Audit follow-up: gap-ledger join and dated verdict fields in sweep inputs; durable-memory note; landed #575 citations and four trading acceptance classes

[seathatflowsinourveins]

Scope

• What this PR changes: the follow-up to the Codex catalog lane's audit of its 32 frozen sweep inputs (2026-10-01), which found 14 winner pins behind the stack and gap texts that committed receipts had already answered. Five parts, each its own commit, with the hot-file commit last.
  1. tools/sota-convergence/landscape-sweep/build_inputs.py: every repository-layer sweep input gains verdict_checked_at, verdict_note and open_gaps_followup, a join of the committed gap-wave owner ledgers onto the shown open gaps. Existing fields keep their shape; no template changes, so the pinned prompt hashes stay.
  2. catalogs/landscape/foundation.json: one dated 2026-10-01 limitations entry on the durable-memory row. No verdict field changes.
  3. docs/decisions/2026-10-01-definitive-sota-wsl-program.md, decision 5: how a comparison's result reaches the landscape catalog.
  4. catalogs/foundation/new-wsl-architecture-20261001.json and its record: the 39 citations of merged Review all replacement WSL architecture layers and retained gates #575 become landed sources, and four trading verdict winners get the acceptance class their committed receipts support.
  5. The trading lane owner's verbatim note on the identity-provenance row, from its acknowledgement comment on this pull request (one sentence; the manifest is unchanged by it).
• Base commit: d6d3bf7d6e28e6ff1c557a07806f757d16b71ccd
• Lane: lane:shared (the edition's three trading rows change; the trading lane owner asked for the correction and re-reads the rows before merge)
• Owned paths touched: the seven files above and, in the last commit, manifests/evidence.json.

Why the stale inputs happened: winners, alternatives and open_gaps of every landscape row are the sealed 2026-09-22 verdict wave. record_verdicts.py --write refuses that run id, build_verdicts.py --check fails when such a row differs from the hash-registered wave document, and tests/test_layer_verdicts.py pins it. The sweep's input builder copied those fields without their date and did not join what the repository recorded afterwards. Each input already carried the current stack pin in components_vs_upstream; nothing said which pin was which.

What the join gives on main's catalogs (the Codex lane's scope and freshness manifest, a scratch work directory):

32 layers; 20 foundation; ...; gap follow-ups: 133 joined from 3 ledger(s), 4 dropped (no sealed open gap with that index and text), 190 for gaps beyond the 5 shown

323 of the 327 ledger entries match a sealed open gap by index and exact text; the 4 that do not are in ci-supply-chain and are counted, not joined. 110 of the 152 gap snippets the inputs show get a follow-up. What the ledgers do not record, such as the 2026-09-25 memory run, still reaches a sweep through --seeds.

SOTA sources

• Ledger reading: this repository's reviewed implementation, tools/sota-convergence/lane_packets.py at d6d3bf7d (GAP_LEDGER_GLOB, gap_receipts_index: the same ledgers, the same rule that a receipt path counts only when it is a file in the checkout). The ledgers' own rule text (catalogs/landscape/gap-wave2-20260923--agent-lab-17.json, key rule) defines the statuses that are passed through unchanged.
• Sealed-wave rule: tools/sota-convergence/README.md at d6d3bf7d, "Record verdicts" (a grandfathered wave is never re-recorded) and its practice on this row in Grand catalogs carry each winner's evidence basis; ai-memory is the retained control, not a measured best #269 (89cf253c) and Pin ai-memory 2.4.1 and mcporter 0.14.1 on Linux; workstation SOTA refresh record (headroom 0.39.0 qualified, not switched) #307 (ff20470b): dated limitations entries.
• Adopted-only winners: tools/sota-convergence/lane_packets.py build_candidate (adopted is a disposition of selected or conditional) and scripts/landscape.py lane_winner_components (rejects a non-adopted winner key), both at d6d3bf7d.
• Memory comparison facts: evidence/artifacts/memory-stack-20260925/convergence.json (arms C3 and D2, amendment A14) and blueprints/memory-stack/longmemeval/v4/PREREGISTRATION.md (arms D2h, K1, M1, M2). Upstream currency: https://github.com/akitaonrails/ai-memory/releases/tag/v2.5.2, published 2026-10-01T07:40:18Z (gh api repos/akitaonrails/ai-memory/releases/latest).
• The four corrected acceptance classes: the committed receipts under evidence/artifacts/gap-wave2-20260923/ (us-equities__identity-provenance/0-dvc-repro-restore-contract.json and 4-four-store-comparison.json, us-equities__data-quality-orchestration/5-duckdb-gate-input.json, us-equities__evaluation-experiments/1-five-package-installs.json and 2-matched-promptfoo-inspect-comparison.json), classed by docs/acceptance-evidence-policy.md.
• The merit rule: the user's message of 2026-10-01, quoted in decision 5 of the program record.

Evidence-class table

Claim	Evidence class	Command / receipt
The builder joins the ledgers and dates the sealed fields; the old fields keep their shape	local_integration	python3 -m unittest tests.test_landscape_sweep_harness tests.test_landscape_sweep_skills (three new tests, written failing first); a real build on main's catalogs, summary line above
The row note passes the landscape validator and leaves the sealed wave untouched	local_integration	python3 scripts/landscape.py --root . rc 0; python3 tools/sota-convergence/build_verdicts.py --check rc 0
A label-withheld packet keeps one neutral sentence of the note and none of its selection statements	local_integration	lane_packets.build_all_packets(..., withhold=True) in both builds of RealLedgerIncumbentBlindnessTests; tests.test_lane_packets
DVC, pandera, MLflow and Inspect AI each have a recorded run	source_review of committed receipts	the five receipt files named above, read on d6d3bf7d; no run was repeated
Nothing was run on a model, a memory system or a provider	none	n/a

Local commands run

$ python3 scripts/landscape.py --root .                              # rc 0
$ python3 tools/sota-convergence/build_verdicts.py --check          # rc 0
$ python3 scripts/build_ecosystem.py --check                        # rc 0, status passed
$ python3 scripts/validate.py                                       # rc 0: 69 components, 8901 hashed files, 183 receipts
$ python3 -m unittest -q tests.test_landscape_sweep_harness tests.test_landscape_sweep_skills tests.test_ecosystem_manifest tests.test_lane_packets tests.test_blind_checkout tests.test_landscape tests.test_layer_verdicts tests.test_record_verdicts tests.test_component_matrix tests.test_practice_references tests.test_handbook_summary tests.test_verdict_review_gate tests.test_saturation_ledger tests.test_platform_status tests.test_adjudicate tests.test_blind_layer_decisions tests.test_verdict_lane_vendoring tests.test_agentsview_qualification
Ran 1305 tests in 138.230s
OK (skipped=46)
$ git diff --check origin/main HEAD                                 # rc 0

Manifest: 4 of the 7 changed files are registered and re-registered; files rows 8,901 before and after, receipts 183, convergence records 26, none lost.

Decision record

docs/decisions/2026-10-01-definitive-sota-wsl-program.md, decision 5 (extended here), and docs/decisions/2026-10-01-new-wsl-architecture-edition.md (counts and the lesson that a none_recorded is a claim that needs the same receipt search as a stronger class). Alternatives considered for the stale inputs: a hand edit of the verdict fields (refused by the tooling); a Claude-only single-lane re-record (one model family, and it can only re-affirm an adopted candidate); a sentence in the worker templates (the fit template is shared with the skills modality and both prompt hashes are pinned, so the note travels inside each repository-layer input instead); a two-lane re-record of all rows, which is program unit U4 and stays queued. Overturn: a recorded head-to-head on the new distribution changes a candidate's disposition, and a new wave re-records the row.

Host evidence

Not relevant: no file under evidence/hosts/ changes.

Checklist

• No workflow or action changes.
• No secrets are printed, logged or committed.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[seathatflowsinourveins]

Trading lane: acknowledged at 339d57c6 (base d6d3bf7d). I read the 12 us-equities rows in code against the base.

• identity-provenance is now local_integration_check. DVC is cited to 0-dvc-repro-restore-contract.json as native operation, with its text and the four-store comparison reason.
• data-quality-orchestration is now structural_validation. pandera is cited to 5-duckdb-gate-input.json as a local integration check.
• evaluation-experiments is now structural_validation. MLflow and Inspect AI are cited to 1-five-package-installs.json as native operation, and the overturn-comparison reason is added.
• Every other trading row changes only in citation form: pending_source becomes source_path (Review all replacement WSL architecture layers and retained gates #575 landed), with the same text.
• No change to any verdict, gate or closure cell.

One note change, verbatim. In identity-provenance's notes, replace "no gate blocks it, and it needs a recorded DVC install first. The verdict's own winner, DVC 3.67.1, has no pin of record." with "no gate blocks it. DVC 3.67.1 already ran natively on synthetic fixtures (2026-09-23), so the comparison needs a fresh DVC install on the target host and the point-in-time fixture. DVC's pin is the verdict's baseline 3.67.1; it has no stack pin, and the new distribution installs the current upstream release."