ORIGINAL LAB · ENGINE-DERIVED RESULTS

16 PDF extraction decisions measured against exact fixture bytes

Stable digital tables and safe exports are checked beside encryption, scans, unsupported filters, rotation, merged geometry, hidden optional content, opaque overlays and incremental revisions. Encoded PDF names and closepath fills must not bypass suppression.

Measured result

Worked synthetic benchmark; not a live customer result. Every PDF is generated from checked-in fixture code and processed at the fixed clock 2026-08-04T00:00:00.000Z. No OCR, remote upload, decryption, accounting or universal PDF support claim is made.

16 / 16expected decisions matched

5bounded PASS scenarios

1 + 10BLOCK + INDETERMINATE

0false PASS or sensitive-cell leaks

Engine source SHA-256:

37800bda34774cdbefcd1cc2fe662dc375195534051e4dd53653db9d5bf27ddb

Fixture source SHA-256:

81bd026e0da0ca0fe991a163f8665cd9f63bdb2828ece76b8d9e020718c9d9ff

Bar chart: five PASS, one BLOCK and ten INDETERMINATE scenarios matched; false PASS and sensitive-cell leaks are zero — Counts measure expected decision agreement only. They are not an accuracy percentage for arbitrary PDFs.

Scenario-by-scenario observations

Scenario	Policy	Expected	Observed	Evidence
Stable digital table	Automatic, page 1	PASS	PASS	1 table; 3 data rows
Two-page document	Automatic, pages 1-2	PASS	PASS	2 tables; 6 data rows
Page 2 only	Automatic, page 2	PASS	PASS	Only page 2 enters artifacts and policy hash
Irregular anchors	Manual boundaries 200,400	PASS	PASS	1 table; 3 data rows
Formula and HTML-like strings	Automatic	PASS	PASS	CSV/XLSX strings neutralized; raw values remain reviewable
Encrypted PDF	Automatic	BLOCK	BLOCK	`encrypted_pdf_not_decrypted`
Image-only scan	Automatic	INDETERMINATE	INDETERMINATE	No OCR; image/XObject coverage warning
Unsupported LZW stream	Automatic	INDETERMINATE	INDETERMINATE	Content stream decode incomplete
Rotated coordinate table	Automatic	INDETERMINATE	INDETERMINATE	Candidate rows exist, but rotation prevents PASS
Merged-like spanning text	Manual boundaries 200,400	INDETERMINATE	INDETERMINATE	Text crosses separator and sparse structure is visible
Hidden optional-content table	OCG is OFF in the catalog	INDETERMINATE	INDETERMINATE	Visibility is unresolved; candidate cells are suppressed
Opaque rectangle over table text	Filled overlay follows text	INDETERMINATE	INDETERMINATE	Covered candidate cells are suppressed
Incremental page revision	Active xref requires /Prev resolution	INDETERMINATE	INDETERMINATE	Stale and active revision cells are not exported
Hex-escaped OCG names	`/OCPropert#69es` and `/T#79pe`	INDETERMINATE	INDETERMINATE	PDF names are decoded before the visibility gate
Closepath opaque overlay	`re h f`	INDETERMINATE	INDETERMINATE	Covered candidate cells are suppressed
Hex-escaped incremental key	`/Pr#65v`	INDETERMINATE	INDETERMINATE	Stale revision cells remain suppressed

Method and reproducibility

Generate valid bounded PDF byte streams from the public fixture builders.
Run the production-candidate engine with a fixed clock, exact source identity and extraction policy.
Compare expected and observed verdicts, table counts, selected pages and coverage reasons.
Require hidden OCG, covered and incremental-revision candidate values to be absent from every artifact.
Recompute exact SHA-256 for JSON, CSV, XLSX, SVG, Markdown and the receipt.
Open representative fixture files with Poppler pdfinfo, render pages with pdftoppm, and compare rendered layout separately from parser output.
Repeat fixed-clock runs and require byte-identical artifacts.

node scripts/pdf_table_product_unit_20260804.mjs
node scripts/pdf_table_product_browser_20260804.mjs

Download exact benchmark evidence

Machine results

benchmark-results.json

ed92b602e1ac35c79b9f474cc1a67b345d71126ccdfd1ea82d1bf071b92be0b6

Formula-safe ledger

benchmark-results.csv

2a881f902086c107b50db0e6f3d78fbd6df753caef9403c03125004a2a88fd40

Methodology

benchmark-methodology.md

5f375e1d9459f9ef763099337e7a82e955e8c5de0f837818c02700907e4c2044

Coverage visual

benchmark-coverage.svg

0fa1df095411c33aeb6cb87b517431f73b10274f07365aa7dc17b8d22e5839ae

Receipt

benchmark-receipt.json

1ef10a11b0a6091830d201f9829dea137e387cc523ec51927eb0ac37d4b8d638

Receipt core: a4ac778bd9a5e650444a29aa1388ac04c8061bcecb6dea1dafacc3cc8a0a0028

What the benchmark does and does not establish

Established for these bytes

All sixteen expected decisions matched.
Selected-page and manual-boundary policies changed the receipt.
Rotation, merged-like, hidden, covered and stale revision candidates did not become false PASS.
Encoded names and closepath fills did not bypass the same suppression gate.
Hidden, covered and stale candidate cell markers were absent from every artifact.
Every produced artifact was hashable as exact bytes.

Not established

No universal PDF coverage, OCR or decryption.
No accounting, legal or semantic correctness.
No promise that a visual table maps cleanly to PDF text operators.
No safety claim after a source, policy, engine or artifact changes.

Sources and corrections

Reviewed 4 August 2026 against PDF / ISO 32000 resources, RFC 4180, OWASP CSV Injection and ECMA-376.

Report a fixture mismatch, stale hash, unsafe export or unsupported claim to [email protected]. Reference the scenario ID and exact source hash; do not attach confidential PDFs.