# PDF ANATOP Benchmark Result

- document_id: `sovrient_market_intelligence_integration_strategy`
- generated_utc: `2026-03-17T20:40:11Z`

## Mode Summary

| Mode | Pass | Partial | Fail | Success Rate | Artifact Bytes | Approx Input Tokens |
| :--- | ---: | ---: | ---: | ---: | ---: | ---: |
| `raw_pdf_only` | 3 | 0 | 3 | 0.50 | 44147 | 2362 |
| `pdf_plus_plain_text` | 3 | 0 | 3 | 0.50 | 9448 | 2362 |
| `pdf_plus_anatop_twin` | 6 | 0 | 0 | 1.00 | 28556 | 7140 |

## Main Finding

For this document, raw PDF and plain text recover metadata and direct boundary claims, but fail on stable table structure and machine-declared claim topology. The ANATOP twin passes all six fixed tasks.

## Notes

- `raw_pdf_only` and `pdf_plus_plain_text` can recover metadata and some direct claims, but they do not preserve the two matrix tables as stable structured rows.
- `pdf_plus_anatop_twin` passes all six tasks because it exposes normalized tables, claims, and anchors directly.
- For this six-page document, the main gain is structural correctness and anchorability rather than dramatic byte reduction.