2026-10-09 Before protocol freeze or training: numeric dataset validation caught a centre-distance failure for teacher example3427. Birth spacing had been checked on half-integer coordinates before final integer rounding. Corrected the authored generator to check the actual integer coordinates. No test model was evaluated and no trained weights existed. Full dataset validation rerun required.
2026-10-09 After model evaluation: the two first interpreter-return packaging commands failed at provenance-path serialization because their supplied relative paths were compared with an absolute root. Inference had completed and written raw fields, but no inference record. Corrected the artwork utility to resolve its input path. Reran fixed-weight inference; retained raw outputs were identical. This was an I/O correction, not a model, dataset or score change.
2026-10-09 Artwork recovery helper initially used Node deepStrictEqual to compare a Uint8Array packet with a Buffer, rejecting identical bytes due to container type. Replaced that helper assertion with Buffer byte equality. No packet, raw field, image or normative decoder was changed. The entire artwork recovery batch was rerun.
2026-10-09 Publication import caught unescaped literal paragraph newlines inside a reader's prose JSON string. Frozen first-reading source bytes and hashes are preserved. A derived reading.normalized.json escapes those line breaks without changing the prose. Permissive parsing applies only to this prose import, never to field packets, input grammar, datasets or test results. Also moved Matplotlib's cache to a writable job-local cache after its default user cache produced a permission warning.
2026-10-09 Final integration review caught structured prose anchors displayed as object strings and reply-only editing still targeting the old input. Display anchors were formatted into readable text; original structured records remain unchanged. Added an explicit exact-byte "Use reply as next field" transition and blocked input editing in reply-only view. The final inert UI suite contains25 passing checks, including a real generated n22 case. No learned weights, benchmark labels or model scores changed.