DAAF-Contribution-Community/daaf / .claude/skills/pdf-reading
pdf-reading
VerifiedReads research PDFs with page-anchored provenance. Covers DAAF's deterministic pdf_ingest CLI (unaltered poppler text as the quote anchor, a second extractor, page flags, OCR on scanned pages, 2000 px renders), the lit_verify card gate that checks each verbatim quote and number on its cited page, the vision cross-check policy and its dated vision-validated model list, two-column, table and scan handling, and the arXiv HTML lane (section-cited cards). Use whenever an agent must read, quote, or extract numbers, tables or figures from a PDF already on disk: OA papers saved by web_fetch.sh (exit 8), user-supplied papers, codebooks, methodology reports. For research reading it overrides the synced anthropic-skills:pdf, whose pypdf and pip-install recipes DAAF blocks. NOT for finding or downloading papers (scholarly-search, web-retrieval), or creating, merging or filling PDFs.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 36.0 KB |
| references/arxiv-and-html.md | markdown | 14.5 KB |
| references/evidence-card.md | markdown | 94.4 KB |
| references/extraction-strategies.md | markdown | 9.6 KB |
| references/long-papers.md | markdown | 3.0 KB |
| references/pdf-ingest.md | markdown | 28.7 KB |
| references/toolchain.md | markdown | 8.9 KB |
| references/vision-validation.md | markdown | 47.2 KB |
references/evidence-card.md
View on GitHub# The evidence card and the `lit_verify card` gate
How to write an evidence card that `scripts/lit_verify.sh card` can verify,
what every check and status means, and how to fix a failing card. Source of
truth: the header, `validate_card()`, `check_card()` and the selftest in
`scripts/lit_verify.py` (`TOOL_VERSION = "1.17.0-daaf"`; card
schemas `lit-verify-card-v1`, `lit-verify-card-v2` and `lit-verify-card-v3`,
result schema `lit-verify-card-check-v3`, re-derivation `rederive-v2`, quant
fields `quant-v1`, page labels `page-label-v1`, figure pages `vision-figure-v1`) and the ladder,
number and column-label rules in `scripts/pdf_text_common.py` (`ladder-v4`,
`numbers-v6`, `column-check-v1`, `normalize-v1`). `lit_verify report`, which
checks a review deliverable against the claims built from these cards, is
documented in the literature-review skill (`references/lit-verify-report.md`).
The card schema lives on this page: the Literature Review output contract
(`agent_reference/LIT_REVIEW_OUTPUT_CONTRACT.md`) covers what a review
promises its consumers (the claims table built from the cards, the audit and
settlement records) and points here for the card. If this page and the code
disagree, the code wins.
**Contents:** Syntax · Card anatomy · Front matter · The lit-verify block ·
Card schema v2 and section locators · Card schema v3: typed claims, quant
fields, page labels · Checks in order · Re-derivation ·
The match ladder · The quote-number rule · Identifying quotes · Page
classes · Coverage statuses · Numbers · Vision consistency · arXiv HTML
statuses · Integrity and warning statuses · Status index · Verdicts and
grades · card_check JSON · MANIFEST row · Exit codes · Fixing a failing card
· Selftest · Worked examples
## Syntax
```
bash /daaf/scripts/lit_verify.sh card <PAPER_DIR> <CARD_PATH> [--ingest-suffix S] [--timeout S]
bash /daaf/scripts/lit_verify.sh selftest <SCRATCH_DIR>
```
The orchestrator's `audit` and `route` subcommands, which check who read a
card and on which install route, and `settle`, which records a person's
comparison of the readings an audit lists with their PDF pages, are
documented in `vision-validation.md` § Gating.
- `PAPER_DIR` = `papers/{paper_key}/` written by `pdf_ingest`. It and its
parent (the papers dir, where MANIFEST.jsonl lives) must pass the
destination guard; its basename must be a paper key.
- `CARD_PATH` must resolve to a file **directly inside** `PAPER_DIR`, at most
5 MB, UTF-8. Name it `evidence_card.md`, then `evidence_card_a.md`,
`evidence_card_b.md` for revisions.
- `--ingest-suffix` pins an ingest version (`''` or `_a`…`_z`); default: the
latest complete ingest (the latest `ingest_manifest{S}.json`).
- `--timeout` (1.7.0): seconds per subprocess when re-deriving the anchors
from the source (default 120, as `pdf_ingest`); exceeded →
`SOURCE_REDERIVE_ERROR`, a recorded fail.
- No network calls, no model calls. It never reads the fenced reading copies
(`pages/`, `fulltext.md`); it matches the byte-for-byte anchor, after
re-deriving it from the stored source (§ Re-derivation).
- **The source must stay where `pdf_ingest` found it** (the manifest's
`source.path`, normally web_fetch's raw artifact in `web_fetches/`, or the
file in `user_pdfs/`): a moved or deleted source now fails closed
(`SOURCE_MISSING`).
## Card anatomy
`evidence_card{S}.md` = YAML front matter + free prose sections (the
Literature Review card's: coverage ledger, question and contribution, design and identification,
sample, measures, findings table, heterogeneity, limitations, appraisal,
open questions, relevance and directness, reader confidence, and the
**text/vision discrepancy log**) + **exactly one** fenced block whose info
string is `lit-verify` (three or more backticks or tildes), holding strict
JSON. The gate reads only that block. Zero blocks, two blocks, an unclosed
block, invalid JSON, duplicate keys, NaN/Infinity, unknown fields or wrong
types → exit 8. JSON allows no comments, so commentary goes in prose.
## Front matter
Free-form for the gate. The Literature Review card lists `paper_key`, `doi`, `openalex_id`,
`citekey`, `title`, `authors`, `year`, `venue`, `evidence_grade`,
`source_pdf_sha256`, `extractors`, `vision_mode`, `reader_model`, `read_at`,
`review_type`. A v2 card writes `source_sha256` (in place of
`source_pdf_sha256`) and, on arXiv HTML, `source_version` (`v4`) and the
source as "arXiv HTML vN". This skill adds the vision-gating fields that the block does
**not** define (putting them in the block would be an unknown field, exit 8):
| Field | Value | Who sets it |
|-------|-------|-------------|
| `reader_model_source` | `self-report` | The reader. The audit result (`transcript-confirmed`, `mismatch` or `unknown`) is recorded outside the card by `lit_verify audit`, in `vision_audit{S}.json`, because editing a checked card breaks the SHA-256 its `card_check` records (`vision-validation.md` § Gating) |
| `reader_route` | One of the route names in `vision-validation.md` § Routes, copied from the dispatch's `ROUTE:` line (which the orchestrator copies from `lit_verify route`) | The reader, copying the dispatch. `lit_verify audit` compares it with the route `lit_verify route` recorded |
| `vision` | `validated` or `unavailable` (the wording of CLAUDE.md § User Preferences, "PDF page-image cross-checking"); `untested-check` or `untested-accepted` when the reader, not on the list, read page pictures at the researcher's recorded choice (the dispatch's `VISION:` line) | The reader, from the list check and the `VISION:` line |
`lit_verify` does **not** check any front-matter field.
## The lit-verify block
| Field | Required | Rule |
|-------|----------|------|
| `schema_version` | yes | `"lit-verify-card-v1"` (or `"lit-verify-card-v2"` or `"lit-verify-card-v3"`, next sections); anything else → exit 8 |
| `paper_key` | yes | Non-empty; must equal the PAPER_DIR basename (else `PAPER_KEY_MISMATCH`, a fail) |
| `source_pdf_sha256` | yes | 64 lowercase hex; must equal the ingest manifest's `source.sha256` (else `SOURCE_HASH_MISMATCH`) |
| `ingest_suffix` | no | `""` or `_a`…`_z`; a mismatch is only a warning (hashes decide integrity) |
| `reader_model` | yes | Non-empty string, recorded, not judged |
| `vision_mode` | yes | `auto`, `all-pages` or `off` |
| `vision_validated` | yes | JSON boolean |
| `vision_list_version` | yes | String or `null` (write the list version from SKILL.md) |
| `vision_override` | no (v1, v2 and v3) | `check` or `accept`: the researcher's recorded choice to let a reader that is not on the list read page pictures, copied from the dispatch's `VISION:` line (`override-check`, `override-accept`). Needs `vision_validated: false`, a `vision_mode` other than `off` and a PDF card (else exit 8). Omit it otherwise. The gate only reads the declaration; `lit_verify audit` binds it to the record (`vision-validation.md` § The researcher's choice for an untested reader) |
| `coverage` | yes | Array, one object per page: `page` (int ≥ 1), `anchors` (array of strings), `summary` (string), optional `sparse` (bool, advisory), optional `vision_only` (bool). A page listed twice → exit 8. A page absent from 1..N is a recorded fail (`MISSING_PAGE`), not a parse error |
| `claims` | yes | Array of objects: `claim_id` (unique, `[A-Za-z0-9][A-Za-z0-9_.:-]{0,63}`), `page` (int), optional `page_end` (exactly `page + 1`), `quote` (string; `null` or empty only when `source` is `vision`), `numbers` (array of strings, as printed), `source` (`text`, `vision`, `both` or `ocr`), optional `table_ref` (string or null), optional `claim` (string) |
That is schema v1, which stays **PDF-only** and checks exactly as before.
## Card schema v2 and section locators (1.4.0)
`lit-verify-card-v2` is accepted beside v1. It names the source kind, so one
schema covers both anchors. Use v2 for every arXiv HTML card; a v2 card with
`anchor_kind: "pdf"` checks exactly as the same v1 card.
| Field | `anchor_kind: "pdf"` | `anchor_kind: "arxiv_html"` | Rule |
|-------|----------------------|-----------------------------|------|
| `schema_version` | `"lit-verify-card-v2"` | same | — |
| `anchor_kind` | `"pdf"` | `"arxiv_html"` | Required; must equal the ingest manifest's `anchor_kind` (a manifest without the field is `pdf`), else `ANCHOR_KIND_MISMATCH` (fail; nothing else is checked) |
| `source_sha256` | required | required | 64 lowercase hex = the manifest's `source.sha256` (the raw HTML's hash on HTML). v1's `source_pdf_sha256` is an **unknown field** in v2 (exit 8) |
| `source_version` | absent or `null` (a value → exit 8) | **required**, `v` + digits (`"v4"`) | Must equal the manifest's `arxiv.version`, else `SOURCE_VERSION_MISMATCH` (fail) |
| `paper_key`, `ingest_suffix`, `reader_model`, `vision_mode`, `vision_validated`, `vision_list_version` | as v1 | as v1, but `vision_mode` `"off"` or `vision_validated` false (both declared available → exit 8) | — |
| `coverage` | per page, as v1 | one object per **section**: `section`, `anchors`, `summary`, optional `sparse`; `page` and `vision_only` are unknown fields; a section listed twice → exit 8 | — |
| `claims` | as v1 | `claim_id`, `section`, optional `section_end`, `quote` (non-empty string), `numbers`, `source` (`"text"` only, else exit 8), optional `table_ref`, `claim` | `section_end` must differ from `section` (exit 8) and be the **next** unit in document order, else `SECTION_END_NOT_NEXT` (fail; the quote is then checked in its section alone) |
**Section locators.** A `section` is a unit label exactly as `pdf_ingest`
prints it (manifest `units[].label`; the fulltext marker `section=…`):
`front`, a heading number (`3.2`, `A.1`, `S4.1`, thesis `1.3.1`), a LaTeXML
id fallback (`S1.SSx1`, `bib`), `unitN`, or a de-duplicated `1#2`. Pattern
`^[A-Za-z0-9][A-Za-z0-9._:#-]{0,79}$` (anything else → exit 8). A label the
ingest does not have is `SECTION_NOT_FOUND` (fail). Every unit 1..N needs a
coverage entry (`MISSING_SECTION`, fail). Uniqueness of a quote is judged
across all units and their boundaries. How units and labels are built:
`arxiv-and-html.md` § Units and labels.
## Card schema v3: typed claims, quant fields, page labels (1.9.0)
`lit-verify-card-v3` is accepted beside v1 and v2 (both check exactly as
before). It is v2 (`anchor_kind`, `source_sha256`, `source_version`; PDF or
arXiv HTML) plus typed claims and an optional per-paper note, so the
Literature Review corpus (`lit_corpus claims build`) can read a paper's
claims, limitations, checklist answers and estimates as data. Use it for
Literature Review reading (`literature-review` skill, `references/paper-notes.md`
and `references/formal-review.md`).
| Where | Field | Req | Rule (structure errors → exit 8) |
|-------|-------|-----|------|
| claim | `claim_type` | yes | `finding`, `method`, `limitation`, `author_open_question`, `definition`, `measure`, `appraisal`, `citation_use` |
| claim | `limitation_kind` | iff `limitation` | `author_stated` or `reader_assessed`. Replaces the interim claim-text prefix "Limitation (author-stated): …" |
| claim | `appraisal` | iff `appraisal` | `{guide, item, answer}`: `guide` one of `rob2`, `robins-i-v2`, `robins-e`, `wwc5`, `amstar2`, `jbi-qual`, `casp-qual`, `jbi-xs`, `axis`, `theory`, `other`; `item` and `answer` non-empty, at most 40 characters, in the tool's own numbering and scale (`"1.2"`, `"PY"`). Replaces the interim prefix "rob2 1.2 PY: …". No question tables, answer scales or ratings: the researcher makes the calls |
| claim | `citation_use` | iff `citation_use` | `{cites_key, relation}`: `cites_key` a paper key or `"external:<doi or W-id>"`; `relation` one of `builds_on`, `uses_method`, `extends_result`, `contrasts`, `disputes`, `background`. The quote is the citing sentence |
| claim | `quant` | no | `{estimate?, se?, ci_low?, ci_high?, n?, ci_level?, metric?}`, every value a non-empty string (numbers as printed, at most 40 characters; `metric` at most 80); another key → exit 8 |
| claim | `page_label` | no, PDF only | The page number as **printed** on the cited page (`"7"`, `"iv"`, `"S3"`; pattern `^[A-Za-z0-9][A-Za-z0-9.\-]{0,11}$`). On an arXiv HTML card → exit 8 |
| claim | `vision_reading` | no, PDF only (1.15.0) | How sure the reader is of the numbers it read from a page picture: `exact` (the default when absent), `estimated` (for example read off a chart's axis) or `uncertain`. Allowed only with `source` `vision` or `both` (else exit 8). It applies to every number of the claim, so numbers read with different confidence go in separate claims. The figure-page rule reads it (§ Vision consistency) |
| block | `paper_note` | no | `{design, study_weight, weight_reason?, limitation_claim_ids, directness?}`: `design` one of the paper-notes design labels (`rct` … `other`); `study_weight` `substantial`, `moderate`, `limited` or `null` (a `weight_reason` is required with a weight); every id in `limitation_claim_ids` must be a claim of this card with `claim_type: limitation`; `directness` parts `population`, `intervention`, `comparator`, `outcome`, `setting`, `time`, each `direct`, `partly` or `indirect` |
Typed claims are claims: every quote, whatever its type, goes through the
ladder, the identifying rule, the quote-number rule and the number checks.
A checklist answer or a citing sentence is therefore page-checked like a
finding. A v1 or v2 card with any of these fields is exit 8 (unknown field).
**Quant fields (check 7, `quant-v1`).** Each numeric field (`estimate`, `se`,
`ci_low`, `ci_high`, `n`) must be one of the claim's own `numbers` (same value
and digits; parentheses must agree, except for `se`), so no estimate reaches
the corpus without the number check having run on it. The item (kind
`claim_quant`, with `quant_field` and `matched_number`) takes the outcome of
that number:
| Status | Severity | When |
|--------|----------|------|
| `QUANT_OK` | pass | the matched number is `NUMBER_OK` |
| `QUANT_NUMBER_PARTIAL` | partial | the matched number is partial (OCR, OCR layer, vision only, conversion error) |
| `QUANT_NUMBER_REVIEW` | review | the matched number needs a person (not near the quote, other precision, sign-ambiguous) |
| `QUANT_NUMBER_FAILED` | fail | the matched number failed |
| `QUANT_NOT_IN_NUMBERS` | fail | the value is none of the claim's numbers, or not one number |
| `QUANT_N_NOT_INTEGER` | fail | `n` is not a whole, non-negative count |
| `QUANT_CI_ORDER` | review | `ci_low` is larger than `ci_high` |
| `QUANT_ESTIMATE_OUTSIDE_CI` | review | the estimate lies outside its own interval |
**Printed page label (check 8, `page-label-v1`, PDF only).** `PAGE_LABEL_OK`
(pass) when the label prints as a whole word in the first or last two
non-empty lines of the cited page's re-derived text (its header or footer);
otherwise `PAGE_LABEL_UNCONFIRMED` (warning: the label is kept but marked
unconfirmed, because some PDFs print no folio). A label found only in body
text ("Table 7") does not confirm it. Cite "p. 7" from a confirmed label;
otherwise check the page first.
The reasons on these items are plain language for the researcher, for
example: "The AI reader recorded 0.09 as this claim's estimate, but 0.09 is not
one of the numbers the claim lists, so DAAF has not checked it against the
page. Add it to the claim's numbers exactly as printed, or correct the
estimate."
The card_check (`lit-verify-card-check-v3`, additive) gives `claim_quote`
items a `claim_type` (null on v1 and v2 cards), adds the `claim_quant` and
`page_label` item kinds, `counts.claims_by_type` and `counts.quant_fields`,
and `rule_versions.quant` and `rule_versions.page_label`.
## Checks in order
1. **Integrity.** Card `source_pdf_sha256` = manifest; card `paper_key` =
manifest `paper_key` = PAPER_DIR basename; each anchor file's SHA-256 =
the manifest record and the page_index record; page_index.parquet = its
manifest record; each OCR text file used = its record (a mismatching OCR
file is not used); each lane B file = its manifest and page_index record
(1.3.0; a mismatch is `LANE_B_HASH_MISMATCH`, a fail, and that page's
lane B is not used); the ingest manifest = the hash in the pdf_ingest
MANIFEST row (no row: `NO_INGEST_ROW`, a fail from 1.7.0). An ingest made with another
`NORMALIZE_VERSION`: warning. v2 (1.4.0): card `anchor_kind` = ingest
(`ANCHOR_KIND_MISMATCH`), `source_version` = `arxiv.version`
(`SOURCE_VERSION_MISMATCH`), every complete ingest of the key has one
anchor kind (`MIXED_ANCHOR_KINDS`); on HTML each `rows{S}/uNNN.txt` = its
manifest and page_index record (`TABLE_ROWS_HASH_MISMATCH`; the file is
then not used) and the `HTML_EXTRACT_VERSION` is one the tool knows
(else warning `HTML_EXTRACT_VERSION_DIFFERS`). The statuses:
§ Integrity and warning statuses.
1b. **Re-derivation** (1.7.0): the anchor text, lane B and table rows are
re-derived from the stored source and compared byte-for-byte; from 1.8.0
also the header/footer lines, the grade-relevant page flags and OCR
acceptance, from the same hashed bytes (§ Re-derivation).
2. **Quote ladder** for every coverage anchor and claim quote, then the
**quote-number rule** on every claim quote that passed (`ladder-v4`).
3. **Identifying-quote rule** on every quote that matched.
4. **Coverage** of every page 1..N, then the paper-level OCR-layer cap.
5. **Numbers** of every claim.
6. **Vision consistency** with what the card declares and the embedded
vision list, and (1.8.0 part 5) a vision reading on a scanned page goes
to human review (`VISION_ON_SCAN_PAGE`).
## Re-derivation (1.7.0 `rederive-v1`; 1.8.0 `rederive-v2`)
Hashes alone prove only that the manifest, page_index and anchor files agree
with **each other**: in review cycle 1 a hand-written ingest set with a
fabricated anchor sentence and consistent hashes (and a MANIFEST row) passed.
So before any quote is matched, `lit_verify` re-derives the text it is about
to trust from the source the ingest recorded, with `pdf_ingest`'s own code
(`pdf-ingest.md` § Functions lit_verify reuses):
| Source | What is re-derived and compared byte-for-byte | Mismatch |
|--------|-----------------------------------------------|----------|
| The stored source (`source.path`) | Must resolve under the DAAF root and outside the forbidden prefixes (`.claude`, `.git`, `/tmp`, …), be a file of at most `SOURCE_MAX_BYTES` (100 MiB, `pdf_ingest`'s default `--max-bytes`) and hash to the manifest's `source.sha256` | `SOURCE_MISSING` (not recorded, outside the root, moved, deleted, oversize or unreadable) or `SOURCE_FILE_HASH_MISMATCH` (another file at that path): fail; nothing is re-derived, so no anchor is trusted |
| PDF anchor | `pdftotext` page by page with `pdf_ingest.anchor_argv()` (`LC_ALL=C`), **every** page (the uniqueness rule reads them all), 4 in parallel (`REDERIVE_WORKERS`); the page count from `pdfinfo` | `ANCHOR_REDERIVE_MISMATCH` (fail) |
| PDF lane B | `pdf_ingest.lane_b_page_text()` (pypdfium2) on every page that has a lane B file | `ROW_SOURCE_REDERIVE_MISMATCH` (fail; that page's lane B is then not used, so no number takes the row path there) |
| arXiv HTML | `pdf_ingest.html_units()` / `unit_evidence()` (`HTML_EXTRACT_VERSION`): unit count, labels, unit text, unit flags and the watermark's arXiv id and version; each `rows{S}/uNNN.txt` | `ANCHOR_REDERIVE_MISMATCH`; rows: `ROW_SOURCE_REDERIVE_MISMATCH` (the file is not used) |
- A `pdftotext` or pypdfium2 version other than the one the ingest recorded
is `EXTRACTOR_VERSION_DIFFERS` (warning): the byte comparison still decides.
- `pdftotext`, `pdfinfo` or (1.8.0) `pdfimages` missing → exit 10. One of
them failing, or exceeding `--timeout`, or HTML that no longer decodes →
`SOURCE_REDERIVE_ERROR` (fail; raise `--timeout` if it timed out).
pypdfium2 missing or failing → `ROW_SOURCE_NOT_REDERIVED` (warning; lane B
is not used, so no number takes the row path).
- **Read once (1.8.0, `rederive-v2`):** the source file is read
once and those bytes are hashed; every lane re-derives from the same
bytes, never from the path again (poppler reads a sealed memfd as
`/dev/fd/N`; pypdfium2, pdfplumber and the HTML lane read the bytes). A
file swapped in at the path after the hash changes nothing. Without
memfd sealing (non-Linux) → `SOURCE_REDERIVE_ERROR` (fail; no fallback to
the path).
- **Header/footer, page flags and OCR (1.8.0):** from
the re-derived anchors, with `pdf_ingest`'s own rules (its
`header_footer_records()`, `text_flag_evidence()`, `caption_evidence()`,
`image_evidence()`, `ocr_layer_evidence()`):
- the header/footer lines a declared span may skip; a manifest list that
differs → `HEADER_FOOTER_REDERIVE_MISMATCH` (fail), and only the
re-derived lines are skipped (a forged record had let a span quote skip
a printed line);
- the page flags that move a grade (`GRADE_FLAGS`: `table`, `numeric`,
`scanned`, `ocr_layer_suspected`) and the truly-scanned image evidence,
from the text, `pdfimages -list` and `pdfinfo` on the same bytes;
pdfplumber only where needed (`find_tables` on a caption-less `table`
page a numbered claim cites; elsewhere `table` changes no item and is
kept as recorded; `page.images` on a scanned page of a document
pdfimages lists no image for, and on an `ocr_layer_suspected` page
pdfimages alone does not confirm). A flag or page class that differs →
`FLAGS_REDERIVE_MISMATCH` (fail); the gate uses the re-derived flags
(a forged removal of 70 `ocr_layer_suspected` flags had lifted the
paper-level OCR-layer cap). pdfplumber unavailable → `FLAGS_NOT_REDERIVED` (warning) and
a conservative reading: an unconfirmed `table` closes the row path, a
recorded `ocr_layer_suspected` is kept, and a scanned page in a document
pdfimages lists no image for counts as truly scanned (so a manifest made
with pdfplumber can read such a page as sparse, and that class
difference is a `FLAGS_REDERIVE_MISMATCH`; the DAAF image ships
pdfplumber). Residual: a page where only pdfplumber's image box reaches
80% of the page area (pdfimages' computed area stays under 80%, whether
or not it lists the image), with what that area raised
(`ocr_layer_suspected`, or the truly-scanned class) removed from the
manifest: the pdfplumber probe does not run on such a page in a document
pdfimages lists images for (measured 2026-10-27: no such page in the
five bake-off PDFs);
- OCR text on a page that re-derives with ≥ 200 non-whitespace anchor
chars (not scanned, so `pdf_ingest` never OCRs it) → `OCR_ON_TEXT_PAGE`
(fail; the OCR text is not used).
- **Not re-derived:** OCR text itself (Tesseract on a 300-DPI render costs
seconds per page). Its files are hash-checked only, and a paper whose OCR
text is used carries one `OCR_NOT_REDERIVED` warning; OCR evidence is
already capped at PARTIAL. So forged OCR text on a page whose re-derived
anchor has fewer than `SCANNED_MAX_CHARS` (200) non-whitespace characters
is graded at most PARTIAL (`PASS_OCR`, with the `OCR_NOT_REDERIVED`
warning): the paper never reaches `FULL_TEXT_VERIFIED` on OCR evidence (a
residual the user accepted on 2026-10-27). The display-only
flags (figure, garbled, rotated, cjk_risk, low_agreement, ocr) move no
verdict.
- `rederive.source_kind` (recorded only; no verdict effect): `web_fetch`
(a web_fetch `MANIFEST.jsonl` beside the source names it with its hash;
`web_fetch_row` gives `manifest_line`, `final_url` and `fetch_timestamp`),
`web_fetch_row_conflict` (rows name the file with another hash),
`user_supplied` (no such row), `synthetic` (selftest only; the CLI never
produces it) or `unknown` (that MANIFEST could not be read); null when no
source was read (`SOURCE_MISSING`, `SOURCE_FILE_HASH_MISMATCH`).
`source_kind_ingest` is what `pdf_ingest` recorded at ingest time.
- **What the gate does not prove (the whole-chain residual).** Re-derivation
proves that the card and the ingest agree with the bytes on disk at
`source.path`, not that those bytes came from the publisher. An agent with
Write can author a source file, a matching web_fetch MANIFEST row, the
ingest set and the card, and that self-consistent forgery passes.
Re-derivation (the anchors, lane B and, from 1.8.0, the header/footer
lines, OCR acceptance and the `GRADE_FLAGS` `table`, `numeric`, `scanned`
and `ocr_layer_suspected`) catches a forged ingest over a genuine source;
OCR text stays hash-checked only (above). It cannot catch a forged source.
The orchestrator's audit can: compare `source_kind` and `web_fetch_row` in
the `card_check` with its own record of fetches (its transcript and the
web_fetch MANIFEST rows of the fetches it dispatched). A `web_fetch` source
whose line, URL and timestamp
match no dispatched fetch, any `web_fetch_row_conflict`, a `user_supplied`
source the user did not supply, and `synthetic` or `unknown` on a real card
are reported, not trusted. `lit_verify audit` runs this check (step 6,
`vision-validation.md` § Gating): a `web_fetch` source must be the very
row (line, file name, hash, final URL) of a fetch under the project's
`web_fetches/`, a `user_supplied` source must sit under its `user_pdfs/`,
and the file must still hash to the recorded value; anything else is the
outcome `chain_unconfirmed`. A forger who can write the project's own
`web_fetches/` record still passes: the check narrows the residual, it
does not close it.
- `card_check` records the outcome in `rederive` (§ card_check JSON) and the
MANIFEST row a summary. On the nine bake-off cards every re-derivation
matched, in 0.06–0.97 s under 1.7.0 (0.968 s for the 192-page report)
and 0.065–1.67 s under 1.8.0 (1.665 s for that report).
- Selftest and unit-test ingests are built in code and have no source file:
`build_synthetic_ingest()` registers the planted text (and, 1.8.0, what
the image and table tools would report for it) in an in-process registry
(`SYNTHETIC_SOURCES`) that the CLI never fills, so no file or argument can
reach it.
**If a card fails here, do not edit the card.** Report it: the ingest or its
source changed after the ingest, and the orchestrator re-ingests (or restores
the source to its recorded path).
## The match ladder (`ladder-v4`)
Evaluated for a quote cited to page P (or P..P+1 with `page_end`); first hit
wins.
| Tier | Meaning | Severity |
|------|---------|----------|
| `REDACTED_SPAN` | The quote holds a scrub token, `[REDACTED:`, or fence text: it was copied from a reading copy, not the page | review |
| `PASS` | Exact substring of NORMALIZE(page text) on a cited page | pass |
| `PASS_SKELETON` | Exact on SKELETON (absorbs fused compound hyphens, small-caps splits) | pass |
| `PASS_SPAN` | `page_end` = P+1 declared and the quote runs from the end of P into the start of P+1 (recorded running footers/headers between them skipped) | pass |
| `PASS_OCR` | Exact or skeleton match in a cited page's OCR text only | partial |
| `NEAR` | Best windowed rapidfuzz `partial_ratio_alignment` score on a cited page ≥ 95 (anchor or OCR text) | review, **never a pass** |
| `WRONG_PAGE` | Exact or skeleton match only on other pages, other pages' OCR, or across a page boundary not declared with `page_end` | fail |
| `NOT_FOUND` | None of the above | fail |
Long quotes (over 64 chars) are scored in 64-char windows (step 32); the
score is the minimum window score, and the windows must align in order
within 1.25 × the quote length + 10 chars, or the quote is not NEAR
(stitched fragments are not a near match).
The tiers are those of `ladder-v1`; `ladder-v2` (1.6.0) adds the
quote-number rule below on top of them, `ladder-v3` (1.8.0) widens it, and
`ladder-v4` (1.8.0, part 3) makes its dash check symmetric.
## The quote-number rule (`ladder-v2`, 1.6.0; `ladder-v3` and `ladder-v4`, 1.8.0)
SKELETON deletes `-`, so without this rule a claim quote "slope was 0.164"
passed on a page printing "slope was −0.164", and "510%" on "5–10%". Now a
**claim quote** that passes on any tier (`PASS`, `PASS_SKELETON`,
`PASS_SPAN`, `PASS_OCR`, `PASS_OCR_LAYER`, `PASS_CONVERSION_ERROR`) must
print, **in order**, the same number tokens, sign included, as the page span
it matched (`pdf_text_common.quote_number_check()`; `quote_numbers_agree()`
is its 3-tuple form, the `ladder-v2` name). The quote is parsed as
written; the page side is the page's *number view* (NORMALIZE coordinates
with superscripts, thin-space groups and a sign lost to a line-break rejoin
put back; the whole tokens that overlap the matched span). A plain space
between digit groups is read both ways; agreement under either reading
counts. Otherwise the quote is `QUOTE_NUMBERS_DIFFER` (review): the item
keeps the pass it replaced in `status_if_confirmed`, and carries
`quote_numbers`, `page_numbers` and `quote_numbers_agree`. A passing quote
whose matched span cannot be located is graded the same way (fail closed).
Coverage anchors are not checked (they carry no claim). A
`QUOTE_NUMBERS_DIFFER` quote also leaves the row path (§ Numbers).
`ladder-v3` (1.8.0;
`pdf_text_common.quote_number_check()`) adds three things:
- **Marks.** Each token's superscript and footnote letter must agree too, so
"p < 10³" over a printed "p < 10⁻³" (the superscript minus dropped) and
"0.164a" over "−0.164a" are caught.
- **Sign-shaped dashes, position by position.** Every `-` the span prints
directly (or after one space) before a digit or `.digit` must stand at the
same place in the quote, compared on the SKELETON characters of the two
(equal by construction). Every dash and minus that NORMALIZE folds counts,
the superscript minus included, and so does a dash the line-break rejoin
deleted (`K-` + line break + `12`). This catches a deleted dash even where
no number token is parsed ("0.164ab" over "−0.164ab"). A dash outside the
span (before its first character) is not checked; under `ladder-v3` an
extra dash in the quote was allowed (see `ladder-v4` below).
- **Spaced dashes compare without the sign.** When the page token is
sign-ambiguous (§ Numbers), or a spaced dash precedes the quote's token,
the two tokens compare digits without the sign. The dash itself is still
checked by the position rule.
`ladder-v4` (1.8.0, part 3) makes the dash
check **symmetric**. A dash the quote prints before a number where the page
prints no dash at that SKELETON position is `QUOTE_NUMBERS_DIFFER` too, when
it can read as a sign: a spaced dash with no digit, `%` or dash before it
("rose by – 0.30" over "rose by 0.30"), or a dash glued to the digit with no
letter or digit before it ("(Girls)−0.30" over "(Girls) 0.30"). Without it
such a quote was certified verbatim while reading as a sign flip (its
declared numbers were still checked). Compound and range dashes stay
allowed: "K-12" over a rejoined "K12", "5-10" over "5 10", "[26 - 28]"
over "[26 28]" (the number rules never read them as a sign either). On the
page side any dash at that position counts as printed: inside the span,
just before it, or deleted by the line-break rejoin.
**Fixing `QUOTE_NUMBERS_DIFFER`.** Compare `quote_numbers` with
`page_numbers` in the item; the reason says whether the digits, a mark
(superscript or footnote letter) or a sign-shaped dash differ, and whether
the page prints the dash ("the matched page span prints a sign-shaped dash")
or the quote added it ("the quote prints a sign-shaped dash"). Usually a
minus, a range dash, a superscript or a digit was lost, or a dash added,
when the quote was retyped. Re-copy the quote from `pages/pNNN.txt` with
every number exactly as printed (`−0.164`, `5–10%`, `10⁻³`). If the page
itself prints the dash as something that is not a sign (a binary minus in a
formula, an exponent), leave the item for the human check. Measured on 2,407
genuine quotes from the eight bake-off papers (five PDFs and three arXiv HTML
papers, 2026-10-23 to 2026-10-27: 1,879 quote windows sampled from their
ingests plus the 528 claim quotes and coverage anchors of their nine cards),
the rule raised 2 false
alarms in 2,407 genuine quotes under `ladder-v2` (both quotes starting with
such a minus) and 0 under `ladder-v3` (the binary minus there is now
sign-ambiguous) and `ladder-v4`. Of 217 planted minus deletions it
flagged 208 under `ladder-v2` and 217 under `ladder-v3` and `ladder-v4`; of
planted added dashes, `ladder-v4` flags 999 of 999 spaced after a word and
18 of 18 glued after a bracket (`ladder-v3`: 0 of each).
## Identifying quotes
A matched quote must be **identifying**: at least `MIN_QUOTE_CHARS = 40`
normalized characters AND `MIN_QUOTE_WORDS = 6` words AND its SKELETON occurs
**exactly once** across all pages' anchor text and page boundaries (plus all
OCR text when it matched in OCR). Both thresholds are **PROVISIONAL**,
kept after the S3 bake-off (2026-10-19): on five real papers the length
floor almost never decided the outcome; uniqueness did. Random 6-word,
40-character spans were non-unique 3–11% of the time in four papers and 41%
in a 192-page report full of repeated table headers and notes ("Sample
Control Mean Impact SE p-value Effect" recurs 88 times), and 20-word spans
were still non-unique 32% of the time there. So on such papers, quote a
distinctive sentence, never a table header or boilerplate note. Otherwise the
status is `NON_IDENTIFYING` (a claim quote fails; a NEAR on a short quote is
NON_IDENTIFYING too).
**Duplicated pages (1.1.0).** When a page has no unique line of its own
because its text recurs on another page (a scan that repeats a page), no
anchor can be unique. A full-length anchor (at least 40 characters and 6
words) that matches exactly on that page then covers it as
`COVERED_DUPLICATE`. Claim quotes get no such waiver: cite a claim from a
duplicated page as a quote that includes the page's distinctive context, or
flag it in prose.
**Sparse waiver (Q-B).** On a sparse or truly scanned page, a coverage anchor
that is too **short** but matches exactly (`PASS`, `PASS_SKELETON` or
`PASS_OCR`) on that page is `COVERED_SPARSE`, if it is unique or the page has
no unique line of its own. On a normal page a non-identifying anchor does not
cover the page.
## Page classes (user decision Q-B, 2026-10-14)
| Class | Rule | Effect |
|-------|------|--------|
| `normal` | Not flagged `scanned` | — |
| `scanned` (truly scanned) | Flagged `scanned` AND one embedded image covers at least `TRULY_SCANNED_IMAGE_AREA_MIN = 0.50` of the page area (**PROVISIONAL**; from 1.8.0 the flag and the image evidence are re-derived from the source bytes, § Re-derivation; before, they were parsed from pdf_ingest's flag evidence); also when image evidence is unavailable (pdfimages lists no image in the document and pdfplumber is unavailable; a missing pdfimages is exit 10), the conservative reading | Adds a `SCANNED_PAGE` partial item: the card is at best PARTIAL |
| `sparse` | Flagged `scanned` but not mostly one image (a figure page with a short caption) | No cap; the sparse waiver applies. The card's `sparse` field is advisory (a mismatch is a warning) |
**OCR-layer pages (1.2.0, user decision 2026-10-19).** A page flagged
`ocr_layer_suspected` (an image covers at least 80% of the page and the
anchor has text: someone else's OCR, which the gate never compared with the
image) is recorded per page as `ocr_layer: true` and treated like our own
OCR (Q-A) **for claims only**: a claim quote that passes there becomes
`PASS_OCR_LAYER` and each number found inside it `NUMBER_OCR_LAYER`
(`ocr-layer`), or `NUMBER_OCR_LAYER_VISION` (`ocr-layer+vision`) when the
claim's `source` is `both`; all are partial, so the card is at best PARTIAL.
Coverage anchors on such a page do **not** cap: the flag also fires on a
born-digital cover image (bake-off: one false positive in 342 pages), and a
cited-claims rule keeps that harmless. On the bake-off scan ED083756 (70 of
71 pages flagged) every claim is now partial.
**Paper-level cap (1.6.0, user decision 2026-10-23).** A paper with at least
`OCR_LAYER_PAPER_CAP_PAGES = 2` pages flagged `ocr_layer_suspected` gets one
`OCR_LAYER_PAPER` item (partial) and is at best PARTIAL even when no claim
cites those pages: much of its text is someone else's OCR. One flagged page
alone does not cap the paper (the false-positive cover image above). PDF
only; `counts.pages_ocr_layer` records the number of flagged pages.
## Coverage statuses (per page)
| Status | When | Severity |
|--------|------|----------|
| `PASS` | An identifying text-layer anchor passes | pass |
| `COVERED_SPARSE` | The sparse waiver applied | pass on a sparse text page; partial on a scanned page or via OCR |
| `COVERED_DUPLICATE` | A full-length exact anchor on a page with no unique line (its text recurs on another page) | pass; partial on a scanned page or via OCR |
| `PASS_OCR` | Best anchor matched in OCR text | partial |
| `VISION_ONLY` | A `vision_only` digest on a truly scanned page with no text in any lane, with validated vision | partial |
| `COVERED_EMPTY` | A page with no text in any lane that is not truly scanned, given an empty `anchors` list or a validated `vision_only` digest | pass |
| `NEEDS_REVIEW` | No passing anchor; the best is NEAR or REDACTED_SPAN | review |
| `NOT_COVERED` | No anchor passes | fail |
| `MISSING_PAGE` | No coverage entry | fail |
Every anchor is checked: a `NOT_FOUND` or `WRONG_PAGE` anchor always fails
(a fabricated quote), even on a page another anchor covers. An extra
NON_IDENTIFYING or NEAR anchor on a covered page is a warning.
**A truly scanned page with empty OCR text** can only be covered by a
validated `vision_only` digest (`VISION_ONLY`, partial). Without a listed
reader it is `NOT_COVERED` and the card FAILs: report such a paper to the
orchestrator rather than forcing the card. A listed reader's digest covers
the page, but no model is validated on scanned pages (the list covers table
pages without a `scanned` or `ocr_layer_suspected` flag,
`vision-validation.md` § The validation probe), so from 1.8.0 part 5 the
gate sends it to a person: `VISION_ON_SCAN_PAGE` (review, § Vision
consistency), and the card is NEEDS_REVIEW, not PARTIAL. The same holds for
any vision reading (a digest, or a claim with `source` `vision` or `both`)
of a page the gate re-derives as truly scanned or as carrying an OCR text
layer. Write `?` for digits you cannot read and never infer them from
context or trends, mark the reading as unvalidated in the card, log every
difference from the OCR or text lane, and tell the orchestrator.
## Numbers (`numbers-v6`)
Versions: `numbers-v2` locality, `numbers-v3` the unique-row path,
`numbers-v4` the spaced sign, superscript marks and row guards (1.6.0),
`numbers-v5` the anchor-cell guard (1.7.0), `numbers-v6` footnote letters,
the line-start sign and sign-ambiguous dashes (1.8.0). Write each number exactly as
printed. It must be **local** to the claim's
quote by one of two paths (recorded per number as `locality`): inside the
quote (`in-quote`), or in the one lane B row the quote's leading row label
identifies (`row`, 1.3.0). What to quote for a table number:
§ What to quote for a table number below. Rules applied to both the card
string and the page text:
- U+2212, U+2013, U+FE63, U+FF0D directly before a digit → `-`; then NFKC.
- `-` is a sign only at the start, after whitespace or after one of
`( [ { = < > : ; , / ~`, so "2019-2020" is two positive numbers. From
`numbers-v4` a dash separated from the number by spaces (and at most one
line break) is a sign after one of those characters or at the start
(`Est = - 0.041`, `(- 0.044)`); after a digit it stays a range or
subtraction (`28 - 75`). From `numbers-v6` (1.8.0) it is also
a sign at a **line start** (only spaces or tabs back to the line break: a
column-major table cell `– 0.30`). After a letter or a closing bracket
with a space before the dash (`of – 0.30`, `(n − 1)`, `(Girls) – 0.30`)
the sign is **ambiguous**: the dash may be a minus or a subtraction or
range dash. A claim of either sign that only such a token prints is
`sign-ambiguous` (review, never `ok`); zero excepted. "K-" + line break +
"12" (no space before the dash) is a compound, not a sign.
- Superscript and subscript characters (`¹`, `²`, `⁻`, `ᵃ`, `₂`) are marks,
never digits (`numbers-v4`: NFKC used to fold `0.62¹` into `0.621`). A
superscript the **claim** writes must be printed (it can be an exponent,
`10⁵`); a claim without one matches a page value that has one (a footnote
mark left out). Stars and daggers are not compared.
- One **footnote letter** a–g right after a number that ends the word
(`−0.164a`, `1.20b`, `2017a`, `(7b)`) is read as the number with the letter
as a mark (`numbers-v6`); write the number without it (`-0.164`), or with
it, which must then match. Two letters, a digit, or `.`/`-`/`+`/`'` and a
letter or digit after it keep the old reading (`5mg`, `3rd`, `1e-5`,
`5a.m.`), and other letters stay non-numbers (`1990s`, `5m`, `2x`).
- `,`, U+2009, U+202F and a plain space are thousands separators only
between groups of exactly 3 digits (`1,204` = `1 204` = 1204). A plain space
is ambiguous (two table cells "1" "204"), so the page is searched under
both readings.
- `.05` → `0.05`. Trailing `*`, `†`, `‡` are kept as marks. `%` is an
attribute. Parentheses mean an SE or t statistic, **never** a negative.
- Same value and same digit string → exact; same value, other digits
(`0.5` vs `0.50`) → precision differs; sign is part of the value.
| Status | Label | When | Severity |
|--------|-------|------|----------|
| `NUMBER_OK` | `ok` | Same value and digits in the cited page's anchor text, and local: inside the claim's quote where it matched (`locality: in-quote`) or in the one lane B row its leading label identifies (`locality: row`) | pass |
| `NUMBER_NOT_NEAR_QUOTE` | `not-near-quote` | Printed on the cited page (in the lane it was found in) but not local, or the claim has no quote located there (a `vision` claim without a quote) | review |
| `NUMBER_OCR_LAYER` | `ocr-layer` | As `ok`, on a page flagged `ocr_layer_suspected` | partial |
| `NUMBER_OCR_LAYER_VISION` | `ocr-layer+vision` | As `ocr-layer`, claim `source` `both` | partial |
| `NUMBER_OCR_ONLY` | `ocr-only` | Only in that page's OCR text, inside the quote's OCR match | partial |
| `NUMBER_OCR_VISION` | `ocr+vision` | Only in OCR text, claim `source` `both` (vision corroborates) | partial (the paper still relies on OCR, Q-A) |
| `NUMBER_VISION_ONLY` | `vision-only` | In no text lane, claim `source` `vision` | partial |
| `NUMBER_PRECISION_DIFFERS` | `precision-differs` | Same value, other digits | review |
| `NUMBER_SIGN_AMBIGUOUS` | `sign-ambiguous` | (`numbers-v6`, 1.8.0) Inside the claim's quote, the only token printing the claim's digits (either sign) follows a spaced dash after a word or a closing bracket (`of – 0.30`): minus or subtraction is undecided | review |
| `NUMBER_MISSING` | `missing` | Anything else (a changed digit, a dropped minus, the spaced form `= - 0.041` included from `numbers-v4`, a line-start cell `– 0.30` from `numbers-v6`; `both` declares a text reading, so a number in no text lane fails). A minus deleted *inside the quote* is the quote's `QUOTE_NUMBERS_DIFFER` (review, `ladder-v3`), not this | fail |
| `NUMBER_UNPARSEABLE` | `unparseable` | The card string is not one number | fail |
**Locality (`numbers-v2`, 1.2.0, user decision 2026-10-19).** Under
`numbers-v1` a number passed when its value was printed anywhere on the
cited page; in the S3 bake-off a planted last-digit change still passed 44%
of the time on dense table pages (351 of 800) and a sign flip 5% (42 of
800). Now a page token printing the value must overlap the place where the
claim's quote matched (`NUMBER_NEAR_CHARS = 0`: inside the quote; page
tokens are used, so a minus the SKELETON tier ignores cannot slip through;
a declared `page_end` span is matched across the page boundary). A value
printed elsewhere on the page is `not-near-quote`: a **human check, not a
fail**; a value absent from the page is still `missing`. Measured on the
bake-off ingests (2026-10-19; 453 table claims from B and C, quote =
12 words from the row label as a reader copying from `pages/` gets them;
119 prose numbers from A, B, F), share graded ok, genuine / last digit
changed / sign flipped:
| Rule | Table numbers | Prose numbers |
|------|---------------|---------------|
| Page-level (`numbers-v1`) | 1.000 / 0.110 / 0.009 | 1.000 / 0.034 / 0.000 |
| **Inside the quote (chosen)** | **0.305 / 0.013 / 0.000** | **1.000 / 0.000 / 0.000** |
| Within 40 chars of the quote | 0.322 / 0.026 / 0.000 | 1.000 / 0.000 / 0.000 |
| Within 160 chars | 0.470 / 0.035 / 0.000 | 1.000 / 0.000 / 0.000 |
| Lane B row (the number on one lane B line that shares ≥ 6 letters with the quote) | 1.000 / 0.053 / 0.000 | 0.941 / 0.000 / 0.000 |
The rule with the fewest fabricated passes was chosen (gate integrity
first). *Correction (2026-10-22):* that measurement quoted each table claim
from the **first** occurrence of its row label on the page. On report-style
pages a label such as "Full Sample" recurs in several panels, so that quote
often belonged to another row, and the 0.305 above understates the in-quote
rule (re-measured below with each quote taken from the number's own row).
**The unique-row rule (`numbers-v3`, 1.3.0, user decision 2026-10-22).** The user
adopted lane B's row-by-row text as locality evidence, because human review
time matters. A number printed on the anchor page also clears (`locality:
row`) when all of these hold:
1. the claim quote passed on the **anchor** (`PASS`, `PASS_SKELETON`,
`PASS_SPAN`, `PASS_OCR_LAYER` or, on arXiv HTML, `PASS_CONVERSION_ERROR`)
and the quote-number rule held (a `QUOTE_NUMBERS_DIFFER` quote never takes
the row path). Lane B never anchors a quote;
2. the quote's **leading label** identifies **exactly one** lane B row on
the page(s) where the quote matched. The leading label is the quote's
letters before its first number that follows a letter, or the whole quote
when it holds no such number. A row is a lane B line whose own label (its
letters before its first number after a letter, at least
`ROW_LABEL_MIN_CHARS = 3` letters from `numbers-v4`; 2 before) is
followed by at least one number in a cell-like tail (letters at most
`ROW_TAIL_LETTER_SHARE_MAX = 0.25` of the tail's non-space characters).
From `numbers-v4` (1.6.0) the row path also needs a page flagged
`table` or `numeric` (`ROW_PAGE_FLAGS`) and a quote whose anchor match
begins at a row start (no letter before it on its line), on a single page
(a declared two-page span never takes the row path). From `numbers-v5`
(1.7.0) the identified label must also **stand as a cell** on the raw
anchor line where the quote's match begins: what follows the label on that
line is cell-like (letters at most 25% of its non-space characters; an
empty rest counts), so a column-major label alone on its line
(`Huber-White (HW)`) or a row-major label followed by its cells passes,
while a prose sentence that merely starts with a label word
("ORGANIZATION and other models; …") does not
(`pdf_text_common.row_label_is_cell()`). Otherwise `row_match` is
`ineligible`, with the reason.
The quote's leading label must *start with* the row's label and the label
must end at a word boundary ("No" does not match "Notes"). The longest
matching label wins. Two or more lines with that label are `ambiguous`,
none is `none`. Letters only, so `co-teaching` = `coteaching`, but rows
`Grade 3` and `Grade 4` share the label `grade` and are ambiguous (the
safe side);
3. that row prints the claimed value with the same digits and sign, after
its label (a row index such as `7.` is not a cell).
An `ambiguous` or `none` label falls back to the in-quote path. Otherwise
the number is `not-near-quote` (review). Every number item records
`locality` (`in-quote`, `row` or null) and `row_match` (`unique`,
`ambiguous`, `none`, or null when the row path was not tried) so a reviewer
sees which path cleared it. Measured on the bake-off ingests, with each table
quote taken from the number's own row (2026-10-22: 449 claims from B and
C; 60 DOI/URL and footnote rows excluded). Shares graded ok, genuine / last
digit changed / sign flipped:
| Rule | Table numbers, pooled | Column-major rows (31) |
|------|-----------------------|------------------------|
| Inside the quote (`numbers-v2`) | 0.976 / 0.018 / 0.000 | 0.645 / 0.000 / 0.000 |
| Lane B, any line sharing ≥ 6 letters (B6) | 1.000 / 0.058 / 0.000 | 1.000 / 0.097 / 0.000 |
| **Unique row OR inside the quote (`numbers-v3`, shipped in 1.3.0)** | **0.996 / 0.018 / 0.000** | **0.935 / 0.000 / 0.000** |
On the 2026-10-19 sample (430 table claims after removing 23 noise rows),
the `numbers-v3` rule cleared 0.277 genuine and B6 1.000, with 0.016 and 0.056
fabricated last digits. Those quotes started at a recurring label: the
`numbers-v3` rule calls 98% of C's quotes `ambiguous`, and B6 clears them by
matching any line that shares the label, including the look-alike rows. Prose
numbers (51 claims from A, B, F): 1.000 genuine and 0.000 fabricated under
both in-quote and the `numbers-v3` rule.
**Residuals (documented, not caught):**
- (a) *same row, another column.* In a row `Overage 0.58 0.57 0.59 …`, a
claim of 0.59 for the first column clears: the gate knows the row, not the
column. The same holds inside a quote that holds the whole row. A vision
cross-check by a listed reader can catch it, as the reader's own check:
the gate does not compare vision readings with columns. Since 1.5.0
such a number carries the visible label `column-not-checked` (below).
- (b) *look-alike rows* are closed: under B6, "Time III (Inclusive with
co-teaching)" matched the "… without co-teaching" row and "Full Sample"
matched another table's row. Now the label must identify one row, so such
claims go to review. *Remote rows* (found in review, 2026-10-23) were not
closed by `numbers-v3`: a prose quote named a prose lane B line ("which …",
"of …") and cleared its numbers (69 prose windows in four bake-off papers).
`numbers-v4`'s guards (above) cut that to 6 (measured 2026-10-23); those
left start at a row start on a table page with a label a table row shares.
`numbers-v5`'s anchor-cell guard closed the two holes a re-check still found
(a prose sentence starting with a row-label word; a lane B fragment row
"the 1" that any quote starting with "the" identified), at the cost of one
genuine row-path quote on the bake-off papers (now review).
**The column label (`column-check-v1`, 1.5.0, user decision 2026-10-23).** The
user asked for residual (a) to be visible "for future reference, and then it
can go to human review if required later on down the line". So a number
that clears now says whether the gate could tell which cell it came from.
This is a **label only**: no number status, verdict or grade changes. Every
number item carries:
| Field | Value |
|-------|-------|
| `column_checked` | `false` when the clearing text prints two or more distinct comparable values; `true` when the claimed value is the only one; `null` when the number did not clear (no `locality`: `missing`, `not-near-quote`, `precision-differs`, `vision-only`, `unparseable`) |
| `column_label` | `"column-not-checked"` when `column_checked` is false, else null |
| `column_values` | The number of distinct comparable values counted (null with `column_checked` null) |
| `column_severity` | `COLUMN_UNCHECKED_SEVERITY` (`"label"`) on a labeled item, else null |
The rule (`pdf_text_common.column_values`):
- **Clearing text.** For `locality: in-quote`, the claim quote's matched span
in the page text the number was found in (the anchor, or the OCR text for
`ocr-only` and `ocr+vision`); for `locality: row`, the identified lane B
row (on HTML, the `rows/` row) after its label, as the row rule reads it.
Every cleared status gets the label: `ok`, `ocr-layer`, `ocr-layer+vision`,
`ocr-only`, `ocr+vision`, `conversion-error`.
- **Number token.** The number parser above, one reading only (a plain space
between digit groups splits). A token that overlaps the token printing the
claim is the claim itself, so `1 204` read as 1204 is one number, not the
cells 1 and 204.
- **Comparable.** Same shape as the printed claim: integer vs decimal, and
with vs without `%`. A count `1,204` is not a candidate for the coefficient
`0.05`. Parentheses and sign do not split shapes: a coefficient and its SE
can be confused, and so can a point estimate and its interval bounds.
- **Not counted.** Years (a bare 4-digit integer from `COLUMN_YEAR_MIN =
1900` to `COLUMN_YEAR_MAX = 2099`, unless the claim is itself a year) and
reference numbers (an integer right after one of the 32
`COLUMN_REFERENCE_WORDS`, case-insensitive, an optional `.` after the word:
`table`, `tables`, `tab`, `figure`, `figures`, `fig`, `figs`, `section`,
`sections`, `sec`, `chapter`, `ch`, `appendix`, `panel`, `model`, `models`,
`column`, `columns`, `col`, `equation`, `eq`, `eqs`, `page`, `pages`, `p`,
`pp`, `note`, `notes`, `footnote`, `step`, `wave`, `hypothesis`; no other
plural is listed, so "Appendices 2" or "Panels 3" counts). So "In 2019 Table 4 listed 41
boats on page 3" is one data number, and an ordinary prose sentence with one
number is not labeled.
- **Values, not tokens.** Distinct digit strings, the claim's own value
counted once however often it is printed: a row `0.57 0.57` with nothing
else comparable makes the claim true in either column (`column_checked:
true`); `0.57 0.59 0.57` does not.
What it costs and catches: on the bake-off cards (2026-10-23) the label
fell on the 5 numbers C quotes from one five-cell row, on one HTML number
inside a confidence interval (`7.2% (6.5% - 11.2%)`), and on one HTML number
in a short math expression beside a heading number; no prose sentence with
one number was labeled. The residual it marks was measured at 1.8% of
last-digit fabrications on PDF true-row quotes (the unique-row table above)
and about 9% on whole-row arXiv HTML quotes (`arxiv-and-html.md` § Measured
on real papers). `counts` gains
`numbers_column_not_checked`, the JSON lists the labeled items in
`column_not_checked_items`, and the stdout summary line prints the count
(`… 15 number(s), 5 column-not-checked; …` on C).
**Severity `label` and the review hook.** `label` is a sixth severity level
that no verdict rule reads, and it is never an item's own `severity` (only
its `column_severity`). It is a constant, `COLUMN_UNCHECKED_SEVERITY` in
`lit_verify.py`, recorded in `card_check` `constants`, so a later decision
can route labeled numbers to a human. Set it to `"review"` and every labeled
item's `severity` is raised to review (never lowered), its reason says
"column not checked", and the card becomes NEEDS_REVIEW with no other code
change. Until then, a reviewer or the synthesis stage can read the label and
check those cells by hand. The card needs no change: v1, v2 and v3 cards get
the label alike, and a labeled number is not an error to fix.
### What to quote for a table number
- **Row-major anchor** (`pages/` shows the row's label and cells on one
line): quote from the row label through the cell, for example
`North Quay (with tide gauge) 0.62 0.63`. This clears in-quote.
- **Column-major anchor** (`pages/` lists the row labels, then each column):
**start the claim quote at the row label** and copy forward from `pages/`,
usually a run of row labels such as `North Quay (with tide gauge) North Quay
(without tide gauge) All Harbours`. The gate finds the row in lane B. Open
`pages/pNNN_rows.txt` (the reading aid, `pdf-ingest.md` § Reading aid) to
see which value belongs to which row. Never quote the aid itself: its lines
carry a `lane-B| ` prefix, and lane B is not the anchor, so a quote copied
from it fails (`NOT_FOUND`) unless the anchor prints the same text.
- The quote must still be identifying. When a label recurs on the page (two
`Full Sample` rows), the row path is `ambiguous`: quote the row's cells
in-quote if the anchor allows it, or accept the review item.
- No card change: v1 cards are unchanged and still valid, and their numbers
can only gain the row path. `table_ref` stays optional and is not checked.
Known limits: a plain-space thousands separator reads both ways (by design);
European decimals (`1.204,5`) are not handled, so write such a number as a
quote instead and flag it; parenthesis or percent differences between card
and page are notes on an `ok`, not failures.
## Vision consistency
With `vision_mode: "off"` or `vision_validated: false`, every claim with
`source` `vision` or `both` and every `vision_only` digest is a fail
(`VISION_UNAVAILABLE_CLAIM`), unless the block declares
`vision_override` (`check` or `accept`): the researcher chose to let a
reader that is not on the list read page pictures, so its readings are
checked as a listed reader's are, and the scan and figure rules below apply
unchanged. The gate does not bind the declaration to the researcher's
record; `lit_verify audit` does (`vision-validation.md` § The researcher's
choice for an untested reader), and a declaration with no matching record
sends every reading to a person there.
**Fail-closed on the list (1.6.0).** `lit_verify` embeds the vision-validated
list (`VISION_VALIDATED_LISTS`, one entry per `vision_list_version`; the
current `vision-list-2026-10-30` holds eight model IDs, five on the two
Anthropic routes and three GPT models on the two GPT shim routes, and the
older `vision-list-2026-10-19` (empty) and `vision-list-2026-10-28` (the five
Anthropic models) are kept so older cards stay interpretable) and records it
in `card_check`
`constants`. `vision_validated: true` for a `reader_model` that is not on the
entry for the card's `vision_list_version`, or for a list version the tool
does not know, is `VISION_NOT_ON_LIST` (fail), and vision then counts as
unavailable, so the card's vision claims also fail as
`VISION_UNAVAILABLE_CLAIM`. Only the block is read: the route is front
matter, so the gate checks the model ID against the version's rows for any
route. Whether the reader really ran as that model, on that route, stays
with the orchestrator's audit (`vision-validation.md` § Gating).
**A vision reading on a scan goes to a person (1.8.0 part 5, user decision
2026-10-28).** The list covers table pages without a `scanned` or
`ocr_layer_suspected` flag. On the probe's one scan, a listed model filled in
digits it could not read from the column's trend in 2 of 3 runs: numbers that
look plausible and often have no clean text to check them against. So every
claim with `source` `vision` or `both`, and every `vision_only` digest the
gate uses, that cites a page the gate classes as truly scanned
(`pages[].class: "scanned"`) or as carrying an OCR text layer
(`pages[].ocr_layer: true`) gets one `VISION_ON_SCAN_PAGE` item per cited
page (kind `vision`, severity **review**). A card with one is NEEDS_REVIEW,
and its `evidence_grade_if_review_confirmed` is `FULL_TEXT_PARTIAL`: a
number a person checked against a scan by eye is still not verified by a
text layer. The item's reason is written for the researcher, for example:
> The AI reader took the number 88 from a picture of PDF page 2 (its
> position in the file; the printed page number may differ). This page
> looks like a scan. DAAF has tested AI reading of page pictures only on
> clean digital pages; on a scan, a model can guess digits it cannot make
> out. Please compare 88 with PDF page 2.
The reason says the page *looks like* a scan: the flags also fire on a
born-digital cover printed over a full-page picture. On a page with an OCR
text layer it adds "its text was machine-recognized (OCR) when the PDF was
made, so the text may contain errors"; a `both` claim was read "from a
picture of PDF page N … and from the page's machine-read text"; a claim
with no numbers ends "Please compare the claim with PDF page N", a digest
"Please compare the summary with PDF page N"; a page whose type could not
be checked says "DAAF could not reopen the original PDF to check this page,
so it treats it as a possible scan."
- **Review, not fail.** Scans are where vision helps most; a fail would push
readers to avoid them. Nothing else about the reading changes: its number
and quote statuses stay as they are.
- **The page type comes from the source.** The class and the OCR-layer flag
are the ones the gate re-derives from the hashed source bytes
(§ Re-derivation), never the card's `sparse` field or the manifest's
flags, so a forged flag cannot hide a scan, except in the documented
residual (§ Re-derivation: a scan or OCR layer whose only image evidence
is a pdfplumber image box, removed from a forged manifest in a document
where pdfimages lists images; no such page in the five bake-off PDFs). A
page whose facts could not be re-derived (the source could not be read,
or does not give that page) is treated as a possible scan, and its reason
says so; the card already fails then. Without pdfplumber
(`FLAGS_NOT_REDERIVED`) the conservative reading keeps a recorded OCR
layer and counts an unconfirmed scan as truly scanned.
- **Only the scan part of the list's scope.** A vision reading on a sparse
page (flagged `scanned`, little text, not mostly one image) is outside the
list too, but raises nothing here: it is still unvalidated, the reader
labels and reports it, and `lit_verify audit` lists it for a person
(`vision-validation.md` § Gating, step 5). Figure and chart pages have
their own warning (below). Measured on the five bake-off PDFs (342 pages):
2 sparse pages, both near-blank text pages; all 10 born-digital figure
pages class `normal`.
- **An ignored digest raises nothing.** A `vision_only` digest on a page
with text in any lane is `VISION_ONLY_IGNORED` (an anchor must cover the
page), so it gets no review item.
- **Whatever the reader's list status.** An unlisted reader's vision
readings also fail (`VISION_NOT_ON_LIST`, `VISION_UNAVAILABLE_CLAIM`), so
that card is FAIL, with the review item listed beside the fails. A
revised card that keeps the vision reading gets the item again.
- PDF only: arXiv HTML units have no page images.
**A number read from a chart or figure that nothing in the paper confirms
is a warning (1.15.0, by the user's decision, `vision-figure-v1`).** No model is
validated on figure pages (the probe scored tables only). The gate raises
`VISION_FIGURE_UNCORROBORATED` (kind `vision`, severity **warning**: no
verdict or grade change) once per claim and cited page when all of these
hold:
- the claim's `source` is `vision` or `both`;
- the cited page is a figure page of class `normal` (the `figure` flag, or
figure evidence the gate re-derives from the source: a figure caption or
a pictured image), and not a scan page (those get `VISION_ON_SCAN_PAGE`);
- a number of the claim is **in doubt**: the claim's `vision_reading` is
`estimated` or `uncertain`, or the number is in no text lane of that page
(anchor, lane B or OCR text);
- and **no page of the paper prints it** in its re-derived text (anchor or
lane B). A number the text states raises nothing: the text corroborates
the reading.
The item lists the numbers in doubt (`numbers_in_doubt`, and the claim's
`vision_reading`), and its reason is written for the researcher:
> The AI reader took the number 42 from a picture of a chart or figure on
> PDF page 3 (its position in the file; the printed page number may differ)
> and marked the reading as an estimate. The paper's text does not print 42
> anywhere, so nothing in the paper confirms it, and DAAF has not tested AI
> reading of numbers from charts. If it matters to the review, please
> compare 42 with the figure on PDF page 3.
`lit_verify audit` lists these warnings under `should_look` (outcome
`confirmed_with_checks` when nothing else is open). A vision claim on a
figure page with no number raises nothing.
## arXiv HTML statuses (1.4.0)
On an HTML ingest the page checks run on units (class `sparse` when
`pdf_ingest` flagged the unit sparse, else `normal`; never `scanned`), with
these statuses beside the PDF ones:
| Status | When | Severity |
|--------|------|----------|
| `ANCHOR_KIND_MISMATCH` | Card `anchor_kind` (v1 = `pdf`) differs from the ingest's | fail (nothing else checked) |
| `SOURCE_VERSION_MISMATCH` | Card `source_version` differs from the ingest's arXiv version | fail |
| `MIXED_ANCHOR_KINDS` | The key's complete ingests mix `pdf` and `arxiv_html` | fail |
| `TABLE_ROWS_HASH_MISMATCH` | A `rows{S}/uNNN.txt` hash differs from its records | fail |
| `SECTION_NOT_FOUND` | A coverage or claim `section`/`section_end` label the ingest does not have | fail |
| `MISSING_SECTION` | A unit with no coverage entry (the `MISSING_PAGE` analogue) | fail |
| `WRONG_SECTION` | The quote matches exactly only in other units, or across an undeclared unit boundary (the `WRONG_PAGE` analogue) | fail |
| `SECTION_END_NOT_NEXT` | `section_end` is not the unit right after `section` | fail |
| `PASS_CONVERSION_ERROR` | A claim quote passed in a unit flagged `conversion_error` (LaTeXML `ltx_ERROR`) | partial |
| `NUMBER_CONVERSION_ERROR` (`conversion-error`) | A number that would be `ok`, in such a unit | partial |
| `HTML_EXTRACT_VERSION_DIFFERS` | The ingest used an extraction version the tool was not built for | warning |
Coverage anchors in a conversion-error unit do not cap. Every other status
(the ladder tiers, `NON_IDENTIFYING`, `COVERED_SPARSE`, `COVERED_DUPLICATE`,
the number statuses) means what it means on a page; the OCR and vision
statuses cannot occur.
## Integrity and warning statuses
Statuses about the ingest and the card as a whole (item `kind` `integrity`,
`card`, `coverage`, `claim` or `paper`), beside the re-derivation statuses
(§ Re-derivation) and the HTML ones above:
| Status | When | Severity |
|--------|------|----------|
| `SOURCE_HASH_MISMATCH` | Card `source_pdf_sha256` (v2: `source_sha256`) differs from the ingest manifest's `source.sha256` | fail |
| `PAPER_KEY_MISMATCH` | Card `paper_key`, or the manifest's `paper_key`, differs from the PAPER_DIR basename | fail |
| `ANCHOR_HASH_MISMATCH` | An anchor file's SHA-256 differs from its manifest record, or page_index's `anchor_sha256` differs from that record | fail |
| `PAGE_INDEX_HASH_MISMATCH` | `page_index{S}.parquet` differs from its manifest record | fail |
| `OCR_HASH_MISMATCH` | An OCR text file differs from its record (the file is then not used) | fail |
| `LANE_B_HASH_MISMATCH` | A lane B file differs from its manifest or page_index record (not used) | fail |
| `INGEST_MANIFEST_HASH_MISMATCH` | The ingest manifest differs from the hash in the pdf_ingest MANIFEST row that wrote it | fail |
| `NO_INGEST_ROW` | No pdf_ingest MANIFEST row in the papers dir records the ingest manifest (a fail from 1.7.0; a warning before) | fail |
| `PAGE_OUT_OF_RANGE` | A coverage entry or a claim cites a page beyond the ingest's page count (the claim is not checked further) | fail |
| `INGEST_SUFFIX_DIFFERS` | The card's `ingest_suffix` is not the suffix checked (hashes decide integrity) | warning |
| `INCOMPLETE_INGEST_IGNORED` | A suffix has artifacts but no ingest manifest (pdf_ingest did not finish); it is never read | warning |
| `NORMALIZE_VERSION_DIFFERS` | The ingest recorded a `NORMALIZE_VERSION` other than the tool's | warning |
| `SPARSE_DECLARATION_MISMATCH` | The card says `sparse: true` for a page (unit) `pdf_ingest` did not flag `scanned` (`sparse`); the class is computed, the field is advisory | warning |
| `VISION_ONLY_IGNORED` | A `vision_only` digest on a page with text: an anchor must cover it | warning |
| `OCR_LAYER_PAPER` | ≥ 2 pages flagged `ocr_layer_suspected` (§ Page classes) | partial |
| `SCANNED_PAGE` | A truly scanned page (§ Page classes) | partial |
An extra `NON_IDENTIFYING`, `NEAR` or `REDACTED_SPAN` anchor on a page that
another anchor covers is downgraded to `warning`.
## Status index
What `lit_verify card` can write (`lit_verify` 1.15.0 or later), with the
severity and the section that explains each: the 82 keys of `SEVERITY` in
`lit_verify.py` (one of them,
`COLUMN_NOT_CHECKED`, is a label, never a status: last row), plus
`COVERED_SPARSE`†. `COVERED_SPARSE` and `COVERED_DUPLICATE` take their
severity from the page (§ Coverage statuses).
† Not a `SEVERITY` key: a coverage status whose severity is set per page.
| Severity | Statuses |
|----------|----------|
| pass | `PASS`, `PASS_SKELETON`, `PASS_SPAN` (§ The match ladder); `COVERED_EMPTY`, `COVERED_DUPLICATE`, `COVERED_SPARSE`† (§ Coverage statuses); `NUMBER_OK` (§ Numbers); `QUANT_OK`, `PAGE_LABEL_OK` (§ Card schema v3) |
| partial | `PASS_OCR`, `VISION_ONLY`, `SCANNED_PAGE`, `PASS_OCR_LAYER`, `OCR_LAYER_PAPER` (§ Page classes, § Coverage statuses); `NUMBER_OCR_ONLY`, `NUMBER_OCR_VISION`, `NUMBER_VISION_ONLY`, `NUMBER_OCR_LAYER`, `NUMBER_OCR_LAYER_VISION` (§ Numbers); `PASS_CONVERSION_ERROR`, `NUMBER_CONVERSION_ERROR` (§ arXiv HTML statuses); `QUANT_NUMBER_PARTIAL` (§ Card schema v3) |
| review | `NEAR`, `REDACTED_SPAN` (§ The match ladder); `QUOTE_NUMBERS_DIFFER` (§ The quote-number rule); `NEEDS_REVIEW` (coverage, § Coverage statuses); `NUMBER_PRECISION_DIFFERS`, `NUMBER_NOT_NEAR_QUOTE`, `NUMBER_SIGN_AMBIGUOUS` (§ Numbers); `VISION_ON_SCAN_PAGE` (§ Vision consistency); `QUANT_NUMBER_REVIEW`, `QUANT_CI_ORDER`, `QUANT_ESTIMATE_OUTSIDE_CI` (§ Card schema v3) |
| fail | `WRONG_PAGE`, `NOT_FOUND` (§ The match ladder); `NON_IDENTIFYING` (§ Identifying quotes); `NOT_COVERED`, `MISSING_PAGE` (§ Coverage statuses); `NUMBER_MISSING`, `NUMBER_UNPARSEABLE` (§ Numbers); `VISION_UNAVAILABLE_CLAIM`, `VISION_NOT_ON_LIST` (§ Vision consistency); `ANCHOR_KIND_MISMATCH`, `SOURCE_VERSION_MISMATCH`, `SECTION_NOT_FOUND`, `MISSING_SECTION`, `WRONG_SECTION`, `SECTION_END_NOT_NEXT`, `TABLE_ROWS_HASH_MISMATCH`, `MIXED_ANCHOR_KINDS` (§ arXiv HTML statuses); `SOURCE_HASH_MISMATCH`, `PAPER_KEY_MISMATCH`, `ANCHOR_HASH_MISMATCH`, `PAGE_INDEX_HASH_MISMATCH`, `OCR_HASH_MISMATCH`, `LANE_B_HASH_MISMATCH`, `INGEST_MANIFEST_HASH_MISMATCH`, `NO_INGEST_ROW`, `PAGE_OUT_OF_RANGE` (§ Integrity and warning statuses); `ANCHOR_REDERIVE_MISMATCH`, `ROW_SOURCE_REDERIVE_MISMATCH`, `SOURCE_MISSING`, `SOURCE_FILE_HASH_MISMATCH`, `SOURCE_REDERIVE_ERROR`, `HEADER_FOOTER_REDERIVE_MISMATCH`, `FLAGS_REDERIVE_MISMATCH`, `OCR_ON_TEXT_PAGE` (§ Re-derivation); `QUANT_NUMBER_FAILED`, `QUANT_NOT_IN_NUMBERS`, `QUANT_N_NOT_INTEGER` (§ Card schema v3) |
| warning | `INGEST_SUFFIX_DIFFERS`, `INCOMPLETE_INGEST_IGNORED`, `NORMALIZE_VERSION_DIFFERS`, `SPARSE_DECLARATION_MISMATCH`, `VISION_ONLY_IGNORED` (§ Integrity and warning statuses); `VISION_FIGURE_UNCORROBORATED` (§ Vision consistency); `HTML_EXTRACT_VERSION_DIFFERS` (§ arXiv HTML statuses); `EXTRACTOR_VERSION_DIFFERS`, `OCR_NOT_REDERIVED`, `ROW_SOURCE_NOT_REDERIVED`, `FLAGS_NOT_REDERIVED` (§ Re-derivation); `PAGE_LABEL_UNCONFIRMED` (§ Card schema v3) |
| label | `COLUMN_NOT_CHECKED` (a number item's `column_severity` only, never its `severity`; § Numbers, the column label) |
## Verdicts and grades
Every item carries a severity `pass`, `partial`, `review`, `fail` or
`warning`. A sixth level, `label` (1.5.0), appears only as a number item's
`column_severity` and is ignored by every verdict rule (§ Numbers, the
column label).
| Verdict | When | `evidence_grade_suggested` | Exit |
|---------|------|-----------------------------|------|
| `FAIL` | Any fail item (all listed) | null | 11 |
| `NEEDS_REVIEW` | No fail; a NEAR, REDACTED_SPAN, NEEDS_REVIEW (coverage: a page's best anchor NEAR or REDACTED_SPAN), QUOTE_NUMBERS_DIFFER, precision-differs, not-near-quote, sign-ambiguous or VISION_ON_SCAN_PAGE item | null; `evidence_grade_if_review_confirmed` = `FULL_TEXT_VERIFIED`, or `FULL_TEXT_PARTIAL` when partial items or a VISION_ON_SCAN_PAGE item exist | 0 |
| `PARTIAL` | No fail or review; an OCR_LAYER_PAPER, PASS_OCR, PASS_OCR_LAYER, VISION_ONLY, ocr-only, ocr+vision, ocr-layer, ocr-layer+vision, vision-only or SCANNED_PAGE item, a COVERED_SPARSE or COVERED_DUPLICATE on a truly scanned page or matched in OCR text, or (arXiv HTML) a PASS_CONVERSION_ERROR or conversion-error item | `FULL_TEXT_PARTIAL` | 0 |
| `PASS` | Everything anchored in the text layer | `FULL_TEXT_VERIFIED` | 0 |
What synthesis may do with each grade: `FULL_TEXT_VERIFIED` supports any
claim, effect sizes included; `FULL_TEXT_PARTIAL` supports claims from
verified pages only, with the gate's exceptions listed.
## `card_check{S}.json`
Written in `PAPER_DIR`, named after the card version: `evidence_card.md` →
`card_check.json`, `evidence_card_a.md` → `card_check_a.json`, any other
card name → `card_check_<slug>.json`. **Never overwritten**: a re-check of the
same card version gets `_1`, `_2` (`card_check_a_1.json`). Written only when a
verdict exists (exit 0 or 11). Holds `kind`, `schema_version`,
`tool_version`, `checked_at`, `paper_key`, `card` (path, bytes, SHA-256,
`reader_model`, `vision_mode`, `vision_validated`, `vision_list_version`),
`ingest` (suffix, manifest name and SHA-256, page count, source SHA-256,
pdf_ingest version, normalize version), `rule_versions`, `constants` (with
the PROVISIONAL list), `verdict`, the grades, `vision_available`,
`fuzzy_tier_run`, `counts`, `failures`, `review_items`, `partial_items`,
`warnings`, `pages` (class, image area, `ocr_layer`, `lane_b` (lane B
available and used), coverage status per page) and `items` (number items
also carry `locality` and `row_match`, 1.3.0; constants gain
`ROW_LABEL_MIN_CHARS`). Card quotes in it are injection-scrubbed and secret-redacted, with
the original quote's SHA-256 kept. **It holds no page text.**
**Result schema `lit-verify-card-check-v2` (1.4.0)** is v1 plus additive
fields, written for PDF and HTML cards alike: top-level `anchor_kind`;
`card.schema_version`, `card.anchor_kind`, `card.source_version`;
`ingest.anchor_kind`, `unit_count`, `source_sha256`, `source_version`,
`arxiv_id_version`, `html_extract_version` (`ingest.source_pdf_sha256` is
null on HTML); every item carries `section` and `section_end` (the unit
labels on HTML, where `page`/`page_end` are null; null on PDF); `sections`
(one summary per unit on HTML: `unit`, `section`, `latexml_id`, `class`,
`basis`, `conversion_error`, `table_rows`, `anchors`, `coverage_status`;
`[]` on PDF) beside `pages` (`[]` on HTML); `counts` gains `sections`,
`sections_covered`, `sections_sparse`, `sections_conversion_error` on HTML
(in place of the page counts); constants gain `HTML_EXTRACT_VERSIONS_KNOWN`.
**1.5.0 (additive; result schema unchanged):** number items also carry
`column_checked`, `column_label`, `column_values` and `column_severity`;
`counts` gains `numbers_column_not_checked` (also in the MANIFEST row's
`counts`); `column_not_checked_items` lists the labeled items; `rule_versions`
gains `column_check` (`column-check-v1`); `constants` gain
`COLUMN_UNCHECKED_SEVERITY`, `COLUMN_YEAR_MIN`, `COLUMN_YEAR_MAX` and
`COLUMN_REFERENCE_WORDS`. `numbers` stayed `numbers-v3`: the label is not a
number rule, so it is versioned apart.
**1.6.0 (additive; result schema unchanged):** `claim_quote` items carry
`quote_numbers`, `page_numbers`, `quote_numbers_agree` and
`status_if_confirmed` (null until the quote-number rule runs on a passing
quote); number items can carry `row_match: "ineligible"` with the reason;
`counts` gains `pages_ocr_layer`; `rule_versions.match_ladder` is
`ladder-v2` and `numbers` `numbers-v4`; `constants` gain
`ROW_TAIL_LETTER_SHARE_MAX`, `ROW_PAGE_FLAGS`, `OCR_LAYER_PAPER_CAP_PAGES`,
`VISION_LIST_VERSION` and `VISION_VALIDATED_LISTS`.
**1.7.0 (additive; result schema unchanged):** a top-level `rederive`
block: `version` (`rederive-v1`), `method`, `source_path`, `source_sha256`,
`extractor`, `extractor_version` (now) and `extractor_version_ingest`,
`anchors_checked`, `anchor_mismatches` and `row_source_mismatches` (lists of
pages or unit labels; `"count"` for a page-count mismatch),
`row_sources_checked`, `ocr` (a note when OCR text was hash-checked only),
`outcome` (`match`, `mismatch`, `source missing`, `source hash mismatch` or
`error`) and `seconds` (1.8.0 adds, still additive: `version`
`rederive-v2`, `read`, `source_kind`, `source_kind_ingest`, `web_fetch_row`,
`header_footer` {records, outcome}, `flags` {pages_checked, mismatches,
table_probed, table_not_rederived_uncited, not_rederived},
`ocr_on_text_pages`; `constants` gain `GRADE_FLAGS`); `rule_versions` gains `rederive` and `numbers` is
`numbers-v5`; `constants` gain `SOURCE_MAX_BYTES`, `REDERIVE_TIMEOUT_S` (the
`--timeout` used) and `REDERIVE_WORKERS`.
**1.8.0, part 2 (result schema unchanged):** `rule_versions.match_ladder`
is `ladder-v3` and `numbers` `numbers-v6`; a number item can carry the
status `NUMBER_SIGN_AMBIGUOUS` (label `sign-ambiguous`, review, `locality`
null); a `QUOTE_NUMBERS_DIFFER` reason now says whether the digits, a mark
or a sign-shaped dash differ.
**1.8.0, part 3 (result schema unchanged):** `rule_versions.match_ladder`
is `ladder-v4`; a `QUOTE_NUMBERS_DIFFER` reason can also say that the quote
prints a sign-shaped dash the matched span does not. No new status.
**1.8.0, part 5 (result schema unchanged):** an item can carry the status
`VISION_ON_SCAN_PAGE` (kind `vision`, review, with `page` and, for a claim,
`claim_id`; § Vision consistency); `rule_versions` gains `vision_scope`
(`vision-scope-v1`), so a result shows whether the rule ran (one made
before it has no such key).
**1.15.0 (result schema unchanged):** an item can carry the status
`VISION_FIGURE_UNCORROBORATED` (kind `vision`, warning, with `page`,
`claim_id`, `number`, `numbers_in_doubt` and `vision_reading`; § Vision
consistency); `rule_versions` gains `vision_figure` (`vision-figure-v1`).
## MANIFEST row (papers dir, kind `lit_verify`)
One row per attempt with a valid papers dir: `kind`, `tool_version`,
`command`, `paper_key`, `parameters`, `timestamp`, `requesting_context`,
`card` (path, bytes, SHA-256), `ingest` (suffix, manifest, manifest SHA-256,
page count, and since 1.4.0 `anchor_kind` and `unit_count` (null on PDF)),
`rederive` (1.7.0: `version`, `method`, `outcome`, `anchors_checked`,
`row_sources_checked`, `seconds`, and the counts `anchor_mismatches` and
`row_source_mismatches`; 1.8.0 adds `source_kind`, the `header_footer`
outcome and the counts `flag_mismatches` and `ocr_on_text_pages`),
`artifacts.card_check` (name, bytes, SHA-256),
`verdict`, `evidence_grade_suggested`, `counts`, `redaction_count`,
`secret_redactions`, `exit_code`, `note`. A usage error or a refused papers
dir writes no row.
## Exit codes
| Exit | Meaning |
|------|---------|
| 0 | Verdict PASS, PARTIAL or NEEDS_REVIEW (in the card_check JSON); `audit` outcome `confirmed`, `confirmed_with_checks` or `confirmed_untested_reader`; `route` detected; `settle` items listed or settlement recorded; `vision-override` choice recorded; selftest: every expectation held |
| 1 | Usage error (bad flags; PAPER_DIR or PROJECT_DIR not a directory; CARD_PATH, `--card` or `--audit` not a file; `settle` with `--list` and a decision flag, or with neither `--list` nor `--items` and `--decision`; `vision-override` `--level check` or `accept` without `--model-id`, `--model-id` with `--level withdraw`, or a `--model-id` that is not a model ID) |
| 2 | Refused path (PAPER_DIR / papers dir / PROJECT_DIR / SCRATCH_DIR guard; CARD_PATH, `--card` or `--audit` not directly inside PAPER_DIR; `audit` PAPER_DIR not under PROJECT_DIR/papers/; `settle --items` outside the DAAF root; basename not a paper key; an ingest artifact path resolving outside PAPER_DIR; a secret value in an argument) |
| 3 | Reserved by `run_with_capture.sh`, never used |
| 8 | Card cannot be parsed (missing, duplicate or invalid lit-verify block; oversize or non-UTF-8 card; a v2 structural rule broken: `source_version` missing on HTML or present on PDF, a non-`text` source or vision declared available on HTML, a bad or repeated section label, `section_end` = `section`; a v3 `vision_reading` on a claim that is not a PDF vision reading; a `vision_override` that is not `check` or `accept`, or one beside `vision_validated: true`, `vision_mode: "off"` or an arXiv HTML card); `audit`: an agent ID outside the hook's id allowlist or given twice, or a `--transcript` that is not that agent's file; `settle`: not an audit record name, an unreadable record, or an items file that is oversize, not UTF-8, empty or holds a word that is not an item id |
| 9 | `audit`: an input fails its hash check (no card_check recorded for the card, the card_check differs from its MANIFEST row, or the card from its card_check), or an audit record for the card already exists; `settle`: the audit record differs from its MANIFEST row, the card or card_check changed or was checked again since the audit, or an item id does not re-derive |
| 10 | Ingest artifacts missing or incomplete (no manifest for the suffix, a missing anchor, OCR, table-rows or page_index file, an unreadable manifest, a manifest with an unknown `anchor_kind`, an HTML manifest without unique unit labels or `arxiv.version`), or a required dependency missing (rapidfuzz; polars or pyarrow; from 1.7.0 pdftotext and pdfinfo, which re-derive PDF anchors; from 1.8.0 pdfimages, which re-derives page flags); `route` or `vision-override`: the route detector cannot be loaded |
| 11 | FAIL as a recorded result (card_check written); `audit` outcome `person_must_check`, `mismatch` or `chain_unconfirmed` (record written); `settle` refused, nothing written (empty user words, user words over 4000 characters, `does_not_match` without a rationale, an item id twice, not this audit's, or the chain's); `vision-override` refused, nothing written (empty user words, `--user-words` or `--rationale` over 4000 characters, `check` or `accept` without a rationale, a model already on the vision-validated list for this route, nothing to withdraw); selftest: an expectation did not hold |
| 12 | Internal error (no verdict; see the MANIFEST note) |
## Fixing a failing card
At most **two** self-revisions, each a new card file
(`evidence_card_a.md`, `evidence_card_b.md`), then return with the latest
`card_check` path and the open items.
| Status | Usual cause | Fix |
|--------|-------------|-----|
| `NOT_FOUND` | Paraphrase, a quote from memory, or a word changed | Re-copy the sentence from `pages/pNNN.txt` |
| `WRONG_PAGE` | Wrong page number, or a quote that crosses a page break | Correct `page`; for a crossing quote, declare `page_end: page + 1` |
| `NON_IDENTIFYING` | Too short or a phrase that recurs | Quote a longer, distinctive sentence |
| `REDACTED_SPAN` | Copied a scrub token or fence line from a reading copy | Quote the page text around it; if the page itself holds injection-shaped text, flag it |
| `NEAR` | One character off (an OCR error, a typo in the card) | Re-copy; if the page really differs from what you read, log it |
| `NUMBER_MISSING` | Number rewritten (rounded, sign dropped, cell from another page) | Write it as printed, on the page where it is printed |
| `NUMBER_PRECISION_DIFFERS` | `0.5` for `0.50` | Copy the printed digits |
| `NUMBER_NOT_NEAR_QUOTE` | The quote does not hold the number, and its leading label identifies no single lane B row printing it (`row_match` says why) | Start the quote at the row label (§ What to quote for a table number), or quote the anchor run that holds the number; else leave it for the human check |
| `VISION_UNAVAILABLE_CLAIM` | A `vision`/`both` claim with vision unavailable | Use `text` (or `ocr`) with a text quote, or drop the claim and note it |
| `MISSING_PAGE` / `MISSING_SECTION` / `NOT_COVERED` | A page or unit skipped, or only short anchors on a normal one | Add an identifying anchor for it (a heading-only `sparse` unit: its heading, exactly) |
| `*_HASH_MISMATCH` | The ingest or card changed after it was written | Do not "fix" the card: report it; the orchestrator re-ingests |
| `WRONG_SECTION` | Wrong section label, or a quote that runs into the next unit | Correct `section`; for a quote that crosses into the next unit, declare `section_end` with that unit's label |
| `SECTION_END_NOT_NEXT` | `section_end` names a unit that is not the next one | Use the next unit's label (manifest `units`, in order), or split the quote |
| `SECTION_NOT_FOUND` | A label typed from the paper's printed numbering, or reconstructed | Copy the label from the manifest's `units[].label` or the fulltext marker `section=…` |
| `PASS_CONVERSION_ERROR` (partial) | The claim quotes a unit where LaTeXML could not convert a macro | Quote another unit that states the claim, or accept PARTIAL (or read the paper from the PDF under its own key) |
| `ANCHOR_KIND_MISMATCH` | A v1 or `pdf` card on an HTML ingest, or the reverse | Rewrite the card as `lit-verify-card-v2` with the ingest's `anchor_kind` and its locators (sections or pages) |
| `SOURCE_VERSION_MISMATCH` | The card names another arXiv version than the one ingested | Write the ingested version (manifest `arxiv.version`); if you read another version, that version needs its own fetch, key and ingest |
| `MIXED_ANCHOR_KINDS` | A PDF and an HTML ingest under one paper key | Report it: the orchestrator moves one source to its own key |
| `QUOTE_NUMBERS_DIFFER` (review) | A minus, range dash, superscript or digit lost when the quote was retyped, or (`ladder-v4`) a sign-shaped dash added that the page does not print | Re-copy the quote with every number as printed (§ The quote-number rule); a non-sign dash on the page stays a human check |
| `NUMBER_SIGN_AMBIGUOUS` (review) | The page prints the number after a spaced dash that follows a word or bracket (`of – 0.30`) | Nothing to fix in the card if the number is copied right: a human reads the page (or its render) and decides minus or subtraction |
| `VISION_NOT_ON_LIST` | `vision_validated: true` for a model not on the card's list version | Set `vision_validated: false`, `vision: unavailable`, and re-read flagged pages from the text layer |
| `VISION_ON_SCAN_PAGE` (review) | The AI reader took a number, a claim or a page summary from a picture of a page that looks like a scan (including a scan whose text was machine-recognized when the PDF was made). DAAF has not tested AI picture-reading on scans | Nothing to fix if the reading was copied carefully (`?` for digits you could not read, the reading labelled unvalidated, disagreements logged): a person compares it with the page. If the page's own text or OCR text prints the same thing, cite that lane instead (`source` `text` or `ocr`). Never drop the vision source to hide a disagreement: log it |
| `SOURCE_MISSING` / `SOURCE_FILE_HASH_MISMATCH` | The ingest's source was moved, deleted or replaced after the ingest | Do not edit the card: report it; restore the source to its recorded path or re-ingest |
| `ANCHOR_REDERIVE_MISMATCH` / `ROW_SOURCE_REDERIVE_MISMATCH` / `HEADER_FOOTER_REDERIVE_MISMATCH` / `FLAGS_REDERIVE_MISMATCH` / `OCR_ON_TEXT_PAGE` | The ingest's files (anchors, lane B or table rows, header/footer records, grade-moving page flags, OCR text on a text page) do not re-derive from its source (edited or forged after the ingest; for a page class, also a check run without pdfplumber, § Re-derivation) | Do not edit the card: report it; the orchestrator re-ingests |
| `SOURCE_REDERIVE_ERROR` | A poppler tool (`pdftotext`, `pdfinfo`, `pdfimages`) failed or timed out on the source, the HTML no longer decodes, or memfd sealing is unavailable (non-Linux) | Re-run with a larger `--timeout` if it timed out; otherwise report it |
| `NO_INGEST_ROW` / `INGEST_MANIFEST_HASH_MISMATCH` | The ingest was not written by a recorded `pdf_ingest` run, or its manifest changed afterwards | Report it; the orchestrator re-ingests |
## Selftest
`lit_verify.sh selftest <SCRATCH_DIR>` builds synthetic ingests in a work dir
under `SCRATCH_DIR`, plants fabrications in cards, runs the real card gate on
each and checks every expected verdict: a genuine card (PASS), one digit
changed, a quote cited to the adjacent page, an invented 12-word quote, a
dropped minus sign, a non-identifying short quote, a tampered anchor file, a
vision claim on a `vision_validated: false` card (all FAIL), a synonym-swap
paraphrase (NEEDS_REVIEW via NEAR), an OCR quote on a truly scanned page
(PARTIAL via PASS_OCR), the fused compound hyphen (PASS_SKELETON), a
sparse non-image page (COVERED_SPARSE), a number printed on the page but
not in its quote (NEEDS_REVIEW via NUMBER_NOT_NEAR_QUOTE, 1.2.0), a claim
on an existing OCR text layer page (PARTIAL via PASS_OCR_LAYER, 1.2.0), and
on a column-major table page (1.3.0): a number found through its one lane B
row (PASS), a fabricated value the same row prints in another column (PASS:
the documented residual), a fabricated value from a look-alike row and a row
label shared by two rows (both NEEDS_REVIEW via NUMBER_NOT_NEAR_QUOTE), the
column label (1.5.0: the same-row residual number is `column-not-checked`
and a one-number quote on that page is column-checked, verdict still PASS),
and (1.4.0) on a synthetic arXiv HTML paper ingested through `pdf_ingest`'s real
`--source arxiv-html` lane (units `front`, `1`, `1.1` with a table, `2` with
an `ltx_ERROR` span, `bib`): a genuine v2 card with a number found in-quote
and one through its `rows/` row (PASS via NUMBER_OK), a quote cited to
another section (FAIL via WRONG_SECTION), a claim in the conversion-error
section (PARTIAL via PASS_CONVERSION_ERROR), a card pinned to another arXiv
version (FAIL via SOURCE_VERSION_MISMATCH) and a v1 card on the HTML ingest
(FAIL via ANCHOR_KIND_MISMATCH); (1.6.0) on a sign-and-row page: a genuine
spaced minus `= - 0.041` (PASS) and its sign-flipped claim (FAIL via
NUMBER_MISSING), a minus deleted inside a quote (NEEDS_REVIEW via
QUOTE_NUMBERS_DIFFER), a prose quote naming a prose lane B line
(NEEDS_REVIEW via NUMBER_NOT_NEAR_QUOTE) and `vision_validated: true` for a
model not on the list (FAIL via VISION_NOT_ON_LIST); (1.7.0) a forged ingest
set with consistent hashes and a MANIFEST row (FAIL via
ANCHOR_REDERIVE_MISMATCH), an ingest without a pdf_ingest MANIFEST row (FAIL
via NO_INGEST_ROW), and on a column-major page a row label alone on its
anchor line (PASS via NUMBER_OK) and a prose quote starting with that label
(NEEDS_REVIEW via NUMBER_NOT_NEAR_QUOTE); (1.8.0) three forged manifest
fields, each with a forged MANIFEST row: a header/footer record that lets a
span quote skip a printed line (FAIL via HEADER_FOOTER_REDERIVE_MISMATCH),
removed `scanned` / `ocr_layer_suspected` flags (FAIL via
FLAGS_REDERIVE_MISMATCH) and OCR text on a text page (FAIL via
OCR_ON_TEXT_PAGE); (1.8.0, part 2) on a number-parsing page: a genuine
footnote-lettered negative `−0.164a` (PASS), its minus deleted in the quote
and a superscript minus deleted from `10⁻³` (both NEEDS_REVIEW via
QUOTE_NUMBERS_DIFFER), a spaced minus after a word (NEEDS_REVIEW via
NUMBER_SIGN_AMBIGUOUS), and a line-start cell `– 0.25`, genuine (PASS) and
sign-flipped (FAIL via NUMBER_MISSING); (1.8.0 part 3, `ladder-v4`) a quote
that adds a spaced dash before the positive `1.20b` (NEEDS_REVIEW via
QUOTE_NUMBERS_DIFFER); (1.8.0 part 5) a listed reader's vision reading on a
page with an existing OCR text layer (NEEDS_REVIEW via VISION_ON_SCAN_PAGE);
(1.15.0) on a figure page, a listed reader's number from a chart that the
paper prints nowhere (PARTIAL plus the warning
VISION_FIGURE_UNCORROBORATED) and one the paper prints on another page
(PARTIAL, no warning); the page-3 vision claim of an unlisted reader under
the researcher's choice (`vision_override: check`), checked like a listed
reader's reading (PARTIAL, no VISION_UNAVAILABLE_CLAIM); plus one check that
each card run wrote one MANIFEST row with a matching `card_check` hash: **56
card expectations**. Eighteen `route`, `audit`, `settle` and
`vision-override` expectations follow (the ROUTE line and a row
with environment keys by presence only; a selftest card's synthetic source
is `chain_unconfirmed`; a second audit of the same card is refused, exit 9;
a glob character in the agent ID is refused, exit 8; a card edited after its
check is refused, exit 9; a figure reading by a model DAAF cannot confirm is
listed for a person; `settle --list` lists it and marks the chain not
settleable; `settle` refuses empty user words, an id not of this audit, the
chain's id and a `does_not_match` without a rationale, exit 11, and records
a matching reply with its hash-bound row, exit 0; `vision-override` refuses a
withdrawal with no choice in effect, empty user words and a `check` without
a rationale, exit 11, and records a `check`, after which `route
--reader-model` prints `VISION: override-check` with the warning, and
`VISION: text-only` once the choice is withdrawn). The same selftest adds 24
`report` expectations (a synthetic review built by `lit_corpus`, memos with
planted violations; literature-review `references/lit-verify-report.md`),
**98 in all**. It needs
no PDF tools, but it needs rapidfuzz, polars or pyarrow, **and polars** (the
HTML case is ingested by `pdf_ingest`, which writes its page_index with
polars; without it, exit 10). It removes its work dir and prints
`lit_verify: SELFTEST PASS: 98/98 expectations held; work dir removed`.
## Worked examples
`tests/python/fixtures/lit_verify/example_evidence_card.md` is a complete
card for the synthetic fixture `tests/python/fixtures/pdf_ingest/pdf/multipage.pdf`
(every word invented): a PASS_SKELETON anchor on a fused hyphen, a sparse
figure page, a truly scanned page read through OCR (PASS_OCR plus
SCANNED_PAGE), an SE in parentheses and a U+2212 minus. Its expected result
is verdict PARTIAL, `FULL_TEXT_PARTIAL`. A minimal block:
```lit-verify
{"schema_version": "lit-verify-card-v1", "paper_key": "user_00000001",
"source_pdf_sha256": "<64 hex from ingest_manifest source.sha256>",
"reader_model": "claude-opus-5-5", "vision_mode": "auto",
"vision_validated": true, "vision_list_version": "vision-list-2026-10-30",
"coverage": [{"page": 1, "anchors": ["<a verbatim, distinctive sentence from page 1>"],
"summary": "one line"}],
"claims": [{"claim_id": "C1", "page": 1, "quote": "<verbatim sentence>",
"numbers": ["-0.087", "(0.044)"], "source": "text"}]}
```
Here `vision_validated` is `true` because `claude-opus-5-5` is on the list and
the dispatch in this example named the route `anthropic-subscription`
(front matter `reader_route`). A reader whose route was not given, or whose
model and route have no row, writes `false`.
(Placeholders shown in angle brackets are for illustration; a real card has
none, and a card holding this block verbatim would fail its hash check.)
**arXiv HTML.** `tests/python/fixtures/lit_verify/arxiv_html_evidence_card.md`
is a v2 card for the synthetic LaTeXML fixture
`tests/python/fixtures/pdf_ingest/html/synthetic_latexml.html` (every word
invented), ingested with `--source arxiv-html`: section locators including
fallback ids (`S1`, `S1.SSx1`), a composed thesis number (`2.3.1`) and a
de-duplicated heading, heading-only `sparse` units covered by their heading
(COVERED_SPARSE), a table claim quoted through the row, a quote that spans
into the next unit with `section_end`, and a claim in a conversion-error
unit (PASS_CONVERSION_ERROR). Expected: verdict PARTIAL,
`FULL_TEXT_PARTIAL`. A minimal block:
```lit-verify
{"schema_version": "lit-verify-card-v2", "paper_key": "user_00000002",
"anchor_kind": "arxiv_html", "source_sha256": "<64 hex: ingest_manifest source.sha256>",
"source_version": "v4", "reader_model": "claude-opus-5-5", "vision_mode": "off",
"vision_validated": false, "vision_list_version": "vision-list-2026-10-30",
"coverage": [{"section": "front", "anchors": ["<a verbatim, distinctive sentence>"],
"summary": "title and abstract"}],
"claims": [{"claim_id": "C1", "section": "3.2", "quote": "<verbatim sentence>",
"numbers": ["0.41"], "source": "text"}]}
```
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| canvas-designanthropics/skillsCreate beautiful visual art in .png and .pdf documents using design philosophy. You should use this skill when the user asks to create a poster, piece of art, design, or other static piece. Create original visual designs, never copying existing artists' work to avoid copyright violations. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| docxanthropics/skillsUse this skill whenever the user wants to create, read, edit, or manipulate Word documents (.docx files) or Word templates (.dotx files). Triggers include: any mention of 'Word doc', 'word document', '.docx', '.dotx', or requests to produce professional documents with formatting like tables of contents, headings, page numbers, or letterheads. Also use when extracting or reorganizing content from .docx or .dotx files, inserting or replacing images in documents, performing find-and-replace in Word files, working with tracked changes or comments, or converting content into a polished Word document. If the user asks for a 'report', 'memo', 'letter', 'template', or similar deliverable as a Word or .docx file, use this skill. Do NOT use for PDFs, spreadsheets, Google Docs, or general coding tasks unrelated to document generation. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| pdfanthropics/skillsUse this skill whenever the user wants to do anything with PDF files. This includes reading or extracting text/tables from PDFs, combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks, creating new PDFs, filling PDF forms, encrypting/decrypting PDFs, extracting images, and OCR on scanned PDFs to make them searchable. If the user mentions a .pdf file or asks to produce one, use this skill. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| pptxanthropics/skillsUse this skill any time a .pptx or .potx file is involved in any way — as input, output, or both. This includes: creating slide decks, pitch decks, or presentations; reading, parsing, or extracting text from any .pptx or .potx file (even if the extracted content will be used elsewhere, like in an email or summary); editing, modifying, or updating existing presentations; combining or splitting slide files; working with templates (.potx), layouts, speaker notes, or comments. Trigger whenever the user mentions "deck," "slides," "presentation," or references a .pptx or .potx filename, regardless of what they plan to do with the content afterward. If a .pptx or .potx file needs to be opened, created, or touched, use this skill. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |