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.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
docs(partners): write the GTC 2027 poster submissions and re-check every partner row on 2026-10-08
nirholas39118f65eb87
diff --git a/docs/nvidia-gtc-2027-poster.md b/docs/nvidia-gtc-2027-poster.mdnew file mode 100644index 000000000..095b1a64d--- /dev/null+++ b/docs/nvidia-gtc-2027-poster.md@@ -0,0 +1,150 @@+# NVIDIA GTC 2027 poster submissions++Paste-ready submissions for the [GTC 2027 Poster Call for Submissions](https://www.nvidia.com/en-us/gtc/present/call-for-submissions/)+(San Jose, 2027-03-15 to 03-18). This is the abstract work that+[nvidia-visibility-map.md](./nvidia-visibility-map.md) asked to have written before the form+opened. The form is now open.++| | |+|---|---|+| **Deadline** | **Tuesday 2026-11-10, 5 p.m. PT** |+| **Submit at** | `https://register.nvidia.com/flow/nvidia/gtc27/cfsposters/form/submittername/login` |+| **Notification** | Rolling, from mid-December 2026 |+| **Obligation if accepted** | One author attends the Poster Reception (between 2027-03-14 and 03-18) and answers questions in person |+| **Who submits** | Owner. The form is tied to the presenting author's identity |++Checked 2026-10-08. The call states the bar plainly: posters "must present original technical+or research-based work", and anything "promotional, business use cases, conceptual, or lack+technical depth will not be considered". Selection weighs proven technical outcomes, innovation+enabled by accelerated computing, and a research framework (problem, NVIDIA technology,+benchmarking method, results). Both submissions below are written to that frame: no product+pitch, no token, no pricing, every number measured on production and reproducible from a+command in this repo.++The guidelines page (`register.nvidia.com/flow/nvidia/gtc27/cfsresources/page/posters`) renders+client-side, so its field limits could not be read in advance. Each abstract is therefore given+at up to three lengths. Use the longest one the form accepts.++Submit **A** first. It is the stronger fit: original method, NVIDIA ACE technology at its center,+and a coverage result nobody else has published. **B** is a valid second submission if the form+allows more than one per author.++---++## A. Driving Audio2Face-3D onto avatars nobody authored for it++**Title (95 characters)**++> Audio2Face-3D on Arbitrary User Rigs: Direct and Derived Blendshape Mapping for Browser Avatars++**Topic:** Content Creation / Rendering (secondary: Agentic AI). **Industry:** Media and Entertainment.+**Technologies:** NVIDIA ACE Audio2Face-3D, Riva ASR, Magpie TTS, Nemotron (NIM), NVCF gRPC.++**Abstract, short (about 500 characters)**++> Audio2Face-3D emits ARKit blendshape tracks, but avatars on the open web use ARKit, VRM vowels, Oculus visemes, or no face morphs at all. We present a browser runtime that maps one A2F-3D track onto any of them: direct canonicalization of ARKit spellings, plus a derived path that reconstructs vowel and viseme expressions from ARKit weights. On 467 real user avatars, 98.0% of rigs with any morph targets were face-driven by A2F. Text to speech to 30 fps animation completes in a median 2.96 s for 6 s of speech.++**Abstract, standard (about 1,250 characters)**++> NVIDIA Audio2Face-3D produces high-quality facial animation as an ARKit-52 blendshape track, and is usually demonstrated on characters authored for it in a game engine. Web avatars are different: they arrive from text-to-3D generators, photo reconstruction, avatar creators, and arbitrary uploads, and expose ARKit morphs, VRM vowel expressions, Oculus visemes, or nothing usable. We built a browser runtime that drives any of these from one A2F-3D track. A direct path canonicalizes the many spellings of ARKit shapes; a derived path inverts a VRM/Oculus-to-ARKit component map so an expression's activation is reconstructed from the ARKit frame, and is suppressed on meshes that already have direct mouth coverage to prevent lip overshoot. Speech comes from Magpie TTS at 44.1 kHz; A2F-3D consumes a 16 kHz resample while the browser plays the original audio and samples the 30 fps track by the audio clock. Measured on production: across 467 distinct user avatar files, 98.0% of rigs carrying any morph targets were face-driven (194 direct, 6 derived), and text in to speech plus animation out completed in a median 2.96 s (2.29 to 3.90 s) for 5.9 to 7.0 s of speech. We report the mapping tables, failure modes, and coverage by avatar source.++**Abstract, long (about 300 words)**++> **Problem.** NVIDIA Audio2Face-3D (A2F-3D) converts speech into an ARKit-52 blendshape track and is typically shown driving characters authored for it in Unreal or Omniverse. On the open web the character is not authored for it: avatars are generated from a text prompt, reconstructed from a selfie, exported from third-party creators, or uploaded from anywhere. Their faces expose ARKit morphs under inconsistent names, VRM vowel expressions, Oculus visemes, or no morph targets at all.+>+> **Method.** We present a zero-install browser runtime that drives any such rig from one A2F-3D track. The voice loop is Riva ASR in, Nemotron (NIM) reasoning, Magpie TTS out, and A2F-3D animation, all over NVCF gRPC. Magpie's 44.1 kHz output is downmixed and resampled to the 16 kHz A2F-3D contract, while the browser plays the original audio and samples the 30 fps, 55-shape track against the audio clock with inter-frame interpolation. Per mesh, a direct path canonicalizes ARKit spellings; a derived path inverts a VRM/Oculus-to-ARKit component map, reconstructing each expression's activation as the normalized sum of its ARKit components. The derived path is suppressed when a mesh already has direct jaw or mouth coverage, which removes double-driven lip overshoot. Unmappable rigs degrade to amplitude lipsync rather than a frozen face.+>+> **Benchmark.** We classified 469 distinct avatar files made by real users (stratified by source, at most 60 per source; 467 readable) with the production player, and measured end-to-end latency on the production endpoint.+>+> **Results.** 98.0% of rigs that carry any morph targets are face-driven by A2F-3D (194 direct, 6 derived, 4 unmapped). Source matters: 57 of 60 selfie reconstructions and 53 of 60 studio avatars are driven directly, while third-party creator exports often ship without morphs. Text in to synchronized speech and animation out completes in a median 2.96 s for about 6 s of speech, faster than real time.++### Evidence behind every number++| Claim | Value | How it was measured | Reproduce |+|---|---|---|---|+| Rig coverage | 98.0% of morphed rigs driven (200 of 204); 42.8% of all files (200 of 467) | 469 distinct avatar files from human accounts, at most 60 per source, classified by the shipping `A2FPlayer.attach()` on each file's real morph target names, 2026-10-08 | `node scripts/a2f-rig-coverage.mjs` |+| Direct vs derived | 194 direct ARKit mouth, 6 derived from VRM/Oculus expressions | same run | same |+| End-to-end latency, text in | median 2.96 s, range 2.29 to 3.90 s, for 5.9 to 7.0 s of speech; 6 of 6 succeeded | six consecutive `POST https://three.ws/api/a2f` with a 20-word line, 2026-10-08 | `curl -X POST https://three.ws/api/a2f -H 'content-type: application/json' -d '{"text":"..."}'` |+| End-to-end latency, audio in | 1.29 to 2.19 s for 5.5 s of audio | three consecutive calls with a WAV body, 2026-09-22 | [forum post](./nvidia-forum-browser-digital-human.md#measured-against-production-2026-09-22) |+| Track shape | 30 fps, 55 blendshapes (ARKit-52 plus three tongue shapes) | every response above | response `animation.fps`, `animation.blendShapeNames` |++Coverage by source, from the same run (counts of distinct files):++| Source | Direct | Derived | Unmapped | No morphs |+|---|---|---|---|---|+| Selfie reconstruction | 57 | 0 | 0 | 3 |+| Studio | 53 | 0 | 0 | 7 |+| Direct upload | 27 | 0 | 2 | 0 |+| Fork | 26 | 2 | 0 | 32 |+| Auto-rig | 19 | 0 | 0 | 0 |+| Import | 8 | 4 | 1 | 47 |+| Upload | 4 | 0 | 1 | 55 |+| Avaturn | 0 | 0 | 0 | 59 |+| Forge (text to 3D props) | 0 | 0 | 0 | 59 |++Say plainly in the poster: forge outputs are mostly props and have no face, and the Avaturn+exports stored on three.ws carry no morph targets at all (verified on the raw files), so those+avatars use the amplitude fallback. The 98.0% figure is coverage of rigs that have a face to+drive, and the poster must label it that way.++### Poster panels++1. Problem: one A2F-3D track, five face conventions (a grid of the same line of dialogue on an ARKit, VRM, and Oculus rig).+2. Pipeline: Riva, Nemotron, Magpie, A2F-3D over NVCF gRPC; the 44.1 kHz playback versus 16 kHz inference split.+3. Mapping: the direct canonicalizer and the inverted derived map, with the overshoot suppression rule.+4. Results: the coverage table above and the latency series.+5. Lessons: function ids that rotate and fail closed; interpolating a 30 fps track for 60 and 120 Hz displays.++Source files to cite: `src/voice/a2f-player.js`, `src/voice/arkit-blendshapes.js`,+`api/_lib/a2f-nvidia.js`, `api/_lib/tts-nvidia.js`, `api/_lib/asr-nvidia.js`. Live demo for+reviewers: `https://three.ws/demos/audio2face`.++---++## B. Serving image-to-3D on NVIDIA L4 and RTX PRO 6000 Blackwell++**Title (93 characters)**++> Cost and Latency per Successful Mesh: Serving Image-to-3D on Serverless L4 and Blackwell GPUs++**Topic:** AI Infrastructure (secondary: Content Creation / Rendering). **Industry:** Media and Entertainment.+**Technologies:** NVIDIA L4, NVIDIA RTX PRO 6000 Blackwell, CUDA 12, NVIDIA Kaolin, nvdiffrast, NIM.++**Abstract, short (about 500 characters)**++> We serve open image-to-3D models (TRELLIS on L4, Hunyuan3D 2.1 on RTX PRO 6000 Blackwell) as serverless GPU workers for a public text-to-3D product. Over 31 days the fleet produced 19,270 successful meshes. We report the metric that matters to an operator: list-price GPU cost per successful output ($0.18 on L4 TRELLIS, $0.44 on Blackwell Hunyuan3D), p50/p95 job latency (124/325 s and 135/304 s), and how cross-lane failover recovered 810 failed attempts before a user saw an error.++**Abstract, standard (about 1,100 characters)**++> Open image-to-3D models are now good enough for production, but published results report quality, not what it costs to serve them to real users. We operated a fleet of serverless GPU workers on Cloud Run behind a public text-to-3D product for 31 days: Microsoft TRELLIS on NVIDIA L4 across two regions, Hunyuan3D 2.1 on NVIDIA RTX PRO 6000 Blackwell, and an auto-rigging worker on L4, with NVIDIA NIM-hosted models for reasoning and vision QA. The fleet produced 19,270 successful meshes. We define cost per successful output as list-price GPU, vCPU, and memory time from Cloud Monitoring billable instance time, divided by successful meshes on that lane, and measure $0.18 for L4 TRELLIS, $0.44 for Blackwell Hunyuan3D, and $0.30 for rigging. Job latency was p50 124 s / p95 325 s on L4 TRELLIS and p50 135 s / p95 304 s on Blackwell Hunyuan3D. Warm minimum instances dominate cost at this volume, so throughput per billable hour (7.8 and 7.3 meshes) is the lever. We also report two worker incidents, their root causes, and a cross-lane failover that recovered 810 failed attempts before a user saw a failure.++### Evidence behind every number++Every figure comes from [the 30-day proof brief](./partners/proof-brief-2026-10.md) (window+2026-08-31 to 2026-09-30), which lists the query behind each one. Regenerate it on the submission+date with `npm run partners:proof -- --from <start> --to <end>` and update any number that moved.+The latency figures for L4 TRELLIS are from the catalog seeder driving the same workers (maker+latency was only stamped from 2026-09-24); the brief's caveats section says so, and the poster+must too.++---++## Content Interest Survey: closes 2026-10-09, 5 p.m. PT++A separate, lighter form that signals topic fit before sessions are planned:+`https://forms.gle/iwGUwc5YQfw6f7X27`. Its questions are not published in advance; these answers+cover what it asks for (topics you want, and novel work you are doing with NVIDIA technology).++- **Themes:** Agentic AI; Content Creation / Rendering; AI Infrastructure.+- **Industry:** Media and Entertainment.+- **Novel work (paste):** "Driving NVIDIA Audio2Face-3D onto arbitrary user-generated avatars in a browser tab, with no install: Riva ASR, Nemotron, Magpie TTS, and A2F-3D over NVCF gRPC, with a direct-plus-derived blendshape mapping that face-drives 98% of real user rigs that carry morph targets. Also: serving TRELLIS and Hunyuan3D image-to-3D on serverless L4 and RTX PRO 6000 Blackwell, measured as cost and latency per successful mesh."+- **Sessions we want to see:** ACE and Audio2Face on the web and on non-authored rigs; serverless GPU inference economics; 3D generative model serving.++---++## Related++- [NVIDIA visibility map](./nvidia-visibility-map.md): every NVIDIA surface, and why GTC is the largest+- [Browser digital-human forum post](./nvidia-forum-browser-digital-human.md): the published technical basis for poster A+- [30-day proof brief](./partners/proof-brief-2026-10.md): the evidence behind poster B+- [Partnership pipeline](./partners/opportunities.md): where this sits among the other partner actionsdiff --git a/docs/nvidia-visibility-map.md b/docs/nvidia-visibility-map.mdindex 2da3bdcbf..1b8a70081 100644--- a/docs/nvidia-visibility-map.md+++ b/docs/nvidia-visibility-map.md@@ -151,12 +151,21 @@ these dates are NVIDIA-stated yet. Nothing to submit to this week. Keep theweekly watch running; the abstracts in step 2 below are what should be writtenwhile the window is shut.+**Re-checked 2026-10-08: the window is open.** NVIDIA now states the dates,+**2027-03-15 to 03-18, San Jose**. The [poster call](https://www.nvidia.com/en-us/gtc/present/call-for-submissions/)+closes **2026-11-10 at 5 p.m. PT** and the Content Interest Survey closes+**2026-10-09 at 5 p.m. PT**. No open speaker call is posted. Both poster+submissions and the survey answers are written in+[nvidia-gtc-2027-poster.md](./nvidia-gtc-2027-poster.md); what remains is the+owner submitting them as the presenting author.+Actions, in order:1. **Put [nvidia.com/gtc/call-for-submissions](https://www.nvidia.com/gtc/call-for-submissions/)on a weekly watch.** Missing this window costs a full year. It is the one itemon this page with a hard, externally-set deadline.-2. **Write the poster and session abstracts before the form opens.** The strongest+2. **Write the poster and session abstracts before the form opens.** Done+ 2026-10-08: [nvidia-gtc-2027-poster.md](./nvidia-gtc-2027-poster.md). The strongestsubmission we have is the browser-native digital human: Audio2Face-3D plus Rivaplus Nemotron driving a rigged avatar in a tab with no install, on a fleet ofL4s and one Blackwell. Second-strongest is the L4-and-Blackwell engineeringdiff --git a/docs/partners/opportunities.md b/docs/partners/opportunities.mdindex fd4448675..9e93e6613 100644--- a/docs/partners/opportunities.md+++ b/docs/partners/opportunities.md@@ -16,7 +16,9 @@ what unblocks it.**Last reconciled against the source docs and live program pages on 2026-09-15. Startup programs from the CloudCredits sweep were added on 2026-09-18 and carry "checked 2026-09-18". On 2026-09-16 theexternal directories, PRs, and program windows that were unverified or stale were re-checkedagainst their live sources; only rows that say "checked 2026-09-16" carry that re-check, and each-states what the live page showed and links the proof.+states what the live page showed and links the proof. On **2026-10-08** every external row was+re-checked again; rows that say "checked 2026-10-08" carry that pass, and the ones whose status+changed are called out in place.---@@ -27,11 +29,12 @@ for a cycle, not delayed. Nothing here is on anyone's calendar yet.| By when | Opportunity | The one action | Owner ||---|---|---|---|-| **Now** | GitHub Open Source Friday. [Issue #254](https://github.com/githubevents/open-source-friday/issues/254) is open with the `approved` label. Checked 2026-09-16: no date has been posted; the only comment is GitHub's 2026-09-01 request to pick a slot | Select a Friday at `https://gh.io/osf-booking`, then reply on issue #254 with the date | Owner |-| **Apply as soon as a venue is real** | [Hacktoberfest 2026 Fest](https://hacktoberfest.com/host/). Checked 2026-09-16: "Applications are now open and Fests are confirmed on a rolling basis", decisions typically within a week, and no application deadline is published. Apply at `https://hacktoberfest.com/my/` | Name the venue, city, capacity, and host, then submit either a one-day Hack Day or Meet Up application | Owner |-| **2026-09-26** | IBM TechXchange early-bird registration. [IBM's FAQ](https://www.ibm.com/events/techxchange/faq), checked 2026-09-16: early bird is USD 1279.20 from 13 July to **26 September** 2026 (not 29 September as previously recorded), then USD 1599. The event runs 2026-10-26 to 10-29 at the Georgia World Congress Center, Atlanta; the speaking CFP closed in May, but sponsorship, the Sandbox Expo and the colocated user-group route are open, and we moderate an IBM Community user group | Register, or decide not to attend | Owner |-| **Not open as of 2026-09-16** | IBM Champions nomination. Checked 2026-09-16: the [programme page](https://www.ibm.com/community/champions-program/) shows no nomination window and no class-of-2027 call has been announced. The previous cycle [opened 2025-09-15](https://community.ibm.com/community/user/blogs/libby-ingrassia/2025/09/15/ibm-champions-nominations) and closed 2025-11-21. The page now also points to an IBM Rising Champions Advocacy badge route marked "new details coming soon", so the intake may change. Our record already matches all four of IBM's published criteria (spoke at an IBM event, moderate a user group, published on their community, built on Granite). Caveat checked 2026-09-16: production has no `WATSONX_*` credentials and `/api/ibm/galaxy` answers `watsonx_not_configured`, so Granite is not running in production today; restore the credentials before any nomination says it is. Draft: [ibm-champions-nomination.md](../../marketing/partner-packets/ibm-champions-nomination.md). We have never entered | Check the programme page weekly; when the call opens, nominate a person (the program is for individuals, not companies) | Owner |-| **Watch weekly** | NVIDIA GTC 2027 call for speaking/poster submissions. This is separate from the Startup Pavilion, which NVIDIA confirmed is invite-only | Keep the CFP page on a weekly check; maintain the demo and profile for Pavilion consideration | Either |+| **2026-10-09, 5 p.m. PT** | NVIDIA GTC 2027 Content Interest Survey ([form](https://forms.gle/iwGUwc5YQfw6f7X27)). Checked 2026-10-08: open since 2026-09-16, closes tomorrow. Answers are written | Paste the answers from [nvidia-gtc-2027-poster.md](../nvidia-gtc-2027-poster.md#content-interest-survey-closes-2026-10-09-5-pm-pt) | Owner |+| **2026-11-10, 5 p.m. PT** | NVIDIA GTC 2027 poster call (San Jose, 2027-03-15 to 03-18). Checked 2026-10-08: [open](https://www.nvidia.com/en-us/gtc/present/call-for-submissions/), original technical work only, nothing promotional. Two submissions are written with measured production evidence, including a new rig-coverage benchmark (`node scripts/a2f-rig-coverage.mjs`). No speaker call is posted; the Startup Pavilion stays invite-only | Submit poster A from [nvidia-gtc-2027-poster.md](../nvidia-gtc-2027-poster.md) as the presenting author (acceptance means attending the poster reception) | Owner |+| **Now** | GitHub Open Source Friday. [Issue #254](https://github.com/githubevents/open-source-friday/issues/254) is open with the `approved` label. Checked 2026-10-08: still no date posted, and no activity since GitHub's 2026-09-01 request to pick a slot | Select a Friday at `https://gh.io/osf-booking`, then reply on issue #254 with the date | Owner |+| **Before 2026-10-31** | [Hacktoberfest 2026 Fest](https://hacktoberfest.com/host/). Checked 2026-10-08: still "Applications are now open and Fests are confirmed on a rolling basis" with no published deadline, but Hacktoberfest ends 2026-10-31 and decisions take about a week, so a Fest application is only useful in roughly the next two weeks. Apply at `https://hacktoberfest.com/my/` | Name the venue, city, capacity, and host, then submit either a one-day Hack Day or Meet Up application | Owner |+| **2026-10-26** | IBM TechXchange attendance. [IBM's FAQ](https://www.ibm.com/events/techxchange/faq), checked 2026-10-08: the early bird (USD 1279.20) **closed 2026-09-26 unused**; the standard pass is USD 1599, the hotel block closed 2026-10-01, and full refunds ended 2026-09-22. The event runs 2026-10-26 to 10-29 at the Georgia World Congress Center, Atlanta; the speaking CFP closed in May, but sponsorship, the Sandbox Expo and the colocated user-group route are open, and we moderate an IBM Community user group | Register, or decide not to attend | Owner |+| **Not open as of 2026-10-08** | IBM Champions nomination. Checked 2026-10-08: the [programme page](https://www.ibm.com/community/champions-program/) still shows no nomination window and no class-of-2027 call, which is now three weeks later than last year's opening. The previous cycle [opened 2025-09-15](https://community.ibm.com/community/user/blogs/libby-ingrassia/2025/09/15/ibm-champions-nominations) and closed 2025-11-21. The page now also points to an IBM Rising Champions Advocacy badge route marked "new details coming soon", so the intake may change. Our record already matches all four of IBM's published criteria (spoke at an IBM event, moderate a user group, published on their community, built on Granite). Caveat checked 2026-09-16 and again 2026-10-08 (not on `three-ws-api`, not in Secret Manager, not in either env file): production has no `WATSONX_*` credentials and `/api/ibm/galaxy` answers `watsonx_not_configured`, so Granite is not running in production today; restore the credentials before any nomination says it is. Draft: [ibm-champions-nomination.md](../../marketing/partner-packets/ibm-champions-nomination.md). We have never entered | Check the programme page weekly; when the call opens, nominate a person (the program is for individuals, not companies) | Owner || Spring 2027 | IBM TechXchange 2027 call for speakers | Submit when it opens | Owner |Dates and intake routes for the IBM rows are verified in@@ -50,7 +53,7 @@ work already done and is waiting on a single human step.| 1 | **OKX.AI agent #2632** | Submit the on-chain `agent update`, then activate | A built listing is sitting rejected over a payment bug **that was fixed and shipped on 2026-09-05**. OKX's own `agent x402-check` now reads `valid: true` on all four paid rows. Nothing is being built here, a finished listing is simply not live. || 2 | **IBM Agent Connect catalog listing** | Email `[email protected]` for the BYOL `APP_ID` | Everything else is built and verified ([submission pack](../../marketing/ibm-partner-plus/agent-connect-listing.md)). One email opens an IBM-backed enterprise channel. || 3 | **AWS Marketplace** | Create the product in the AWS Marketplace Management Portal | The SaaS integration is built, deployed, and conformant. The listing was simply never created, so an enterprise procurement channel we already paid the engineering cost for returns nothing. |-| 4 | **NVIDIA Accelerated Apps Catalog** | Complete the existing Shipping product record exactly as NVIDIA requested | NVIDIA answered the request: the portal record is the periodically reviewed application. Copy, technical evidence, logo, brand color, and screenshots are ready; only the authenticated save remains. On 2026-09-16 three.ws was absent from all 1,846 public catalog records. |+| 4 | **NVIDIA Accelerated Apps Catalog** | Complete the existing Shipping product record exactly as NVIDIA requested | NVIDIA answered the request: the portal record is the periodically reviewed application. Copy, technical evidence, logo, brand color, and screenshots are ready; only the authenticated save remains. On 2026-09-16 three.ws was absent from all 1,846 public catalog records. Checked 2026-10-08: the catalog URL now redirects to `marketplace.nvidia.com/en-us/enterprise/applications/`, which did not load for an automated check, and a web search finds no three.ws listing. || 5 | **NVIDIA NGC Catalog** | One container build plus the partner legal agreement | The only NVIDIA directory with self-serve intake. The EULA prerequisite is already closed. || 6 | **OpenAI Plugin Directory** | Submit | We already meet the gating requirement (public OAuth 2.1 MCP server), and the output-quality blocker that held this back is closed: the forge quality gate measured **10/10 verdicts** on production 2026-09-11, up from 0/10 two days earlier. It was closed by fixing the vision failover chain, not by clearing the GCP billing hold. || 7 | **Second IBM community event** | Pick a date | A fully costed proposal ([ibm-next-event.md](../ibm-next-event.md)) has been waiting on nothing but a calendar decision. Cheapest unblock on this page. |@@ -71,7 +74,7 @@ No action required. Listed so nobody re-pitches something we already have.| Alibaba Cloud Marketplace | live; product page, storefront, and blog feature all render three.ws (checked 2026-09-16) | [listings.md](../listings.md) || BNB Chain Dappbay | live | [listings.md](../listings.md) || Hugging Face org account | live; one Space, one model, two articles, latest article 2026-08-29 (checked 2026-09-16) | [huggingface.md](../huggingface.md) |-| Official MCP Registry | live; 72 servers under `io.github.nirholas/*`, including `io.github.nirholas/three.ws` 1.0.1 (published 2026-06-12) (checked 2026-09-16) | [registry search](https://registry.modelcontextprotocol.io/v0/servers?search=three.ws) |+| Official MCP Registry | live; 72 servers under `io.github.nirholas/*`, including `io.github.nirholas/three.ws` 1.0.1 (published 2026-06-12) (checked 2026-09-16). The staged batch has since landed: `herald-mcp`, `knock-mcp`, and `home-mcp` 0.1.0 all resolve (checked 2026-10-08) | [registry search](https://registry.modelcontextprotocol.io/v0/servers?search=three.ws) || PulseMCP | live, ingested from the official registry, 215 visitors that week (checked 2026-09-16) | [pulsemcp.com/servers/nirholas-three-ws](https://www.pulsemcp.com/servers/nirholas-three-ws) || Glama MCP directory | live as connectors ingested from the official registry, but **unclaimed** (checked 2026-09-16); see the claim row below | [three.ws connector](https://glama.ai/mcp/connectors/io.github.nirholas/three.ws) || LobeHub MCP marketplace | live; `@three-ws/portal`, `@three-ws/avatar-mcp`, `@three-ws/x402-mcp` and other `nirholas-*` servers are indexed (checked 2026-09-16). The chat plugin is not; see Open, unclaimed | [@three-ws/portal](https://lobehub.com/mcp/nirholas-three-ws-portal) |@@ -88,7 +91,7 @@ The work is done or nearly done. A human on our side has to act.|---|---|---|---|| OKX.AI agent #2632 | Rejected 2026-09-03, cause fixed 2026-09-05 | On-chain `agent update`, then activate. The stored listing copy is stale, so it is update-then-activate, not activate alone | Owner || IBM Agent Connect (BYOL MCP listing) | Fully prepared | `APP_ID` request email, then Concierge | Owner |-| AWS Marketplace | Integration code deployed, listing never created. Checked 2026-09-16: `three-ws-api` carries **no `AWS_MP_*` variable** (none in `.env`/`.env.local` either), so `POST /api/aws-marketplace/subscription` answers `503 not_configured` and a buyer arriving from a new listing would hit the error page | From seller account `155407237916`, create the metering IAM key and SNS/EventBridge secret listed in [aws-marketplace.md](../aws-marketplace.md), set them with `gcloud run services update --update-env-vars`, then create the product in the portal | Owner |+| AWS Marketplace | Integration code deployed, listing never created. Checked 2026-09-16 and again 2026-10-08: `three-ws-api` carries **no `AWS_MP_*` variable** (none in `.env`/`.env.local` either), so `POST /api/aws-marketplace/subscription` answers `503 not_configured` and a buyer arriving from a new listing would hit the error page | From seller account `155407237916`, create the metering IAM key and SNS/EventBridge secret listed in [aws-marketplace.md](../aws-marketplace.md), set them with `gcloud run services update --update-env-vars`, then create the product in the portal | Owner || NVIDIA NGC Catalog | Prerequisites cleared | One build plus partner legal agreement | Owner + eng || NVIDIA Accelerated Apps Catalog | NVIDIA answered 2026-09-14; complete Shipping records are periodically reviewed, and the current record still understates the runtime stack | Complete the technology fields, Product Description, Technical Details, logo, and brand color in the portal; then check the public catalog monthly | Owner || OpenAI Plugin Directory | Eligible | Submit | Owner |@@ -97,9 +100,9 @@ The work is done or nearly done. A human on our side has to act.| Second IBM community event | Proposal written and costed | Pick a date | Owner || IBM G2 review | Draft ready in [ibm-g2-review.md](../../marketing/partner-packets/ibm-g2-review.md), written in the past tense because production has had no `WATSONX_*` credentials since at least 2026-07-29 (checked 2026-09-16) | The actual watsonx.ai user rewrites it in their own words with their own console screenshot, or restores the credentials first | Owner || CoinGecko market addition for $THREE | Form researched; only one of our live venues qualifies | Submit for that venue | Owner |-| HackerNoon author page | Checked 2026-09-16: [hackernoon.com/u/three-ws](https://hackernoon.com/u/three-ws) is live but shows **no published stories** ("eagerly awaiting @three-ws's next masterpiece"). RSS import lands in the drafts queue; nothing has gone out from it | Open the HackerNoon drafts queue and submit the imported announcements for editorial review | Owner |+| HackerNoon author page | Checked 2026-10-08: [hackernoon.com/u/three-ws](https://hackernoon.com/u/three-ws) still shows **no published stories**. RSS import lands in the drafts queue; nothing has gone out from it. **New 2026-10-08:** the profile carries stray data that is not ours: the job title, the "Read My Stories" button, and the "Portfolio" link all describe or point at someone else's profile. The bio itself is correct | First clean the profile (title, stories link, portfolio link). Then open the drafts queue and submit the imported announcements for editorial review | Owner || DEXTools, reciprocal 3D embed | Our side is shipped: DEXTools charts are already a provider in every three.ws chart switcher, and as of 2026-09-20 `/coin3d` is framable and keyed by pair address, which is the identifier a DEXTools pair page already holds. As of 2026-10-08 proposal pillars 3 and 4 are built too: DEXTools leads every terminal row, launches get a Social Boost card, and every link we send is counted at `/api/coin/dextools-stats` ([surface map](../listings.md#dextools)). The relationship has public receipts in both directions: three $THREE Social Boost wins with DEXTools buybacks behind them | Send [dextools-proposal.md](./dextools-proposal.md) to a DEXTools partnerships or front-end contact. Nothing else is blocked; the embed is live and free whether or not they answer | Owner |-| Glama connector claim | Checked 2026-09-16: at least seven three.ws connectors are listed. The main `three.ws` connector is unclaimed and reports **Unhealthy** because Glama has no OAuth test credentials for `https://three.ws/api/mcp` | Claim ownership on the connector page (maintainer email `[email protected]` is already declared in `public/.well-known/glama.json`), then add a test profile under Admin | Owner |+| Glama connector claim | Checked 2026-10-08: still unclaimed and **Unhealthy, 0.0% uptime over 51 days**, last tested 2026-10-07; Glama's own note blames missing test credentials. That is correct behaviour on our side, not a bug: `/api/mcp` deliberately answers an MCP client's `initialize` with the 401 that starts OAuth, so no unauthenticated health check can pass. Glama now also offers a domain claim through `https://three.ws/.well-known/glama.json` (already served) or a `_glama-claim.three.ws` DNS record | Claim ownership on the connector page (maintainer email `[email protected]` is already declared in `public/.well-known/glama.json`), then add a test profile under Admin | Owner |### Anthropic, which this page was missing@@ -112,11 +115,10 @@ on it, and the cheapest to move.| Opportunity | State | Next action | Owner ||---|---|---|---|-| Official MCP Registry, remaining batch | 72 servers are already live (see Live and working). Checked 2026-09-16: the three new servers in the staged batch (`herald-mcp`, `knock-mcp`, `home-mcp`) are still absent from the registry | `mcp-publisher login github`, review the staged batch, run it | Owner || Claude plugin marketplace | Ready, install evidence captured | Submit | Owner || Claude Connectors Directory | Ready. The old blocker (auth-gated discovery) is resolved | Org role plus a reviewer-credential decision, then submit | Owner |-| Smithery, mcp.so | Checked 2026-09-16: **not listed**. [Smithery's registry API](https://registry.smithery.ai/servers?q=three.ws) and its rendered search return no three.ws or `nirholas` server; [mcp.so search](https://mcp.so/search?q=three.ws) shows "0 results" for both `three.ws` and `nirholas`. Listing content is generated and committed. (PulseMCP, previously grouped here, is already live.) | Create each account and submit | Owner |-| punkpeye/awesome-mcp-servers, AxiomeCG/awesome-threejs | Checked 2026-09-16: **absent** from both READMEs, and neither repo has a PR or issue mentioning three.ws. Copies are ready in `marketing/growth/submissions/` | Open one PR per list from the prepared copy | Owner |+| Smithery, mcp.so | Checked 2026-09-16 and again 2026-10-08: **not listed**. [Smithery's registry API](https://registry.smithery.ai/servers?q=three.ws) and its rendered search return no three.ws or `nirholas` server; [mcp.so search](https://mcp.so/search?q=three.ws) shows "0 results" for both `three.ws` and `nirholas`. Listing content is generated and committed. (PulseMCP, previously grouped here, is already live.) | Create each account and submit | Owner |+| punkpeye/awesome-mcp-servers, AxiomeCG/awesome-threejs | Checked 2026-09-16 and again 2026-10-08: **absent** from both READMEs, and neither repo has a PR or issue mentioning three.ws. Copies are ready in `marketing/growth/submissions/` | Open one PR per list from the prepared copy | Owner |## In flight, blocked on them@@ -124,7 +126,7 @@ Nothing for us to do but keep the thread warm. Do not re-scope these into our ow| Opportunity | State | Note ||---|---|---|-| OpenAI Cookbook PR #2874 | Open since 2026-07-21. Checked 2026-09-16: still open and unmerged, review required, 24 comments, last activity our own comment on 2026-09-14 | Revival steps are in [openai-listing-channels.md](../openai-listing-channels.md) |+| OpenAI Cookbook PR #2874 | Open since 2026-07-21. Checked 2026-10-08: still open and unmerged, review required, 24 comments, no activity since our own comment on 2026-09-14 | Revival steps are in [openai-listing-channels.md](../openai-listing-channels.md) || Dedicated three.ws page on the IBM domain | Promised 2026-06-18, not live | **Do not link it or describe it as public until IBM ships it** || IBM co-promotion on X | Promised 2026-06-18, no recorded delivery | Reopen alongside the Agent Connect conversation |@@ -157,10 +159,10 @@ Real opportunities with no owner and no motion. Each needs a decision before it| Opportunity | Why it is plausible | What it needs first ||---|---|---|| Google Cloud Marketplace | Production already runs entirely on Google Cloud Run and Vertex AI. The technical story is finished before the conversation starts | A decision to pursue co-listing and joint GTM |-| Microsoft Azure Marketplace | Third leg of the enterprise procurement story after AWS and Alibaba. Checked 2026-09-18: [Microsoft for Startups](https://www.microsoft.com/en-us/startups) is open (up to USD 150,000 in Azure credits plus customer-network access) and is the entry route | Sequencing decision. Do not start before the AWS listing exists |-| LobeHub chat plugin index | `public/.well-known/lobehub-plugin.json` (identifier `3d-agent`) is shipped, dated 2026-04-17. Checked 2026-09-16: the [plugin index](https://chat-plugins.lobehub.com/index.json) holds 40 plugins and none is three.ws. Our MCP servers are already listed separately | A decision: submit the plugin to the LobeHub plugin index, or retire the manifest in favour of the MCP listings that already exist |-| PostHog for Startups | PostHog is the production analytics stack (`src/analytics.js`). Checked 2026-09-18: USD 50,000 in credits for 12 months plus a startup spotlight | Apply from the production org; confirm the age and funding rules on the form |-| OpenAI Codex for Open Source | A new program inside an existing partner: six months of ChatGPT Pro with Codex and up to USD 25,000 in API credits for maintainers of widely used public projects. Checked 2026-09-18: open | Submit the maintainer application |+| Microsoft Azure Marketplace | Third leg of the enterprise procurement story after AWS and Alibaba. Checked 2026-10-08: [Microsoft for Startups](https://www.microsoft.com/en-us/startups) is still open (up to USD 150,000 in Azure credits plus customer-network access) and is the entry route | Sequencing decision. Do not start before the AWS listing exists |+| LobeHub chat plugin index | `public/.well-known/lobehub-plugin.json` (identifier `3d-agent`) is shipped, dated 2026-04-17. Checked 2026-10-08: the [plugin index](https://chat-plugins.lobehub.com/index.json) still holds 40 plugins and none is three.ws. Our MCP servers are already listed separately | A decision: submit the plugin to the LobeHub plugin index, or retire the manifest in favour of the MCP listings that already exist |+| PostHog for Startups | PostHog is the production analytics stack (`src/analytics.js`). Checked 2026-10-08: still open, USD 50,000 in credits for 12 months plus a startup spotlight | Apply from the production org; confirm the age and funding rules on the form |+| OpenAI Codex for Open Source | A new program inside an existing partner: six months of ChatGPT Pro with Codex and up to USD 25,000 in API credits for maintainers of widely used public projects. Checked 2026-10-08: open, reviewed on a rolling basis | Submit the maintainer application || Cerebras startup deal | Cerebras is a live free-tier rung in the LLM chain and the deal includes co-marketing. Checked 2026-09-18: the public non-YC page returns 404 | Ask Cerebras whether the non-YC deal still exists || Deepgram Startup Program | Up to USD 100,000 in speech credits and customer stories. Not in the stack: `/api/asr` runs on NVIDIA Riva | A decision to add a second ASR rung, which onboards a new external API || A DeFi protocol partnership | A five-pillar proposal was drafted 2026-07-08 and lives in this directory. It is a point-in-time pitch with figures that are now over a year stale | Decide whether to revive. If yes, the numbers need refreshing before it is sent |diff --git a/scripts/a2f-rig-coverage.mjs b/scripts/a2f-rig-coverage.mjsnew file mode 100644index 000000000..90f6e436b--- /dev/null+++ b/scripts/a2f-rig-coverage.mjs@@ -0,0 +1,218 @@+#!/usr/bin/env node+// Audio2Face-3D rig coverage across the avatars people actually brought.+// ---------------------------------------------------------------------------+// A2F-3D emits an ARKit-52 track (plus tongue). Our avatars come from the+// studio, the forge, selfie reconstruction, Avaturn, and arbitrary uploads, so+// the face on the other end may expose ARKit morphs, VRM vowels, Oculus+// visemes, or nothing. src/voice/a2f-player.js decides per mesh whether to+// drive morphs DIRECTLY (the name canonicalizes to ARKit) or to DERIVE a vowel+// or viseme expression from the ARKit frame, and falls back to amplitude+// lipsync when neither applies.+//+// This script measures how often each path fires on real avatars, using the+// shipping A2FPlayer itself: it reads each GLB's morph target names (the JSON+// chunk only, by HTTP range request), hands them to A2FPlayer.attach() as the+// morph dictionaries three.js would build, and classifies the result. It is the+// evidence behind the coverage figure in docs/nvidia-gtc-2027-poster.md.+//+// Usage:+// node scripts/a2f-rig-coverage.mjs sample up to 60 distinct files per source+// node scripts/a2f-rig-coverage.mjs --per-source 200+// node scripts/a2f-rig-coverage.mjs --json machine-readable report+//+// Read-only. Reads DATABASE_URL and S3_PUBLIC_DOMAIN from the environment, then+// .env.local, then .env. Seed-cron accounts, QA logins, and synthetic test+// registrations are excluded with the same predicate as partners:proof.++import { readFileSync } from 'node:fs';+import { fileURLToPath } from 'node:url';+import path from 'node:path';+import { neon } from '@neondatabase/serverless';+import { A2FPlayer } from '../src/voice/a2f-player.js';+import { ARKIT_GROUPS, canonicalARKitName } from '../src/voice/arkit-blendshapes.js';++const REPO_ROOT = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..');+for (const envFile of ['.env.local', '.env']) {+ try {+ const raw = readFileSync(path.resolve(REPO_ROOT, envFile), 'utf8');+ for (const line of raw.split('\n')) {+ const m = line.match(/^\s*([A-Z0-9_]+)\s*=\s*(.*)\s*$/);+ if (!m || process.env[m[1]]) continue;+ process.env[m[1]] = m[2].replace(/^(['"])(.*)\1$/, '$2');+ }+ } catch {+ // Optional: the shell environment may already carry the vars.+ }+}++const args = process.argv.slice(2);+const flag = (name) => args.includes(`--${name}`);+const opt = (name, fallback) => {+ const i = args.indexOf(`--${name}`);+ return i >= 0 && args[i + 1] ? args[i + 1] : fallback;+};+const PER_SOURCE = Number(opt('per-source', 60));+const CONCURRENCY = 8;+// Production's public R2 domain, used only when the env does not name one.+const PUBLIC_DOMAIN = process.env.S3_PUBLIC_DOMAIN || 'https://pub-2534e921bf9c4314addcd4d8a6e98b7b.r2.dev';++if (!process.env.DATABASE_URL) {+ console.error('missing DATABASE_URL (looked in the environment, .env.local, .env)');+ process.exit(1);+}+const sql = neon(process.env.DATABASE_URL);++const HUMAN_USER = `not u.service_account+ and u.email::text !~* '(@qa\\.three\\.ws$|\\.invalid$|^(qa|gdqa|e2e|test|probe|audit|home-e2e)[-_.0-9])'`;++const MOUTH = new Set([...ARKIT_GROUPS.jaw, ...ARKIT_GROUPS.mouth]);++function urlFor(key) {+ if (/^https?:\/\//i.test(key)) return key;+ return `${PUBLIC_DOMAIN}/${key.split('/').map(encodeURIComponent).join('/')}`;+}++// The glTF JSON of a .glb or .vrm, fetching as few bytes as the server allows.+async function fetchGltfJson(url) {+ const head = await fetch(url, { headers: { range: 'bytes=0-19' } });+ if (!head.ok) throw new Error(`HTTP ${head.status}`);+ let buf = Buffer.from(await head.arrayBuffer());+ if (buf.length < 20 || buf.readUInt32LE(0) !== 0x46546c67) {+ // Not binary glTF (or the server ignored the range and sent a .gltf).+ const full = head.status === 200 ? buf : Buffer.from(await (await fetch(url)).arrayBuffer());+ if (full.readUInt32LE(0) !== 0x46546c67) return JSON.parse(full.toString('utf8'));+ buf = full;+ }+ const jsonLen = buf.readUInt32LE(12);+ if (buf.length < 20 + jsonLen) {+ const res = await fetch(url, { headers: { range: `bytes=20-${19 + jsonLen}` } });+ if (!res.ok) throw new Error(`HTTP ${res.status}`);+ const body = Buffer.from(await res.arrayBuffer());+ // A server that ignores Range answers 200 with the whole file.+ return JSON.parse((res.status === 206 ? body : body.subarray(20, 20 + jsonLen)).toString('utf8'));+ }+ return JSON.parse(buf.subarray(20, 20 + jsonLen).toString('utf8'));+}++// The object graph A2FPlayer.attach() walks, built the way three.js's+// GLTFLoader names morph targets: mesh.extras.targetNames, else the index.+function sceneFromGltf(gltf) {+ const meshes = [];+ for (const [mi, mesh] of (gltf.meshes || []).entries()) {+ const targets = mesh.primitives?.[0]?.targets;+ if (!targets?.length) continue;+ const names = mesh.extras?.targetNames || [];+ const dict = {};+ targets.forEach((_, i) => {+ dict[names[i] ?? String(i)] = i;+ });+ meshes.push({+ isMesh: true,+ name: mesh.name || `mesh${mi}`,+ morphTargetDictionary: dict,+ morphTargetInfluences: new Array(targets.length).fill(0),+ });+ }+ return { meshes, root: { traverse: (fn) => meshes.forEach(fn) } };+}++function classify(gltf) {+ const { meshes, root } = sceneFromGltf(gltf);+ if (!meshes.length) return { path: 'no-morphs', mouthShapes: 0 };+ const player = new A2FPlayer();+ player.attach(root);+ const d = player.describe();+ const mouth = new Set(+ d.direct.map((s) => s.slice(s.lastIndexOf(':') + 1)).filter((n) => MOUTH.has(n)),+ );+ if (mouth.size) return { path: 'direct', mouthShapes: mouth.size };+ if (d.expression.length) return { path: 'derived', mouthShapes: 0 };+ // Morphs exist but none the player can drive: report what they were called,+ // which is how the next mapping table gets written.+ const sample = meshes.flatMap((m) => Object.keys(m.morphTargetDictionary)).filter((n) => !canonicalARKitName(n));+ return { path: 'unmapped-morphs', mouthShapes: 0, sample: sample.slice(0, 6) };+}++async function pool(items, n, fn) {+ const out = new Array(items.length);+ let next = 0;+ await Promise.all(+ Array.from({ length: n }, async () => {+ while (next < items.length) {+ const i = next++;+ out[i] = await fn(items[i]);+ }+ }),+ );+ return out;+}++// One row per distinct file: stock avatars imported by many people count once,+// and a source contributes at most PER_SOURCE files so no single source decides+// the headline number.+const rows = await sql.query(+ `select source, storage_key, owners from (+ select a.source, a.storage_key, count(distinct a.owner_id)::int as owners,+ row_number() over (partition by a.source order by md5(a.storage_key)) as rn+ from avatars a join users u on u.id = a.owner_id+ where a.deleted_at is null and ${HUMAN_USER}+ and (a.storage_key ilike '%.glb' or a.storage_key ilike '%.vrm' or a.content_type ilike '%gltf%')+ group by a.source, a.storage_key+ ) t where rn <= $1 order by source, storage_key`,+ [PER_SOURCE],+);++const results = await pool(rows, CONCURRENCY, async (r) => {+ try {+ return { ...r, ...classify(await fetchGltfJson(urlFor(r.storage_key))) };+ } catch (err) {+ return { ...r, path: 'unreadable', error: String(err.message || err).slice(0, 80) };+ }+});++const PATHS = ['direct', 'derived', 'unmapped-morphs', 'no-morphs', 'unreadable'];+const bySource = {};+for (const r of results) {+ bySource[r.source] ??= Object.fromEntries(PATHS.map((p) => [p, 0]));+ bySource[r.source][r.path]++;+}+const totals = Object.fromEntries(PATHS.map((p) => [p, results.filter((r) => r.path === p).length]));+const readable = results.length - totals.unreadable;+const faceDriven = totals.direct + totals.derived;+const unmappedNames = {};+for (const r of results) for (const n of r.sample || []) unmappedNames[n] = (unmappedNames[n] || 0) + 1;++const report = {+ measured_at: new Date().toISOString(),+ per_source_cap: PER_SOURCE,+ files: results.length,+ readable,+ totals,+ a2f_face_driven: faceDriven,+ a2f_face_driven_ratio: readable ? faceDriven / readable : 0,+ morphed_files: totals.direct + totals.derived + totals['unmapped-morphs'],+ by_source: bySource,+ top_unmapped_morph_names: Object.entries(unmappedNames)+ .sort((a, b) => b[1] - a[1])+ .slice(0, 15),+};++if (flag('json')) {+ console.log(JSON.stringify(report, null, 2));+} else {+ const pct = (n, d) => (d ? `${((100 * n) / d).toFixed(1)}%` : 'n/a');+ console.log(`A2F rig coverage, ${report.measured_at}`);+ console.log(`${results.length} distinct avatar files (cap ${PER_SOURCE} per source), ${readable} readable\n`);+ console.log(['source'.padEnd(14), ...PATHS.map((p) => p.padStart(16))].join(''));+ for (const [src, c] of Object.entries(bySource)) {+ console.log([src.padEnd(14), ...PATHS.map((p) => String(c[p]).padStart(16))].join(''));+ }+ console.log(['total'.padEnd(14), ...PATHS.map((p) => String(totals[p]).padStart(16))].join(''));+ console.log(`\nface driven by A2F: ${faceDriven} of ${readable} readable (${pct(faceDriven, readable)})`);+ console.log(` direct ARKit mouth: ${totals.direct} derived from VRM/Oculus expressions: ${totals.derived}`);+ console.log(`of files that carry any morph targets: ${pct(faceDriven, report.morphed_files)} driven`);+ if (report.top_unmapped_morph_names.length) {+ console.log('\nmost common morph names the player cannot map yet:');+ for (const [n, c] of report.top_unmapped_morph_names) console.log(` ${String(c).padStart(4)} ${n}`);+ }+}diff --git a/scripts/audit-docs.mjs b/scripts/audit-docs.mjsindex b3c57f7d9..cf0be4dea 100644--- a/scripts/audit-docs.mjs+++ b/scripts/audit-docs.mjs@@ -313,6 +313,7 @@ const UNPUBLISHED_DOCS = new Map([['threejs-forum-post', 'internal: forum post drafted for the three.js forum'],['nvidia-forum-gpu-fleet-post', 'internal: forum post drafted for the NVIDIA developer forums'],['nvidia-forum-model-retirement-post', 'internal: forum post drafted for the NVIDIA developer forums'],+ ['nvidia-gtc-2027-poster', 'internal: poster submissions drafted for the NVIDIA GTC 2027 call'],['huggingface-agent-feedback-loop', 'internal: article drafted for the Hugging Face community blog'],['devto-mcp-fleet-post', 'internal: article drafted for dev.to'],['google-cloud-community-post', 'internal: article drafted for the Google Cloud community'],