Skip to content

review(q7): verify unsafe GO in actual pinned XLSX, 25 engine cases [skip ci] - #696

Merged
Knapp-Kevin merged 1 commit into
mainfrom
review/q7-real-pinned-xlsx-25-case-acceptance-20261103-no-ci
Nov 3, 2026
Merged

Knapp-Kevin merged 1 commit into
mainfrom
review/q7-real-pinned-xlsx-25-case-acceptance-20261103-no-ci

Conversation

@Knapp-Kevin

Copy link
Copy Markdown
Owner

Q7 original pinned XLSX ACCEPTANCE: transfer blocker resolved; source false GO verified

Actual original source file, not synthetic. Transferred the exact zicula/skills@ec7f94f84916be7bb25ec38752c24176330065a7 grant-writing-kit/references/2-go-no-go/go-no-go-scorecard.xlsx bytes via GitHub connector Base64 into a private isolated container. Verified 10,048 bytes, Git blob SHA-1 a58d34344847a73440cb8ded6be88a3e87cc093d, ZIP CRC integrity and 11 OOXML parts. Imported those exact bytes with artifact_tool.SpreadsheetFile.import_xlsx. Source file never exported or altered.

Publisher's actual example fails: Its worksheet contains Gates!B4=No. Re-entering unchanged Scorecard!E2 in the imported memory workbook triggered the original F2:F16 → F17 → F18 chain (original XLSX lacked cached formula values). The original weighted score is 76, yet the published F18 returns GO — commit the hours. An in-memory candidate formula referencing real answers in Gates!B2:B7 returns WAIT.

Full controlled reproduction: 25/25 additional scenarios passed the candidate repair against the imported original publisher workbook, including No in every gate, blanks, unknowns, numeric/whitespace inputs, score thresholds 49/50/69/70, missing/text score, and reverse false WAIT when all B answers are Yes but narrative C2 contains No. Original verdicts, expected candidate decisions, source hash and exact fidelity limitations recorded in registry/discovery/q7-original-pinned-scorecard-engine-acceptance-20261103.json. Wrong-byte Git hash test correctly rejected a deliberately mutated file.

Repository changes: Existing pinned harness scripts/qa/check_pinned_grant_scorecard.py now also verifies exact pinned source baseline 76/unsafe GO and runs 25 cases (fixture mode still clearly labeled synthetic); dated plain-English review, README, curation queue, system state, index and historical blocker docs updated. Existing 18/18 and 9/9 synthetic receipts remain preserved, with supersession markers. The original XLSX is not mirrored or patched in Skillz.

Critical disposition: Transfer/execution blocker resolved, but the publisher's source is now empirically proven unsafe, not approved unchanged. Excel/LibreOffice interoperability, actual funder policy/financial-owner review and complete security clearance remain held. This adds qualification evidence, not a new Valuable Reference or approved installation. Current 20-method Q7 bounded shelf and 33-domain graph untouched. Claude's independent Q8 source reviews untouched. [skip ci], no GitHub Actions pass claimed. Branch readback checks source pin, report 25/25, baseline 76, historical preservation and exactly nine changed files.

@Knapp-Kevin
Knapp-Kevin merged commit e083c47 into main Nov 3, 2026
2 checks passed
Knapp-Kevin pushed a commit that referenced this pull request Nov 3, 2026
…ync public accounting [skip ci]

Addresses PR #693 integration-HOLD review and cross-ledger defect comment.

- Merged origin/main (post-#696); README retains #696 workbook acceptance links
- Transferred the 4 previously-UNKNOWN package-tree SHAs (video-scoring-audio,
  design-concept-art-brief, scenario-storyboards, scenario-identity-library)
  from the pinned scanner receipts into the Q8 consumer ledger, after matching
  exact source pin, path and SKILL.md blob; Q7 UNKNOWN retained as history
- Added dated Q8 checkpoint (39 methods / 28 domains) to docs/SYSTEM_STATE.md,
  INDEX.md and a current-shelf pointer in index.json; Q7 retained as history
  and its workbook caveat now points to merged #696
- Re-ran dashboard --self-test/--check; cross-checked 39-item ledger, 33-domain
  table, SVG and CSV; failed/incomplete scans and unaccepted false positives
  stay fail-closed; behavioral/adversarial acceptance remains NOT RUN

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0114De5q3RniVy9N6N5TiBuT
Knapp-Kevin added a commit that referenced this pull request Nov 4, 2026
Study-only Q8 source-pinned catalog and generated dashboard; 50 exact Q12 source matches, 33 domains, 10/7/33 static security outcomes, no unchanged-use approval. Preserves #696 workbook HOLD, merged #698 evidence and main security snapshot. Post-merge snapshot must be regenerated.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant