Repairing C after ordinary access

The request-preservation and subject-resolution changes bring the unchanged expert graph from 7/15 to 10/15 on the exposed ordinary development screen. All 21 standalone routes now select the expected expert. No selection, decomposition or assembly failures remain; all seven retained cases pass. Six owners agree on complete transcripts, and automatic and forced replay pass. The five remaining failures come from three incorrect C answers. This closes the bounded access repair, not the full learning milestone or public readiness.

The first preservation attempt scored 9/15 and failed one reference repair. Both attempts are retained locally, with CPU verification, source archives, all-owner evidence and resource retirement records. No expert was trained in either inference attempt. See the access record.

The first committed prescription tested one training-data intervention. C's 192 original conversations contain 144 distinct unwrapped questions: 128 occur under only one wrapper, and 16 under four. The new preparer crosses every question with all five existing wrappers and bare wording, producing 864 conversations. Targets come exclusively from the original training records. The question strings from development and the opened final are used only for exclusion, before tokenization; no scoring answers enter training. No fact, evaluation wording or serving answer lookup is added.

The hypothesis is that independent coverage of wording and prompt format reduces C's input sensitivity. This is a hypothesis about this model, not an established explanation of all three errors. Prior work supports systematic coverage when injecting facts through SFT, but used different models and data: Mecklenburg et al..

Only C's existing 134,225,920-parameter tail receives updates, starting from 618e3eb0… with fresh Adam. The fixed recipe is 108 steps, batches of 16, microbatches of four, learning rate 0.00005, four warmup steps and response-only cross-entropy. Two full dataset passes cover every conversation twice. The backbone, A/B experts, selector, question handling and generation rules remain fixed. The current C weights are not permanently protected from improvement; the accepted behavior is what the quality gate protects.

Only the terminal checkpoint is evaluated. Acceptance requires all 15 ordinary cases and all 13 distinct forced controls correct, preservation of all 13 previously correct C standalone development answers under their exact earlier prompts, exact replay and complete-call metering. No new final opens, no next cohort starts, no NEURO is issued, and this experiment cannot promote serving. The actual full-system result decides whether the method helped.

The completed result improved ordinary answering 10/15 → 14/15, forced answers 10/13 → 12/13, and C's earlier standalone development questions 13/16 → 15/16. No previously correct answer in either preservation inventory was lost. All four training owners agreed; all six serving transcripts agreed. A fresh process reproduced updates 104–108 from the input boundary, and both inference replays passed. The 108 updates took 985 seconds. All four instances, volumes and the temporary security group were retired; estimated compute was at most $2.28, with storage and transfer separate. The candidate still failed the unchanged all-answer gate. These are exposed development results, not a fresh final or new cohort.

The remaining error confuses a single claim's window with a complete job: the expert answers 4096 where the window limit is 16. Source inspection establishes the relationship: planner_work.apply passes claim_planner.window to planner_window.validate, which bounds that window to 16 updates; the full recipe may prescribe 4096. Existing training variations described a window without connecting it to a submitted claim.

The separately frozen scope continuation starts from terminal C f1aa7624…. Its training variations add six claim-window questions and six contrasting whole-job questions. The same original training inventory supplies every target; all other topics remain. Crossing with the six input contracts gives 936 conversations. Development and opened-final question strings remain excluded. This intervention is explicitly informed by development failures and cannot establish independent generalization, even if the diagnostic passes.

The continuation freezes 64 updates: all 59 balanced batches once, followed by the first five balanced batches. Learning rate, optimizer, microbatch, question handling, selector, A/B experts and backbone use the previous recipe. Only the terminal C checkpoint is tested. It must preserve all 14 currently correct ordinary cases and all 15 currently correct C development answers, as well as pass the original all-answer, replay and metering gates. This is a small repair within the same cohort, not completion of TODO items 1 or 2.

The first scope allocation stopped during object download, before training: publication receipts lacked the fetch inventory's folder field. Preparation now derives the owned objects destination from the checkpoint, validates hash and size against that checkpoint, and rejects conflicting destinations. The recovered allocation uses identical training bytes, recipe, schedule and retention inventory. Its freeze changes only to bind the corrected handoff. The original attempt remains recorded and is not counted as a quality run.

The corrected allocation passed the original development gate: ordinary conversations 14/15 → 15/15, unique forced controls 12/13 → 13/13, and C's standalone preservation inventory 15/16 → 16/16. No previously correct tested answer was lost. The claim-window answer changed from 4096 to 16; the complete-job answer stayed correct. All seven retained A/B/general conversations kept identical neural outputs, including the mixed requests. The 64 updates took 575.6 seconds; serving evaluation took 124.8 seconds. Four training owners and six complete serving transcripts agree. A fresh process reproduced updates 60–64 exactly, and automatic/forced inference replays passed. CPU rescoring agrees with the frozen scorer.

The resulting research service is 3b3d5a2160eb1bec670a03beb85d14df23fbf24d192741416b4658c08c9d0ddb, with C checkpoint 7d976925e8ff9d3c7588c2eed5ac9852a02d454318842093fcc353e7fa2338fb. The result manifest binds the source, local evidence archive, object catalog and retirement record. The setup failure is also retained. Both final numerical boundaries are available by immutable public object hashes; metadata remains local and committed by archive digest. Both allocations are fully retired, including their volumes and temporary security groups; the three protected hosts are unchanged. The successful scope run's conservative compute estimate is $1.68 and its failed setup's is $0.67. Including both earlier inference repairs and the 108-step C repair, this recorded sequence totals $6.64 in estimated compute. Storage and transfer are separate. No native issuance or serving promotion occurred.

This closes the measured ordinary-serving obstruction on the existing exposed development inventory. It does not count as a third admitted cohort or an independent holdout pass. The next learning proof must use prospectively frozen ordinary questions for new data, protect prior behavior, and run through native admission and separate quality promotion. This is the remaining work in TODO items 1 and 2, whose completion criteria are unchanged.

One four-host allocation performs sharded prefix production, C training, replay of the final four-update window from its input state, and six-owner serving. Its absolute deadline is two hours, training deadline one hour and compute planning cap $25. The final two boundaries are retained in the existing immutable project object store before retirement. Earlier window metadata alone does not establish payload availability for settling all those windows.