artifacts/patent_phase0_sample50_ai_review.md

# Patent Phase-0 sample-50 AI review

Follow-up to `artifacts/patent_phase0_zero_yield_investigation.md`.

This run increased the retrieval target from the earlier observed 10 patents to
50 patents and applied an AI review pass to the high-attention candidate spans.
No Poucher targets, model-ready targets, training data, or publish artifacts were
changed.

## Commands

```text
python scripts/patent_substantivity_phase0.py \
  --sample-size 50 \
  --pages-per-query 8 \
  --delay 0.2 \
  --timeout 30 \
  --retrieval-source auto \
  --sample-out data/patent_phase0_sample_50.jsonl \
  --candidates-out data/patent_substantivity_phase0_candidates_50.jsonl \
  --report-out artifacts/patent_substantivity_phase0_report_50.md

jq -c 'select(.confidence >= 0.7 and ((.compound_ref_verbatim != null) or (.resolved_cas != null)))' \
  data/patent_substantivity_phase0_candidates_50.jsonl \
  > data/patent_substantivity_phase0_candidates_50_high_attention.jsonl

python scripts/review_patent_substantivity_tuples.py \
  --candidates data/patent_substantivity_phase0_candidates_50_high_attention.jsonl \
  --tuples-out data/patent_substantivity_reviewed_tuples_50_high_attention_hermes.jsonl \
  --audit-out artifacts/patent_substantivity_tuple_review_50_high_attention_hermes.jsonl \
  --summary-out artifacts/patent_substantivity_tuple_review_50_high_attention_hermes.json \
  --md-out artifacts/patent_substantivity_tuple_review_50_high_attention_hermes.md \
  --backend hermes \
  --batch-size 5 \
  --sleep 0.1 \
  --timeout 180
```

An Anthropic review attempt was also made against all 191 candidate spans, but
all calls failed with account credit errors. The successful AI review below used
the Hermes backend.

## Expanded retrieval

- Sample patents fetched: 50.
- Retrieval status: ok.
- Patents with numeric keyword spans: 36.
- Numeric keyword spans: 191.
- High-attention spans: 35.
- Spans with printed CAS references: 0/191.
- Spans with printed compound/example reference or CAS: 35/191.

Construct inventory:

- `fragrance_intensity`: 68.
- `longevity`: 42.
- `odor_intensity`: 37.
- `substantivity`: 16.
- `headspace_concentration_or_intensity`: 15.
- `duration`: 9.
- `persistence`: 2.
- `unknown_numeric_measure`: 1.
- `retention`: 1.

## AI review result

- Candidate file reviewed: `data/patent_substantivity_phase0_candidates_50_high_attention.jsonl`.
- Backend: Hermes.
- Processed spans: 35.
- Accepted single-molecule measured tuples: 0.
- Accepted tuples with Poucher CAS overlap: 0.
- Failures: 0.

The high-attention AI review rejected all spans. The main rejection pattern was
not a downstream extraction bug; it was that the stronger-looking spans still
described composition-level examples, fragrance mixtures, fixative/modulator
comparisons, analytical method setup, or procedural timepoints rather than one
named molecule with one measured substantivity/longevity/tenacity outcome.

## Current verdict

Increasing sample size improved retrieval coverage substantially, but did not
produce a usable Poucher-aligned patent tuple set. The bottleneck remains source
targeting: the current Google Patents query/fetch path mostly surfaces broad
formula examples, methods, figure captions, and mixture claims.

Recommended next retrieval change is to target table-bearing patents directly:
query and fetch examples/tables where a named material row and a measured
intensity, persistence, residual percentage, or evaporation profile are present
in the same table. The current span-level probe should not be merged into
training data.