CortexLM · echobt · Sep 3, 2026 · Sep 3, 2026 · Sep 3, 2026
@@ -41,13 +41,13 @@ eval_image = "ghcr.io/cortexlm/relearn-eval"
 # deploy/env/relearn-challenge.env.example). Check GET /v1/status ->
 # can_score, live_harvest_wired, champion_baseline_recorded.
 #
-# EMPTY ON PURPOSE. Every digest built so far fails the harvest, and an empty
-# digest is the fail-closed state: submissions answer 503 "eval image digest
-# not pinned" instead of renting a B200 that cannot score. Do not guess a
-# digest here to make the 503 go away.
+# Proven CUDA scoring image from CortexLM/relearn publish-eval-image on
+# 8ffbe8a0481cc62c1e2d6ed80a2211befa98dd3f (run 33730126042). The runtime job
+# pulled this digest, ran the harvest-PATH check, and proved
+# `import vllm, torchvision` plus `relearn-eval selftest` on these bytes.
+# Not the slim contract image. A live host still needs harvest + champion.
 #
-# The next pin needs a CortexLM/relearn image whose CUDA scoring layer ships
-# **vLLM and torchvision**. Do not re-pin or re-harvest any of these:
+# Do not re-pin or re-harvest any of these:
 #
 #   sha256:cbc4bbb8… (f3cfa69) exit 1, no EVAL_OK: no vLLM on the rented GPU
 #                     image, transformers fallback, then Qwen3VLVideoProcessor
@@ -56,11 +56,10 @@ eval_image = "ghcr.io/cortexlm/relearn-eval"
 #   sha256:303c6357… printed no RELEARN_EVAL_OK
 #   sha256:00839671… exit 127, no /usr/bin/relearn-eval
 #   sha256:86240d76… pre-CUDA
-eval_image_digest = ""
+eval_image_digest = "sha256:4806db4bd6650415c3705290117fd0190cd2a6d387cef871dd3d9ff1a8bb6d7e"
 relearn_git = "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/CortexLM/relearn"
-# CortexLM/relearn commit that built the digest above. Empty while unpinned:
-# f3cfa69 built cbc4bbb8, which never printed RELEARN_EVAL_OK.
-relearn_git_sha = ""
+# CortexLM/relearn commit that built the digest above.
+relearn_git_sha = "8ffbe8a0481cc62c1e2d6ed80a2211befa98dd3f"
 
 # Public split ids (miners may train on these). Holdout selection must exclude
 # every one of them. The scored holdout records stay off git.

@@ -75,36 +75,34 @@ fn pin_says_its_commitment_is_not_the_live_seal() {
     );
 }
 
-/// Every digest built so far failed the live harvest, so the committed pin is
-/// empty and live scoring is refused. An empty digest is a 503 on submit — the
-/// fail-closed state — not a fallback to the sim harness.
+/// The committed pin is the CUDA scoring digest publish-eval-image proved
+/// on CortexLM/relearn@8ffbe8a (run 33730126042): harvest-PATH, import
+/// vllm/torchvision, and selftest against the pulled bytes. Not the slim
+/// contract image, and not any digest that already failed a live harvest.
 #[test]
-fn committed_pin_is_fail_closed_until_a_working_image_ships() {
+fn committed_pin_is_the_proven_runtime_eval_image() {
     let p = pin();
-    assert!(
-        p.eval_image_digest.is_empty(),
-        "no CortexLM/relearn image has printed RELEARN_EVAL_OK on a rented pod; \
-         an empty digest 503s, a guessed one rents a B200 that cannot score"
+    assert_eq!(
+        p.eval_image_digest,
+        "sha256:4806db4bd6650415c3705290117fd0190cd2a6d387cef871dd3d9ff1a8bb6d7e"
     );
-    assert!(!p.can_rent(), "an unpinned host must not rent");
-    assert!(
-        p.relearn_git_sha.is_empty(),
-        "no digest, no commit that built it"
+    assert!(p.can_rent(), "the proven CUDA scoring digest is rentable");
+    assert_eq!(
+        p.relearn_git_sha,
+        "8ffbe8a0481cc62c1e2d6ed80a2211befa98dd3f"
     );
 }
 
-/// What the committed pin actually does to a live host, not just what it says.
-/// The digest is the root cause, so it has to be the reported one.
+/// Pinning the image is not the whole job: a live host still needs a wired
+/// harvest. Sim stays opt-in.
 #[test]
-fn the_committed_pin_refuses_live_scoring_and_says_why() {
+fn the_committed_pin_still_needs_a_wired_harvest() {
     let err = relearn_eval::scoring_readiness(&pin(), relearn_eval::EvalBackend::Lium, None)
-        .expect_err("an unpinned host must not be ready to score");
+        .expect_err("a pinned host without a harvest must not be ready to score");
     assert!(
-        matches!(err, relearn_eval::EvalError::EvalImageUnpinned),
+        matches!(err, relearn_eval::EvalError::LiveHarvestUnavailable),
         "{err}"
     );
-    assert!(err.to_string().contains("eval image digest not pinned"));
-    // Sim is still opt-in only, so this is a 503 rather than a sim verdict.
     assert!(relearn_eval::scoring_readiness(&pin(), relearn_eval::EvalBackend::Sim, None).is_ok());
     assert!(!relearn_eval::force_sim(), "sim is never implicit");
 }

@@ -56,7 +56,7 @@ Removed as **live products**. Shared rails (`prism-lium*`, `prism-competition` p
 | Crates (`crates/relearn-*`) | **done** | task (holdout commitment + contamination fingerprints), score (public–holdout gap, contamination evidence, vision shuffle, off-path general-bench canary), store, eval, http, challenge. |
 | Binary (`bins/relearn-challenge`) | **done** | HTTP API on `:8095`. |
 | Compose / images | **done** | Default compose + `images.yml` target `relearn-challenge`. |
-| Eval pin | **blocked** | `config/relearn-pin.toml` — `eval_image_digest` empty on purpose, so submissions 503 `eval image digest not pinned`. Needs a `CortexLM/relearn` image whose CUDA scoring layer ships **vLLM + torchvision**. Every digest tried so far failed on a rented pod and is named in the pin: `cbc4bbb8…` (exit 1, transformers fallback then `Qwen3VLVideoProcessor` wanted torchvision), `201cc5d2…`, `303c6357…`, `00839671…`, `86240d76…`. Do not re-pin or re-harvest any of them, and do not guess a digest. |
+| Eval pin | **done** | `config/relearn-pin.toml` — `eval_image` `ghcr.io/cortexlm/relearn-eval`, digest `sha256:4806db4b…`, `relearn_git_sha` `8ffbe8a0…` ([`CortexLM/relearn`](https://github.com/CortexLM/relearn) `8ffbe8a`, [publish-eval-image run 33730126042](https://github.com/CortexLM/relearn/actions/runs/33730126042)). CUDA scoring image: harvest-PATH + `import vllm, torchvision` + selftest proved on the pulled digest. Not the slim contract image. Do not re-pin `cbc4bbb8…` / `201cc5d2…` / `303c6357…` / `00839671…` / `86240d76…`. A live host still 503s until harvest + champion baseline are recorded. |
 | Holdout | **done** | Commitment in git, records operator-side (`RELEARN_HOLDOUT_FILE`) and verified at boot. Mismatch → submissions 503. |
 | Teacher | **v0** | Weights `incoai/GLM-5.3-NVFP4` served from `RELEARN_TEACHER_LOCAL_DIR` (never pass the Hugging Face repo id to vLLM). HTTP wire `glm-5.3`. Missing `RELEARN_TEACHER_API_URL` → `can_score: false` and 503 before rent. Judge-only. |
 | Emission | **4000 bps** | Default share (sum across all four challenges is `10000`). |
@@ -175,17 +175,17 @@ Agent/operator contracts: root [`AGENTS.md`](../AGENTS.md), [`deploy/AGENTS.md`]
 |-----|--------|
 | DCAP verify holds the attest mutex | A cold Intel PCS fetch (up to 20 s) serialises attestation submissions. |
 | DCAP error classification | Matches on `anyhow` message text; re-run `cargo test -p attest-policy --features dcap` after any `dcap-qvl` bump. |
-| Relearn eval image digests | `relearn-image-eval` `sha256:81c40dc6…` and `relearn-agent-eval` `sha256:4db52b13…` are pinned (PR #3, `54d3537`). **`relearn-eval` is not pinned**: the last candidate, `sha256:cbc4bbb8…` (`f3cfa69`), exited 1 on a rented pod without `RELEARN_EVAL_OK` — no vLLM on the CUDA scoring image, transformers fallback, then `Qwen3VLVideoProcessor` crashed for want of torchvision. `relearn` submissions 503 until [`CortexLM/relearn`](https://github.com/CortexLM/relearn) ships an image with vLLM + torchvision. A live host still 503s after that until harvest + champion baseline are recorded. |
+| Relearn eval image digests | `relearn-eval` `sha256:4806db4b…` (`8ffbe8a`, [run 33730126042](https://github.com/CortexLM/relearn/actions/runs/33730126042)), `relearn-image-eval` `sha256:81c40dc6…` and `relearn-agent-eval` `sha256:4db52b13…` (PR #3, `54d3537`). Digest-only, no floating tag. Do not re-pin `cbc4bbb8…`. A live host still 503s until harvest + champion baseline are recorded. |
 | Relearn Image / Agent holdout salts | Both committed commitments use documented **dev** salts so local and staging work out of the box. Production must rotate to a private salt **and** a private catalogue for each, replace the commitments, and re-sign. |
 | Relearn holdout salt | The committed `holdout_commitment` is the CI / local one — a documented dev salt over a synthetic catalog so the stack boots without operator secrets. It is **not** the live seal. Production must rotate to a private salt **and** a private catalog, replace the commitment in `config/relearn-pin.toml`, and re-sign the trust root ([`../config/CEREMONY.md`](../config/CEREMONY.md)). |
-| Relearn live scoring | Blocked on the eval image (no working digest). Behind that, the remaining blockers are operator state: the harvest (`live_harvest_wired`) and the champion baseline (`champion_baseline_recorded`). Each has its own **503** and its own boot-log line. Sim is opt-in (`RELEARN_FORCE_SIM=1`, CI / local only), reported as `eval_backend` on `/v1/status` and on the submit row — never a fallback. Refusals persist no row. |
+| Relearn live scoring | Eval image is pinned. Remaining blockers are operator state: the harvest (`live_harvest_wired`) and the champion baseline (`champion_baseline_recorded`). Each has its own **503** and its own boot-log line. Sim is opt-in (`RELEARN_FORCE_SIM=1`, CI / local only), reported as `eval_backend` on `/v1/status` and on the submit row — never a fallback. Refusals persist no row. |
 | Relearn live harvest | Control-plane client is **done** (`crates/relearn-lium-harvest`): boot the digest-pinned image, deliver the request, read `RELEARN_METRICS=`, verify against pin + run identity, terminate with verification. Wired on the Lium path from `LIUM_API_KEY` + `LIUM_SSH_PUBLIC_KEY_FILE`; `/v1/status` reports `live_harvest_wired`. The **scoring code** ships in [`CortexLM/relearn`](https://github.com/CortexLM/relearn) and must implement the contract in [`RELEARN.md`](RELEARN.md) § Eval image contract — an image that does not print a bound `RELEARN_METRICS=` document is a 503, never a sim score. |
 | Relearn holdout on rented pods | The harvest request carries the holdout items, so a rented pod sees the private split for the run. Mitigated by the digest-pinned image, `/tmp/relearn_eval` delivery, post-run scrub, and verified termination — not eliminated. Rotate salt + catalog and re-sign on suspicion. An in-enclave design would be needed to remove the exposure. |
 | Relearn teacher / judge keys on rented pods | `RELEARN_TEACHER_API_KEY` and `RELEARN_T2I_JUDGE_API_KEY` are forwarded into the pod when set, because a Lium `InstanceSpec` carries no env and the image judges over HTTP. Delivered over stdin (never a command line) and scrubbed after the run. A missing URL refuses **before** rent. A miner-controlled pod could still spend the operator's quota — scope and rate-limit those credentials, or leave them unset when the pod reaches the API without auth. |
 | Relearn champion baseline | Live hosts need an operator-recorded measurement (`RELEARN_BASE_CHAMPION_FILE`, verified against the pin's `eval_image_digest` + `holdout_commitment`). Unset means `champion_baseline_recorded: false` and every submission 503s before the gates. Sim hosts seed the sim baseline. |
 | Bounty severity on the backend feed | Scoring credits a `valid` row only when the backend publishes a `severity`. Until CortexLM/backend emits it, valid rows land as `valid_unpriced`, no miner can be crowned, and the share burns. Fail-closed by design: an unpriced bug cannot be paid for. |
 | Bounty scoring backend | The CortexLM/backend public feed is the only scorer. Without a readable `BOUNTY_BACKEND_PUBLIC_URL`, `POST /v1/reports` answers **503** rather than collecting bug-hunting work the host could never pay for, and the emitter pays nobody — it covers `E` with `ChallengeInternal` so the 3000 bps burns to uid 0 without 409ing every other challenge's seal. `BOUNTY_FORCE_SIM` is retired: a local scorer here would pay on adjudications no validator could reproduce. |
-| Relearn public repo | [`CortexLM/relearn`](https://github.com/CortexLM/relearn) exists; `relearn_git_sha` is empty in `config/relearn-pin.toml` because no image from it has scored on a rented pod. Bump the SHA and the digest together, from a commit reachable on that repo's default branch. Seed mirror: `docs/external-miner/relearn-seed/`. |
+| Relearn public repo | [`CortexLM/relearn`](https://github.com/CortexLM/relearn) exists; `relearn_git_sha` `8ffbe8a0…` is pinned with `eval_image_digest` `sha256:4806db4b…`. Seed mirror: `docs/external-miner/relearn-seed/`. |
 | Mainnet (netuid 100) | Owner wallet not yet on this machine, so prod runs with `BASE_GATEWAY_REQUIRE_OWNER=0`. |
 | Prod pin placeholders | `deploy/pins/prod.json` still ships zero-digests until the first successful promote; registry mode rejects placeholders. |
 | Spaces backup secrets | First prod promote is fail-closed without `BASE_BACKUP_ENDPOINT` + `SPACES_ACCESS_KEY_ID` / `SPACES_SECRET_ACCESS_KEY` (or AWS_* fallbacks) in GitHub. |

@@ -42,16 +42,17 @@ one of these holds:
 `live_harvest_wired`, and `champion_baseline_recorded`, and every submit row
 echoes `eval_backend`, so a sim run is never mistaken for a real verdict.
 
-`eval_image_digest` is **empty on purpose**. Every image built so far failed on
-a rented pod, most recently `sha256:cbc4bbb8…` (`f3cfa69`), which exited 1
-without printing `RELEARN_EVAL_OK`: the CUDA scoring layer shipped no vLLM, fell
-back to transformers, and then crashed because `Qwen3VLVideoProcessor` needs
-torchvision. The next pin waits on a `CortexLM/relearn` image that carries
-**vLLM and torchvision**. Do not guess a digest to clear the 503 — an unpinned
-host refuses for free, a wrongly pinned one rents a B200 that cannot score.
-
-Even once the image is pinned, a live host must still wire the harvest and
-record a champion baseline before `can_score` turns true; each has its own 503.
+`eval_image_digest` is pinned to the CUDA scoring image
+`sha256:4806db4bd6650415c3705290117fd0190cd2a6d387cef871dd3d9ff1a8bb6d7e`
+(`CortexLM/relearn` `8ffbe8a`,
+[publish-eval-image run 33730126042](https://github.com/CortexLM/relearn/actions/runs/33730126042)).
+The runtime job pulled that digest, ran the harvest-PATH check, and proved
+`import vllm, torchvision` plus `relearn-eval selftest` on those bytes. It is
+**not** the slim contract image. Do not re-pin `sha256:cbc4bbb8…` (`f3cfa69`)
+or the other named dead digests in `config/relearn-pin.toml`.
+
+A live host must still wire the harvest and record a champion baseline before
+`can_score` turns true; each has its own 503.
 
 A refusal is not a submission: nothing is persisted unless scoring produced a
 verdict, so a 503 leaves no row behind.

@@ -20,7 +20,7 @@
 | `rejected` with `PerturbationEvidenceMissing` | No perturbed rerun in the eval document | Fail-closed, like an empty `manifest`. The brittleness floor is not skipped by omitting the series |
 | `rejected` with `BaseCanaryEvidenceMissing` | No known-answer canaries in the eval document | Fail-closed. The base-competence floor is not skipped by omitting the series |
 | `503` on submit | Holdout file missing or mismatched | Operator: `RELEARN_HOLDOUT_FILE` must match the pin commitment |
-| `503 eval image digest not pinned` | Host has no `sha256:` eval image and did not opt into sim | Expected on `relearn` right now: no `relearn-eval` image has scored on a rented pod yet, so the digest is deliberately empty and the subnet refuses rather than rent hardware that cannot score. Image and Agent digests are pinned (`relearn` PR #3), so a 503 there is an operator pin/file mismatch. `GET /v1/status` shows `can_score: false` |
+| `503 eval image digest not pinned` | Host has no `sha256:` eval image and did not opt into sim | Operator pin/file mismatch: `relearn-eval` is pinned (`sha256:4806db4b…`, `8ffbe8a`). Image and Agent digests are also pinned (`relearn` PR #3). `GET /v1/status` shows `can_score: false` |
 | `503 … no in-process sim` | Digest is pinned but the live harvest is not wired on that host | Operator issue; `/v1/status` shows `live_harvest_wired: false`. The control plane refuses to substitute sim numbers |
 | `503 backend: lium …` | The eval pod could not be rented, reached, or torn down | Transient. Retry; the run is not banked and no verdict was recorded |
 | `503 backend: RELEARN_TEACHER_API_URL not set …` | Operator has not configured the judge the eval image needs | Operator issue; `/v1/status` shows `can_score: false`. No pod was rented |