Record the Claude-family half of the model-routes initial sweep - #1188

Scope

• What this PR changes: it records the Claude-family half of the model-routes initial sweep as retained evidence. This is a single-family initial sweep under the CC's 2026-10-15T19:52Z ruling (A). It does not edit any catalog, convergence or ledger file. The layer stays reopened until this half and the GPT-family half (foundation: retain GPT-only initial model-routes sweep #1177) are merged per MERGE-PLAN.md. The merged result supports TRIAL labels on NativeHorizonNY only, never FINAL.
• Base commit: a93ba0b8c (the tools used are unchanged since 8053f14e7; see RUN-RECORD.md).
• Lane: lane:foundation
• Owned paths touched: evidence/artifacts/model-routes-sweep-claude-20261015/, evidence/artifacts/model-routes-sweep-claude-20261015-attempts/, manifests/evidence.json (registration rows only), tests/test_provenance_digest_custody.py (one new test class, appended).

SOTA sources

• Recipe: recipes/saturation-sweep.md at 8053f14e7 (sections 2-3: worker labels, retained returns shape, survival rule).
• Harness: tools/sota-convergence/landscape-sweep/ at 8053f14e7. build_args.py staged the run (frozen templates.json and schemas/; prompts_sha256 f8263a7b7cce54784278ac909ecb1ab5e83376411891b3e4880fd134df516a7d). The stage scripts reproduce sweep.js's Claude prompt code verbatim (claudeDiscoverPrompt; the refuter and critic prompts in runLayer and the critic block). usage_record.py (the vendored examples/claude-native/workflows/child-usage.mjs) measured usage, and convert.py produced returns.json.
• Protocol: docs/decisions/2026-10-11-layer-resolution-protocol.md, steps 1-3. Recorded in RUN-RECORD.md.
• Requirement hash rule: scripts/saturation_ledger.py:524 (requirement_sha256) at 8053f14e7.
• Carried candidates: catalogs/landscape/angles/foundation-model-routes.json at 8053f14e7 (14 candidates).
• Worker role: adoption/agents/claude/landscape-sweep-worker.md (Opus at effort max, read-only tool list), as installed.
• List prices for the cost estimate: anthropics/skills@dbd4588f9e1033efb41dad4bef2f7947c8993d44, skills/claude-api/shared/models.md:79 and shared/prompt-caching.md:144. The cache TTL is the assumption: prompt-caching.md:144 gives cache writes at 2x input for the 1-hour TTL and 1.25x for the 5-minute TTL; the estimate assumes the 1-hour TTL (about $21.48 at the 5-minute rate).
• Authority, CC records:
  • the CC ruling of 2026-10-15T19:52Z (A) (merge rule and $40 cap);
  • the CC direction record direction-cc-20261015T1912Z-new-wsl-harness-first.md (research-cc-role-20261008/; cited by name only, ruling 2, Models: model-routes);
  • the CC's full read of the first head, MICRO-full-afcc08cb.md (by name), whose corrections this head carries (RUN-RECORD.md, "Corrections").
• Vendor primary sources the workers read: listed per proposal in returns.json, each with its dated URL (dev.meta.ai docs, developers.openai.com Codex models, code.claude.com model configuration, platform.claude.com model overview, GitHub REST).

Evidence-class table

Local commands run

$ python3 tools/sota-convergence/landscape-sweep/build_args.py --work-dir <W> --sweep-id model-routes-claude-20261015 --date 2026-10-15 --layers model-routes --no-stars
prompts_sha256 f8263a7b7cce54784278ac909ecb1ab5e83376411891b3e4880fd134df516a7d
$ python3 tools/sota-convergence/landscape-sweep/usage_record.py --transcript-dir <run> --out child-usage-<run>.json   (x3)
rc 0, status complete
$ python3 tools/sota-convergence/landscape-sweep/convert.py --workflow-output synthesized-return.json --work-dir <W> --scope scope-model-routes.json --out out --limit "<single-family note>"
rc 0   (first conversion; a control rerun reproduced it byte for byte)
$ python3 <A>/join_usage.py <A> <A>/usage-joined.json
rc 0, 4 children
$ python3 tools/sota-convergence/landscape-sweep/convert.py --workflow-output <W>/synthesized-return.json --work-dir <W> --scope <S>/scope-model-routes-published.json --usage <A>/usage-joined.json --out <S>/out --limit "<same single-family note>"
rc 0   (77 values changed: the requirement hash, and model and effort on the 38 Claude votes; counts unchanged)
$ python3 -m unittest tests.test_provenance_digest_custody tests.test_published_digest_privacy tests.test_evidence_manifest
OK (43 tests, 1 skipped); the 4 new ModelRoutesClaudeHalfCustodyTests FAIL on the first head afcc08cbc and pass here
$ python3 scripts/validate.py
{"components": 71, "hashed_files": 12156, "profiles": 4, "receipts": 240, "status": "passed"}

Decision record

No new decision record. The CC's ruling of 19:52Z (A) governs. MERGE-PLAN.md states the merge rule and what would overturn a candidate: survival needs all three votes (Claude facts, Claude fit, GPT fit). The ledger-countable sweep is the documented two-family launch after registration (#1179).

Follow-ups (from the CC's full read of this head, MICRO-full-3c7a586f.md, cited by name)

• RUN-RECORD.md:157-158 says the custody test "checks every digest" in the two directories. That overclaims: bare 7-63-hex prefixes, SHA-1 values and subdirectories are not covered. A later correction narrows the sentence or widens the test.
• MERGE-PLAN.md:37-38 routes its PROPOSAL (vendor-document capture instead of a per-survivor source review) to a CC ruling. It should route through the layer-resolution protocol instead.
• join_usage.py:3-5 drops child_usage.web_search when it joins the three stage records. The measured WebSearch counts stay in the per-stage child-usage-*.json files; the joined file should carry them too.
• From the GPT full read of this head (GPT-VERDICT-3c7a586f8.md, cited by name; ACK, its first three P3s match the CC's above): an adjacent recovery path, where missing refuter results appear as empty refutations in the critic summary. Recorded for the harness owner; not changed in this evidence PR.

Host evidence

Not applicable: nothing under evidence/hosts/ changes.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a
  version comment (no floating tags). (No workflow change.)
• New/changed workflows declare top-level permissions: {} and grant each job
  only what it needs (contents: read, or an explicitly justified addition). (No workflow change.)
• No secrets are printed, logged or committed; no new required secret was
  added without a documented owner.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved (not deleted,
  moved or overwritten).

Rebuild: under the CC's ruling (2026-10-15), the branch was rebuilt as ONE commit on a93ba0b8c holding the corrected tree and lease-pushed over the first head afcc08cbc. GitHub may keep the earlier commit in the PR timeline, and this description's edit history may keep its earlier revision; both are recorded residuals. <A> is the attempts directory, <W> the private work directory and <S> the co-op scratch directory.

Independence: this half stayed blind to #1177 (disclosure in RUN-RECORD.md). Both families read this PR.

🤖 Generated with Claude Code