Planning-capability evidence publication record

• Public copy date: 2026-07-22
• Canonical public source revision: fakoli/anvil-serving@31d95adaf68157b81318325356516cef9569b10f
• Bundle tree at that revision: 6e26ebc80b853e151075e8807c0bba084480f823
• Notes mirror revision inspected: fakoli/anvil-serving-notes@7b46ceb6ae62252f8f808f6c065706a24e7970bb
• Notes import revision: 71b0124f541840f3276e118db00aeb4a4126f99a
• Canonical public-source bundle before publication edits: 21 files, 159,272 bytes
• Admission decision: publish the complete bounded bundle; every file is below 1 MiB and the bundle is below ADR-0027's 5 MiB aggregate limit
• Sanitization scan: no user-home paths, email addresses, API keys, authorization values, session identifiers, or secret-like values matched
• Publication normalization and edits: restore the imported checkout to canonical LF Git bytes; make local generation transactional and fail closed; bind aggregation to the six independently judged output hashes; add Ruff file directives to the historical offline calculators; and remove trailing spaces from four Markdown model outputs so repository whitespace gates pass. Non-whitespace output text, prompt text, judge records, metrics, and manifests are otherwise unchanged; the source hashes preserve the exact originals
• 2026-07-22 verification: python grade_struct.py followed by python aggregate.py reproduced the tracked structural scoreboard, judge aggregate, and CSV/JSON outputs with a clean Git diff; missing confirmation, partial request failure, and changed-output aggregation probes fail closed
• Not rerun: local/frontier generation and blind judging. Model IDs, endpoints, timings, and quality scores remain historical observations from 2026-06-28.

SOURCE-SHA256SUMS.txt records the canonical Git blob bytes for every source artifact at the public source revision. The three edited scripts and four whitespace-normalized output files will not match those source hashes; those differences are intentional and fully enumerated above.