VFR1: Verified Fault-First Revision

Status: closed negative after the sole frozen train-only quality pilot.

Result

Job 747591 completed all 128 deterministic rows in 1,966.17 seconds. Only 7/128 = 5.47% outputs satisfied the strict two-block parser and all seven extracted revisions verified. There were no reference leaks or boxed answers inside parsed FAULT blocks, but 53/128 = 41.41% outputs exhausted the 1,024 token budget. The other 121 outputs failed tag cardinality.

A single read-only attribution scored the unparsed raw completions directly with the original per-row assessors. Only 67/128 = 52.34% were correct: math 27/67, logic/science 40/55, and code 0/6. Thus the failure is not merely strict-format compliance; the teacher outputs also miss the frozen 90% correctness floor by a large margin. The full generation, capability data, matched fits, and development evaluation were not authorized. Jobs 747593--747600, 747602, and 747603 were canceled before allocation.

Trace SHA-256 is 9e0d7812...8c5d, trace-report SHA-256 is 5d9286c0...cbd5, and quality-report SHA-256 is 96244f9a...38b9. The machine-readable summary is SHOHIN_VFR1_PILOT_RESULT.json. Close exact VFR1 without a teacher-prompt, decoding, context, parser, seed, or threshold rescue.

Measured bottleneck

IDR1's 9,655 train presentations contain 3,294 source_verified_repair targets. Their median target length is only 11 characters; 3,245 are exactly a boxed answer and 3,214 are under 80 characters. Thus 34.1% of the revision curriculum teaches answer emission after a failed draft rather than a reasoning process that locates and repairs the failure. In contrast, verified candidate targets have median length 706 characters.

Hypothesis

A revision owner should externalize a compact fault state before generating the replacement trajectory. Because later autoregressive tokens attend to the fault block, this changes the correction computation rather than selecting among already completed answers. The target format is:

<FAULT>
earliest decisive error or independently checked no-fault statement
</FAULT>
<REVISION>
complete corrected solution
</REVISION>

At inference, only the model owns both blocks. There is no verifier, reference answer, teacher, task router, or correctness bit. A strict parser extracts one complete revision; malformed traces fail closed.

Frozen data-quality pilot

Build exactly 5,824 unique train-only teacher requests from the hash-bound IDR1 identities. The teacher is pinned Qwen3.5-9B at revision c202236235762e1c871ad0ccb60c8ee5ba337b9a with the immutable B1 adapter SHA-256 854a7cc44fbc2b54418f4e5bd09b7efeed0da44fc9ce217b0bb6b1997b722971. It sees source, model-owned draft, and the verified train-only reference only while constructing supervision. Assessor fields are never rendered.

The first gate is a deterministic 128-identity pilot, selected from the first immutable request shard before any full generation. It passes only if:

• at least 95% of outputs parse into exactly one nonempty FAULT and REVISION;
• at least 90% of extracted revisions pass the original answer/code verifier;
• no more than 2% mention a provided/given/verified reference;
• no FAULT block contains a boxed answer;
• at most 10% exhaust 1,024 generated tokens; and
• all source, model, adapter, request, token, memory, and output hashes are complete.

Failure closes this exact teacher format before full corpus generation. A pass allows eight disjoint H100 generation shards, followed by an independent CPU merge/audit. It does not by itself authorize a capability fit.

Conditional capability gate

Only after the full corpus passes the same quality conditions may one treatment and one deterministic within-task/length-stratum shuffled-fault control train from the same B1 warm start for 256 updates. Both use identical complete revision text and the same global target-token multiset; only the source-to-fault assignment differs. Development remains the existing 1,289 source-disjoint identities. Holdout remains sealed.

Treatment promotion requires all of:

• at least 603/1,289 overall;
• at least +10 answers over the matched shuffled-fault control;
• math at least 223, logic/science at least 349, and code at least 17;
• at least 95% strict two-block parse coverage and zero evaluator leakage; and
• complete matched update, token, FLOP, latency, truncation, and protected-hash receipts.

Any miss closes exact VFR1 without teacher-prompt, temperature, seed, update, rank, context, parser, or threshold rescue.