec4db",
+ "revision": "three-ws-api-00475-lq6"
+ },
+ "passed": false,
+ "steps": [
+ {
+ "index": 1,
+ "kind": "goto",
+ "target": "https://three.ws/pose?src=/avatars/michelle.glb&title=Michelle&anim=idle",
+ "detail": "loaded https://three.ws/pose?src=/avatars/michelle.glb&title=Michelle&anim=idle",
+ "ok": true,
+ "ms": 17424,
+ "frames": 0
+ },
+ {
+ "index": 2,
+ "kind": "expect",
+ "target": "Playing",
+ "detail": "appeared after 3.4 s",
+ "ok": true,
+ "ms": 3403,
+ "frames": 0
+ },
+ {
+ "index": 3,
+ "kind": "hold",
+ "target": "1800",
+ "detail": "1800 ms",
+ "ok": true,
+ "ms": 67863,
+ "frames": 0
+ },
+ {
+ "index": 4,
+ "kind": "type",
+ "target": "a person does jumping jacks",
+ "detail": "typed 27 characters",
+ "ok": true,
+ "ms": 98323,
+ "frames": 0
+ },
+ {
+ "index": 5,
+ "kind": "click",
+ "target": "Generate",
+ "detail": "clicked; POST /api/forge-motion answered 202 in 1.4 s",
+ "ok": true,
+ "ms": 16924,
+ "frames": 0
+ },
+ {
+ "index": 6,
+ "kind": "expect",
+ "target": "Playing “a person does jumping jacks",
+ "ok": false,
+ "detail": "locator.waitFor: Timeout 120000ms exceeded.",
+ "ms": 121875,
+ "frames": 0
+ }
+ ],
+ "facts": {},
+ "saw": [
+ "Playing"
+ ],
+ "responses": [
+ {
+ "method": "POST",
+ "path": "/api/forge-motion",
+ "status": 202,
+ "seconds": 1.4
+ }
+ ],
+ "cuts": [],
+ "video": null
+}
diff --git a/data/x-content/proofs/voice-article.json b/data/x-content/proofs/voice-article.json
new file mode 100644
index 000000000..6c75fe9ee
--- /dev/null
+++ b/data/x-content/proofs/voice-article.json
@@ -0,0 +1,292 @@
+{
+ "id": "voice-article",
+ "scenarioHash": "791fc7b832546be2e2d231e3e05d98ef68cfac6f22b7e7bce3526c541479a32c",
+ "ranAt": "2026-10-01T19:21:01.763Z",
+ "target": {
+ "commit": "68c8ec4db",
+ "revision": "three-ws-api-00475-lq6"
+ },
+ "passed": true,
+ "steps": [
+ {
+ "index": 1,
+ "kind": "goto",
+ "target": "https://three.ws/voice",
+ "detail": "loaded https://three.ws/voice",
+ "ok": true,
+ "ms": 9630,
+ "frames": 0
+ },
+ {
+ "index": 2,
+ "kind": "scroll",
+ "target": "Browse voices",
+ "detail": "scrolled 672 px",
+ "ok": true,
+ "ms": 3779,
+ "frames": 27
+ },
+ {
+ "index": 3,
+ "kind": "read",
+ "target": "voices",
+ "detail": "voices = \"357\"",
+ "ok": true,
+ "ms": 3,
+ "frames": 0
+ },
+ {
+ "index": 4,
+ "kind": "hold",
+ "target": "1600",
+ "detail": "1600 ms",
+ "ok": true,
+ "ms": 5567,
+ "frames": 48
+ },
+ {
+ "index": 5,
+ "kind": "type",
+ "target": "Andrew",
+ "detail": "typed 6 characters",
+ "ok": true,
+ "ms": 3627,
+ "frames": 28
+ },
+ {
+ "index": 6,
+ "kind": "wait",
+ "target": "1200",
+ "detail": "1200 ms, not filmed",
+ "ok": true,
+ "ms": 1222,
+ "frames": 0
+ },
+ {
+ "index": 7,
+ "kind": "hold",
+ "target": "1200",
+ "detail": "1200 ms",
+ "ok": true,
+ "ms": 4161,
+ "frames": 36
+ },
+ {
+ "index": 8,
+ "kind": "click",
+ "target": "Use this voice",
+ "detail": "clicked",
+ "ok": true,
+ "ms": 2635,
+ "frames": 21
+ },
+ {
+ "index": 9,
+ "kind": "type",
+ "target": "Take the next exit. There is a charger two minutes past the bridge.",
+ "detail": "typed 67 characters",
+ "ok": true,
+ "ms": 17567,
+ "frames": 150
+ },
+ {
+ "index": 10,
+ "kind": "click",
+ "target": "Speak",
+ "detail": "clicked; POST /api/tts/synthesize answered 200 in 4.6 s",
+ "ok": true,
+ "ms": 6328,
+ "frames": 21
+ },
+ {
+ "index": 11,
+ "kind": "expect",
+ "target": "KB ·",
+ "detail": "appeared after 0.0 s",
+ "ok": true,
+ "ms": 19,
+ "frames": 0
+ },
+ {
+ "index": 12,
+ "kind": "hold",
+ "target": "1600",
+ "detail": "1600 ms",
+ "ok": true,
+ "ms": 5249,
+ "frames": 48
+ },
+ {
+ "index": 13,
+ "kind": "click",
+ "target": "Speak",
+ "detail": "clicked; POST /api/tts/synthesize answered 200 in 2 s",
+ "ok": true,
+ "ms": 3528,
+ "frames": 21
+ },
+ {
+ "index": 14,
+ "kind": "expect",
+ "target": "cached",
+ "detail": "appeared after 0.0 s",
+ "ok": true,
+ "ms": 31,
+ "frames": 0
+ },
+ {
+ "index": 15,
+ "kind": "hold",
+ "target": "1800",
+ "detail": "1800 ms",
+ "ok": true,
+ "ms": 6163,
+ "frames": 54
+ },
+ {
+ "index": 16,
+ "kind": "goto",
+ "target": "https://three.ws/drive?avatar=13f259c7-7024-4d68-b1f0-dbbf52c06209",
+ "detail": "loaded https://three.ws/drive?avatar=13f259c7-7024-4d68-b1f0-dbbf52c06209",
+ "ok": true,
+ "ms": 16297,
+ "frames": 0
+ },
+ {
+ "index": 17,
+ "kind": "expect",
+ "target": "Ready",
+ "detail": "appeared after 1.1 s",
+ "ok": true,
+ "ms": 1133,
+ "frames": 0
+ },
+ {
+ "index": 18,
+ "kind": "hold",
+ "target": "1500",
+ "detail": "1500 ms",
+ "ok": true,
+ "ms": 41330,
+ "frames": 45
+ },
+ {
+ "index": 19,
+ "kind": "click",
+ "target": "Driving",
+ "detail": "clicked",
+ "ok": true,
+ "ms": 16957,
+ "frames": 21
+ },
+ {
+ "index": 20,
+ "kind": "expect",
+ "target": "Parked",
+ "detail": "appeared after 0.0 s",
+ "ok": true,
+ "ms": 10,
+ "frames": 0
+ },
+ {
+ "index": 21,
+ "kind": "click",
+ "target": "Type",
+ "detail": "clicked",
+ "ok": true,
+ "ms": 14566,
+ "frames": 21
+ },
+ {
+ "index": 22,
+ "kind": "type",
+ "target": "Day mode, please",
+ "detail": "typed 16 characters",
+ "ok": true,
+ "ms": 32161,
+ "frames": 48
+ },
+ {
+ "index": 23,
+ "kind": "press",
+ "target": "Enter",
+ "detail": "pressed Enter",
+ "ok": true,
+ "ms": 4205,
+ "frames": 6
+ },
+ {
+ "index": 24,
+ "kind": "wait",
+ "target": "1000",
+ "detail": "1000 ms, not filmed",
+ "ok": true,
+ "ms": 1509,
+ "frames": 0
+ },
+ {
+ "index": 25,
+ "kind": "read",
+ "target": "light",
+ "detail": "light = \"Day\"",
+ "ok": true,
+ "ms": 6,
+ "frames": 0
+ },
+ {
+ "index": 26,
+ "kind": "hold",
+ "target": "2600",
+ "detail": "2600 ms",
+ "ok": true,
+ "ms": 62254,
+ "frames": 78
+ }
+ ],
+ "facts": {
+ "voices": "357",
+ "light": "Day"
+ },
+ "saw": [
+ "KB ·",
+ "cached",
+ "Ready",
+ "Parked"
+ ],
+ "responses": [
+ {
+ "method": "POST",
+ "path": "/api/tts/synthesize",
+ "status": 200,
+ "seconds": 4.6
+ },
+ {
+ "method": "POST",
+ "path": "/api/tts/synthesize",
+ "status": 200,
+ "seconds": 2
+ }
+ ],
+ "cuts": [
+ {
+ "afterStep": 10,
+ "seconds": 3.5
+ },
+ {
+ "afterStep": 16,
+ "seconds": 14.8
+ }
+ ],
+ "video": {
+ "path": "public/x-media/voice-article/reel.mp4",
+ "sha256": "e5e30fae1ee54d035e9ffd7174a57d86aa4ffc45c7efc5082199bc3a59cb864d",
+ "frames": 673,
+ "motion": 0.6,
+ "durationSec": 22.43,
+ "width": 1280,
+ "height": 720,
+ "fps": 30,
+ "videoCodec": "h264",
+ "pixFmt": "yuv420p"
+ }
+}
diff --git a/data/x-content/queue.json b/data/x-content/queue.json
index 24a9a8850..84db236ec 100644
--- a/data/x-content/queue.json
+++ b/data/x-content/queue.json
@@ -12461,6 +12461,1004 @@
}
],
"mentions": {}
+ },
+ {
+ "id": "forge-article",
+ "status": "review",
+ "kind": "article",
+ "tier": 1,
+ "lane": "article",
+ "pattern": "longform",
+ "notBefore": "2026-10-02T00:00:00Z",
+ "covers": [
+ "/forge",
+ "workers/model-trellis",
+ "workers/model-hunyuan3d",
+ "workers/model-triposg"
+ ],
+ "article": {
+ "title": "How a sentence becomes a textured 3D model, and where the job goes when a GPU fails",
+ "body": "data/x-content/articles/forge-article.md",
+ "cover": {
+ "path": "public/x-media/forge-article/cover.png",
+ "alt": "A 3D model generated on the Forge from a text prompt, shown in its viewer on a dark background: a Greek hoplite in a polished bronze cuirass and crested helmet, with a red tunic, bronze greaves, sandals and a round shield held at his side"
+ }
+ },
+ "probes": [
+ {
+ "type": "api",
+ "name": "the self-hosted TRELLIS lane reports healthy",
+ "url": "https://three.ws/api/forge?health=1",
+ "expect": {
+ "status": 200,
+ "contains": "The TRELLIS (self-host) worker is reachable."
+ }
+ },
+ {
+ "type": "api",
+ "name": "the Hunyuan3D worker is loaded and ready on its Blackwell GPU",
+ "url": "https://model-hunyuan3d-21-rtx-93741856042.us-central1.run.app/health",
+ "expect": {
+ "status": 200,
+ "contains": "\"ready\":true"
+ }
+ },
+ {
+ "type": "api",
+ "name": "the Forge page is served",
+ "url": "https://three.ws/forge",
+ "expect": {
+ "status": 200
+ }
+ }
+ ],
+ "posts": [
+ {
+ "text": "A text prompt on the Forge runs through two chains: one paints a reference picture, the other rebuilds it as a textured 3D model on our own NVIDIA GPUs on Cloud Run. When a worker fails mid-job, the job moves to the next engine under the same id. The write-up covers the lanes, the rescue order and the limits: three.ws/forge"
+ }
+ ],
+ "claims": [
+ {
+ "says": "one paints a reference picture, the other rebuilds it as a textured 3D model",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/forge",
+ "contains": "renders a preview image, then builds 3D"
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/forge-reference-image.js",
+ "contains": "The text→3D pipeline is: user prompt → prompt director rewrite → a reference\n// IMAGE is generated → TRELLIS/Hunyuan3D reconstructs a textured GLB from that\n// image."
+ }
+ ]
+ },
+ {
+ "says": "on our own NVIDIA GPUs on Cloud Run",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://model-trellis-lp642k3kpa-uc.a.run.app/health",
+ "contains": "\"gpu_name\":\"NVIDIA L4\""
+ },
+ {
+ "type": "page",
+ "url": "https://model-hunyuan3d-21-rtx-93741856042.us-central1.run.app/health",
+ "contains": "\"gpu_name\":\"NVIDIA RTX PRO 6000 Blackwell Server Edition\""
+ }
+ ]
+ },
+ {
+ "says": "When a worker fails mid-job, the job moves to the next engine under the same id",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/forge.js",
+ "contains": "// configured lane: the client keeps polling this id and sees\n\t\t// status:\"running\" with a new backend instead of a dead end."
+ }
+ ]
+ },
+ {
+ "says": "when an engine fails partway through a job, the job moves to the next engine under the same job id and the progress bar keeps going",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/inside-the-forge",
+ "contains": "You keep watching the same progress bar; you never find out it happened."
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/forge-failover.js",
+ "contains": "The client\n// keeps polling the SAME job id and simply sees status:\"running\" with a\n// new `backend`"
+ }
+ ]
+ },
+ {
+ "says": "a Draft or Standard model costs nothing and needs no account",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/forge",
+ "contains": "Free · no sign-up · yours to download"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/inside-the-forge",
+ "contains": "Draft and standard usually cost us nothing to produce, because they run on free lanes and our own GPUs."
+ }
+ ]
+ },
+ {
+ "says": "The realism of the final mesh is set almost entirely by the realism of that one reference image",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-reference-image.js",
+ "contains": "The realism of the final mesh is set almost entirely by the realism of\n// that one reference image"
+ }
+ ]
+ },
+ {
+ "says": "painted as one centered subject on a plain seamless background, under soft even light, with real-world materials",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-reference-image.js",
+ "contains": "// - a single centered subject on a plain seamless studio background,\n// - soft even shadowless lighting,\n// - real-world materials and surface micro-detail (the photoreal look),"
+ }
+ ]
+ },
+ {
+ "says": "A Gemini image model on Vertex AI leads, because it draws the most photoreal reference",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_mcp3d/text-to-image.js",
+ "contains": "// Quality-first ordering: the Vertex Gemini image model outdraws 4-step\n\t// distilled FLUX for photoreal reference images"
+ }
+ ]
+ },
+ {
+ "says": "falls through to FLUX on NVIDIA's hosted inference, then to further providers, ending in a keyless rung, so no single provider's failure is ever the terminal error",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_mcp3d/text-to-image.js",
+ "contains": "keyed rungs: no single provider's failure is ever the terminal error."
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/forge-reference-image.js",
+ "contains": "if the Vertex lane is unconfigured or fails\n// for any reason, it falls through to the existing reference-image provider"
+ }
+ ]
+ },
+ {
+ "says": "Photos skip this step: your photo is the reference. A sketch skips it too",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/inside-the-forge",
+ "contains": "If you uploaded photos instead, this step is skipped: your photos are the reference. Sketches skip it too and drive the geometry directly."
+ }
+ ]
+ },
+ {
+ "says": "Draft and Standard go to TRELLIS on our own worker. High goes to Hunyuan3D 2.1 on our own worker",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-tiers.js",
+ "contains": "\tdraft: Object.freeze({ image: 'trellis_selfhost' }),\n\tstandard: Object.freeze({ image: 'trellis_selfhost' }),"
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/forge-tiers.js",
+ "contains": "\thigh: Object.freeze({ image: 'hunyuan3d' }),"
+ },
+ {
+ "type": "page",
+ "url": "https://model-hunyuan3d-21-rtx-93741856042.us-central1.run.app/health",
+ "contains": "\"model\":\"hunyuan3d-2.1\""
+ }
+ ]
+ },
+ {
+ "says": "paints a full PBR material set (colour, roughness and normal maps) rather than one baked colour map",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "Hunyuan3D 2.1 runs a shape DiT for geometry, then a separate PBR paint pass"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "Hunyuan3D 2.0 gives you a mesh with a single baked diffuse texture."
+ }
+ ]
+ },
+ {
+ "says": "a prompt that clearly reads as a hard-surface object (a robot, a vehicle, a piece of furniture) swaps TRELLIS ahead of Hunyuan3D",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-tiers.js",
+ "contains": "Only an unambiguously hard-surface subject reorders the lanes."
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/forge-tiers.js",
+ "contains": "const HARDSURFACE_SUBJECT_RE = /\\b(robot|mech|mecha|drone|vehicle|"
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/forge-tiers.js",
+ "contains": "is crisp on hard-surface / mechanical / geometric subjects"
+ }
+ ]
+ },
+ {
+ "says": "Hugging Face Spaces for photos, and NVIDIA's hosted TRELLIS for text, which keeps a text prompt answerable even when every one of our GPU workers is cold",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-tiers.js",
+ "contains": "returns a model when every GPU worker (and HuggingFace) is cold or down"
+ }
+ ]
+ },
+ {
+ "says": "Engines that need your own API key run only when you pick them",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/forge.js",
+ "contains": "BYOK lanes are excluded (never auto-selected)."
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/forge",
+ "contains": "your own key"
+ },
+ {
+ "type": "file",
+ "path": "src/forge.js",
+ "contains": "// Flip the engine picker to a backend by id"
+ }
+ ]
+ },
+ {
+ "says": "It takes the first lane in its order that is healthy or simply unknown; if none is clean, it takes the first that is not confirmed down",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-tiers.js",
+ "contains": "// 1. The first lane that is healthy OR simply unknown."
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/forge-tiers.js",
+ "contains": "// 2. Otherwise the first lane not confirmed 'down'"
+ }
+ ]
+ },
+ {
+ "says": "about 12,000 at Draft, about 30,000 at Standard, and up to 200,000 at High",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/inside-the-forge",
+ "contains": "Draft: about 12,000 polygons, fast."
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/inside-the-forge",
+ "contains": "Standard (the default): about 30,000 polygons"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/inside-the-forge",
+ "contains": "High: up to 200,000 polygons"
+ }
+ ]
+ },
+ {
+ "says": "The poll handler marks the failed lane unhealthy for 90 seconds",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-lane-health.js",
+ "contains": "const LANE_COOLDOWN_SECONDS = 90;"
+ },
+ {
+ "type": "file",
+ "path": "api/forge.js",
+ "contains": "if (platformLaneFailed) await markLaneUnhealthy(failedBackend);"
+ }
+ ]
+ },
+ {
+ "says": "recovers the original prompt and the stored reference picture from the job's record, resubmits them to the next lane, and binds the old job id to the new one",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-failover.js",
+ "contains": "The failed poll recovers the original inputs\n// (prompt + reference image) from the creation row, resubmits to the next\n// configured lane, and binds old-handle → new-handle in Redis."
+ }
+ ]
+ },
+ {
+ "says": "our TRELLIS worker, then our Hunyuan3D worker, then a hosted TRELLIS lane where one is configured, then TripoSG",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-failover.js",
+ "contains": "const ASYNC_REDISPATCH_ORDER = ['trellis_selfhost', 'hunyuan3d', 'trellis', 'triposg'];"
+ }
+ ]
+ },
+ {
+ "says": "A job gets its first lane plus up to three backups",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-failover.js",
+ "contains": "so a request tries at most 1 primary + 3 backup lanes."
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/forge-failover.js",
+ "contains": "export const MAX_FAILOVER_HOPS = 3;"
+ }
+ ]
+ },
+ {
+ "says": "TripoSG closes the chain because it returns an untextured mesh, so it runs only once every textured lane has failed or is marked down",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-failover.js",
+ "contains": "// TripoSG closes the chain. It returns an untextured mesh, so it only runs\n// once every textured lane has failed or is marked down."
+ }
+ ]
+ },
+ {
+ "says": "a worker outage from 18 to 28 September that took Hunyuan3D and TRELLIS down together: without it, roughly 400 photo jobs a day ended in a hard failure while the TripoSG worker sat healthy",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-failover.js",
+ "contains": "// 2026-09-18..28 worker outage, which took Hunyuan3D and TRELLIS self-host\n// down together, ended roughly 400 photo jobs a day in a hard failure while\n// the TripoSG worker sat healthy."
+ }
+ ]
+ },
+ {
+ "says": "The response names the lanes that can serve a fresh retry",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/forge.js",
+ "contains": "...(suggestions.length ? { retryable: true, retry_backends: suggestions } : {}),"
+ }
+ ]
+ },
+ {
+ "says": "TRELLIS runs on an L4. Hunyuan3D 2.1 runs on an RTX PRO 6000 Blackwell, held warm with one instance, because a cold start means loading 14 GiB of weights before any work begins",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://model-trellis-lp642k3kpa-uc.a.run.app/health",
+ "contains": "\"gpu_name\":\"NVIDIA L4\""
+ },
+ {
+ "type": "page",
+ "url": "https://model-hunyuan3d-21-rtx-93741856042.us-central1.run.app/health",
+ "contains": "\"gpu_name\":\"NVIDIA RTX PRO 6000 Blackwell Server Edition\""
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "Our RTX service runs with min instances and max instances both set to 1, held warm."
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "a cold start means loading 14 GiB before the first token of work"
+ }
+ ]
+ },
+ {
+ "says": "Every worker speaks the same contract",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "every image-to-3D worker speaks the same wire contract."
+ }
+ ]
+ },
+ {
+ "says": "The health check reads the second field, so a worker that is up but still loading is reported degraded, not healthy",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-health.js",
+ "contains": "is up but still loading its model; a generation will queue until it is ready."
+ }
+ ]
+ },
+ {
+ "says": "Task state lives in object storage rather than in the instance",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "Task state lives in object storage, not in the instance."
+ }
+ ]
+ },
+ {
+ "says": "18 GiB staged plus 14 GiB loaded is a 32 GiB peak against the L4 tier's 32 GiB ceiling, and the container was killed on every cold start",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "So the peak is not 14 GiB and it is not 18 GiB. It is both at once, against the L4 tier's ceiling of 32 GiB:"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "Every single cold start ended the same way:"
+ }
+ ]
+ },
+ {
+ "says": "The RTX PRO 6000 tier's floor of 80 GiB clears it",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "The platform minimums for nvidia-rtx-pro-6000 on Cloud Run are 20 CPU and 80 GiB, which clears a 32 GiB peak with room to spare"
+ }
+ ]
+ },
+ {
+ "says": "Blackwell needs its GPU code built with NVIDIA's 12.8 toolkit. The worker image compiles its extensions for both GPU generations, so which card runs it is a deploy flag rather than a rebuild",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "Blackwell is sm_120 and needs cu128 wheels; no flag substitutes for the rebuild."
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "We compile the custom CUDA extensions for both architectures, not just the one we deploy to:"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "contains": "That turns \"which card is available in this region today\" from a rebuild into a deploy flag"
+ }
+ ]
+ },
+ {
+ "says": "Hugging Face Spaces block for the whole generation and NVIDIA's hosted lane takes text only, so neither can pick up a job that failed on a poll; both appear only in the retry suggestions",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-failover.js",
+ "contains": "HuggingFace Spaces cannot be auto-redispatched from a poll (its\n// provider blocks through the whole generation), so it appears only in the\n// retry_backends suggestions"
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/forge-failover.js",
+ "contains": "NVIDIA is excluded\n// (text-only, and a poll-time redispatch always reconstructs from a stored\n// reference image)."
+ }
+ ]
+ },
+ {
+ "says": "a job that started from several photos continues from one of them, and says so",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/forge.js",
+ "contains": "multi-view original degrades visibly (views_used: 1), never silently."
+ }
+ ]
+ },
+ {
+ "says": "Each check waits 4 seconds",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/forge-health.js",
+ "contains": "const PROBE_TIMEOUT_MS = 4_000;"
+ }
+ ]
+ }
+ ],
+ "mentions": {}
+ },
+ {
+ "id": "quality-gate-article",
+ "status": "review",
+ "kind": "article",
+ "tier": 1,
+ "lane": "article",
+ "pattern": "longform",
+ "notBefore": "2026-10-02T00:00:00Z",
+ "covers": [
+ "/inspect",
+ "/diff",
+ "@three-ws/glb-diff"
+ ],
+ "article": {
+ "title": "Two checks before a 3D model ships: the catalog gate, and a diff that names what broke",
+ "body": "data/x-content/articles/quality-gate-article.md",
+ "cover": {
+ "path": "public/x-media/quality-gate-article/cover.png",
+ "alt": "Two 3D characters overlaid in the Model Diff viewer on a black background: a translucent pink jointed mannequin, the X Bot sample highlighted as the model before the change, standing beside Michelle, a green-skinned character with leafy green hair and yellow trousers, caught mid-dance as her clip plays"
+ }
+ },
+ "probes": [
+ {
+ "type": "api",
+ "name": "the free diff endpoint reports the X Bot to Michelle change as breaking",
+ "url": "https://three.ws/api/3d/diff?a=https://three.ws/avatars/xbot.glb&b=https://three.ws/avatars/michelle.glb",
+ "expect": {
+ "status": 200,
+ "contains": "\"severity\":\"breaking\""
+ }
+ },
+ {
+ "type": "api",
+ "name": "the inspector page is served",
+ "url": "https://three.ws/inspect",
+ "expect": {
+ "status": 200
+ }
+ }
+ ],
+ "posts": [
+ {
+ "text": "Our catalog turns away a character that arrives as more than 8 meshes, and you can run that exact check on your own .glb without uploading it. The write-up covers the five structural checks, how Model Diff names the joints and clips a change removed, the exit codes for CI, and where its rename detection still guesses wrong: three.ws/inspect"
+ }
+ ],
+ "claims": [
+ {
+ "says": "more than 8 meshes",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "if (category === 'avatar' && m.meshCount > 8) reasons.push('too_many_meshes_for_a_character');"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/inspect",
+ "contains": "This runs the exact quality gate three.ws applies to its own catalog entries against any .glb you give it."
+ }
+ ]
+ },
+ {
+ "says": "you can run that exact check on your own .glb without uploading it",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/inspect",
+ "contains": "It runs entirely in your browser: nothing is uploaded, and the verdict is instant."
+ },
+ {
+ "type": "file",
+ "path": "src/inspect-page.js",
+ "contains": "import { gateMesh, explainMeshGate, SEED_GATE_VERSION } from '../api/_lib/seed-mesh-gate.js';"
+ }
+ ]
+ },
+ {
+ "says": "the five structural checks",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/inspect",
+ "contains": "These checks measure whether a model is built well enough to publish: density, weight, texture, bounds, assembly."
+ }
+ ]
+ },
+ {
+ "says": "how Model Diff names the joints and clips a change removed",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/diff?a=/avatars/xbot.glb&b=/avatars/michelle.glb",
+ "contains": "lost 2 joint(s): mixamorig:LeftEye, mixamorig:RightEye. Clips that target them will not play."
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/diff?a=/avatars/xbot.glb&b=/avatars/michelle.glb",
+ "contains": "5 animations were removed (\"agree\", \"idle\", \"run\" and 2 more)."
+ }
+ ]
+ },
+ {
+ "says": "the exit codes for CI",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "packages/glb-diff/README.md",
+ "contains": "Exit codes are the contract a CI job depends on, so they are deliberate:"
+ }
+ ]
+ },
+ {
+ "says": "where its rename detection still guesses wrong",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/diff?a=/avatars/xbot.glb&b=/avatars/michelle.glb",
+ "contains": "headShake to SambaDance"
+ },
+ {
+ "type": "file",
+ "path": "packages/glb-diff/src/diff.js",
+ "contains": "similarity: (x, y) => 0.65 * jaccard(x.targets, y.targets) + 0.35 * ratio(x.duration, y.duration),"
+ }
+ ]
+ },
+ {
+ "says": "has to clear a structural gate before anyone sees it, and the Model Inspector runs that same gate, the same module with the same thresholds",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/inspect",
+ "contains": "Same module, same thresholds, same decision."
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/seed-quality.js",
+ "contains": "Rejects are never published."
+ }
+ ]
+ },
+ {
+ "says": "The seeding job that fills the catalog imports it, and so does the inspector page",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/inspect",
+ "contains": "The page imports the same module the seeding cron runs, so the rules here cannot drift from the rules in production."
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/seed-quality.js",
+ "contains": "} from './seed-mesh-gate.js';"
+ },
+ {
+ "type": "file",
+ "path": "api/cron/forge-seed-cron.js",
+ "contains": "import { evaluateSeedAsset, inProcessTransport, quarantineReject } from '../_lib/seed-quality.js';"
+ }
+ ]
+ },
+ {
+ "says": "A threshold change lands in both places at once because there is only one place",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "threshold change lands in both places at once because there is only one place."
+ }
+ ]
+ },
+ {
+ "says": "the gate reads nothing but the glTF JSON chunk",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "reads nothing but\n// the glTF JSON chunk, so it can run anywhere"
+ }
+ ]
+ },
+ {
+ "says": "A build check walks the real bundler graph from the page and fails the moment a Node built-in becomes reachable",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "`npm run check:browser-graph` (it walks the real bundler graph from the HTML\n// entries and fails the build the moment a `node:` built-in becomes reachable)"
+ }
+ ]
+ },
+ {
+ "says": "Every verdict carries a gate version, because an accept rate is only comparable within one version of the rules",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "// Bumped whenever a threshold changes, and stored on every verdict"
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "rate is only comparable within one gate version."
+ }
+ ]
+ },
+ {
+ "says": "Geometry density: between 1,500 and 1,500,000 vertices",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "minVertices: numEnv('SEED_GATE_MIN_VERTICES', 1_500),"
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "maxVertices: numEnv('SEED_GATE_MAX_VERTICES', 1_500_000),"
+ }
+ ]
+ },
+ {
+ "says": "File weight: between 20 KB and 80 MB",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "minBytes: numEnv('SEED_GATE_MIN_BYTES', 20_000),"
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "maxBytes: numEnv('SEED_GATE_MAX_BYTES', 80 * 1024 * 1024),"
+ }
+ ]
+ },
+ {
+ "says": "a collapsed mesh passes every triangle count and still draws as a speck",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "The bounding box has no extent, so the mesh draws as a single speck no matter how the camera is framed. It passes triangle counts, which is exactly why it needs its own check."
+ }
+ ]
+ },
+ {
+ "says": "a character may arrive as at most 8 meshes. A prop is exempt",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "bound: category === 'avatar' ? 'at most 8' : 'props exempt',"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/inspect",
+ "contains": "Props skip the character-only checks."
+ }
+ ]
+ },
+ {
+ "says": "A failed check comes with the fix, written next to the rule it explains, so the guidance cannot drift from the rule",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "It lives beside the gate rather than in the page so the guidance\n// cannot drift from the rule it describes"
+ }
+ ]
+ },
+ {
+ "says": "The texture rule is deliberately stricter than the interactive Forge",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-mesh-gate.js",
+ "contains": "This bound is stricter than the interactive forge on purpose: a catalog entry is the platform vouching for itself."
+ }
+ ]
+ },
+ {
+ "says": "It clears the density floor of 1,500 with 1,728 vertices",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/inspect?url=https%3A%2F%2Fthree.ws%2Favatars%2Ffox.glb",
+ "contains": "1,728 vertices"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/inspect?url=https%3A%2F%2Fthree.ws%2Favatars%2Ffox.glb",
+ "contains": "Clears the catalog bar"
+ }
+ ]
+ },
+ {
+ "says": "Fox is the sample labelled \"Passes, barely\"",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/inspect-page.js",
+ "contains": "{ id: 'fox', label: 'Fox', url: '/avatars/fox.glb', note: 'Passes, barely' },"
+ }
+ ]
+ },
+ {
+ "says": "The gate is the first of two stages. On the catalog path a vision model then renders the mesh",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/seed-quality.js",
+ "contains": "2. VISION JUDGE (Vertex Gemini). Renders the mesh and asks the judge two"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/inspect",
+ "contains": "Whether it looks good is judged separately by a vision model on the catalog path."
+ }
+ ]
+ },
+ {
+ "says": "pairs objects across the two the way git pairs renamed files: by name first, then by a hash of the content, then by similarity",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/diff",
+ "contains": "Objects are matched by name, then by content hash, then by similarity, the same three-tier strategy git uses for file renames."
+ }
+ ]
+ },
+ {
+ "says": "Vertex data is compared at a quantum of 1e-4 units, so the float drift of a re-export does not register while a genuine edit does",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/diff",
+ "contains": "Vertex data is compared at a quantum of 1e-4 units, so the float drift a re-export or a Draco round trip introduces does not register while a genuine edit does."
+ }
+ ]
+ },
+ {
+ "says": "Every change set carries one severity: none, cosmetic, minor, major or breaking",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "packages/glb-diff/README.md",
+ "contains": "Every change set carries one severity, the worst of everything it found:"
+ },
+ {
+ "type": "file",
+ "path": "packages/glb-diff/src/severity.js",
+ "contains": "export const SEVERITIES = ['none', 'cosmetic', 'minor', 'major', 'breaking'];"
+ }
+ ]
+ },
+ {
+ "says": "A texture swap is minor because everything still loads and plays",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "packages/glb-diff/src/severity.js",
+ "contains": "A 40 MB texture swap is `minor` because everything still"
+ }
+ ]
+ },
+ {
+ "says": "That is the failure that renders fine and never moves",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/diff",
+ "contains": "That is the failure that renders fine and never moves."
+ }
+ ]
+ },
+ {
+ "says": "The skeleton section names the two eye joints that are gone and says clips addressing them will not play. The animation section lists the five clips that no longer exist",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/diff?a=/avatars/xbot.glb&b=/avatars/michelle.glb",
+ "contains": "lost 2 joint(s): mixamorig:LeftEye, mixamorig:RightEye. Clips that target them will not play."
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/diff?a=/avatars/xbot.glb&b=/avatars/michelle.glb",
+ "contains": "5 animations were removed (\"agree\", \"idle\", \"run\" and 2 more)."
+ }
+ ]
+ },
+ {
+ "says": "The page, a command line tool and a free HTTP endpoint all run the same published package",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/diff",
+ "contains": "so a verdict in CI can never disagree with what you see here."
+ },
+ {
+ "type": "file",
+ "path": "api/3d/diff.js",
+ "contains": "import { describeModel, diffDescriptions, formatMarkdown, formatText } from '@three-ws/glb-diff';"
+ },
+ {
+ "type": "file",
+ "path": "src/model-diff.js",
+ "contains": "from '@three-ws/glb-diff';"
+ }
+ ]
+ },
+ {
+ "says": "Zero means the diff ran and stayed below the threshold",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "packages/glb-diff/README.md",
+ "contains": "| `0` | The diff ran and stayed below `--fail-on` |"
+ },
+ {
+ "type": "file",
+ "path": "packages/glb-diff/README.md",
+ "contains": "| `2` | The tool could not run: bad arguments, unreadable input, unparseable model |"
+ }
+ ]
+ },
+ {
+ "says": "caps each side at 32 MB",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/3d/diff.js",
+ "contains": "const MAX_BYTES = 32 * 1024 * 1024;"
+ }
+ ]
+ },
+ {
+ "says": "The similarity pass is skipped when two models would need more than 40,000 comparisons, and the result says so",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "packages/glb-diff/src/match.js",
+ "contains": "When it trips, the result\n// says so rather than quietly downgrading renames into add/remove noise."
+ },
+ {
+ "type": "file",
+ "path": "packages/glb-diff/src/match.js",
+ "contains": "const MAX_SIMILARITY_PAIRS = 40_000;"
+ }
+ ]
+ },
+ {
+ "says": "Most avatars on the platform ship with compressed geometry",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "CLAUDE.md",
+ "contains": "Most three.ws avatars ship with `EXT_meshopt_compression`"
+ }
+ ]
+ },
+ {
+ "says": "the command line tool needs the optional meshoptimizer package installed next to it",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "packages/glb-diff/src/document.js",
+ "contains": "decoderPromise = import('meshoptimizer')"
+ },
+ {
+ "type": "file",
+ "path": "packages/glb-diff/package.json",
+ "contains": "\"peerDependenciesMeta\": {\n \"meshoptimizer\": {\n \"optional\": true"
+ }
+ ]
+ }
+ ],
+ "mentions": {}
}
]
}
diff --git a/data/x-content/reviews/any-skeleton-article.json b/data/x-content/reviews/any-skeleton-article.json
index 1931cbff2..342c9a5f0 100644
--- a/data/x-content/reviews/any-skeleton-article.json
+++ b/data/x-content/reviews/any-skeleton-article.json
@@ -1,10 +1,9 @@
{
"id": "any-skeleton-article",
"contentHash": "73bc79be1907627aabf54f54c24d06d6ebbcda35dabf5bee98fc943c3c16914b",
- "reviewedAt": "2026-10-01T19:28:44.631Z",
+ "reviewedAt": "2026-10-01T19:32:29.398Z",
"passed": false,
"blockers": [
- "spelling retargeter: did you mean retargeted, retarget, retargets?",
"spelling lowerarm: did you mean lower, lowers, lowery?",
"spelling canonicalizer: did you mean canonicalize, canonicalized, canonicalizes?",
"spelling Miku: did you mean mike, Mike, maik?",
@@ -424,12 +423,6 @@
"ok": true,
"detail": "HTTP 200"
},
- {
- "kind": "spelling",
- "target": "retargeter",
- "ok": false,
- "detail": "did you mean retargeted, retarget, retargets?"
- },
{
"kind": "spelling",
"target": "lowerarm",
@@ -470,17 +463,9 @@
"specificity": 5,
"voice": 5,
"professionalism": 5,
- "visual": 4
+ "visual": 5
},
- "issues": [
- {
- "severity": "minor",
- "area": "visual",
- "quote": "Article cover: public/x-media/any-skeleton-article/cover.png",
- "problem": "The cover shows one avatar on one skeleton, while the post's claim is that the same clips play across Mixamo, VRoid, Unreal and Daz rigs. It is real product, sharp, and does not contradict the copy, so it ships, but a multi-avatar frame (like the article's own one-idle-four-bodies.png) would support the claim more directly.",
- "fix": "Optional: use the four-avatar frame /x-media/any-skeleton-article/one-idle-four-bodies.png as the cover instead; not required to ship."
- }
- ],
+ "issues": [],
"rewrite": {
"posts": [
"Our animation clips were made once, on one reference skeleton, and they play on Mixamo, VRoid, Unreal and Daz avatars built later by other tools. The write-up covers how bone names are matched, why a matching name is not enough, and where the retargeter stops on purpose: three.ws/avatar-artifact"
@@ -489,7 +474,7 @@
"A generated 3D circus ringmaster in a red tailcoat, top hat and tall boots standing at rest with his hands open in front of him, lit in red and violet on a dark stage under a starfield. He is idling on the shared animation library in the three.ws Avatar Artifact viewer."
]
},
- "summary": "Every claim in the post and article passed verification against live pages and source files, and the voice is clean: no hashtags, no emoji, no dashes, mechanism first. The only note is that the cover shows a single avatar rather than the cross-skeleton claim, but it is real product and accurate, so this is ready to publish.",
+ "summary": "Every claim in the post is backed by a passing verification against the live pages and source files, and the copy states mechanisms (one canonical skeleton, bone-name matching, deliberate stop conditions) rather than adjectives. Voice, links, and the cover image with its alt text all conform, so it ships unchanged.",
"model": "nvidia:moonshotai/kimi-k3",
"fallbacks": [
"Vertex 403: {\n \"error\": {\n \"code\": 403,\n \"message\": \"Request had insufficient authentication scopes.\",\n \"status\": \"PERMISSION_DENIED\",\n \"details\": [\n {\n \"@type\": \"type.googleapis.com/goog",
diff --git a/data/x-content/reviews/forge-article.json b/data/x-content/reviews/forge-article.json
new file mode 100644
index 000000000..ab0be8d2f
--- /dev/null
+++ b/data/x-content/reviews/forge-article.json
@@ -0,0 +1,533 @@
+{
+ "id": "forge-article",
+ "contentHash": "12279cb509f8d695672307a5096fabc5dc341b95f913978c3de06001b8821ecc",
+ "reviewedAt": "2026-10-01T19:35:28.679Z",
+ "passed": true,
+ "blockers": [],
+ "texts": [
+ "How a sentence becomes a textured 3D model, and where the job goes when a GPU fails",
+ "A text prompt on the Forge runs through two chains: one paints a reference picture, the other rebuilds it as a textured 3D model on our own NVIDIA GPUs on Cloud Run. When a worker fails mid-job, the job moves to the next engine under the same id. The write-up covers the lanes, the rescue order and the limits: three.ws/forge"
+ ],
+ "lint": [
+ {
+ "where": "article body",
+ "rule": "register",
+ "severity": "major",
+ "message": "\"simply\" is filler; replace it with the fact it stands in for"
+ },
+ {
+ "where": "article body",
+ "rule": "register",
+ "severity": "major",
+ "message": "\"actually\" is filler; replace it with the fact it stands in for"
+ }
+ ],
+ "verification": {
+ "ok": true,
+ "checks": [
+ {
+ "claim": "one paints a reference picture, the other rebuilds it as a textured 3D model",
+ "kind": "evidence:page",
+ "target": "https://three.ws/forge",
+ "ok": true,
+ "detail": "live page shows \"renders a preview image, then builds 3D\""
+ },
+ {
+ "claim": "one paints a reference picture, the other rebuilds it as a textured 3D model",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-reference-image.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "on our own NVIDIA GPUs on Cloud Run",
+ "kind": "evidence:page",
+ "target": "https://model-trellis-lp642k3kpa-uc.a.run.app/health",
+ "ok": true,
+ "detail": "live page shows \"\"gpu_name\":\"NVIDIA L4\"\""
+ },
+ {
+ "claim": "on our own NVIDIA GPUs on Cloud Run",
+ "kind": "evidence:page",
+ "target": "https://model-hunyuan3d-21-rtx-93741856042.us-central1.run.app/health",
+ "ok": true,
+ "detail": "live page shows \"\"gpu_name\":\"NVIDIA RTX PRO 6000 Blackwell Server Edition\"\""
+ },
+ {
+ "claim": "When a worker fails mid-job, the job moves to the next engine under the same id",
+ "kind": "evidence:file",
+ "target": "api/forge.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "when an engine fails partway through a job, the job moves to the next engine under the same job id and the progress bar keeps going",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/inside-the-forge",
+ "ok": true,
+ "detail": "live page shows \"You keep watching the same progress bar; you never find out it happened.\""
+ },
+ {
+ "claim": "when an engine fails partway through a job, the job moves to the next engine under the same job id and the progress bar keeps going",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-failover.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "a Draft or Standard model costs nothing and needs no account",
+ "kind": "evidence:page",
+ "target": "https://three.ws/forge",
+ "ok": true,
+ "detail": "live page shows \"Free · no sign-up · yours to download\""
+ },
+ {
+ "claim": "a Draft or Standard model costs nothing and needs no account",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/inside-the-forge",
+ "ok": true,
+ "detail": "live page shows \"Draft and standard usually cost us nothing to produce, because they run on free lanes and our own GPUs.\""
+ },
+ {
+ "claim": "The realism of the final mesh is set almost entirely by the realism of that one reference image",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-reference-image.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "painted as one centered subject on a plain seamless background, under soft even light, with real-world materials",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-reference-image.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "A Gemini image model on Vertex AI leads, because it draws the most photoreal reference",
+ "kind": "evidence:file",
+ "target": "api/_mcp3d/text-to-image.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "falls through to FLUX on NVIDIA's hosted inference, then to further providers, ending in a keyless rung, so no single provider's failure is ever the terminal error",
+ "kind": "evidence:file",
+ "target": "api/_mcp3d/text-to-image.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "falls through to FLUX on NVIDIA's hosted inference, then to further providers, ending in a keyless rung, so no single provider's failure is ever the terminal error",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-reference-image.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Photos skip this step: your photo is the reference. A sketch skips it too",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/inside-the-forge",
+ "ok": true,
+ "detail": "live page shows \"If you uploaded photos instead, this step is skipped: your photos are the reference. Sketches skip it too and drive the geometry directly.\""
+ },
+ {
+ "claim": "Draft and Standard go to TRELLIS on our own worker. High goes to Hunyuan3D 2.1 on our own worker",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-tiers.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Draft and Standard go to TRELLIS on our own worker. High goes to Hunyuan3D 2.1 on our own worker",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-tiers.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Draft and Standard go to TRELLIS on our own worker. High goes to Hunyuan3D 2.1 on our own worker",
+ "kind": "evidence:page",
+ "target": "https://model-hunyuan3d-21-rtx-93741856042.us-central1.run.app/health",
+ "ok": true,
+ "detail": "live page shows \"\"model\":\"hunyuan3d-2.1\"\""
+ },
+ {
+ "claim": "paints a full PBR material set (colour, roughness and normal maps) rather than one baked colour map",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"Hunyuan3D 2.1 runs a shape DiT for geometry, then a separate PBR paint pass\""
+ },
+ {
+ "claim": "paints a full PBR material set (colour, roughness and normal maps) rather than one baked colour map",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"Hunyuan3D 2.0 gives you a mesh with a single baked diffuse texture.\""
+ },
+ {
+ "claim": "a prompt that clearly reads as a hard-surface object (a robot, a vehicle, a piece of furniture) swaps TRELLIS ahead of Hunyuan3D",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-tiers.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "a prompt that clearly reads as a hard-surface object (a robot, a vehicle, a piece of furniture) swaps TRELLIS ahead of Hunyuan3D",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-tiers.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "a prompt that clearly reads as a hard-surface object (a robot, a vehicle, a piece of furniture) swaps TRELLIS ahead of Hunyuan3D",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-tiers.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Hugging Face Spaces for photos, and NVIDIA's hosted TRELLIS for text, which keeps a text prompt answerable even when every one of our GPU workers is cold",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-tiers.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Engines that need your own API key run only when you pick them",
+ "kind": "evidence:file",
+ "target": "api/forge.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Engines that need your own API key run only when you pick them",
+ "kind": "evidence:page",
+ "target": "https://three.ws/forge",
+ "ok": true,
+ "detail": "live page shows \"your own key\""
+ },
+ {
+ "claim": "Engines that need your own API key run only when you pick them",
+ "kind": "evidence:file",
+ "target": "src/forge.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "It takes the first lane in its order that is healthy or simply unknown; if none is clean, it takes the first that is not confirmed down",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-tiers.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "It takes the first lane in its order that is healthy or simply unknown; if none is clean, it takes the first that is not confirmed down",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-tiers.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "about 12,000 at Draft, about 30,000 at Standard, and up to 200,000 at High",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/inside-the-forge",
+ "ok": true,
+ "detail": "live page shows \"Draft: about 12,000 polygons, fast.\""
+ },
+ {
+ "claim": "about 12,000 at Draft, about 30,000 at Standard, and up to 200,000 at High",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/inside-the-forge",
+ "ok": true,
+ "detail": "live page shows \"Standard (the default): about 30,000 polygons\""
+ },
+ {
+ "claim": "about 12,000 at Draft, about 30,000 at Standard, and up to 200,000 at High",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/inside-the-forge",
+ "ok": true,
+ "detail": "live page shows \"High: up to 200,000 polygons\""
+ },
+ {
+ "claim": "The poll handler marks the failed lane unhealthy for 90 seconds",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-lane-health.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "The poll handler marks the failed lane unhealthy for 90 seconds",
+ "kind": "evidence:file",
+ "target": "api/forge.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "recovers the original prompt and the stored reference picture from the job's record, resubmits them to the next lane, and binds the old job id to the new one",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-failover.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "our TRELLIS worker, then our Hunyuan3D worker, then a hosted TRELLIS lane where one is configured, then TripoSG",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-failover.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "A job gets its first lane plus up to three backups",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-failover.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "A job gets its first lane plus up to three backups",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-failover.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "TripoSG closes the chain because it returns an untextured mesh, so it runs only once every textured lane has failed or is marked down",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-failover.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "a worker outage from 18 to 28 September that took Hunyuan3D and TRELLIS down together: without it, roughly 400 photo jobs a day ended in a hard failure while the TripoSG worker sat healthy",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-failover.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "The response names the lanes that can serve a fresh retry",
+ "kind": "evidence:file",
+ "target": "api/forge.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "TRELLIS runs on an L4. Hunyuan3D 2.1 runs on an RTX PRO 6000 Blackwell, held warm with one instance, because a cold start means loading 14 GiB of weights before any work begins",
+ "kind": "evidence:page",
+ "target": "https://model-trellis-lp642k3kpa-uc.a.run.app/health",
+ "ok": true,
+ "detail": "live page shows \"\"gpu_name\":\"NVIDIA L4\"\""
+ },
+ {
+ "claim": "TRELLIS runs on an L4. Hunyuan3D 2.1 runs on an RTX PRO 6000 Blackwell, held warm with one instance, because a cold start means loading 14 GiB of weights before any work begins",
+ "kind": "evidence:page",
+ "target": "https://model-hunyuan3d-21-rtx-93741856042.us-central1.run.app/health",
+ "ok": true,
+ "detail": "live page shows \"\"gpu_name\":\"NVIDIA RTX PRO 6000 Blackwell Server Edition\"\""
+ },
+ {
+ "claim": "TRELLIS runs on an L4. Hunyuan3D 2.1 runs on an RTX PRO 6000 Blackwell, held warm with one instance, because a cold start means loading 14 GiB of weights before any work begins",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"Our RTX service runs with min instances and max instances both set to 1, held warm.\""
+ },
+ {
+ "claim": "TRELLIS runs on an L4. Hunyuan3D 2.1 runs on an RTX PRO 6000 Blackwell, held warm with one instance, because a cold start means loading 14 GiB of weights before any work begins",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"a cold start means loading 14 GiB before the first token of work\""
+ },
+ {
+ "claim": "Every worker speaks the same contract",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"every image-to-3D worker speaks the same wire contract.\""
+ },
+ {
+ "claim": "The health check reads the second field, so a worker that is up but still loading is reported degraded, not healthy",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-health.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Task state lives in object storage rather than in the instance",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"Task state lives in object storage, not in the instance.\""
+ },
+ {
+ "claim": "18 GiB staged plus 14 GiB loaded is a 32 GiB peak against the L4 tier's 32 GiB ceiling, and the container was killed on every cold start",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"So the peak is not 14 GiB and it is not 18 GiB. It is both at once, against the L4 tier's ceiling of 32 GiB:\""
+ },
+ {
+ "claim": "18 GiB staged plus 14 GiB loaded is a 32 GiB peak against the L4 tier's 32 GiB ceiling, and the container was killed on every cold start",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"Every single cold start ended the same way:\""
+ },
+ {
+ "claim": "The RTX PRO 6000 tier's floor of 80 GiB clears it",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"The platform minimums for nvidia-rtx-pro-6000 on Cloud Run are 20 CPU and 80 GiB, which clears a 32 GiB peak with room to spare\""
+ },
+ {
+ "claim": "Blackwell needs its GPU code built with NVIDIA's 12.8 toolkit. The worker image compiles its extensions for both GPU generations, so which card runs it is a deploy flag rather than a rebuild",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"Blackwell is sm_120 and needs cu128 wheels; no flag substitutes for the rebuild.\""
+ },
+ {
+ "claim": "Blackwell needs its GPU code built with NVIDIA's 12.8 toolkit. The worker image compiles its extensions for both GPU generations, so which card runs it is a deploy flag rather than a rebuild",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"We compile the custom CUDA extensions for both architectures, not just the one we deploy to:\""
+ },
+ {
+ "claim": "Blackwell needs its GPU code built with NVIDIA's 12.8 toolkit. The worker image compiles its extensions for both GPU generations, so which card runs it is a deploy flag rather than a rebuild",
+ "kind": "evidence:page",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "live page shows \"That turns \"which card is available in this region today\" from a rebuild into a deploy flag\""
+ },
+ {
+ "claim": "Hugging Face Spaces block for the whole generation and NVIDIA's hosted lane takes text only, so neither can pick up a job that failed on a poll; both appear only in the retry suggestions",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-failover.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Hugging Face Spaces block for the whole generation and NVIDIA's hosted lane takes text only, so neither can pick up a job that failed on a poll; both appear only in the retry suggestions",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-failover.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "a job that started from several photos continues from one of them, and says so",
+ "kind": "evidence:file",
+ "target": "api/forge.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Each check waits 4 seconds",
+ "kind": "evidence:file",
+ "target": "api/_lib/forge-health.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "kind": "probe:api",
+ "target": "the self-hosted TRELLIS lane reports healthy",
+ "ok": true,
+ "detail": "HTTP 200, expectations met"
+ },
+ {
+ "kind": "probe:api",
+ "target": "the Hunyuan3D worker is loaded and ready on its Blackwell GPU",
+ "ok": true,
+ "detail": "HTTP 200, expectations met"
+ },
+ {
+ "kind": "probe:api",
+ "target": "the Forge page is served",
+ "ok": true,
+ "detail": "HTTP 200, expectations met"
+ },
+ {
+ "kind": "link",
+ "target": "https://three.ws/forge",
+ "ok": true,
+ "detail": "HTTP 200"
+ },
+ {
+ "kind": "link",
+ "target": "https://three.ws/api/forge?health=1",
+ "ok": true,
+ "detail": "HTTP 200"
+ },
+ {
+ "kind": "link",
+ "target": "https://three.ws/blog/image-to-3d-on-nvidia-l4-and-blackwell",
+ "ok": true,
+ "detail": "HTTP 200"
+ }
+ ]
+ },
+ "editor": {
+ "verdict": "publish",
+ "scores": {
+ "accuracy": 5,
+ "clarity": 5,
+ "specificity": 5,
+ "voice": 5,
+ "professionalism": 5,
+ "visual": 5
+ },
+ "issues": [],
+ "rewrite": {
+ "posts": [
+ "A text prompt on the Forge runs through two chains: one paints a reference picture, the other rebuilds it as a textured 3D model on our own NVIDIA GPUs on Cloud Run. When a worker fails mid-job, the job moves to the next engine under the same id. The write-up covers the lanes, the rescue order and the limits: three.ws/forge"
+ ],
+ "altText": [
+ "A 3D model generated on the Forge from a text prompt, shown in its viewer on a dark background: a Greek hoplite in a polished bronze cuirass and crested helmet, with a red tunic, bronze greaves, sandals and a round shield held at his side"
+ ]
+ },
+ "summary": "Every factual claim in the head post and article body is backed by a passing verification result against live pages, worker health endpoints, or source files, and the link resolves. The voice contract is fully respected (mechanism lead, no hashtags, emoji, dashes or hype) and the cover image matches its alt text and shows real Forge output, so this ships unchanged.",
+ "model": "nvidia:moonshotai/kimi-k3",
+ "fallbacks": [
+ "Vertex 403: {\n \"error\": {\n \"code\": 403,\n \"message\": \"Request had insufficient authentication scopes.\",\n \"status\": \"PERMISSION_DENIED\",\n \"details\": [\n {\n \"@type\": \"type.googleapis.com/goog",
+ "openrouter 402: {\"error\":{\"message\":\"Insufficient credits. Add more using https://openrouter.ai/settings/credits\",\"code\":402,\"metadata\":{\"limit_source\":\"openrouter_credits\",\"remedy_hint\":\"Add credits at https://openr",
+ "openai 429: {\n \"error\": {\n \"message\": \"Your account is not active, please check your billing details on our website.\",\n \"type\": \"billing_not_active\",\n \"param\": null,\n \"code\": \"billi"
+ ],
+ "rewriteFindings": [
+ "claims: claim \"when an engine fails partway through a job, the job moves to the next engine under the same job id and the progress bar keeps going\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"a Draft or Standard model costs nothing and needs no account\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The realism of the final mesh is set almost entirely by the realism of that one reference image\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"painted as one centered subject on a plain seamless background, under soft even light, with real-world materials\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"A Gemini image model on Vertex AI leads, because it draws the most photoreal reference\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"falls through to FLUX on NVIDIA's hosted inference, then to further providers, ending in a keyless rung, so no single provider's failure is ever the terminal error\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Photos skip this step: your photo is the reference. A sketch skips it too\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Draft and Standard go to TRELLIS on our own worker. High goes to Hunyuan3D 2.1 on our own worker\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"paints a full PBR material set (colour, roughness and normal maps) rather than one baked colour map\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"a prompt that clearly reads as a hard-surface object (a robot, a vehicle, a piece of furniture) swaps TRELLIS ahead of Hunyuan3D\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Hugging Face Spaces for photos, and NVIDIA's hosted TRELLIS for text, which keeps a text prompt answerable even when every one of our GPU workers is cold\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Engines that need your own API key run only when you pick them\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"It takes the first lane in its order that is healthy or simply unknown; if none is clean, it takes the first that is not confirmed down\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"about 12,000 at Draft, about 30,000 at Standard, and up to 200,000 at High\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The poll handler marks the failed lane unhealthy for 90 seconds\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"recovers the original prompt and the stored reference picture from the job's record, resubmits them to the next lane, and binds the old job id to the new one\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"our TRELLIS worker, then our Hunyuan3D worker, then a hosted TRELLIS lane where one is configured, then TripoSG\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"A job gets its first lane plus up to three backups\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"TripoSG closes the chain because it returns an untextured mesh, so it runs only once every textured lane has failed or is marked down\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"a worker outage from 18 to 28 September that took Hunyuan3D and TRELLIS down together: without it, roughly 400 photo jobs a day ended in a hard failure while the TripoSG worker sat healthy\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The response names the lanes that can serve a fresh retry\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"TRELLIS runs on an L4. Hunyuan3D 2.1 runs on an RTX PRO 6000 Blackwell, held warm with one instance, because a cold start means loading 14 GiB of weights before any work begins\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Every worker speaks the same contract\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The health check reads the second field, so a worker that is up but still loading is reported degraded, not healthy\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Task state lives in object storage rather than in the instance\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"18 GiB staged plus 14 GiB loaded is a 32 GiB peak against the L4 tier's 32 GiB ceiling, and the container was killed on every cold start\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The RTX PRO 6000 tier's floor of 80 GiB clears it\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Blackwell needs its GPU code built with NVIDIA's 12.8 toolkit. The worker image compiles its extensions for both GPU generations, so which card runs it is a deploy flag rather than a rebuild\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Hugging Face Spaces block for the whole generation and NVIDIA's hosted lane takes text only, so neither can pick up a job that failed on a poll; both appear only in the retry suggestions\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"a job that started from several photos continues from one of them, and says so\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Each check waits 4 seconds\" does not appear in the copy; claims must quote the post"
+ ]
+ },
+ "editorError": null
+}
diff --git a/data/x-content/reviews/quality-gate-article.json b/data/x-content/reviews/quality-gate-article.json
new file mode 100644
index 000000000..6db7aea3c
--- /dev/null
+++ b/data/x-content/reviews/quality-gate-article.json
@@ -0,0 +1,475 @@
+{
+ "id": "quality-gate-article",
+ "contentHash": "2f98cb70150ab228b36694f99c7f2b22e3d1934a78c2abb6b6609f4bf4c66ce9",
+ "reviewedAt": "2026-10-01T19:33:53.189Z",
+ "passed": true,
+ "blockers": [],
+ "texts": [
+ "Two checks before a 3D model ships: the catalog gate, and a diff that names what broke",
+ "Our catalog turns away a character that arrives as more than 8 meshes, and you can run that exact check on your own .glb without uploading it. The write-up covers the five structural checks, how Model Diff names the joints and clips a change removed, the exit codes for CI, and where its rename detection still guesses wrong: three.ws/inspect"
+ ],
+ "lint": [],
+ "verification": {
+ "ok": true,
+ "checks": [
+ {
+ "claim": "more than 8 meshes",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "more than 8 meshes",
+ "kind": "evidence:page",
+ "target": "https://three.ws/inspect",
+ "ok": true,
+ "detail": "live page shows \"This runs the exact quality gate three.ws applies to its own catalog entries against any .glb you give it.\""
+ },
+ {
+ "claim": "you can run that exact check on your own .glb without uploading it",
+ "kind": "evidence:page",
+ "target": "https://three.ws/inspect",
+ "ok": true,
+ "detail": "live page shows \"It runs entirely in your browser: nothing is uploaded, and the verdict is instant.\""
+ },
+ {
+ "claim": "you can run that exact check on your own .glb without uploading it",
+ "kind": "evidence:file",
+ "target": "src/inspect-page.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "the five structural checks",
+ "kind": "evidence:page",
+ "target": "https://three.ws/inspect",
+ "ok": true,
+ "detail": "live page shows \"These checks measure whether a model is built well enough to publish: density, weight, texture, bounds, assembly.\""
+ },
+ {
+ "claim": "how Model Diff names the joints and clips a change removed",
+ "kind": "evidence:page",
+ "target": "https://three.ws/diff?a=/avatars/xbot.glb&b=/avatars/michelle.glb",
+ "ok": true,
+ "detail": "live page shows \"lost 2 joint(s): mixamorig:LeftEye, mixamorig:RightEye. Clips that target them will not play.\""
+ },
+ {
+ "claim": "how Model Diff names the joints and clips a change removed",
+ "kind": "evidence:page",
+ "target": "https://three.ws/diff?a=/avatars/xbot.glb&b=/avatars/michelle.glb",
+ "ok": true,
+ "detail": "live page shows \"5 animations were removed (\"agree\", \"idle\", \"run\" and 2 more).\""
+ },
+ {
+ "claim": "the exit codes for CI",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/README.md",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "where its rename detection still guesses wrong",
+ "kind": "evidence:page",
+ "target": "https://three.ws/diff?a=/avatars/xbot.glb&b=/avatars/michelle.glb",
+ "ok": true,
+ "detail": "live page shows \"headShake to SambaDance\""
+ },
+ {
+ "claim": "where its rename detection still guesses wrong",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/src/diff.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "has to clear a structural gate before anyone sees it, and the Model Inspector runs that same gate, the same module with the same thresholds",
+ "kind": "evidence:page",
+ "target": "https://three.ws/inspect",
+ "ok": true,
+ "detail": "live page shows \"Same module, same thresholds, same decision.\""
+ },
+ {
+ "claim": "has to clear a structural gate before anyone sees it, and the Model Inspector runs that same gate, the same module with the same thresholds",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-quality.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "The seeding job that fills the catalog imports it, and so does the inspector page",
+ "kind": "evidence:page",
+ "target": "https://three.ws/inspect",
+ "ok": true,
+ "detail": "live page shows \"The page imports the same module the seeding cron runs, so the rules here cannot drift from the rules in production.\""
+ },
+ {
+ "claim": "The seeding job that fills the catalog imports it, and so does the inspector page",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-quality.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "The seeding job that fills the catalog imports it, and so does the inspector page",
+ "kind": "evidence:file",
+ "target": "api/cron/forge-seed-cron.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "A threshold change lands in both places at once because there is only one place",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "the gate reads nothing but the glTF JSON chunk",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "A build check walks the real bundler graph from the page and fails the moment a Node built-in becomes reachable",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Every verdict carries a gate version, because an accept rate is only comparable within one version of the rules",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Every verdict carries a gate version, because an accept rate is only comparable within one version of the rules",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Geometry density: between 1,500 and 1,500,000 vertices",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Geometry density: between 1,500 and 1,500,000 vertices",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "File weight: between 20 KB and 80 MB",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "File weight: between 20 KB and 80 MB",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "a collapsed mesh passes every triangle count and still draws as a speck",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "a character may arrive as at most 8 meshes. A prop is exempt",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "a character may arrive as at most 8 meshes. A prop is exempt",
+ "kind": "evidence:page",
+ "target": "https://three.ws/inspect",
+ "ok": true,
+ "detail": "live page shows \"Props skip the character-only checks.\""
+ },
+ {
+ "claim": "A failed check comes with the fix, written next to the rule it explains, so the guidance cannot drift from the rule",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "The texture rule is deliberately stricter than the interactive Forge",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-mesh-gate.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "It clears the density floor of 1,500 with 1,728 vertices",
+ "kind": "evidence:page",
+ "target": "https://three.ws/inspect?url=https%3A%2F%2Fthree.ws%2Favatars%2Ffox.glb",
+ "ok": true,
+ "detail": "live page shows \"1,728 vertices\""
+ },
+ {
+ "claim": "It clears the density floor of 1,500 with 1,728 vertices",
+ "kind": "evidence:page",
+ "target": "https://three.ws/inspect?url=https%3A%2F%2Fthree.ws%2Favatars%2Ffox.glb",
+ "ok": true,
+ "detail": "live page shows \"Clears the catalog bar\""
+ },
+ {
+ "claim": "Fox is the sample labelled \"Passes, barely\"",
+ "kind": "evidence:file",
+ "target": "src/inspect-page.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "The gate is the first of two stages. On the catalog path a vision model then renders the mesh",
+ "kind": "evidence:file",
+ "target": "api/_lib/seed-quality.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "The gate is the first of two stages. On the catalog path a vision model then renders the mesh",
+ "kind": "evidence:page",
+ "target": "https://three.ws/inspect",
+ "ok": true,
+ "detail": "live page shows \"Whether it looks good is judged separately by a vision model on the catalog path.\""
+ },
+ {
+ "claim": "pairs objects across the two the way git pairs renamed files: by name first, then by a hash of the content, then by similarity",
+ "kind": "evidence:page",
+ "target": "https://three.ws/diff",
+ "ok": true,
+ "detail": "live page shows \"Objects are matched by name, then by content hash, then by similarity, the same three-tier strategy git uses for file renames.\""
+ },
+ {
+ "claim": "Vertex data is compared at a quantum of 1e-4 units, so the float drift of a re-export does not register while a genuine edit does",
+ "kind": "evidence:page",
+ "target": "https://three.ws/diff",
+ "ok": true,
+ "detail": "live page shows \"Vertex data is compared at a quantum of 1e-4 units, so the float drift a re-export or a Draco round trip introduces does not register while a genuine edit does.\""
+ },
+ {
+ "claim": "Every change set carries one severity: none, cosmetic, minor, major or breaking",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/README.md",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Every change set carries one severity: none, cosmetic, minor, major or breaking",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/src/severity.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "A texture swap is minor because everything still loads and plays",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/src/severity.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "That is the failure that renders fine and never moves",
+ "kind": "evidence:page",
+ "target": "https://three.ws/diff",
+ "ok": true,
+ "detail": "live page shows \"That is the failure that renders fine and never moves.\""
+ },
+ {
+ "claim": "The skeleton section names the two eye joints that are gone and says clips addressing them will not play. The animation section lists the five clips that no longer exist",
+ "kind": "evidence:page",
+ "target": "https://three.ws/diff?a=/avatars/xbot.glb&b=/avatars/michelle.glb",
+ "ok": true,
+ "detail": "live page shows \"lost 2 joint(s): mixamorig:LeftEye, mixamorig:RightEye. Clips that target them will not play.\""
+ },
+ {
+ "claim": "The skeleton section names the two eye joints that are gone and says clips addressing them will not play. The animation section lists the five clips that no longer exist",
+ "kind": "evidence:page",
+ "target": "https://three.ws/diff?a=/avatars/xbot.glb&b=/avatars/michelle.glb",
+ "ok": true,
+ "detail": "live page shows \"5 animations were removed (\"agree\", \"idle\", \"run\" and 2 more).\""
+ },
+ {
+ "claim": "The page, a command line tool and a free HTTP endpoint all run the same published package",
+ "kind": "evidence:page",
+ "target": "https://three.ws/diff",
+ "ok": true,
+ "detail": "live page shows \"so a verdict in CI can never disagree with what you see here.\""
+ },
+ {
+ "claim": "The page, a command line tool and a free HTTP endpoint all run the same published package",
+ "kind": "evidence:file",
+ "target": "api/3d/diff.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "The page, a command line tool and a free HTTP endpoint all run the same published package",
+ "kind": "evidence:file",
+ "target": "src/model-diff.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Zero means the diff ran and stayed below the threshold",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/README.md",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Zero means the diff ran and stayed below the threshold",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/README.md",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "caps each side at 32 MB",
+ "kind": "evidence:file",
+ "target": "api/3d/diff.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "The similarity pass is skipped when two models would need more than 40,000 comparisons, and the result says so",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/src/match.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "The similarity pass is skipped when two models would need more than 40,000 comparisons, and the result says so",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/src/match.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "Most avatars on the platform ship with compressed geometry",
+ "kind": "evidence:file",
+ "target": "CLAUDE.md",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "the command line tool needs the optional meshoptimizer package installed next to it",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/src/document.js",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "claim": "the command line tool needs the optional meshoptimizer package installed next to it",
+ "kind": "evidence:file",
+ "target": "packages/glb-diff/package.json",
+ "ok": true,
+ "detail": "file supports the claim"
+ },
+ {
+ "kind": "probe:api",
+ "target": "the free diff endpoint reports the X Bot to Michelle change as breaking",
+ "ok": true,
+ "detail": "HTTP 200, expectations met"
+ },
+ {
+ "kind": "probe:api",
+ "target": "the inspector page is served",
+ "ok": true,
+ "detail": "HTTP 200, expectations met"
+ },
+ {
+ "kind": "link",
+ "target": "https://three.ws/inspect",
+ "ok": true,
+ "detail": "HTTP 200"
+ },
+ {
+ "kind": "link",
+ "target": "https://three.ws/diff",
+ "ok": true,
+ "detail": "HTTP 200"
+ },
+ {
+ "kind": "link",
+ "target": "https://www.npmjs.com/package/@three-ws/glb-diff",
+ "ok": true,
+ "detail": "HTTP 200 via the npm registry"
+ }
+ ]
+ },
+ "editor": {
+ "verdict": "publish",
+ "scores": {
+ "accuracy": 5,
+ "clarity": 5,
+ "specificity": 5,
+ "voice": 5,
+ "professionalism": 5,
+ "visual": 5
+ },
+ "issues": [],
+ "rewrite": {
+ "posts": [
+ "Our catalog turns away a character that arrives as more than 8 meshes, and you can run that exact check on your own .glb without uploading it. The write-up covers the five structural checks, how Model Diff names the joints and clips a change removed, the exit codes for CI, and where its rename detection still guesses wrong: three.ws/inspect"
+ ],
+ "altText": [
+ "Two 3D characters overlaid in the Model Diff viewer on a black background: a translucent pink jointed mannequin, the X Bot sample highlighted as the model before the change, standing beside Michelle, a green-skinned character with leafy green hair and yellow trousers, caught mid-dance as her clip plays"
+ ]
+ },
+ "summary": "Every claim in the post maps to a passing verification against the live pages or the gate and diff source, including the 8-mesh character limit, the browser-only run, the five checks, named joints and clips, CI exit codes, and the known rename-detection weakness. The post leads on the mechanism, follows the voice contract (no hashtags, emoji, dashes, or hype), keeps the single correct link, and the cover frame plus alt text accurately describe the Model Diff viewer with X Bot beside Michelle.",
+ "model": "nvidia:moonshotai/kimi-k3",
+ "fallbacks": [
+ "Vertex 403: {\n \"error\": {\n \"code\": 403,\n \"message\": \"Request had insufficient authentication scopes.\",\n \"status\": \"PERMISSION_DENIED\",\n \"details\": [\n {\n \"@type\": \"type.googleapis.com/goog",
+ "openrouter 402: {\"error\":{\"message\":\"Insufficient credits. Add more using https://openrouter.ai/settings/credits\",\"code\":402,\"metadata\":{\"limit_source\":\"openrouter_credits\",\"remedy_hint\":\"Add credits at https://openr",
+ "openai 429: {\n \"error\": {\n \"message\": \"Your account is not active, please check your billing details on our website.\",\n \"type\": \"billing_not_active\",\n \"param\": null,\n \"code\": \"billi"
+ ],
+ "rewriteFindings": [
+ "claims: claim \"has to clear a structural gate before anyone sees it, and the Model Inspector runs that same gate, the same module with the same thresholds\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The seeding job that fills the catalog imports it, and so does the inspector page\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"A threshold change lands in both places at once because there is only one place\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"the gate reads nothing but the glTF JSON chunk\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"A build check walks the real bundler graph from the page and fails the moment a Node built-in becomes reachable\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Every verdict carries a gate version, because an accept rate is only comparable within one version of the rules\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Geometry density: between 1,500 and 1,500,000 vertices\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"File weight: between 20 KB and 80 MB\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"a collapsed mesh passes every triangle count and still draws as a speck\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"a character may arrive as at most 8 meshes. A prop is exempt\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"A failed check comes with the fix, written next to the rule it explains, so the guidance cannot drift from the rule\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The texture rule is deliberately stricter than the interactive Forge\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"It clears the density floor of 1,500 with 1,728 vertices\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Fox is the sample labelled \"Passes, barely\"\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The gate is the first of two stages. On the catalog path a vision model then renders the mesh\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"pairs objects across the two the way git pairs renamed files: by name first, then by a hash of the content, then by similarity\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Vertex data is compared at a quantum of 1e-4 units, so the float drift of a re-export does not register while a genuine edit does\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Every change set carries one severity: none, cosmetic, minor, major or breaking\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"A texture swap is minor because everything still loads and plays\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"That is the failure that renders fine and never moves\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The skeleton section names the two eye joints that are gone and says clips addressing them will not play. The animation section lists the five clips that no longer exist\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The page, a command line tool and a free HTTP endpoint all run the same published package\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Zero means the diff ran and stayed below the threshold\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"caps each side at 32 MB\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"The similarity pass is skipped when two models would need more than 40,000 comparisons, and the result says so\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"Most avatars on the platform ship with compressed geometry\" does not appear in the copy; claims must quote the post",
+ "claims: claim \"the command line tool needs the optional meshoptimizer package installed next to it\" does not appear in the copy; claims must quote the post"
+ ]
+ },
+ "editorError": null
+}
diff --git a/data/x-content/reviews/wardrobe-article.json b/data/x-content/reviews/wardrobe-article.json
index 087f49be2..1be16b426 100644
--- a/data/x-content/reviews/wardrobe-article.json
+++ b/data/x-content/reviews/wardrobe-article.json
@@ -1,10 +1,9 @@
{
"id": "wardrobe-article",
- "contentHash": "23d43cc2802248845ae62a042fb368d4ddc0588161d7e9b74b24b55fa1f5510b",
- "reviewedAt": "2026-10-01T19:29:24.184Z",
+ "contentHash": "693ba7cbae45608e4e247702ab492412af94d94f8d68ff703d44c44f9036e80a",
+ "reviewedAt": "2026-10-01T19:31:32.401Z",
"passed": false,
"blockers": [
- "spelling mixamorig: did you mean mixamo?",
"spelling canonicalizer: did you mean canonicalize, canonicalized, canonicalizes?"
],
"texts": [
@@ -376,12 +375,6 @@
"ok": true,
"detail": "HTTP 200"
},
- {
- "kind": "spelling",
- "target": "mixamorig",
- "ok": false,
- "detail": "did you mean mixamo?"
- },
{
"kind": "spelling",
"target": "canonicalizer",
@@ -400,31 +393,16 @@
"professionalism": 5,
"visual": 5
},
- "issues": [
- {
- "severity": "minor",
- "area": "voice",
- "quote": "the arm bones it actually sits on",
- "problem": "Mechanical check flags \"actually\" as filler in the article body; the sentence carries the fact without it.",
- "fix": "Change to \"from the arm bones it sits on\". Same for \"bones a body really has\" (\"bones a body has\") and \"very different proportions\" (state the tolerance or say \"different proportions\")."
- },
- {
- "severity": "minor",
- "area": "accuracy",
- "quote": "the live 3D viewer",
- "problem": "The cover alt text and body caption present static renders as garments \"turned in its live 3D viewer\"; the stills cannot show interactivity, and the cover image shows three renders on a dark background with no viewer chrome. This is defensible but the wording slightly overstates what a still depicts.",
- "fix": "In the cover alt text, drop \"in their live viewers\" or recapture with the visible viewer UI (3D button or orbit controls) so the image supports the caption."
- }
- ],
+ "issues": [],
"rewrite": {
"posts": [
"A shirt has no skeleton, so our garment forge rigs each piece while a reference body is wearing it, then removes the body and keeps the weights the cloth picked up from the joints underneath. The write-up covers the seven stages, the gates a piece has to pass, and what binding to a skeleton does not fix yet: three.ws/wardrobe"
],
"altText": [
- "Three generated 3D garments from the three.ws wardrobe catalog side by side on a dark background: a yellow rain slicker with the hood up, a black leather biker jacket with silver zippers, and a charcoal grey hooded sweatshirt with a front pocket."
+ "Three generated 3D garments from the three.ws wardrobe catalog shown side by side in their live viewers on a dark background: a yellow rain slicker with the hood up, a black leather biker jacket with silver zippers, and a charcoal grey hooded sweatshirt with a front pocket."
]
},
- "summary": "Every claim in the post and article traces to passing verification against the repo and the live wardrobe page, and the X post itself is clean: mechanism lead, one link, no banned structure. The only notes are polish: three filler words in the article body flagged by the mechanical check, and tightening the cover alt text so a still image is not described as a live viewer interaction.",
+ "summary": "Every claim in the post and its article is backed by a passing verification against the repository files or the live wardrobe page, and the cover image matches the alt text and supports the copy. The text follows the house voice exactly, leads with the mechanism, and links the surface once; ship it as is.",
"model": "nvidia:moonshotai/kimi-k3",
"fallbacks": [
"Vertex 403: {\n \"error\": {\n \"code\": 403,\n \"message\": \"Request had insufficient authentication scopes.\",\n \"status\": \"PERMISSION_DENIED\",\n \"details\": [\n {\n \"@type\": \"type.googleapis.com/goog",
diff --git a/data/x-content/stories/agents-board-article.json b/data/x-content/stories/agents-board-article.json
new file mode 100644
index 000000000..0a2da4574
--- /dev/null
+++ b/data/x-content/stories/agents-board-article.json
@@ -0,0 +1,674 @@
+{
+ "id": "agents-board-article",
+ "status": "draft",
+ "kind": "article",
+ "tier": 1,
+ "lane": "article",
+ "pattern": "longform",
+ "notBefore": "2026-10-02T00:00:00Z",
+ "covers": [
+ "/monitor",
+ "/spotlight",
+ "/api/agents/public",
+ "/api/spotlight/list"
+ ],
+ "article": {
+ "title": "What puts an agent at the top of our board, and why Spotlight trending counts days",
+ "body": "data/x-content/articles/agents-board-article.md",
+ "cover": {
+ "path": "public/x-media/agents-board-article/cover.png",
+ "alt": "Three public three.ws agents standing side by side in 3D, each captured live on its own Agent Spotlight entry page: a woman with white hair and glasses in a navy top, a woman in a black hoodie and headphones, and a man in a lightning-print sweater and jeans"
+ }
+ },
+ "scenario": {
+ "format": "landscape",
+ "steps": [
+ {
+ "goto": "https://three.ws/monitor",
+ "settle": 8000
+ },
+ {
+ "read": "fleet",
+ "selector": "#fleet-badge",
+ "match": "/\\s*([\\d,]+)"
+ },
+ {
+ "read": "active",
+ "selector": "#fleet-badge",
+ "match": "([\\d,]+)\\s*active"
+ },
+ {
+ "type": "writing",
+ "into": "Search agents",
+ "awaits": "/api/agents/public",
+ "caption": "Search the board: newest logged action first"
+ },
+ {
+ "hold": 1800,
+ "caption": "Matches keep the order of their latest action"
+ },
+ {
+ "click": "Simply Sage",
+ "awaits": "/api/agents/42534db3"
+ },
+ {
+ "expect": "creative writing guide",
+ "within": 30000
+ },
+ {
+ "wait": 6000
+ },
+ {
+ "hold": 2200,
+ "caption": "Pick a row and the agent's own 3D body loads beside it"
+ },
+ {
+ "goto": "https://three.ws/spotlight",
+ "settle": 7000
+ },
+ {
+ "read": "showcased",
+ "match": "(\\d+)\\s+agents showcased"
+ },
+ {
+ "read": "upvotes",
+ "match": "(\\d+)\\s+upvotes"
+ },
+ {
+ "goto": "https://three.ws/spotlight/8fecd65a-4ab9-484a-9714-691d4040e9e5",
+ "settle": 8000
+ },
+ {
+ "expect": "Machine translation solved meaning years ago",
+ "within": 30000
+ },
+ {
+ "wait": 4000
+ },
+ {
+ "hold": 2600,
+ "caption": "The Spotlight: a written pitch on a live agent"
+ }
+ ]
+ },
+ "probes": [
+ {
+ "type": "scenario"
+ },
+ {
+ "type": "api",
+ "name": "the live board still holds more than 2,500 agents",
+ "url": "https://three.ws/api/agents/public?sort=live&limit=1",
+ "expect": {
+ "json": {
+ "path": "total",
+ "min": 2500
+ }
+ }
+ },
+ {
+ "type": "api",
+ "name": "more than 600 agents on the board have logged an action",
+ "url": "https://three.ws/api/agents/public?sort=live&limit=1",
+ "expect": {
+ "json": {
+ "path": "active_total",
+ "min": 600
+ }
+ }
+ },
+ {
+ "type": "api",
+ "name": "the Spotlight list answers",
+ "url": "https://three.ws/api/spotlight/list?limit=1",
+ "expect": {
+ "contains": "trending_score"
+ }
+ },
+ {
+ "type": "command",
+ "name": "the trending formula, its day clock and the SQL that mirrors it",
+ "argv": [
+ "npx",
+ "vitest",
+ "run",
+ "tests/spotlight-ranking.test.js"
+ ]
+ }
+ ],
+ "posts": [
+ {
+ "text": "An agent reaches the top of the public board by doing something: the board sorts more than 2,500 agents on the newest entry in each one's action log. The article covers who can write to that log, how default-named agents with no activity are kept off, and why the Agent Spotlight's trending clock counts days instead of hours.",
+ "media": [
+ {
+ "path": "public/x-media/agents-board-article/reel.mp4",
+ "probe": {
+ "durationSec": 8.3,
+ "width": 1280,
+ "height": 720,
+ "fps": 30,
+ "videoCodec": "h264",
+ "pixFmt": "yuv420p"
+ },
+ "reel": true
+ }
+ ]
+ }
+ ],
+ "claims": [
+ {
+ "says": "more than 2,500 agents",
+ "evidence": [
+ {
+ "type": "proof",
+ "fact": "fleet",
+ "min": 2500
+ },
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "total = t?.total ?? null;"
+ }
+ ]
+ },
+ {
+ "says": "sorts more than 2,500 agents on the newest entry in each one's action log",
+ "evidence": [
+ {
+ "type": "proof",
+ "fact": "fleet",
+ "min": 2500
+ },
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "order by ac.last_action_at desc nulls last, i.created_at desc"
+ },
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "select max(created_at) as last_action_at, count(*)::int as action_count"
+ }
+ ]
+ },
+ {
+ "says": "puts them in one order: the most recent action each agent logged",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "order by ac.last_action_at desc nulls last, i.created_at desc"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/monitor",
+ "contains": "Every agent ranked by its most recent real action"
+ }
+ ]
+ },
+ {
+ "says": "Sign-up date, popularity and chat counts do not decide who leads",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "order by ac.last_action_at desc nulls last, i.created_at desc"
+ }
+ ]
+ },
+ {
+ "says": "both need no account",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/monitor",
+ "contains": "Nothing on this page needs a session"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "Public reads, no key"
+ }
+ ]
+ },
+ {
+ "says": "Each agent keeps an append-only action log",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agent-actions.js",
+ "contains": "Append-only history of every agent action."
+ }
+ ]
+ },
+ {
+ "says": "its avatar finishes loading, it speaks, a skill finishes",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/agent-protocol.js",
+ "contains": "LOAD_END: 'load-end', // model/asset loading finished"
+ },
+ {
+ "type": "file",
+ "path": "src/agent-protocol.js",
+ "contains": "SPEAK: 'speak', // agent says something"
+ },
+ {
+ "type": "file",
+ "path": "src/agent-protocol.js",
+ "contains": "SKILL_DONE: 'skill-done', // a skill finished"
+ }
+ ]
+ },
+ {
+ "says": "viewing someone else's agent adds nothing to its log",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/agent-identity.js",
+ "contains": "if (!this._owned) return; // viewing someone else's agent : backend would 403"
+ }
+ ]
+ },
+ {
+ "says": "the server would refuse it with a 403",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agent-actions.js",
+ "contains": "if (agentRow.user_id !== auth.userId) return error(res, 403, 'forbidden', 'not your agent');"
+ }
+ ]
+ },
+ {
+ "says": "The board sorts on the newest row in that log, then on creation date",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "order by ac.last_action_at desc nulls last, i.created_at desc"
+ }
+ ]
+ },
+ {
+ "says": "An agent with no logged action sorts after the agents that have one",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "ac.last_action_at desc nulls last"
+ }
+ ]
+ },
+ {
+ "says": "Chats are shown on each row, but they do not move an agent up",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/monitor",
+ "contains": "with action count, chat count"
+ },
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "order by ac.last_action_at desc nulls last, i.created_at desc"
+ }
+ ]
+ },
+ {
+ "says": "More than 600 of the agents on the board have logged at least one action",
+ "evidence": [
+ {
+ "type": "proof",
+ "fact": "active",
+ "min": 600
+ },
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "count(*) filter (where exists ("
+ }
+ ]
+ },
+ {
+ "says": "The board treats six names as placeholders (My First Agent, Agent, Avatar, My Avatar, Untitled Agent and New Agent) along with the names a selfie import generates",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "const PLACEHOLDER_NAMES = ['My First Agent', 'Agent', 'Avatar', 'My Avatar', 'Untitled Agent', 'New Agent'];"
+ },
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "and i.name not like 'Selfie avatar%'"
+ }
+ ]
+ },
+ {
+ "says": "left off the board when it also has no logged action, no chat and no on-chain identity",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "matches": "ac\\.last_action_at is not null\\s+or ch\\.chat_count > 0\\s+or i\\.erc8004_agent_id is not null\\s+or i\\.meta->'onchain' is not null"
+ }
+ ]
+ },
+ {
+ "says": "It stays in the full agent directory, where search and browsing still reach it",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "kept out of the live wall (it's still reachable via search and /agents)"
+ }
+ ]
+ },
+ {
+ "says": "A My First Agent that has done something keeps its row",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "matches": "and \\(\\s+ac\\.last_action_at is not null\\s+or ch\\.chat_count > 0"
+ }
+ ]
+ },
+ {
+ "says": "several of them sit near the top of the board right now",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/api/agents/public?sort=live&limit=24",
+ "contains": "\"name\":\"My First Agent\""
+ }
+ ]
+ },
+ {
+ "says": "a recency stamp that turns green under 15 minutes",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/monitor",
+ "contains": "a recency stamp that turns green under 15 minutes"
+ }
+ ]
+ },
+ {
+ "says": "narrows the fleet on the server, by name and description",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/monitor",
+ "contains": "Server-side search on name and description"
+ },
+ {
+ "type": "proof",
+ "responded": "/api/agents/public"
+ }
+ ]
+ },
+ {
+ "says": "loads that agent's own 3D avatar, live, in the panel beside it",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/monitor",
+ "contains": "The selected agent's live 3D avatar, rendered in the page"
+ },
+ {
+ "type": "proof",
+ "responded": "/api/agents/42534db3"
+ }
+ ]
+ },
+ {
+ "says": "The fleet panel polls every 60 seconds",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/monitor.md",
+ "contains": "| **Fleet** | `/api/agents/public?sort=live` | 60s |"
+ }
+ ]
+ },
+ {
+ "says": "cached at the edge for 30 seconds",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "'Cache-Control': 'public, max-age=30, stale-while-revalidate=120'"
+ }
+ ]
+ },
+ {
+ "says": "the rows that never logged an action fall to the bottom even when they have a chat",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/api/agents/public?sort=live&q=writing&limit=10",
+ "contains": "\"name\":\"Orion #23\",\"description\":\"A research analyst"
+ },
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "ac.last_action_at desc nulls last"
+ }
+ ]
+ },
+ {
+ "says": "An entry is a pitch attached to a live agent, never a copy of it",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "pitch attached to a live agent, never a copy of it"
+ }
+ ]
+ },
+ {
+ "says": "It stores a headline, a one-liner, an optional write-up of up to 4,000 characters, a category, up to six tags and an optional demo link",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "The entry stores a headline, a one-liner, an optional write-up, a category, up to six tags and an optional demo link"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "a write-up of up to 4,000 characters"
+ }
+ ]
+ },
+ {
+ "says": "read from the agent itself on each request, so renaming the agent or swapping its avatar updates the card",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "is read from agent_identities on every request"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "Rename your agent, swap its avatar, register it on-chain: the card updates with it"
+ }
+ ]
+ },
+ {
+ "says": "Each agent gets one entry",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "One entry per agent."
+ }
+ ]
+ },
+ {
+ "says": "carry a Curated badge, so a visitor is never told a builder said something they did not write",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "Some entries carry a Curated badge"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "so a visitor is never told a builder said something they did not write"
+ }
+ ]
+ },
+ {
+ "says": "the builder's own write-up replaces ours the moment they submit one",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "replaces the curated one and the badge goes away"
+ }
+ ]
+ },
+ {
+ "says": "an entry published minutes ago with no votes still scores 1.0 and lands on the first page",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/spotlight.md",
+ "contains": "still scores `1.0` and lands on the first page"
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/spotlight-store.js",
+ "contains": "return (Number(voteCount) + 1) / Math.pow(ageDays + 1, GRAVITY);"
+ }
+ ]
+ },
+ {
+ "says": "At this volume it would put a day-old entry at about 2% of a fresh one",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/spotlight.md",
+ "contains": "that curve puts a 24-hour-old entry at about 2% of a fresh one"
+ }
+ ]
+ },
+ {
+ "says": "a day-old entry still carries a bit under half the weight of a fresh one, a week-old entry with twenty upvotes stays above a fresh empty one, and a month-old entry finally rotates off the front page",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/spotlight.md",
+ "contains": "a day-old entry still\ncarries a bit under half the weight of a fresh one, a week-old entry with twenty\nupvotes stays above a fresh empty one, and a month-old entry finally rotates off\nthe front page"
+ },
+ {
+ "type": "file",
+ "path": "tests/spotlight-ranking.test.js",
+ "contains": "it('keeps a week-old entry with real support above a fresh empty one'"
+ },
+ {
+ "type": "file",
+ "path": "tests/spotlight-ranking.test.js",
+ "contains": "it('leaves a day-old entry within reach of a fresh one'"
+ },
+ {
+ "type": "file",
+ "path": "api/_lib/spotlight-store.js",
+ "contains": "export const GRAVITY = 1.2;"
+ }
+ ]
+ },
+ {
+ "says": "a test fails if the two ever disagree",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "tests/spotlight-ranking.test.js",
+ "contains": "it('computes the same expression in the ORDER BY as trendingScore does'"
+ }
+ ]
+ },
+ {
+ "says": "ties break on the agent's own activity and then on its id, and the same query returns the same page each time",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/_lib/spotlight-store.js",
+ "contains": "Break the tie on the agent's"
+ },
+ {
+ "type": "file",
+ "path": "tests/spotlight-ranking.test.js",
+ "contains": "it('orders totally, so the same query cannot return a different page'"
+ }
+ ]
+ },
+ {
+ "says": "Each account gets one upvote per entry, and votes are counted on read rather than kept as a running total, so a lost write cannot leave a count wrong",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "One upvote per account per entry"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "Votes are counted on read rather than stored as a running total, so a lost write can never leave a counter permanently wrong"
+ }
+ ]
+ },
+ {
+ "says": "a chat with a visitor does not count toward rank at all",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "api/agents/public.js",
+ "contains": "order by ac.last_action_at desc nulls last, i.created_at desc"
+ }
+ ]
+ },
+ {
+ "says": "It holds 15 entries, most of them curated by us, and no entry has an upvote yet",
+ "evidence": [
+ {
+ "type": "proof",
+ "fact": "showcased",
+ "equals": "15"
+ },
+ {
+ "type": "proof",
+ "fact": "upvotes",
+ "equals": "0"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/api/spotlight/list?sort=new&limit=48",
+ "contains": "\"source\":\"curated\""
+ }
+ ]
+ },
+ {
+ "says": "Both surfaces are open now without an account",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/monitor",
+ "contains": "Nothing on this page needs a session"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/spotlight",
+ "contains": "Public reads, no key"
+ }
+ ]
+ }
+ ],
+ "mentions": {}
+}
diff --git a/data/x-content/stories/google-cloud-production.json b/data/x-content/stories/google-cloud-production.json
new file mode 100644
index 000000000..5385e9c72
--- /dev/null
+++ b/data/x-content/stories/google-cloud-production.json
@@ -0,0 +1,64 @@
+{
+ "id": "google-cloud-production",
+ "status": "draft",
+ "kind": "post",
+ "tier": 2,
+ "lane": "community",
+ "pattern": "walkthrough",
+ "notBefore": "2026-10-02T00:00:00Z",
+ "covers": [
+ "/pose",
+ "/api/forge-motion",
+ "workers/model-text2motion"
+ ],
+ "scenario": {
+ "format": "landscape",
+ "steps": [
+ {
+ "goto": "https://three.ws/pose?src=/avatars/michelle.glb&title=Michelle&anim=idle",
+ "settle": 9000
+ },
+ {
+ "expect": "Playing",
+ "within": 30000
+ },
+ {
+ "hold": 1800,
+ "caption": "An avatar in the Animation Studio"
+ },
+ {
+ "type": "a person does jumping jacks",
+ "into": "Generate an animation from a text description",
+ "caption": "Describe a movement in plain words"
+ },
+ {
+ "click": "Generate",
+ "awaits": "/api/forge-motion",
+ "caption": "A motion model on our GPU worker writes it"
+ },
+ {
+ "expect": "Playing “a person does jumping jacks",
+ "within": 120000
+ },
+ {
+ "hold": 6000,
+ "caption": "The new clip plays on her own skeleton"
+ }
+ ]
+ },
+ "probes": [
+ {
+ "type": "scenario"
+ }
+ ],
+ "posts": [
+ {
+ "text": "Placeholder head."
+ },
+ {
+ "text": "Placeholder reply three.ws/pose"
+ }
+ ],
+ "claims": [],
+ "mentions": {}
+}
diff --git a/data/x-content/stories/voice-article.json b/data/x-content/stories/voice-article.json
new file mode 100644
index 000000000..88d248cef
--- /dev/null
+++ b/data/x-content/stories/voice-article.json
@@ -0,0 +1,740 @@
+{
+ "id": "voice-article",
+ "status": "draft",
+ "kind": "article",
+ "tier": 1,
+ "lane": "article",
+ "pattern": "longform",
+ "notBefore": "2026-10-02T00:00:00Z",
+ "covers": [
+ "/voice",
+ "/drive",
+ "/api/tts/catalog",
+ "/api/tts/synthesize"
+ ],
+ "article": {
+ "title": "Picking an agent's voice by ear, and the car page that handles its own commands",
+ "body": "data/x-content/articles/voice-article.md",
+ "cover": {
+ "path": "public/x-media/voice-article/cover.png",
+ "alt": "Michelle, a stylised 3D agent with puffed black hair, red headphones and round red sunglasses, standing at idle on the dark night palette of the three.ws Drive car screen"
+ }
+ },
+ "scenario": {
+ "format": "landscape",
+ "steps": [
+ {
+ "goto": "https://three.ws/voice",
+ "settle": 7000
+ },
+ {
+ "scroll": "Browse voices",
+ "ms": 900
+ },
+ {
+ "read": "voices",
+ "match": "(\\d+) voices across \\d+ providers"
+ },
+ {
+ "hold": 1600,
+ "caption": "Hundreds of voices in one grid, signed out"
+ },
+ {
+ "type": "Andrew",
+ "into": "Search by name, accent, or use case...",
+ "caption": "Search by name, accent or use case"
+ },
+ {
+ "wait": 1200
+ },
+ {
+ "hold": 1200
+ },
+ {
+ "click": "Use this voice",
+ "caption": "Pick one and press Speak"
+ },
+ {
+ "click": "Speak",
+ "awaits": "/api/tts/synthesize",
+ "caption": "Hear it before you choose it"
+ },
+ {
+ "expect": "KB ·"
+ },
+ {
+ "hold": 1600
+ },
+ {
+ "click": "Speak",
+ "awaits": "/api/tts/synthesize",
+ "caption": "The same line twice comes back from the cache"
+ },
+ {
+ "expect": "cached"
+ },
+ {
+ "hold": 1800
+ },
+ {
+ "goto": "https://three.ws/drive?avatar=13f259c7-7024-4d68-b1f0-dbbf52c06209",
+ "settle": 8000
+ },
+ {
+ "expect": "Ready"
+ },
+ {
+ "hold": 1500,
+ "caption": "Drive: the agent on a car screen"
+ },
+ {
+ "click": "Driving",
+ "caption": "Parked, the keyboard unlocks"
+ },
+ {
+ "expect": "Parked"
+ },
+ {
+ "click": "Type"
+ },
+ {
+ "type": "Day mode, please",
+ "into": "Ask your agent something"
+ },
+ {
+ "press": "Enter",
+ "caption": "A command is answered by the page itself"
+ },
+ {
+ "wait": 1000
+ },
+ {
+ "read": "light",
+ "selector": "#dr-light",
+ "match": "(Day|Night)"
+ },
+ {
+ "hold": 2600
+ }
+ ]
+ },
+ "probes": [
+ {
+ "type": "scenario"
+ },
+ {
+ "type": "api",
+ "name": "the catalog still offers more than 300 voices",
+ "url": "https://three.ws/api/tts/catalog?limit=1",
+ "expect": {
+ "json": {
+ "path": "total",
+ "min": 300
+ }
+ }
+ },
+ {
+ "type": "api",
+ "name": "a free lane still speaks for a signed-out visitor",
+ "url": "https://three.ws/api/tts/catalog?limit=1",
+ "expect": {
+ "contains": "\"anonymous\":true"
+ }
+ },
+ {
+ "type": "api",
+ "name": "the Drive page is up",
+ "url": "https://three.ws/drive",
+ "expect": {
+ "contains": "three.ws Drive"
+ }
+ },
+ {
+ "type": "command",
+ "name": "the command matcher answers the phrases and refuses what it must not swallow",
+ "argv": [
+ "npx",
+ "vitest",
+ "run",
+ "tests/drive-surface.test.js"
+ ]
+ }
+ ],
+ "posts": [
+ {
+ "text": "Some things a driver says are control, not conversation. On Drive, \"repeat that\", \"stop talking\", \"louder\" and \"night mode\" are matched on the page before anything reaches a server, while \"stop at the next charger\" still goes to the agent. The article also covers hearing more than 300 voices in the Voice Lab before you pick one.",
+ "media": [
+ {
+ "path": "public/x-media/voice-article/reel.mp4",
+ "probe": {
+ "durationSec": 22.43,
+ "width": 1280,
+ "height": 720,
+ "fps": 30,
+ "videoCodec": "h264",
+ "pixFmt": "yuv420p"
+ },
+ "reel": true
+ }
+ ]
+ }
+ ],
+ "claims": [
+ {
+ "says": "more than 300 voices",
+ "evidence": [
+ {
+ "type": "proof",
+ "fact": "voices",
+ "min": 300
+ },
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "`GET /api/tts/catalog` reports the exact `total`"
+ }
+ ]
+ },
+ {
+ "says": "you can hear any of them before you choose, without signing in",
+ "evidence": [
+ {
+ "type": "proof",
+ "responded": "/api/tts/synthesize"
+ },
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "Edge, Gemini, and NVIDIA all work **signed out**"
+ }
+ ]
+ },
+ {
+ "says": "are answered by the page itself before anything is sent to a server",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "so the page answers them itself, before\n// the network is touched"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "['repeat', ['repeat', 'repeat that'"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "['hush', ['stop', 'stop talking'"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "['louder', ['louder'"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "['night', ['night mode'"
+ }
+ ]
+ },
+ {
+ "says": "are matched on the page before anything reaches a server, while \"stop at the next charger\" still goes to the agent",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "so the page answers them itself, before\n// the network is touched"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "\"Stop\" is a command; \"stop at the next\n// charger\" is a question for the agent"
+ }
+ ]
+ },
+ {
+ "says": "three synthesis lanes that work signed out: Microsoft Edge, Google Gemini and NVIDIA Magpie",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/api/tts/catalog?limit=1",
+ "contains": "\"label\":\"Microsoft Edge\""
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/api/tts/catalog?limit=1",
+ "contains": "\"label\":\"NVIDIA Magpie\""
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/api/tts/catalog?limit=1",
+ "contains": "\"label\":\"Google Gemini\""
+ },
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "Edge, Gemini, and NVIDIA all work **signed out**"
+ }
+ ]
+ },
+ {
+ "says": "The Edge list is fetched live from Microsoft and cached for six hours, so it is not a hardcoded snapshot, and it covers more than 100 locales",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "The Edge catalog is fetched live from Microsoft and cached 6 hours per instance"
+ },
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "Nothing is a hardcoded snapshot"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/api/tts/catalog?limit=1",
+ "contains": "every Microsoft neural voice · 100+ locales"
+ }
+ ]
+ },
+ {
+ "says": "Gemini adds 30 prebuilt voices that take a written direction",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/api/tts/catalog?limit=1",
+ "contains": "30 voices · style-directable"
+ },
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "| Google Gemini | 30 prebuilt |"
+ }
+ ]
+ },
+ {
+ "says": "Magpie adds 11 personas built for the real-time avatar lane",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "| NVIDIA Magpie | 11 personas | Free | The real-time avatar lane, 9 languages. |"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/api/tts/catalog?limit=1",
+ "contains": "the avatar real-time lane"
+ }
+ ]
+ },
+ {
+ "says": "Two more lanes, OpenAI and ElevenLabs, appear in the same grid and need an account",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "The metered lanes (OpenAI, ElevenLabs), cloning, and the Voice Library return `401 unauthorized` without a session or bearer token."
+ }
+ ]
+ },
+ {
+ "says": "Search matches a voice's name, id, locale and labels, and the language filter is built from the locales actually in the list",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "Search matches name, id, locale, and every label; the language filter is built from the locales actually present."
+ }
+ ]
+ },
+ {
+ "says": "Preview plays a voice's own sample clip when it has one, and otherwise renders one short line through the synthesis endpoint",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "\"Preview\" plays a shared voice's own sample clip when it has one (free, no synthesis) and otherwise renders one short line through `/api/tts/synthesize`."
+ }
+ ]
+ },
+ {
+ "says": "Use this voice moves the pick into the playground, where you type up to 1,000 characters, set the speed and press Speak",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "\"Use this voice\" pushes the pick into the playground."
+ },
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "up to 1,000 characters of text"
+ },
+ {
+ "type": "proof",
+ "saw": "KB ·"
+ }
+ ]
+ },
+ {
+ "says": "Each clip is cached against a hash of the whole request: lane, voice, text, model, direction, speed and settings",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "Every clip is cached in R2 keyed on a hash of the full request (provider, voice, text, model, direction, speed, and settings)"
+ }
+ ]
+ },
+ {
+ "says": "Ask for the same line twice and the second answer comes back from storage with no new synthesis",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "so an identical request is served from storage with no upstream call"
+ },
+ {
+ "type": "proof",
+ "saw": "cached"
+ }
+ ]
+ },
+ {
+ "says": "In the capture below the line had already been spoken once, so Speak returned it from the cache",
+ "evidence": [
+ {
+ "type": "proof",
+ "saw": "cached"
+ }
+ ]
+ },
+ {
+ "says": "starts in Driving mode with the keyboard hidden and four controls: Hands free, Repeat, Stop and Type",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/carplay",
+ "contains": "The page starts in Driving and hides text entry"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/carplay",
+ "contains": "Four controls, no more. Hands free, Repeat, Stop, Type."
+ },
+ {
+ "type": "proof",
+ "saw": "Ready"
+ }
+ ]
+ },
+ {
+ "says": "Hold Talk and speak",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "pages/drive.html",
+ "contains": "data-placeholder=\"Hold Talk and speak.\""
+ }
+ ]
+ },
+ {
+ "says": "The agent is told to reply in one or two short sentences that sound right out loud, and never to ask the driver to read anything, tap through a list or look at the screen",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/drive/index.js",
+ "contains": "Reply in one or two short sentences that sound right out loud."
+ },
+ {
+ "type": "file",
+ "path": "src/drive/index.js",
+ "contains": "Never ask them to read anything, tap through a list, or look at the screen."
+ }
+ ]
+ },
+ {
+ "says": "The reply is clamped to two lines on screen and spoken in full",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/carplay",
+ "contains": "The reply is clamped to two lines and spoken in full"
+ }
+ ]
+ },
+ {
+ "says": "Night is the default palette, because a bright panel at speed is glare",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/carplay",
+ "contains": "Night is the default because a bright panel at speed is glare"
+ }
+ ]
+ },
+ {
+ "says": "If the browser reports a real ground speed, moving above walking pace locks the keyboard again, and the Parked switch will not unlock it while the wheels are turning",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/carplay",
+ "contains": "If the browser reports a real ground speed, moving above walking pace locks it back"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/index.js",
+ "contains": "// A live speed signal outranks the button: the driver cannot declare\n\t\t// themselves parked while the wheels are turning."
+ }
+ ]
+ },
+ {
+ "says": "microphone into speech recognition on NVIDIA Riva, with the browser recognizer as the fallback lane, transcript into the chat model, the reply into speech, and the audio into lip sync on a rigged avatar",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/carplay",
+ "contains": "The loop is the platform's existing one, unchanged"
+ },
+ {
+ "type": "file",
+ "path": "docs/carplay.md",
+ "contains": "microphone into `/api/asr` (NVIDIA Riva,\nwith the browser recognizer as the fallback lane)"
+ },
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/carplay",
+ "contains": "audio into the lipsync driver and onto a real rigged avatar"
+ }
+ ]
+ },
+ {
+ "says": "There are eight commands, each with its own set of phrasings: repeat, stop, louder, quieter, night, day, parked and driving",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": " * @returns {'repeat'|'hush'|'louder'|'quieter'|'night'|'day'|'parked'|'driving'|null}"
+ }
+ ]
+ },
+ {
+ "says": "checks each final transcript against it before the network is touched",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": " * Build the interceptor TalkController calls with every final transcript."
+ },
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "so the page answers them itself, before\n// the network is touched"
+ }
+ ]
+ },
+ {
+ "says": "gives up on anything longer than 40 characters",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "if (!phrase || phrase.length > 40) return null;"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": ".replace(/^(hey |ok |okay |please |can you |could you |would you )+/g, '')"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": ".replace(/( please| now| for me)+$/g, '')"
+ }
+ ]
+ },
+ {
+ "says": "That is how \"stop\" ends a reply while \"stop at the next charger\" still reaches the agent",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "The matcher deliberately compares the WHOLE utterance against a phrase set"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "\"Stop\" is a command; \"stop at the next\n// charger\" is a question for the agent"
+ }
+ ]
+ },
+ {
+ "says": "\"Louder\" and \"quieter\" move the volume by 20%",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/drive/index.js",
+ "contains": "louder: () => nudgeVolume(0.2),"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/index.js",
+ "contains": "quieter: () => nudgeVolume(-0.2),"
+ }
+ ]
+ },
+ {
+ "says": "A handled command never reaches the chat model",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/voice/talk-controller.js",
+ "contains": "A handled command never reaches the chat model"
+ }
+ ]
+ },
+ {
+ "says": "Typed text takes the same path as speech",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "src/voice/talk-controller.js",
+ "contains": "Force a turn from text (e.g. typed message). Same downstream path as"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/index.js",
+ "contains": "controller.say(text).catch(onTalkError);"
+ }
+ ]
+ },
+ {
+ "says": "typing \"Day mode, please\" switches the palette to day on the spot, with no request to a server",
+ "evidence": [
+ {
+ "type": "proof",
+ "fact": "light",
+ "equals": "Day"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/index.js",
+ "contains": "day: () => applyLight(true),"
+ },
+ {
+ "type": "file",
+ "path": "src/drive/commands.js",
+ "contains": "['day', ['day mode', 'light mode', 'brighten the screen']]"
+ }
+ ]
+ },
+ {
+ "says": "the provider is refusing this deployment's credentials, so a Gemini preview answers with an error that names the lanes still working",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/ops/x-story-findings-2026-09-30.md",
+ "contains": "Gemini voices fail synthesis with 502 or 503 (\"the provider rejected this deployment's credentials\")"
+ }
+ ]
+ },
+ {
+ "says": "A lane the provider refuses is then marked unavailable for a few minutes and its voices leave the picker",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "is reported `available: false` with the reason for a few minutes, its voices drop out of the picker"
+ }
+ ]
+ },
+ {
+ "says": "the grid's count moves between 327 and 357",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/ops/x-story-findings-2026-09-30.md",
+ "contains": "so the catalog moves between 357 and 327 voices"
+ }
+ ]
+ },
+ {
+ "says": "A voice picked in the Voice Lab reaches an agent once it is assigned in the agent editor",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "To give the voice to an agent, open that agent in the agent editor and assign the saved voice there."
+ }
+ ]
+ },
+ {
+ "says": "Drive answers on a head unit browser or a phone in a cradle today",
+ "evidence": [
+ {
+ "type": "page",
+ "url": "https://three.ws/docs/carplay",
+ "contains": "on a head unit browser, on a phone in a cradle"
+ }
+ ]
+ },
+ {
+ "says": "The native CarPlay version is written and waits on Apple's entitlement grant",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/carplay.md",
+ "contains": "It cannot ship until Apple grants\n`com.apple.developer.carplay-voice-based-conversation`"
+ },
+ {
+ "type": "file",
+ "path": "docs/carplay.md",
+ "contains": "Written, wired, and documented in"
+ }
+ ]
+ },
+ {
+ "says": "there is no wake word, by design, because Apple's voice category does not allow one",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/carplay.md",
+ "contains": "there is **no wake word**"
+ },
+ {
+ "type": "file",
+ "path": "docs/carplay.md",
+ "contains": "**A wake word.** Apple's category forbids it"
+ }
+ ]
+ },
+ {
+ "says": "under heavy load that wait has run long enough to time out",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/ops/x-story-findings-2026-09-30.md",
+ "contains": "timed out at 60 s twice under heavy load"
+ }
+ ]
+ },
+ {
+ "says": "Both work without an account",
+ "evidence": [
+ {
+ "type": "file",
+ "path": "docs/voice-lab.md",
+ "contains": "Browsing and the free lanes work signed out"
+ },
+ {
+ "type": "file",
+ "path": "docs/carplay.md",
+ "contains": "when nobody is signed in (401, no error shown to the user)"
+ }
+ ]
+ }
+ ],
+ "mentions": {}
+}
diff --git a/public/x-media/agents-board-article/cover.png b/public/x-media/agents-board-article/cover.png
new file mode 100644
index 000000000..1bacad5d8
Binary files /dev/null and b/public/x-media/agents-board-article/cover.png differ
diff --git a/public/x-media/agents-board-article/fleet-search.png b/public/x-media/agents-board-article/fleet-search.png
new file mode 100644
index 000000000..da45c9737
Binary files /dev/null and b/public/x-media/agents-board-article/fleet-search.png differ
diff --git a/public/x-media/agents-board-article/reel.mp4 b/public/x-media/agents-board-article/reel.mp4
new file mode 100644
index 000000000..36f4885a4
Binary files /dev/null and b/public/x-media/agents-board-article/reel.mp4 differ
diff --git a/public/x-media/agents-board-article/spotlight-entry.png b/public/x-media/agents-board-article/spotlight-entry.png
new file mode 100644
index 000000000..696fd9c21
Binary files /dev/null and b/public/x-media/agents-board-article/spotlight-entry.png differ
diff --git a/public/x-media/forge-article/cover.png b/public/x-media/forge-article/cover.png
new file mode 100644
index 000000000..023a093c9
Binary files /dev/null and b/public/x-media/forge-article/cover.png differ
diff --git a/public/x-media/forge-article/engines.png b/public/x-media/forge-article/engines.png
new file mode 100644
index 000000000..b76372f48
Binary files /dev/null and b/public/x-media/forge-article/engines.png differ
diff --git a/public/x-media/forge-article/hunyuan-detective.png b/public/x-media/forge-article/hunyuan-detective.png
new file mode 100644
index 000000000..97850afae
Binary files /dev/null and b/public/x-media/forge-article/hunyuan-detective.png differ
diff --git a/public/x-media/quality-gate-article/cover.png b/public/x-media/quality-gate-article/cover.png
new file mode 100644
index 000000000..8b600d34e
Binary files /dev/null and b/public/x-media/quality-gate-article/cover.png differ
diff --git a/public/x-media/quality-gate-article/diff-detail.png b/public/x-media/quality-gate-article/diff-detail.png
new file mode 100644
index 000000000..31e70b6b8
Binary files /dev/null and b/public/x-media/quality-gate-article/diff-detail.png differ
diff --git a/public/x-media/quality-gate-article/inspect-fox.png b/public/x-media/quality-gate-article/inspect-fox.png
new file mode 100644
index 000000000..cf312fd57
Binary files /dev/null and b/public/x-media/quality-gate-article/inspect-fox.png differ
diff --git a/public/x-media/voice-article/cover.png b/public/x-media/voice-article/cover.png
new file mode 100644
index 000000000..25d391e72
Binary files /dev/null and b/public/x-media/voice-article/cover.png differ
diff --git a/public/x-media/voice-article/drive-day-mode.png b/public/x-media/voice-article/drive-day-mode.png
new file mode 100644
index 000000000..2c6b431de
Binary files /dev/null and b/public/x-media/voice-article/drive-day-mode.png differ
diff --git a/public/x-media/voice-article/drive-reply.png b/public/x-media/voice-article/drive-reply.png
new file mode 100644
index 000000000..b93e65fdb
Binary files /dev/null and b/public/x-media/voice-article/drive-reply.png differ
diff --git a/public/x-media/voice-article/playground.png b/public/x-media/voice-article/playground.png
new file mode 100644
index 000000000..ff7312077
Binary files /dev/null and b/public/x-media/voice-article/playground.png differ
diff --git a/public/x-media/voice-article/reel.mp4 b/public/x-media/voice-article/reel.mp4
new file mode 100644
index 000000000..c92380647
Binary files /dev/null and b/public/x-media/voice-article/reel.mp4 differ
diff --git a/public/x-media/voice-article/voice-grid.png b/public/x-media/voice-article/voice-grid.png
new file mode 100644
index 000000000..c5f829694
Binary files /dev/null and b/public/x-media/voice-article/voice-grid.png differ