Optional judge (TypeSafe / Jev)

doc2mark’s rules are deterministic, and a few of their decisions stay open by design: when the evidence is not strong enough the rule keeps the text (verbatim first). Three such decision points have a judge hook: an optional callable asked only about the cases the rule leaves open. The optional TypeSafe add-on answers all three with Jev, TypeSafe’s calibrated yes/no model. Without it, or when it cannot answer, the rules decide exactly as before.

What each hook decides

legibility_judge(page_text) -> Optional[float]: Is this PDF page’s extracted text layer legible? A text layer can be valid Unicode and still be nonsense: a font whose ToUnicode map shifts every letter (Lqyrlfh wrwdo), a subset font whose glyph IDs were taken for characters (6DPSOH), CJK code points substituted for others. No character rule sees these. The judge gets the page text of layers the deterministic garbage detector did not flag: all of it up to 1,500 characters, else a sample of its beginning, middle and end (500 characters each), so a legible opening cannot hide a garbled body. It is consulted only when OCR is on (--ocr <provider> --ocr-images): a page it rates illegible is OCR’d from its render instead of emitting the garbage. See doc2mark.core.strategy.judge_text_layer.
boilerplate_judge(line_text, context) -> Optional[float]: Is this repeated top/bottom line page chrome? Running headers and footers are removed only on strong evidence, and verbatim first: a running header keeps its first copy. A line that repeats but is attached to the page’s content, repeats on few pages, or is a number labelled like the pages, keeps every copy. The judge is asked about exactly those lines, with where and how often they repeat. A line it calls chrome is thinned to its first copy, which always stays, as plain text (or, when the text is already in the body of that page or an earlier one, as on a deck’s cover, that copy is the one left): the brand line printed beside every slide title of a deck goes from once per slide to once. The judge can never remove the last copy of a line, and it is not asked about a running header’s first copy (the rule already removed the others: a letterhead stays once). The question tells the model that titles, per-page labels (Lesson 3), unit notes and disclaimers are content; a judge that answers yes anyway thins them. See PDF text structure for the rule.
non_content_judge(ocr_text) -> Optional[float]: Is this OCR answer only a refusal, an error or a “no readable text” statement? The deterministic check fires only when the whole answer is such a statement; an answer such as “Unable to process the image. Please provide a clearer scan.” reads the same as a support page and is kept. The judge gets short answers (at most 600 characters) the patterns keep. From a TypeSafe probability of 0.95 the answer is no content: in the structured path the image is re-read by the free-form recovery, whose answer is screened the same way (the judge’s verdict is cached, so a re-read that returns the same words is left out too); in the free-form path (--no-structured) the answer is dropped and the result flagged ocr_refusal. From 0.90 up to 0.95 the answer is kept and flagged metadata["non_content_suspected"] and counted in the document’s metadata.extra["ocr_issues"]["suspected"], with its page (or slide, or sheet) in locations. A real answer the judge calls no content is lost, so the threshold sits above the overlap measured on the labelled sets. See doc2mark.ocr.refusal.

Each returns the probability of the “yes” and None when it cannot judge. For legibility_judge and boilerplate_judge, None, an exception or a value outside [0, 1] leaves the rule’s decision in place. For non_content_judge, None, an exception or a value that is not a probability (outside [0, 1], NaN, not a number) keeps the answer (it is then not written to the OCR cache).

Enabling the TypeSafe judge

pip install 'doc2mark[typesafe]'
export TYPESAFE_API_KEY=...            # https://console.typesafe.ai
doc2mark report.pdf --ocr tesseract --ocr-images --judge typesafe

from doc2mark import UnifiedDocumentLoader

loader = UnifiedDocumentLoader(ocr_provider="tesseract", judge="typesafe")
result = loader.load("report.pdf", ocr_images=True)
result.metadata.extra.get("judge")   # questions asked, cached, failed, tokens, cost

DOC2MARK_JUDGE=typesafe does the same when neither judge= nor --judge is given (--judge none / judge="none" switch it off). The default is no judge.

judge= also takes an object: any of the attributes legibility_judge, boilerplate_judge and non_content_judge it has are wired through to the PDF pipeline and the OCR providers (the last one via OCRConfig.non_content_judge). A hook passed explicitly (legibility_judge=, boilerplate_judge=, an OCRConfig that names a non_content_judge) wins over the object’s. TypeSafeJudge is such an object:

from doc2mark import UnifiedDocumentLoader
from doc2mark.judge import TypeSafeJudge

judge = TypeSafeJudge(hooks=("legibility", "non_content"),   # or $DOC2MARK_JUDGE_HOOKS
                      cache_dir="judge-cache",
                      timeout=2.0, max_workers=4)
loader = UnifiedDocumentLoader(ocr_provider="openai", judge=judge)

When the judge cannot answer

The add-on is optional in every sense: without the typesafe extra the SDK is never imported, and without a judge nothing is sent anywhere. When the judge is enabled but cannot answer, every hook returns None and the rules decide:

extra not installed, or no TYPESAFE_API_KEY: one warning (TypeSafe judge unavailable: ...) and the same output as --judge none;
the service rejects the key (401/403): the judge switches itself off for the run;
network errors, timeouts, rate limits (429/529), invalid answers: that question is decided by the rule, and the document logs one warning saying how many questions got no answer. Each HTTP attempt times out after 2 s and a question gets one retry, started only within a 4 s budget, so a question that cannot be answered costs about 4.5 s (up to about 6 s when the service asks to wait). Page and OCR-answer questions of a document are asked concurrently, header/footer lines one at a time. After three failures in a row the judge pauses for a minute (its questions then return at once), so during an outage a document costs at most a few seconds, again after each pause, and once per process with --parallel. HTTP 400, 404, 413 and 422 answers and invalid answers do not count toward the pause.

A conversion never fails because of the judge. A document converted while its judge failed on some question is not written to the document cache (cache_dir), and an OCR answer the judge could not screen is not written to the OCR cache: a later run asks again. A judge that cannot answer at all (no extra, no key, key rejected) is keyed in both caches as no judge, which is what its output is.

Model, questions and determinism

Every question is a TypeSafe Noul over a small named JSON state – page_text; the line with where and font in words (numbers become words: the model reads semantics better than arithmetic); ocr_answer – asked of the pinned model version jev-1.13.0 (never the moving jev-latest alias: the thresholds belong to a version). The question texts are versioned (legibility-v2, boilerplate-v2, non-content-v1, in doc2mark.judge.questions).

Jev is calibrated but not bit-exact (between two runs of the labelled sets most verdicts moved by about 0.01, two by 0.08 and 0.14), so every verdict is cached on disk, keyed by model id, question version and a hash of the state: $DOC2MARK_JUDGE_CACHE, else $XDG_CACHE_HOME/doc2mark/judge, else ~/.cache/doc2mark/judge (cache_dir=False keeps them in memory). A re-run gives the same output and sends nothing. The OCR and document caches key on the judge’s model, question version and threshold, so a result decided by another judge setup is not replayed.

Pages of one PDF are asked about concurrently (at most max_workers requests in flight), as are the OCR answers of one batch; each HTTP attempt times out after timeout seconds (2 by default), with one retry per question started within a 4 s budget.

The SDK logs every request and response body – your document text – at DEBUG level on the typesafe_sdk logger, and the method, URL, status and timing of each request and retry at INFO. doc2mark drops the SDK’s DEBUG records, so doc2mark -v shows the INFO lines but does not write page text to the log, and leaves the logger’s level alone, so a default run (WARNING) prints none of them. To see the wire log, ask for it with TYPESAFE_LOG_LEVEL=debug or set the typesafe_sdk logger’s level yourself.

Privacy

With the judge enabled, text from your documents leaves your machine and is sent to TypeSafe (api.typesafe.ai, the SDK’s default, which TYPESAFE_BASE_URL overrides): up to 1,500 characters of each PDF page whose text layer is judged (all of a short page, else its beginning, middle and end), the repeated header/footer lines that are judged, and short OCR answers. Per TypeSafe’s documentation, Jev is not trained on customer requests or responses and a Data Processing Agreement applies; zero data retention is offered only to enterprise customers (see https://docs.typesafe.ai/legal). Do not enable the judge for documents your agreement with TypeSafe does not cover. doc2mark never logs the API key.

Cost and latency

Jev is priced per input token ($0.042 per million, output free). A question costs about 420 to 470 tokens of fixed overhead plus its state; on the labelled sets a page-text question averaged 684 input tokens ($0.000029), a header-line question 573 ($0.000024) and an OCR-answer question 446 ($0.000019). Latency per request: p50 about 210 ms, p95 about 250 to 270 ms from a Linux host on 2026-09-30 (270 to 300 ms from a laptop for PR #22). Documents converted with the judge record what it did in metadata.extra["judge"] (name, model, asked, cached, fresh, failed, input_tokens, cost_usd), when it asked something; a document replayed from cache_dir carries the first run’s record.

Measured on whole documents (fresh verdict cache, OCR on; eval/judge_eval.py’s generated documents; added time from the 2026-09-30 re-run on a Linux host, the deck row from PR #22 on a laptop, not re-measured because the deck is not public):

document	questions	input tokens	cost	added time
12-page text report	13 (12 pages, 1 header line)	6,790	$0.00029	+1.2 s
8-slide deck with a brand line	10 (8 pages, 2 header lines)	5,051	$0.00021	+1.0 s
3-page letter with a letterhead	3 (pages; the letterhead is not asked about)	1,679	$0.00007	+0.4 s
30-page Traditional-Chinese company deck (image route)	2 (header lines)	1,135	$0.00005	+0.6 s

A second conversion of the same document is answered from the verdict cache: no request, no added time.

How well it works

The labelled sets are in tests/data/judge (see its README.md for sources and labelling); eval/judge_eval.py decides every item with the rule alone and with the rule plus the judge, exactly as the pipeline combines them. Three sets: TRAIN (the only set the legibility and boilerplate thresholds were calibrated on), TEST (held out: split from TRAIN by family, every variant of one template, base text or document on one side, near-duplicates kept together) and EXTERNAL (60 items written by the reviewer of this add-on, never used for calibration). The non-content thresholds (act from 0.95, flag from 0.90) are a fixed policy, set above the overlap of refusals and real answers. The positive class is the hook’s action: the page is illegible (OCR’d), the repeated line is thinned to one copy, the answer is no content.

hook, set (items)	rule only: accuracy / precision / recall	rule + Jev: accuracy / precision / recall
legibility, TRAIN (52)	59.6 % / 88.9 % / 28.6 %	98.1 % / 96.6 % / 100 %
legibility, TEST (46)	67.4 % / 100 % / 37.5 %	100 % / 100 % / 100 %
legibility, EXTERNAL (16)	68.8 % / 100 % / 16.7 %	100 % / 100 % / 100 %
boilerplate, TRAIN (42)	52.4 % / n/a / 0 %	95.2 % / 100 % / 90.0 %
boilerplate, TEST (42)	50.0 % / n/a / 0 %	95.2 % / 100 % / 90.5 %
boilerplate, EXTERNAL (3 asked of 20)	66.7 % / n/a / 0 %	100 % / 100 % / 100 %
non_content, TRAIN (91)	73.6 % / 100 % / 45.5 %	94.5 % / 97.6 % / 90.9 %
non_content, TEST (90)	76.7 % / 100 % / 52.3 %	97.8 % / 97.7 % / 97.7 %
non_content, EXTERNAL (24)	62.5 % / n/a / 0 %	83.3 % / 100 % / 55.6 %

Thresholds: legibility illegible below 0.8, boilerplate chrome from 0.7 (both calibrated on TRAIN), non_content no content from 0.95 and flagged from 0.90. Boilerplate is scored on the questions the pipeline asks (a running header’s first copy is decided by the rule: 17 of the 20 external boilerplate items, and 50 designed lines of the generated documents, are such lines, and the rule already keeps one copy of each). In output terms, on TEST the judge took 62 repeated copies out, none of a line that needed them, and left 2 chrome copies in.

The table is the re-run of 2026-09-30 at becb74b with an empty verdict cache (spark2, eval/docs_audit/run_judge_eval.sh; details in eval/docs_audit/results/judge_eval.md). It matches the PR #22 run except non_content TRAIN with the judge (95.6 / 97.6 / 93.2 % then): one refusal scored 0.94 this time, just under the 0.95 at which the judge acts.

On a real Traditional-Chinese company deck read in place for PR #22 (never committed, not re-measured), the brand line beside every slide title went from 29 copies to 1 (the cover’s), and garbled variants of its page texts were all caught (3 of 6 without the judge).

Margins (2026-09-30 run): legibility’s garbage pages score at most 0.79 and its legible pages at least 0.87 on TRAIN (0.87 on EXTERNAL, a lorem-ipsum page), around the 0.8 threshold; a false alarm costs one OCR pass of a legible page. Boilerplate’s content lines score up to 0.66 on TRAIN against 0.7; a false chrome verdict thins a content line to one copy. Refusals and real answers overlap between about 0.85 and 0.95, which is why only 0.95 and above acts: on TEST one real answer (“no results” from a search page) is still called no content, and on EXTERNAL 4 of 9 refusals are kept (3 of them flagged non_content_suspected). The judge cannot undo a false alarm of the deterministic garbage detector (a legible spec table with 17 replacement characters is still OCR’d): it is asked only about layers the detector keeps. The thresholds belong to jev-1.13.0: re-run eval/judge_eval.py --calibrate before moving the pin.

Current TypeSafe limits for jev-1.13: 100K tokens and 40 requests per second, 32K tokens of state per question (see https://docs.typesafe.ai/models); English is the model’s primary language, CJK is handled with lower accuracy.