Skip to content

Crosswalk the gap-resolution receipts onto the 380 current open gaps (92bb279) - #85

Merged
seathatflowsinourveins merged 3 commits into
mainfrom
claude/gap-crosswalk-92bb279
Sep 23, 2026
Merged

seathatflowsinourveins merged 3 commits into
mainfrom
claude/gap-crosswalk-92bb279

Conversation

@seathatflowsinourveins

Copy link
Copy Markdown
Collaborator

Maps the 40 gap-resolution receipts (#81) onto every open_gaps entry of the 32 landscape rows at 92bb279, which #79 rewrote. Each gap also gets a blocker category, a next check where one is executable, and an owner for the next evidence wave. No verdict changes.

Method

  1. Candidates. Code pairs each gap with every receipt that names its layer.
  2. TypeSafe screen. jev-1.13.0, question revision gap-crosswalk-questions-v1, via tools/sota-convergence/gap_crosswalk.py. The threshold rule was preregistered in evidence/artifacts/gap-crosswalk-92bb279/PREREGISTRATION.md before any call.
    • Eval. On the 315 reviewed bdd04ca gaps, the rule selected T = 0.3, which caught 27 of 29 addressed gaps.
    • Categories. TypeSafe's blocker categories agreed with the reviewed labels on only 46% of gaps, below the 0.60 bar, so they are recorded as advisory only.
    • Current run. 380 gaps, 0 errors; 109 pairs were queued for review.
  3. Review. Nine Opus semantic-evidence-reviewer groups decided every queued pair and categorized every gap. Independent Opus evidence-reviewers re-checked all 30 positive decisions: they agreed with 25 and corrected 5 (3 up to settles, 2 down to not addressed). TypeSafe never sets a status.

Result

  • By existing receipts: 3 gaps settled, 22 advanced, 355 open.
  • Next checks: 258 executable_now gaps have a next check, split by the owners agreed between the two sessions:
    • gap-resolution: 69 gaps in 8 foundation layers;
    • agent-lab-17: 189 gaps in 12 us-equities and 12 foundation layers.
  • Files: the ledger catalogs/landscape/gap-crosswalk-92bb279.json and the page docs/gap-crosswalk-92bb279.md. The retained judgments and review results are hash-checked by tests/test_gap_crosswalk.py.

Verification

  • gap_crosswalk.py build --check reports up to date.
  • python3 -m unittest: 2,134 tests OK (324 environment skips).
  • All validators pass: validate.py, validate_catalogs.py, validate_foundation.py, validate_convergence.py, landscape.py, build_verdicts.py --check, trading_gates.py --check, build_ecosystem.py --check.
  • Guarded gitleaks on the branch commits finds 0 leaks.

Limits

  • Indexes refer to 92bb279. The convergence refresh planned for today will re-record the rows and needs a new crosswalk.
  • The TypeSafe screen missed 2 of 29 addressed gaps on the eval set, and pairs below the threshold were not reviewed.
  • Categories assume one WSL2 workstation, paper orders only, no paid data and no second machine.

🤖 Generated with Claude Code

…2bb279)

- tools/sota-convergence/gap_crosswalk.py: TypeSafe screen (jev-1.13.0, question
  revision gap-crosswalk-questions-v1), eval scoring, and an offline build that
  joins the retained judgments with the retained Opus review.
- Preregistered threshold rule; eval on the 315 reviewed bdd04ca gaps selected
  T=0.3 (27/29 addressed gaps caught). TypeSafe blocker categories agreed 46%
  (below the 0.60 bar), so categories come from the reviewers.
- Review: 9 semantic-evidence-reviewer groups (Opus/high) decided all 109 queued
  pairs and categorized all 380 gaps; 8 evidence-reviewer checks (Opus/high)
  verified the 30 positive decisions (25 agree, 5 corrected).
- Result: 3 gaps settled and 22 advanced by existing receipts, 355 open; 258
  executable_now gaps listed with a next check and an owner per the agreed
  layer split.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
…pt drift and off-rule candidates

The repo ignores *.jsonl, so the retained judgments were never committed (Codex blocker). The build now recomputes each judged state from the current receipts and gap text and compares state_sha256, and requires the judged candidates to equal the ledger's candidate rule; tests cover both guards. eval-score.json reproduces from the retained eval judgments.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
@seathatflowsinourveins
seathatflowsinourveins merged commit 8edf4dc into main Sep 23, 2026
6 checks passed
@seathatflowsinourveins
seathatflowsinourveins deleted the claude/gap-crosswalk-92bb279 branch September 23, 2026 05:22
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