Source Observation 73 EN

Next training material: distinguish behavior from descriptions

Both small models failed to reduce checking cost, so the next step expands development diversity. Three behavior goals cover datasize units and error state, querystring tags and repeated values, and shlex quoting and malformed input.

We sealed24 inputs and expectations before execution, reviewed source and boundaries separately, then ran actual Go originals once. All24 matched; saved-file verification checked large integers, eight expected errors, nil/empty results and partial returns. Whole-child peak RSS was about17.9 MiB, not model-inference cost.

Twenty-four inputs are not24 independent tasks. The next three-parent/three-candidate proposal still has null truth, roles and weights. Source comments or signatures omit error/state conditions; matching observations do not prove sufficient model-input information. Korean and English share each parent. New actual labels, fits and completed final2,400 evaluations remain zero.

Follow-up74 prepared and independently reviewed source-derived descriptions for the same9 candidates. Original A and new B remain the same parents. Eight descriptions have scoped support; query requires a narrower request. Code satisfaction and model-visible evidence are separate. Actual72 uses the same SampleWeights: zero endpoint weights also zero and exclude ranking-loss/gradient contributions. Retained audit pairs are not all active training pairs. Next scopes for logfmt, dataurl, wordwrap and godotenv were audited with full MIT and supplemental Go BSD notices. Description review and correction · Next four sources. New labels, fitting and protected-final counts remain unchanged.

Detailed results, usage scope, source and receipts · 한국어 · Storage experiment