Should have been recognized
carrying on the IES
Measured results · 10 September 2026
Last modified: (dates/navigation repair). Content current as of: (existing dated report). Original content last edited: . Current claims have not been freshly reverified.
What the source shows, and what recognition got wrong
These are retained inputs from completed experiments. Each case shows the exact source image, the observed result, the independently defined expectation, the reason for failure, and the next decision.
Read the incident report archiveRequested bug investigations, stage ownership, causes, and retained evidence (outdated · September 2026)Read the historical recognition gate reportEnglish text conservation, the blocked sequence, and the evidence required to advance (outdated · September 2026)
Current decision
Valid output, rejected recognition
The baseline processed all 66 development pages—WS 13, VNAT 20, and IES 33—and produced structurally valid packages. All three document results were rejected after source comparison. Format validity is not the same as faithful recognition.
3 / 3development PDFs rejected
1 / 55IES notes correctly linked
15 → 11exact VNAT targets after refinement
0 / 23affected VNAT units fully correct
Evidence boundary. VNAT examples below are consumed development evidence. The raised-7 page is a synthetic positive control. The protected Simon Hall holdout has not been opened.
Case 1 · Text refinement
A replacement duplicated its neighbor
A higher-resolution OCR candidate contained the letters of the, but also merged them with neighboring on. The association step then attached that candidate to only one of the two base tokens it overlapped.
Read the separate Case 1 reportExact trace, responsible components, why advanced OCR failed, and the repair (outdated · September 2026)
Open full evaluated PDF22 pages · the document processed by the retained run
Open full raster reference20 pages · the source used for this crop
Recognized after refinement
carrying onontheIES
Why it failed
The candidate onthe overlapped the base tokens on and the, but replaced only the. The untouched on remained, producing on onthe. The same defect also produced a aresult and in inthe.
Current direction
First prove that correct English text survives refinement and canonical emission. The association defect is one known mechanism, but correcting it alone cannot establish reliable text.
A second warning: activity without improvement
“not. The pipeline changed it to “nof, then back to “not: two edits, zero net gain.Case 2 · Scanned appearance
Most italic characters were found, but no complete span was correct
This retained excerpt contains upright paragraphs 97, 98, 100, and 101. Paragraph 99 and its three lettered subparagraphs are italic in the source. Exact span boundaries—not a vague impression of slant—are the acceptance target.
Should have been recognized
4 of 4 exact italic spans, with the neighboring text upright and the source wording unchanged.
What the candidates recognized
Cuneiform: 377 italic characters found, 3 missed, but 0 of 4 exact spans.
Tesseract: 1 italic character found, 386 missed, and 0 of 4 exact spans.
Why Cuneiform failed
It supplied a strong character-level style signal, but changed the text: 17 omissions, 13 substitutions, and 1 addition across the excerpt. Fifteen source italic characters were therefore unavailable for an exact span, and three more were present but not marked italic.
Why Tesseract failed
Its text was closer—5 of 8 evaluated paragraphs were exact—but it almost never asserted italic style. Better text alone could not recover the required appearance spans.
Decision. Automatic reuse of either OCR engine for scanned italics was rejected. The strong partial Cuneiform signal is retained as evidence, but it is not reported as success.
Case 3 · Note identity and linkage
A raised “7” became punctuation, so the note remained ordinary text
This page is a deliberately simple synthetic positive control: one raised call in the body and one matching note at the bottom. It tests the smallest useful footnote relationship without relying on a confidential document.
Should have been recognized
The service gate was secured after the evening patrol7.
A note object named note-7, linked from character offset 53, containing: “Safety staff confirmed the hinge before closure; the damaged seal was entered in the overnight log.”
Recognized output
The service gate was secured after the evening patrol’.
7 Safety staff confirmed the hinge before closure; the damaged seal was entered in the overnight log.
The note was split into ordinary paragraphs. No call identity, note identity, parent, offset, or relation was emitted.
Zz. The lower label read 7 in both modes.Why it failed
Geometry found a plausible raised call and lower label, but geometry cannot prove glyph identity. The reader did not identify the tiny upper mark as 7, and pair similarity of 0.7882 was insufficient to assert a semantic link safely.
Decision
The geometry experiment was rejected: 0 of 3 positive controls were recovered. Thresholds will not be loosened merely to force this example through, because that would create unsupported footnote links elsewhere.
Proposed plan · next bounded unit
Repair one known association defect, then earn the right to rerun
Freeze independent controls
Use two control families plus one isolated VNAT candidate. Define expected outputs before changing the adapter.
Conserve every overlapped token
Implement one rule: a selected OCR candidate must map cleanly to all base tokens it replaces. Ambiguous many-to-one or partial-overlap proposals abstain.
Review the result independently
Check exact output, neighboring tokens, diagnostics, and the no-change path. Stop immediately on any regression or hidden duplication.
Reserve any broader rerun separately
If Unit 28 passes its local controls, authorize a full VNAT run as a new decision. Do not mix that cost or evidence into the adapter repair.
Keep the final holdout closed
WS, VNAT, and IES must all pass their source-based completion gates before Simon Hall is evaluated.
Scope. Italic reuse and note-glyph geometry remain rejected experiments, not parallel implementation work. The immediate proposal addresses the evidenced word-association defect only.
Provenance
These images can be checked byte for byte
The files above are direct copies of retained experiment inputs or source-review crops. They were not resized, redrawn, or recompressed for this page.
Show SHA-256 checksums
- VNAT neighbor-duplication crop
331daca5efbf57155688468c47504bc69294ff4861343a655aa5c2ca1fb2a404- VNAT net-zero crop
7e8d65f4465920d98561b1badf387af9e6c8f349b4b57944d1fa5a7a3b053799- VNAT italic excerpt
d2e348324f1cb2a992195a5050ca9c12f5d34a79c3e581facdf71dae0150d2c7- Raised-7 source page
7288d39a35c11d6e6ff14bc7f4e7a70beb7847ad0f66cc5516a1e8f82ecb0b15- Raised-7 call crop
1b7d60858f7ffaba82c03645a5153b1be98ba4d496a8d99ab31b219b868d0818- VNAT evaluated source PDF
ee83eb5bbb2b05fd762402c8ad8214a54a3b46637925c0dcd8dfe8abc85ecf84- VNAT raster reference PDF
18356059d277b4988000adbae0868d904aaff91fac9c8893f01d3ce75c2565a1- Raised-7 source PDF
a68ed24572e25bafb817e00584a1f69f895e7f0e965b62511e7180ee6d275e40