b70-optimization-lab

DeepSeek V4 REAP/XPU Quality Contract

The primary truth is a fixed subset of official-source teacher logits and task results captured after the Stage 4 source download. Full unpruned IQ3_XXS is a secondary all-expert behavior control; it is quantized and runtime-confounded, so it is not source truth.

spec-eval-contract-v1.json is the separate anti-cheating contract for any new speculative decoder or routing policy. It requires freeze-before-reveal temporal holdouts, paired target/MTP1 controls, actual output-token-ID parity, request-scoped acceptance economics, nonrepetitive short and long contexts, and two independently generated packs. The repeatedly used public 12-prompt suite remains useful for continuity, but cannot by itself promote deeper speculation.

Use ../scripts/freeze-deepseek-spec-candidate.py before materializing either held-out pack. The helper hashes the candidate and both control identities, all supplied patch/policy/draft artifacts, records the only allowed online policy inputs, refuses overwrite, and creates a read-only manifest without a holdout seed. Candidate changes after pack generation spend that pack.

suite-v1.json is frozen as a prompt contract, with prompt text/hashes, categories, scoring-label placeholders, generation settings, tokenizer revision, and critical-case labels. It is not yet an executable scoring suite: the named rubrics, hidden tests, answer keys, scorer revision, request/chat template, stop conditions, and aggregate pass policy still need a separate versioned artifact. No intelligence or promotion decision may use v1 alone. Its file SHA-256 is d0d825c3a4ea4a748864741d767afe1d2a2b375d5407aaec3ee76a8e8246d6e0. Do not change it after candidate results are visible; create a new version.

calibration-v1-plan.json freezes a materializable 8,000-prompt domain mix, including dataset revisions, configs/splits, row IDs, text recipes, a fail-closed deterministic sampler, the true-REAP metric, and nested-candidate rules. The prior invalid request for 1,600 rows from the 164-row HumanEval set was replaced by the 5,000-row APPS training split. The prompt JSONL and observations remain to be materialized. Public 0xSero and sleepyeldrazi mappings are comparison seeds, not final provenance.

The frozen suite must cover:

Promotion requires:

If official teacher evidence cannot be produced and IQ3 cannot run correctly, the intelligence comparison is blocked. Uploader labels, qualitative needle tests, or parameter counts cannot substitute for it.