Independent review of 763bd80234be11942cfde10c30c0e79dcc9f5180 on feat/openenv-arena.

Decision: HOLD integration with submission automation and measured-win claims. These are synthetic, local software reproductions, not model-performance results. No source edits, GPU calls, deployment or Arena submission were made. External implementation remains with ruvultra.

1. P1: production gate signs synthetic promotion

The unmodified checked-in end-to-end test runs evaluator --dry-run, then gate --candidate-budget 1, and asserts promote:true, verified:true, reasons:[] plus successful pinned verification.

node --test integrations/openenv-arena-darwin/test/e2e.test.mjs
# tests 1; pass 1; fail 0

Rerun using committed Darwin dist, with a clean tracked working tree. Test, gate.
Default-budget rejection is incidental statistical rejection, not provenance enforcement. Reject dry-run/mock/proxy-only evidence in the production gate. Signature validation itself works; signatures authenticate bytes, not execution.

2. P1: absent provenance and duplicate seeds pass

Local decide() reproductions promote with only {seedBase:800000,attempts:32} provenance, absent model/source/trajectory hashes, or four duplicate seeds per block. Gate, confirmation.
Require complete typed provenance, frozen cell identities, unique matched seed sets and retained trajectory hashes; recompute classifications and aggregates from validated evidence.

3. P1: implemented fitness differs from the requested fixed-cell criterion

Fitness scores eligible mixed four-episode subsets, not fixed four-attempt cells with 1–3 full successes. Consequences:

• [1,1,1,1,1,.5,.5,truncated] yields primary 3/7, mixedFromReasoning true and regressed false.
• [0,.5,0,.5] yields primary 1 despite zero full successes.
• Full gate reproduction: baseline 36 solved episodes; candidate .5 every fourth episode, otherwise 1, final episode completion_truncated. Matched provenance, confirmation seedBase800000, selection seedBase700000/attempts4/evaluated1. Result: primary 0.6329513623631271, regressed false, promote true, reasons [], eValue 25.62890625 (8 informative of 9 blocks).

Freeze four-attempt cell identities before sampling. Any truncation/format/infra contamination must invalidate that cell while retaining the denominator. Count mixed success only when 0 < solved < 4. Reward variance can be a separate descriptive metric, not a silently substituted objective. Apply identical eligibility rules in selection and confirmation.

Classification also prioritizes reward1 over explicit failure, accepts missing failure, and admits failure:null with finish_reason:length. Contradictions must fail closed. The retained science d2 fixture seeds600201–600203 are the previously diagnosed omitted-quarantined-median ambiguity; exclude that source-bound evidence from reasoning claims pending clarified-source remeasurement.

4. P1: cache and fresh-output evidence are insufficiently validated

Cache, evaluator, runner validation.
A cached block with wrong family/difficulty, duplicate seed7, rewards [1,.5,1,1], zero tokens, and no metadata/plan/receipt produces:
regressed:false, primary:1, costPerWin:0, cachedCells:1, calls:0.
Fresh validation also accepts available:true with receipt:null and wrong model/seeds.
Use one strict validator on fresh output and cache reads/writes, binding cell identity, exact attempts/seeds, runner/environment/model/settings, receipts, trajectory digests and accounting.

5. P1: executed calls can be refunded after malformed output

Runner, evaluator fallback, budget refund.
A synthetic runner increments an invocation marker, writes malformed JSON to --output, and exits0. Three actual invocations under maxTotalNewCells1 were reported as calls [0,0,0], charged budget0.
Carry execution counts through all post-spawn read/parse/cache-write failures. Unknown executed work must never receive a zero-call refund.

6. P1: no enforced preregistration or persistent comparison ledger

Gate configuration can be weakened after outcomes:

• Same evidence, evaluated13: default budget refuses (need14 blocks/have8); candidateBudget1 promotes.
• alpha0.9 permits one winning block.

A truly predeclared single holdout could justify budget1; the implementation does not establish that history. Persist registration before outcomes, binding candidate, source/model/settings, alpha/lambda/budget and confirmation allocation; reject reused confirmations and posthoc weakening. Verification should bind expected rule/configuration/comparison identities, not merely a signing key. Fingerprint all source dependencies, including helpers.

Additional P2 findings

• Budget lease: after owner A fails, B reclaims, but waiting C does not recheck current ownership and proceeds with allowance0 while B runs. Reproduced max_new_cells_exceeded:1>0. Loop and recheck ownership after every wait.
• Cells: RUNNER_KNOBS is empty, so implemented suite_count_delta/sample_count_delta cannot yet evolve.
• Short Trackio completions do not establish trainer thinking mode. Advertised /models metadata does not attest the actual checkpoint/template. Preserve those limitations explicitly.

Acceptance before another readiness review

1. Negative tests reject synthetic, missing/contradictory provenance, duplicate seeds, contaminated cells, all-unsuccessful variance, and malformed cached/fresh trajectories.
2. Executed malformed-output calls exhaust the configured call budget; concurrent waiters cannot bypass ownership.
3. Preregistered durable comparisons reject posthoc policy changes and reused confirmations.
4. Send a new exact SHA and focused test receipts, then independently rerun review before any GPU finalist or scheduler integration.

Existing passing tests encode several of the problematic behaviors. Gate tests10/10 and fitness/cells25passed/1skipped are not evidence these boundaries hold. No TypeScript typecheck or real provider success path was exercised in this review.