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.

docs(backlog): retire the four verified-shipped orders and re-measure the pack against production

nirholasc50037d79c0a
diff --git a/prompts/backlog/00-INDEX.md b/prompts/backlog/00-INDEX.md
index fb2523202..f3f69be53 100644
--- a/prompts/backlog/00-INDEX.md
+++ b/prompts/backlog/00-INDEX.md
@@ -10,25 +10,26 @@ is broken. A work order tells you how to finish it.
---
-## Verification snapshot (2026-08-01, measured, not remembered)
+## Verification snapshot (2026-09-01, measured, not remembered)
Re-measure before you trust any number below. Commands are in each work order.
| Fact | Value | How it was read |
|---|---|---|
-| Production commit | `6cc0370dc`, revision `three-ws-api-00353-tzp` | `curl -s https://three.ws/api/version` |
-| Deploy gap | **none**; prod is at `main` HEAD | compared to `git rev-parse --short HEAD` |
-| `x402_settle` | **down**, 25.9% (504/1948 paid attempts, 3h) | `curl -s https://three.ws/api/healthz` |
-| Settle reject class | `fee_runway_exhausted` 85,331 vs `broadcast_failed` 562 | same, `x402.self_facilitator.settle.fail_reasons` |
-| Solana RPC lanes | `1/3 paid lanes serving` | same, `subsystems[].name == rpc_lanes` |
-| Forge generation | 100% (43/43 finished, 6h) | same |
-| Fact-check benchmark | `ran: false` (honest empty state, never run) | `curl -s https://three.ws/api/fact-check-benchmark` |
-| `optimize?draco=1` | **fixed**, 200 with a 129,972-byte GLB | `curl "https://three.ws/api/avatar/optimize?src=https://three.ws/avatars/cesium-man.glb&draco=1"` |
+| Production commit | `ad7b54c16` (built 2026-08-28), revision `three-ws-api-00404-ph7` | `curl -s https://three.ws/api/version` |
+| Deploy gap | **107 commits**; `main` was `73c8ccbb7` | compared to `git rev-parse --short main` |
+| `x402_settle` | **down**, 5.9% (4/68 paid attempts, 3h), `cause: sponsor_floor` | `curl -s https://three.ws/api/healthz` |
+| Settle reject class | `fee_runway_exhausted` 49,513 of 50,554 failures since boot | same, `x402.self_facilitator.settle.fail_reasons` |
+| Solana RPC lanes | all 4 paid lanes cooling, `recoversIn` per lane exposed | same, `subsystems[].name == rpc_lanes` |
+| Forge generation | 98% (48/49 finished, 6h) | same |
+| Fact-check benchmark | `ran: true`, `source: "database"`, 40% (16/40), published 2026-08-10 | `curl -s https://three.ws/api/fact-check-benchmark` |
| Media CORS at the site edge | `access-control-allow-origin: *` | `curl -I -H 'Origin: https://example.org' https://three.ws/avatars/cesium-man.glb` |
+| `gcloud` auth | dead (token refresh fails non-interactively) | `gcloud run services list` |
-Two of these overturn stale tracker text: the Draco transcode failure closed with
-the 2026-08-01 rebuild, and there is no deploy gap right now, so "ships with the
-next deploy" is no longer an excuse for anything in this pack.
+Two of these overturn text elsewhere in this pack: the benchmark is live from the database
+(order 04's only remaining line closed with the 2026-08-28 deploy), and production is no
+longer at `main`, so "ships with the next deploy" is a real dependency again for anything
+committed after 2026-08-28 06:49 UTC.
---
@@ -40,11 +41,7 @@ observable health, so 01 to 03 come first.
| # | Work order | Blocked on | Owner action needed |
|---|---|---|---|
| 01 | [x402 settle: clear `fee_runway_exhausted`](01-x402-settle-runway.md) | nothing (config-only updates are pre-approved) | one `gcloud auth login` if auth is dead |
-| 02 | [Solana RPC capacity and call-shape routing](02-solana-rpc-capacity.md) | nothing | plan top-up (optional, work order routes around it) |
-| 03 | [Sponsor runway: measure, alert, self-heal](03-sponsor-runway-automation.md) | 01 lands first | ~1 SOL top-up (optional) |
-| 04 | [Publish a real fact-check benchmark run](04-fact-check-benchmark-run.md) | one env var | set `INTERNAL_API_KEY` on the service |
| 05 | [R2 bucket CORS: verify, then fix at the origin](05-r2-bucket-cors.md) | one credential | mint an R2 admin token |
-| 06 | [LLM lanes: kill the dead backstops, close the Claude gap](06-llm-lane-resilience.md) | nothing | Vertex Model Garden entitlement |
| 07 | [BNB testnet: deploy the two finished contracts](07-bnb-testnet-deploys.md) | one funded EOA | fund a throwaway testnet key |
| 08 | [OKX chat bot: move off the codespace](08-okx-chat-bot-always-on.md) | nothing | one email OTP login |
| 09 | [Telegram bots: durable hosting for both feeds](09-telegram-bots-durability.md) | nothing | none |
@@ -52,6 +49,28 @@ observable health, so 01 to 03 come first.
---
+## Retired after verification (2026-09-01)
+
+Four orders were deleted once every agent-doable line of their definition of done was
+verified on disk and in production, not from this log: 02 (Solana RPC capacity: the lane
+probe script, per-method capability routing, `recoversIn` on the ops surface, the runbook),
+03 (sponsor runway: burn measurement, the alert and its formatter tests, the `sponsor_floor`
+sensor that healthz reports today), 04 (the benchmark run: live from the database since the
+2026-08-28 deploy), and 06 (LLM lanes: transport-failure tests for every free rung, the
+metering audit, the opt-in paid mirror, corrected docs). Their files are readable in git
+history; the evidence per line is in [PROGRESS.md](PROGRESS.md) under 2026-09-01.
+
+State of what remains, measured the same day: 01 is code-complete but its outcome line is
+false (settle 5.9%, sponsor wallet under its SOL floor; capital is the owner's), and the
+dry-run reclaim plan still reports sealed wallets as reclaimable. 05 waits on one R2 admin
+token. 07 needs a new funded deployer key: the throwaway key generated on 2026-08-02 no
+longer exists on this machine, so funding the old address would strand the tokens. 08's
+worker is built and committed but has never reported a heartbeat, so it is not running
+anywhere. 09's deliverables live in a sibling repository that is not checked out in this
+workspace, so nothing here can verify them. 10's remaining external step resolved on its own
+(the upstream pull request merged 2026-08-11 and the live discovery endpoint matches it);
+its file is retired as soon as the owner clears the commit gate that its content sits behind.
+
## Shared rules (every work order obeys these)
1. **CLAUDE.md wins.** Read it before you start. The stop-and-ask gates are real:
diff --git a/prompts/backlog/02-solana-rpc-capacity.md b/prompts/backlog/02-solana-rpc-capacity.md
deleted file mode 100644
index c325cbf49..000000000
--- a/prompts/backlog/02-solana-rpc-capacity.md
+++ /dev/null
@@ -1,86 +0,0 @@
-# 02. Solana RPC: route by call shape, survive on 1 of 3 paid lanes
-
-Read [00-INDEX.md](00-INDEX.md) first.
-
-## What is wrong
-
-Healthz reports `rpc_lanes: 1/3 paid lanes serving`. All three paid plans have hit
-their caps at various points in the last two weeks (Helius `-32429 max usage
-reached`, QuickNode `-32003 daily request limit reached`, both Alchemy apps
-`429 Monthly capacity limit exceeded`, and the two Alchemy keys share one account
-quota so the second is not a spare). Production currently runs primarily on free
-lanes, which caps the achievable settle rate and produces `broadcast_failed`,
-`Blockhash not found`, and malformed-response parse errors.
-
-GCP credits cannot solve this: Blockchain Node Engine is Ethereum-only, so there
-is no GCP-hosted Solana RPC to fail over to.
-
-## The trap that makes every naive probe wrong
-
-**Free lanes are not interchangeable, and `getBalance` cannot tell them apart.**
-All the free lanes pass `getBalance`, `getLatestBlockhash`, and
-`getSignatureStatuses`. On `getTokenAccountsByOwner` filtered by `programId`, the
-call every token and USDC balance reader makes constantly:
-
-- PublicNode returns **HTTP 403 `blocked parameter: params.1.programId`**
-- Leo RPC returns **-32603**
-- Only MagicBlock and `api.mainnet-beta.solana.com` serve it
-
-That is why those two hold the top two slots today. Also: never probe with
-`getHealth`. It is unmetered and returns ok on a fully exhausted endpoint.
-
-## The work
-
-1. **Build a capability matrix probe, not a liveness probe.** Write
- `scripts/probe-rpc-lanes.mjs` that runs every lane in `SOLANA_RPC_URL` plus
- `SOLANA_RPC_FALLBACK_URLS` against the **call shapes production actually
- makes**: `getBalance`, `getLatestBlockhash`, `getSignatureStatuses`,
- `getTokenAccountsByOwner` (programId filter), `getProgramAccounts`,
- `sendTransaction` (simulate only), `logsSubscribe` where WS is configured.
- Output a table of lane by method with the exact error code on refusal. Metered
- calls only.
-
-2. **Make the router refuse to promote a lane that cannot serve the shape.** The
- failover chain in `api/_lib/solana/connection.js` currently ranks lanes without
- knowing which methods each supports, which is how a 403 on a `programId` filter
- got sized as an auth failure and parked a whole node for 30 minutes on ordinary
- traffic. Add per-method capability so a method-level refusal demotes the lane
- for that method only, with its own short cooldown, and never counts as an auth
- fault. Cover it with a test that asserts a `blocked parameter` 403 does not
- trigger the auth path.
-
-3. **Cut call volume where it is free to cut.** The ring cadence knobs already
- exist. Audit which pipelines poll on a timer and could be event-driven or
- longer-interval without user-visible change. Land the reductions behind the
- existing config so they can be tuned without a deploy.
-
-4. **Surface quota state honestly.** The exhausted paid lanes re-enter rotation on
- their own when quota resets (`QUOTA_COOLDOWN_MS` is 6h). Show, on the ops
- surface, which lanes are cooling down and when each is expected back, so the
- next reader does not re-diagnose "all lanes dead" during a normal cooldown.
-
-## Verify
-
-```sh
-node scripts/probe-rpc-lanes.mjs # the new matrix, all lanes, all shapes
-curl -s https://three.ws/api/healthz | python3 -c "import json,sys;print([s for s in json.load(sys.stdin)['subsystems']['subsystems'] if s['name']=='rpc_lanes'])"
-npm run gate
-```
-
-## Owner action (optional, the work order does not wait on it)
-
-Restoring one paid plan lifts the ceiling: free lanes cap the achievable settle
-rate at roughly 82 to 95%. Top up or upgrade Helius, QuickNode, or Alchemy. State
-the cost and the measured ceiling in your report and let the owner decide. Do not
-block any of the four tasks above on it.
-
-## Definition of done
-
-- [ ] `scripts/probe-rpc-lanes.mjs` exists, runs offline-safe against live lanes,
- and prints a lane-by-method matrix with error codes.
-- [ ] A method-level refusal no longer parks a lane as an auth failure; test
- committed.
-- [ ] Ops surface shows per-lane quota cooldown with expected recovery time.
-- [ ] `npm run gate` green, `npm run check:rules -- --paths <touched>` clean.
-- [ ] `docs/ops/` documents the capability matrix and the `getHealth` trap.
-- [ ] `data/changelog.json` entry (tag: `infra`).
diff --git a/prompts/backlog/03-sponsor-runway-automation.md b/prompts/backlog/03-sponsor-runway-automation.md
deleted file mode 100644
index 74102eb0e..000000000
--- a/prompts/backlog/03-sponsor-runway-automation.md
+++ /dev/null
@@ -1,81 +0,0 @@
-# 03. Sponsor runway: measure it, alert on it, let it self-heal
-
-Read [00-INDEX.md](00-INDEX.md) first. Run [01](01-x402-settle-runway.md) before
-this one: a widened fee budget changes the burn this work order measures.
-
-## What is wrong
-
-The x402 sponsor wallet (`WwwuGbqHrwF5RG89KhUbmRWEvjnRH9k5kVM5p7T3WwW`, the fee
-payer for platform settles) runs close to its floor. The last measurement put it
-at 0.0318 SOL against a 0.03 floor with a measured burn of 0.060 SOL/day: roughly
-half a day of runway. The number is already computed and rendered by
-`GET /api/ops/payment-outcomes` and the "Payment outcomes" panel on `/admin/ops`
-(see [docs/ops/payment-outcomes.md](../../docs/ops/payment-outcomes.md)), but
-nothing acts on it and nothing warns before it bites.
-
-The failure mode this creates is not a clean outage. Below the floor,
-`sponsorKnownBelowFloor()` makes `buildRequirements()` in
-`api/_lib/x402-paid-endpoint.js` **withdraw the Solana accept from every 402
-challenge**, so the Solana-only ring never attempts a payment and there is nothing
-to reject. Settlements collapse while rail faults stay flat, which is the exact
-opposite of what a reader expects, and the whole thing is invisible unless you
-look at the accepts:
-
-```sh
-curl -s https://three.ws/api/x402/three-intel | jq '.accepts[].network'
-# only eip155:8453 means the Solana accept has been withdrawn
-```
-
-## The work
-
-1. **Publish the runway where a human will see it before it hurts.** The panel
- exists; the alert does not. Wire a threshold alert (days of runway below N,
- default 3) into the existing alerting path, with a message that names the
- wallet, the balance, the floor, the measured burn, and the computed days. Do
- not interpolate the template into its own source text: the bridge-down alert
- shipped exactly that bug and logged its own template instead of the numbers.
-
-2. **Measure burn from data, never from memory.** Derive it from `fee_lamports`
- over successful settles across a 7-day window. Recorded burn has been quoted at
- 10x reality in past triage notes; the alert must carry its own measurement and
- the window it covers.
-
-3. **Close the invisible-failure gap.** Add the withdrawn-Solana-accept condition
- as a first-class degraded reason on the settle sensor, distinct from a rail
- fault, so the outward symptom ("settlements fell, faults flat") maps to the
- right cause without a human remembering this document.
-
-4. **Make reclaim observable.** `reclaimIdleAgentSol` writes **no ledger row on
- failure** (unlike the engine leg's `inflow_failed`), so a failing reclaim is
- invisible. Write a row on failure with the reason. The audit trail that proves
- whether the send loop ran at all is `usage_events` where `tool='economy_reclaim'`.
-
-## Funding is an owner decision, not your call
-
-At the measured burn, roughly 1 SOL covers about 16 days. Render the numbers and
-ask; do not move funds yourself. Any transfer is stop-and-ask gate 1. **Never top
-up per-agent wallets:** that strands SOL in wallets that do not pay fees and kills
-the rail. Top-ups go to the master or the sponsor only.
-
-## Verify
-
-```sh
-# Runway panel, authenticated (ops secret or admin session)
-curl -s -H "x-ops-secret: $OPS_SECRET" https://three.ws/api/ops/payment-outcomes
-
-# The accepts probe: both networks present means the Solana accept is live
-curl -s https://three.ws/api/x402/three-intel | jq '.accepts[].network'
-
-npm run gate
-```
-
-## Definition of done
-
-- [ ] A runway alert fires below the configured threshold and its message contains
- real interpolated numbers, proven by a test over the formatter.
-- [ ] Burn is computed from `fee_lamports` over a stated window and the window is
- shown next to the number.
-- [ ] The settle sensor distinguishes `sponsor_floor` from `rail` causes.
-- [ ] A failed agent reclaim writes a ledger row naming the reason.
-- [ ] `docs/ops/payment-outcomes.md` updated to describe the alert and thresholds.
-- [ ] `data/changelog.json` entry (tag: `infra`), `npm run gate` green.
diff --git a/prompts/backlog/04-fact-check-benchmark-run.md b/prompts/backlog/04-fact-check-benchmark-run.md
deleted file mode 100644
index b1fa6c9fc..000000000
--- a/prompts/backlog/04-fact-check-benchmark-run.md
+++ /dev/null
@@ -1,74 +0,0 @@
-# 04. Publish a real fact-check benchmark run
-
-Read [00-INDEX.md](00-INDEX.md) first.
-
-## Where this stands
-
-Two of the three original problems are closed and verified live:
-
-- Web search leads with Vertex-grounded Google Search. A live call to
- `POST /api/x402/fact-check` returned `supported` at 0.98 confidence with
- `vertexaisearch.cloud.google.com` grounding sources.
-- `GET /api/fact-check-benchmark` returns `ran: false` instead of serving a stale
- run, so `/fact-check` renders its honest "not yet run" state rather than a
- misleading accuracy number. Confirmed again 2026-08-01.
-
-What remains is the measurement itself. `scripts/fact-check-benchmark.mjs` targets
-the **paid** endpoint `https://three.ws/api/x402/fact-check`, which answers a small
-free allowance and then 402s, so a 40-claim run needs a bypass. Neither bypass
-currently exists: there is no token carrying the `x402:bypass` OAuth scope, and
-`INTERNAL_API_KEY` (the `x-api-key` service path in
-[api/_lib/x402/access-control.js](../../api/_lib/x402/access-control.js)) is unset
-both locally and on the `three-ws-api` Cloud Run service.
-
-## The work
-
-1. **Set the bypass.** Config-only, pre-approved:
- ```sh
- gcloud run services update three-ws-api --region us-central1 \
- --project aerial-vehicle-466722-p5 \
- --update-env-vars INTERNAL_API_KEY=<generated>
- ```
- Generate a strong value, store it where the other service secrets live, and
- pass it to the runner as `x-api-key`. `FACT_CHECK_BYPASS_TOKEN` is the
- alternative path if you prefer a scoped token over a service key; pick one and
- document which.
-
-2. **Run it.**
- ```sh
- node scripts/fact-check-benchmark.mjs
- ```
- The runner refuses to publish any run with more than 10% errored claims, so a
- degraded run cannot poison the page. If it refuses, fix the errors rather than
- loosening the threshold: an errored claim usually means an LLM lane or the RPC
- chain is failing, which is work orders [02](02-solana-rpc-capacity.md) and
- [06](06-llm-lane-resilience.md), not this one.
-
-3. **Verify the page tells the truth.** `/fact-check` must render the new run with
- its date, sample size, and methodology link. Check the empty state still works
- by reading the handler, not by deleting the run.
-
-4. **Make it repeatable.** A benchmark that can only be run by the person who set
- the key rots. Document the exact invocation in `docs/`, and consider wiring it
- as a scheduled run so the published number carries a recent date. If you
- schedule it, add the cron to `vercel.json` (which `scripts/create-gcp-scheduler.mjs`
- reads) and note the cron count change against `npm run check:claude`.
-
-## Verify
-
-```sh
-curl -s https://three.ws/api/fact-check-benchmark | python3 -m json.tool
-curl -s -o /dev/null -w '%{http_code}\n' https://three.ws/fact-check
-npm run gate
-```
-
-## Definition of done
-
-- [ ] `GET /api/fact-check-benchmark` returns `ran: true` with a real report,
- sample size, and run timestamp.
-- [ ] `/fact-check` renders that run in a browser with no console errors.
-- [ ] The bypass mechanism (which one, where the key lives, how to rotate it) is
- documented in `docs/`.
-- [ ] Errored-claim rate from the run is stated in [PROGRESS.md](PROGRESS.md).
-- [ ] `data/changelog.json` entry (tag: `feature` or `improvement`) announcing the
- published accuracy number in plain language.
diff --git a/prompts/backlog/06-llm-lane-resilience.md b/prompts/backlog/06-llm-lane-resilience.md
deleted file mode 100644
index da4fbbc25..000000000
--- a/prompts/backlog/06-llm-lane-resilience.md
+++ /dev/null
@@ -1,99 +0,0 @@
-# 06. LLM lanes: retire the dead backstops, close the Claude gap, fix the metering lie
-
-Read [00-INDEX.md](00-INDEX.md) first.
-
-## What is wrong
-
-Three separate defects wear one label ("LLM backstops are dead"). Treat them
-separately.
-
-**1. The paid backstops are unreachable.** `OPENAI_API_KEY` is set on the service
-but the account returns `billing_not_active` 429s, so every OpenAI paid backstop is
-out. The OpenRouter platform key burned its credit on paid-model routing. Traffic
-survives on the free lanes (Groq, NIM) plus the Vertex Gemini credits anchor.
-
-**2. The metering does not show it.** `llm-pricing` records OpenRouter spend as
-**$0**, so the credit burn that killed the key was invisible right up until it was
-gone. A cost lane that reports zero is worse than one that reports nothing.
-
-**3. Claude reachability is split, and the docs describe the design instead of the
-deployed state.** Verified against the live service:
-
-- `/brain` (`api/brain/chat.js`) **does** reach Claude, through the OpenRouter
- mirrors (`anthropic/claude-*`), because `buildPrimary()` falls back to
- `spec.openrouterModel` when the native key is absent.
-- Everything else (`api/chat.js`, `api/_lib/llm.js`, the embed proxy
- `api/llm/anthropic.js`) reaches Anthropic only via `api.anthropic.com` with
- `ANTHROPIC_API_KEY`, which is absent from `.env`, `.env.local`, and the
- `three-ws-api` service. Those paths get no Claude at all.
-- Vertex Claude is off **and** unentitled: `VERTEX_CLAUDE_ENABLED=0`,
- `VERTEX_CLAUDE_PRIMARY=0`, and `rawPredict` returns 404 "project does not have
- access" for every Claude id in both `global` and `us-east5`. Flipping the flag
- alone would 404 every request.
-
-## The work
-
-1. **Make the free-lane path first-class, not a degradation.** Since paid lanes
- are out, the free chain carries production. Audit its failover for the same
- defect class found elsewhere in this repo: a fallback that only catches parse
- errors is bypassed exactly when the provider fails. Prove each rung is
- reachable with a test that simulates the rung above failing at the transport
- level, not just returning bad JSON.
-
-2. **Fix the metering.** OpenRouter spend must be recorded at its real cost in
- `llm-pricing`. A lane whose cost cannot be determined should record `unknown`
- and raise, never `0`. Add a check that fails when any lane with non-zero
- traffic reports exactly zero cost over a window.
-
-3. **Prepare the Claude key rollout so it is one command.** When
- `ANTHROPIC_API_KEY` arrives:
- ```sh
- gcloud run services update three-ws-api --region us-central1 \
- --project aerial-vehicle-466722-p5 --update-env-vars ANTHROPIC_API_KEY=...
- ```
- Never `--set-env-vars`. Fresh accounts start at **Tier 1** rate limits, so keep
- Claude off the x402 ring and any high-QPS lane; chat, brain, and reflection
- backstop traffic only until the tier grows. `claude-fable-5` requires 30-day
- data retention on the org and returns 400 under zero data retention.
- `claude-mythos-5` is deliberately kept out of the `/brain` menu.
-
-4. **Correct the docs that describe Vertex Claude as the live credits-billed
- primary.** It is the design, not the deployed state. Any doc claiming Claude
- traffic bills to GCP credits today is wrong until Model Garden entitlement lands
- and `rawPredict` stops 404ing. Fix those lines in the same change.
-
-5. **Flag, do not silently ship, the OpenRouter-mirror decision.** `api/chat.js`
- and `_lib/llm.js` have no OpenRouter mirror for Claude the way `/brain` does.
- Adding one would work immediately but draws real spend on the platform key for
- ordinary agent traffic. `isPaidModel()` in `_lib/chat-models.js` is the existing
- gate for exactly this shape. Implement it behind that gate, default off, and
- state the cost in your report.
-
-## Owner actions (name them, do not wait on them)
-
-- Reactivate OpenAI billing, or fund the OpenRouter key, or accept free-lane-only
- service.
-- Accept Anthropic terms in Vertex Model Garden for project
- `aerial-vehicle-466722-p5` to bill Claude to the GCP credits. Then re-probe
- `rawPredict` and set `VERTEX_CLAUDE_ENABLED=1`; the Claude 5 ids are already in
- `VERTEX_ANTHROPIC_MODELS`.
-
-## Verify
-
-```sh
-curl -s https://three.ws/api/healthz | python3 -c "import json,sys;print(json.load(sys.stdin)['subsystems']['degraded'])"
-npm run gate
-```
-
-Probe each lane directly rather than trusting a config read: a set key is not a
-working key, which is the entire lesson of defect 1.
-
-## Definition of done
-
-- [ ] Every free-lane rung proven reachable by a transport-level failure test.
-- [ ] No lane with traffic reports exactly `$0`; a check fails if one does.
-- [ ] The key-arrival path is a single documented command with rate-limit guidance.
-- [ ] Every doc claiming Vertex Claude is live today is corrected.
-- [ ] The OpenRouter-mirror option is implemented behind `isPaidModel()`, default
- off, with the cost stated.
-- [ ] `npm run gate` green, `data/changelog.json` entry (tag: `infra`).
diff --git a/prompts/backlog/PROGRESS.md b/prompts/backlog/PROGRESS.md
index 2c958951c..625431375 100644
--- a/prompts/backlog/PROGRESS.md
+++ b/prompts/backlog/PROGRESS.md
@@ -648,3 +648,57 @@ DB row and page wiring surface together with the next deploy, no further action.
Left: one deploy (owner-gated, and gcloud auth is dead again). After it lands,
re-read the endpoint and confirm `source: "database"` with accuracy 0.55.
+
+## 2026-09-01: retirement sweep, every order re-measured against code and production
+
+Production `ad7b54c16` (2026-08-28), `main` `73c8ccbb7`. Nothing below was taken from
+earlier entries in this log; each line was re-read.
+
+Retired (deleted after verification):
+
+- 02 Solana RPC capacity: `scripts/probe-rpc-lanes.mjs` runs and prints the lane x method
+ matrix with verbatim refusals; `api/_lib/solana/connection.js` routes per-method
+ capability (`blockedMethods`) with `tests/solana-rpc-method-capability.test.js`; healthz
+ `rpc_lanes.lanes[]` exposes `recoversAt` / `recoversIn`; `docs/ops/solana-rpc-lanes.md`
+ documents the matrix and the `getHealth` trap; changelog 2026-08-02. This log never had an
+ entry for 02; the code shipped in `893c9f49d` and `801be9857`.
+- 03 Sponsor runway: `api/_lib/x402/sponsor-runway.js` (`measureSponsorBurn`,
+ `formatSponsorRunwayAlert`) with `tests/x402-sponsor-runway.test.js` and
+ `tests/x402-sponsor-runway-monitor.test.js`; `x402-settle-health.js` separates
+ `sponsor_floor` from `rail` and healthz reports `cause: sponsor_floor` today; failed
+ reclaims write `inflow_failed` ledger rows (71,998 present); `docs/ops/payment-outcomes.md`
+ "The runway alert"; changelog 2026-08-01.
+- 04 Benchmark run: `GET /api/fact-check-benchmark` answers `ran: true`,
+ `source: "database"`, 40 claims, generated 2026-08-10; `docs/fact-check.md` documents the
+ bypass; changelog 2026-08-02 (two entries). Observation for the next accuracy pass: the
+ weekly cron's lock key was touched 2026-08-31 but `latest_run` has not advanced since
+ 2026-08-10, and the live score is 40%, not the 55% recorded above.
+- 06 LLM lanes: `tests/api/llm-free-chain-reachability.test.js` injects transport failures
+ on every free rung; `scripts/audit-llm-metering.mjs` + `api/_lib/llm-metering-rule.js`
+ fail a lane that reports $0 with traffic; `docs/ops/llm-lanes.md` carries the key-arrival
+ command and the corrected Vertex Claude status; the OpenRouter Claude mirror is behind
+ `isPaidModel()` and off by default; changelog 2026-08-02. Owner rows unchanged
+ (Model Garden terms, OpenAI billing, OpenRouter funding, revoke fallback key #1).
+
+Kept, with the measured reason:
+
+- 01: `x402_settle` 5.9% with `cause: sponsor_floor`; `fee_runway_exhausted` is 98% of
+ failures since boot. The recurrence guard, fee admission, key rotation and their tests are
+ all live. Still open in code: `reclaimIdleAgentSol` in `api/_lib/economy-sweepback.js`
+ returns the dry-run plan before `recoverSolanaAgentKeypair` runs, so the plan still counts
+ sealed wallets. Owner: fund the sponsor wallet.
+- 05: `ISSUES.md` item 9 still open; `scripts/set-r2-cors.mjs` is ready; needs the R2 admin
+ token in `.env.local`.
+- 07: testnet config endpoint answers `deployed: false`; `contracts/.env` does not exist, so
+ the 2026-08-02 throwaway deployer has no key here and its on-chain balance is zero. Any
+ funding must go to a freshly generated key.
+- 08: the chat-bot worker directory is committed with its Cloud Run config and 29 tests;
+ `bot_heartbeat` has no row for it and healthz reports "no heartbeat reported yet", so the
+ host has never run. Owner: the deploy, the OTP login, an AI-provider key on the service.
+- 09: premise is a sibling repository that is not present in this workspace; unverifiable
+ from here. The 2026-08-01 entry above claims both feeds run on Cloud Run; nothing in this
+ repo can confirm or refute that.
+- 10: upstream pull request merged 2026-08-11 (state MERGED, checks green); live
+ `/discovery/resources` paging and `/supported` match what it registered. Verified shipped.
+ The file's content and this pack's index row sit behind the CLAUDE.md commit gate, so the
+ deletion waits for the owner's yes; nothing else remains.