Preregister latest-local-model qualification on the RTX 4090 (8-K item extraction; Qwen3.8 + past-weeks releases) - #323

Scope

• Lane: lane:foundation. This is the workstation's model-qualification role, and it serves the north-star research input: catalyst extraction from 8-K filings.
• What: a frozen preregistration only. No model download, no SEC fetch and no GPU run happens in this PR. The GPU windows start after it merges.
• User decision (2026-09-26): GPU windows are approved, and only the latest SOTA models are in scope: the Qwen3.8 family and releases from the past weeks.

The protocol (blueprints/convergence-practice/local-inference-latest-20260926/)

Task. 8-K item extraction on the frozen 2020-03-02 daily-index cohort (the same cohort as blueprints/us-equities/catalyst-dataset).

• Filings are re-acquired from EDGAR through the catalyst-provenance EdgarTools SDK. The SEC contact is passed by pointer, fetching stays at no more than 5 requests/s, and the content is kept private.
• Labels are the filings' native header ITEMS. The model input has the SEC header stripped, so the model never sees the answer.

Metrics. Micro-F1 per filing, macro-F1 per item code, exact match, and JSON-valid rate; decode tokens/s, TTFT, peak device memory and failures.

Decision rule. An arm replaces the current production model (C0) only if all of these hold:

1. the paired-bootstrap lower bound (10,000 resamples, seed 20260926) of the F1 difference versus C0 is at least −0.02;
2. its median decode speed is higher than C0's;
3. its JSON-valid rate is at least 0.98;
4. it had no memory, deadline or server failure.

Speed alone does not pass.

GPU window safety.

• window.sh stops only nativestack-generation and always restarts it.
• Admission and reserve limits apply, with bounded monitors and an independent window deadline.
• Each arm runs on a test port under systemd-run memory limits.

Arms, pinned by Hugging Face revision and GGUF sha256:

The literal full-GPU and MTP profiles are recorded as resource_infeasible_on_host, with the arithmetic.

SOTA sources

• llama.cpp b11146 (7fe450e, = v0.5.0): https://github.com/ggml-org/llama.cpp/releases/tag/v0.5.0. --spec-type draft-mtp is in the local --help, and the ggml type table and src/llama-arch.cpp were read at 7fe450e.
• Hugging Face model cards and blobs APIs, read 2026-09-26:
  • unsloth/Qwen3.8-27B-GGUF @4ca72078
  • ggml-org/MiMo-V2.6-Distill-Qwen-9B-GGUF @81baddc3
  • prism-ml/Ternary-Bonsai-2-27B-gguf @b072e1d3
  • XingChen-AGI/Xing4.0-29B-A4B-GGUF @b8cd0a75
  • unsloth/Qwen3.8-Flash-Next-GGUF @38bb39ee
• EdgarTools 5.58.0: FilingHeader.parse_from_sgml_text, used as the label path (see the catalyst-dataset README).
• Pattern reused: blueprints/convergence-practice/gpu-inference/plan.json, with the same admission and reserve limits, frozen sampling and failure rule.

Evidence-class table

Commands

uv run --no-project --with jsonschema --with pyyaml python -m unittest tests.test_local_inference_latest_20260926 -v   # 44 OK
uv run --no-project --with jsonschema --with pyyaml python -m unittest    # full suite 5809 OK (writer run)
python3 scripts/validate_convergence.py --all-recorded --root . --json    # valid, 23 records (this one registered)
python3 scripts/validate.py                                              # passed
bash -n window.sh; shellcheck window.sh                                   # 0 findings

Review

• Two-family review, then one repair round. The findings came from Claude Opus 5.5 at max and GPT-6-Astra at max; the GPT-6 review ran before the account usage limit.
• Every high and medium finding was fixed:
  • the unit allow-list in window.sh;
  • bounded memory queries and cleanup, plus an independent window deadline;
  • server 5xx responses now disqualify an arm;
  • labels are read natively from the header before <FILER>;
  • the C1/C2 profiles were refit to host memory.
• Low findings and the remaining residuals are listed in the experiment record.

🤖 Generated with Claude Code