temporal/README.md

FilingsBench Item 4.02 temporal audit v0.2

This directory freezes 25 primary Form 8-K/8-K/A HTML documents from 25 issuers before running FinStruct. It is a label-free temporal audit set, not human-gold and not a market-prevalence sample.

The deterministic rule was fixed before prediction: query SEC EFTS for the exact phrase "Item 4.02" from 2026-05-01 through 2026-07-27; require the structured SEC items field to contain 4.02; sort by filing date and accession; exclude all issuers in the v0.1 candidate slice; keep the first filing per remaining issuer; take the first 25. manifest.json records the query, exclusions, limitations, record hash, and every downloaded document hash.

Recreate a new snapshot only when intentionally changing the observation window:

finstruct non-reliance-holdout freeze \
  --start-date 2026-05-01 --end-date 2026-07-27 --limit 25 \
  --exclude data/filingsbench_non_reliance_v0.1/records.jsonl \
  --output data/filingsbench_non_reliance_temporal_v0.2

Run system coverage without creating labels:

finstruct non-reliance-holdout run \
  data/filingsbench_non_reliance_temporal_v0.2/records.jsonl \
  --output results/non-reliance-temporal-v0.2-post-fix

The command verifies every local document hash, reports accuracy: null, and labels every saved row unlabeled. See the maintainer audit for the baseline-to-fix analysis and evidence limitations.