Record a cross-family discovery wave on local-model workloads beyond memory and RAG - #223

This records a cross-family discovery wave on one question: which local-model workloads, beyond memory and RAG, are worth running on a 64 GB M5 Pro for the US-equities north star and native agent engineering. It is discovery evidence only. Nothing named here was installed, downloaded or executed, and inclusion selects nothing.

Method

• One frozen packet went to two blind lanes. Neither lane saw the other's return or the catalogs.

  • The Claude lane ran on Opus 5.5 with web research.
  • The Codex lane ran on GPT-6 Astra at high effort with live web search, with --ignore-user-config and hooks disabled.

• The lanes disagreed on one item: tabular foundation models for predicting +200% movers. A blind judge (GPT-6 Sol, high effort, live web search) reviewed the two returns three times, with lane identities scrubbed:

  • original order: skip listed second;
  • letters swapped: skip listed second;
  • listing reversed: skip listed first.

  All three judgments chose skip, so listing position is counterbalanced for this question.

• The evidence directory holds the packet, the sealed lane returns with their original sha256, the adjudication inputs and key, the judgment record, provenance, native usage (null where unknown) and SHA256SUMS.

Outcome

catalogs/us-equities/local-model-workloads-20260924.json holds 9 repository decisions (registered through catalog_decisions.py) and 18 non-repository decisions (Hugging Face models pinned by revision).

Each item records its acceptance gate: a frozen dataset, a baseline and a metric. A trial becomes adoption only through an executed comparison.

Registration and counts

• Identities: 850 → 857. New: GLiNER2, FluidAudio, mlx-audio, granite-tsfm, PIGuard, tabicl, TabPFN.
• README.md, catalogs/us-equities/README.md, saturation-audit.json and manifests/evidence.json are updated (files rehashed, convergence record appended).

Limits

• Single judge, from the same model family as the Codex lane.
• Discovery depth only: no pinned-source review and no execution.
• Pins were taken after the lanes ran.
• The Claude lane's blindness is enforced by its prompt, not by a filesystem boundary.

Verification

• validate.py, validate_catalogs.py, landscape.py, catalog_decisions.py --check, evidence_manifest.py --check, validate_convergence.py --all-recorded (21 records), saturation_ledger.py --check, build_verdicts.py --check, build_ecosystem.py --check and git diff --check: all exit 0.
• The targeted test modules (204 tests) pass, including the blind-export isolation test.
• Full python3 -m unittest: see the CI result.

🤖 Generated with Claude Code