AI evidence in Verify

The Verify view can inspect text and visible layout patterns alongside Content Credentials, integrity, rights and production checks. Available checks start automatically after opening a file. The floating toolbar follows the file you are reading and keeps file, text, URL, inspection and report actions together. On a phone, additional actions are under More. Inspection runs on the device in the web shell and CLI. An MCP request sends the explicitly supplied file bytes to the MCP server.

Reading an assessment

A rounded card with a coloured edge, an eyebrow heading or padded labels in a short list can occur in human work. Useful categories, ordered procedures, quoted examples and brand requirements can explain these patterns. Missing metadata, unavailable models and unread text do not increase evidence strength. No finding establishes human authorship through absence.

Eyebrows count as weak clues whether or not their wording repeats the heading. Explanatory examples remain excluded. Select a finding to see its text spans or page regions, measurements, method and alternative explanations. Repeated occurrences retain their individual locations and measurements. Page navigation and zoom keep region overlays aligned. Filters leave the assessment unchanged.

Review annotations record a required brand pattern, quoted example or misdetected region. An annotation retains the original assessment. Retrying keeps recent assessments and their notes in the JSON export. JSON and readable exports contain recovered text, locations, coverage, versions and file/report SHA-256 digests. Reload checks the report digest against the original file; those hashes establish binding and integrity, not the truth of imported conclusions. Imported probability claims require an installed calibration and are currently refused.

Coverage

The initial browser read covers up to six pages. Inspect more pages to expand the read, within the 100-page session limit. Each page records completed, partial, skipped, unavailable, unsupported, failed or cancelled collectors. A cancelled job retains collected evidence. File replacement and view teardown stop updates to the previous panel.

Image decoding admits at most 40 megapixels and uses a preview edge no larger than 1800 pixels. Accent-card segmentation targets rounded neutral panels with a thin saturated edge; white-on-white panels and other treatments may be missed. Source text is bounded to 65,536 characters per page. Classifier windows use measured token counts, retain context near the document tail and record every inspected range. A bounded sample of a long document reports gaps rather than claiming a complete read. The classifier gate is conservative English prose eligibility, not a general language detector.

Local OCR and the text classifier run when cached. OCR in More uses the existing model download offer. Eligible text without a cached classifier gets Download and Skip in the toolbar. Download opens the standard consent and size offer; Skip retains the assessment with that check marked unavailable. Inspection introduces no file upload or automatic file retention.

PDF reports

Choose Report in the toolbar, or in More on a phone, for a paginated PDF. The report includes the current file preview, metadata, recorded EXIF map, credential observations, completed production checks, and a preview with located findings for each inspected page. Pattern-confidence rings describe detection of features; per-page AI probabilities remain unestimated. Pages beyond the current inspection scope are identified as unread.

The PDF includes the detailed evidence and production JSON as associated data files. Lolly adds Content Credentials and verifies the signature before offering the download. An enrolled device uses its enrolled signing identity; otherwise Lolly uses a local self-signed key. The report states which applies. The signature protects the report and does not authenticate the source file or its authorship. A signing failure produces an error instead of silently saving an unsigned PDF.

Policy in More opens the otherwise hidden Production policy controls. These compare explicit delivery requirements; they do not contribute to AI confidence. See Production checks for requirements and review records.

Calibration status

The development pilot downloaded 1,400 labelled clean scientific abstracts from the RAID dataset, pinned to revision 865cac74188466cb0c3b7574a10204007b57a459. The dataset card declares MIT; upstream human-source licences remain applicable. Source groups were kept together across development, calibration and holdout. This sample does not establish creator, generator-family or domain independence, and pretrained exposure is unknown.

The original CPU-runtime pilot used full graph optimization. Its held-out eligible sample contains 38 human and 164 generated documents. At the operating point selected from calibration for a nominal 1% human false-positive rate, both the old prefix classifier and the revised windows recalled 150 of 164 generated documents, with zero observed human false positives. The 95% upper false-positive bound is still about 9.2%. The candidate fusion's empirical calibration error is 0.0906, above the frozen 0.05 gate. These results support continued evaluation and explicit coverage; they do not qualify a probability release or demonstrate a broad accuracy improvement.

A later browser/Node comparison exposed a quantized CPU graph-optimization discrepancy. Both runners now use basic optimization, and reports distinguish CPU and WebAssembly method versions. The earlier pilot does not calibrate this corrected runtime. A separate development comparison retains the previous prefix runner as its baseline; fresh independent holdout qualification is still required.

Pattern fixtures have geometry labels, not inferred authorship labels. Visual authorship calibration needs reviewed creation history, matched designs with and without the motifs, human templates and generated designs without those motifs. Current-generator, mixed, translated, OCR and broader document populations remain unqualified. Visual clues stay available for review without a numeric authorship claim.

CLI and MCP

lolly inspect file.svg --forensic
lolly inspect document.pdf --forensic --page-cap=12
lolly detect-ai --in=article.txt --json

detect-ai retains its existing estimate fields and adds window ranges, token counts and completeness. probAi is the legacy field name for a raw classifier score. The human-readable output labels that score explicitly.

lolly_inspect accepts forensic: true, a named file and optional forensicPageCap from 1 to 100. Forensic, production and motion inspection use separate requests. All surfaces use the same engine evidence rules and report contract; shell-specific missing collectors remain visible.

Maintainers can reproduce the public development corpus with node scripts/prepare-forensic-corpus.ts, run node scripts/evaluate-forensic-ai.ts for the baseline, freeze preregistration.json, then run the evaluation with --holdout. A frozen evaluation cannot be overwritten by another development run. node scripts/export-forensic-calibration.ts exports the unreleased candidate from those frozen records without refitting.

node scripts/prepare-forensic-design-review.ts creates a review worksheet, provenance manifest and matched SVG, PNG, PDF and native PowerPoint controls. These synthetic controls test localization and exclusions; they are not independent authorship examples. Run node scripts/check-verify-forensic-browser.ts <local-web-origin> against a built web shell to exercise the inspection flow, review export, reload validation, cancellation and mobile layout. Downloads, raw material, private review examples and machine reports stay under plans/. No calibration artifact is installed by these workflows.