Lolly
افتح التطبيق
  1. الثقة
  2. بيانات الاعتماد والتحقق

AI evidence in Lolly Verify

Someone sends you an article, a report or a slide deck, and you wonder whether a model wrote the words. Drop the file on Lolly Verify, or paste the text. Lolly marks the places where the words or the layout look like typical generated work: the exact phrases, headings and page regions, why each one counts and what else could explain the pattern.

You get evidence, not a verdict. The checks run on your device, in the app and in the command line, and start on their own when you open a file or paste text. Nothing is uploaded unless you send a file to the MCP server yourself.

What you get

  • Each clue located in the text or on the page, with its measurement and method.
  • The innocent explanations that could produce the same pattern.
  • How much of the file was actually read, and what was skipped.
  • A signed report you can keep, with your own review notes.

What you never get

  • An AI percentage. No released calibration supports one yet, so the view says Unestimated.
  • A verdict of human authorship. Missing metadata, unavailable models and unread text never count either way.
  • A guess at which model did the writing.

Careful human writing can trip the same patterns. Useful categories, ordered procedures, quoted examples and brand requirements explain many of them, and the view lists those alternatives beside each finding.

Text that Lolly's own reword model wrote is different: it has a public watermark that Lolly Verify can detect. See What Lolly Verify checks.

Reading an assessment

ReadingWhat it means
Evidence strengthA heuristic index of the families of clues observed. Repeats of one pattern count together. This is not an authorship probability.
Pattern confidenceHow sure Lolly is that it recognised or located the particular pattern, with an observed or heuristic basis.
Estimated AI involvementA probability for a declared reference population, available only with a matching released calibration. Currently Unestimated in the view and Not estimable in reports.
Declared originCredential declarations, unsigned metadata and container hints, with their integrity and scope kept separate from inference.

Some of the patterns Lolly looks for:

Select a finding to see its text spans or page regions, measurements, method and alternative explanations. Repeated occurrences have their own locations and measurements. Page navigation and zoom keep the region overlays aligned, and filters leave the assessment unchanged. Rules version forensic-ai/2 invalidates calibrations from earlier rules.

Seeing where: the heat view

The text and the page preview show the evidence as layers, so you can see the location of each observation before you read a list of findings. Toggle the layers beside the page selector.

LayerWhat it shows
HeatEach sentence, and each located region of a page, shaded from amber to red by the findings that touch each one. A specific artifact counts at full strength, a style clue at a little over half, and an excluded match adds nothing. Heat is a reading aid, never a probability.
Matched wordsThe exact characters each finding matched: red for a specific artifact, amber for a style clue, and an underline for a match excluded as quoted or discussed.
Classifier chunksShown only when the local classifier ran. The text is split into groups of sentences of at least 55 words, and each group's raw score is drawn in its own column beside the heat. Scores below 0.85 are not drawn, because they are common on human writing. A group at or above the sentence threshold (0.93 for the current model) is also underlined.

Short passages score more widely than whole documents, so the chunk threshold is higher than the document one, and chunk scores never change the evidence index. On the current development corpus, 1 of 751 human documents had a group at or over 0.93, and the model rarely reaches that level on plain-text chat answers.

The narrow columns at the left of the text map the whole document, one cell per sentence. Select any lit sentence, a cell in the map or a region on the page to open the inspector. It lists every signal there with its guidance level and strength, the classifier's raw score for that group against its threshold, and a link to each full finding. Previous and Next step through the sentences with a signal. Signals about the whole text, such as even sentence rhythm or the overall classifier reading, are listed above the text rather than shading one sentence.

Recording your review

Annotations record a required brand pattern, a quoted example or a misdetected region, and each one keeps the original assessment alongside the note. Retrying keeps recent assessments and their notes in the JSON export.

JSON and readable exports contain the recovered text, locations, coverage, versions and SHA-256 digests of the file and the report. Reload checks the report digest against the original file. Those hashes establish binding and integrity, not that the imported conclusions are true. Imported probability claims need an installed calibration, and are refused today.

What each format covers

FormatAvailable inspection
Text and MarkdownOriginal characters, contextual findings, short numbering sequences and local classifier windows. No pixel geometry.
PNG, JPEG and WebPBounded pixel segmentation and optional local OCR with estimated line boxes. OCR may miss text, so coverage stays partial.
SVGPassive supported shapes and text, with a resource-blocked preview and a pixel fallback. Scripts, external resources, unsupported CSS, filters and masks are excluded.
PDFNative page text and estimated line geometry, supported page rendering and OCR for scans. Font substitutions and renderer limits stay explicit.
PPTXBrowser slide rendering, native text and supported shapes. The Node adapter currently reports native text with layout unavailable.
DOCXAvailable text extraction. Rendered page geometry is unavailable.
Video and audioExisting verification stays available; judging authorship over time is not supported.

Local OCR and the text classifier run when they are cached. OCR, under More, uses the usual model download offer. Eligible text without a cached classifier gets Download and Skip in the toolbar: Download opens the standard consent and size offer, and Skip keeps the assessment with that check marked unavailable. Inspection keeps no file automatically.

The floating toolbar follows the file you are reading and keeps the file, text, URL, inspection and report actions together. On a phone, more actions are under More.

Limits of the read

The first browser read covers up to six pages. Inspect more pages to widen it, within the 100-page session limit. Each page records which collectors completed, ran partially, were skipped, unavailable, unsupported, failed or cancelled. A cancelled job keeps the evidence it collected. Replacing the file or leaving the view stops 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 looks for rounded neutral panels with a thin saturated edge; white-on-white panels and other treatments may be missed. Source text is bounded at 65,536 characters per page. Classifier windows use measured token counts, keep context near the end of the document and record every inspected range. A bounded sample of a long document reports gaps rather than claiming a complete read. The classifier's gate is conservative eligibility for English prose, not a general language detector.

Reports

Choose Report in the toolbar, or under More on a phone, for a paginated PDF. It includes the current file preview, metadata, the recorded EXIF map, credential observations, completed production checks and, for each inspected page, a preview with the located findings. Pattern-confidence rings describe how well a feature was detected; per-page AI probabilities are unestimated. Pages beyond the current inspection are marked as unread.

The PDF has the detailed evidence and production JSON as attached data files. Lolly adds Content Credentials and verifies the signature before offering the download. An enrolled device signs with its enrolled identity; otherwise Lolly uses a local self-signed key, and the report says which. The signature protects the report. It does not authenticate the source file or its authorship. If signing fails, you get an error rather than a silently unsigned PDF.

Policy, under More, opens the otherwise hidden Production policy controls. They compare explicit delivery requirements and do not add to AI confidence. See Production checks for requirements and review records.

From the command line and MCP

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

detect-ai keeps its existing estimate fields and adds window ranges, token counts and completeness. probAi is the older field name for a raw classifier score; the readable output labels that score plainly.

lolly_inspect accepts forensic: true, a named file and an optional forensicPageCap from 1 to 100. Forensic, production and motion inspection use separate requests. Every surface uses the same engine evidence rules and report contract, and a collector missing on one shell is shown as missing.

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; the licences of the upstream human sources still apply. Source groups were kept together across development, calibration and holdout. This sample does not establish independence of creator, generator family or domain, and exposure during pretraining is unknown.

The original CPU-runtime pilot used full graph optimisation. Its held-out eligible sample contains 38 human and 164 generated documents. At the operating point chosen 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 no observed human false positives. The 95% upper bound on false positives 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 show a broad accuracy improvement.

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

Pattern fixtures have geometry labels, not inferred authorship labels. Calibrating visual authorship 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 are available for review without a numeric authorship claim.

For maintainers: measuring rule changes

Rule changes are measured with node scripts/prepare-forensic-corpus-v4.ts, which assembles a wider local corpus (RAID across eight domains, HC3 human and ChatGPT answers, learner essays, presumed-human READMEs and current chat answers, split by source group into development and holdout), and node scripts/evaluate-forensic-rules.ts --compare=<earlier run>, which reports each rule's firing rate on human and AI documents with an upper bound and the movement of evidence bands. Tune on the development split only.

To reproduce the public development corpus, run node scripts/prepare-forensic-corpus.ts, then node scripts/evaluate-forensic-ai.ts for the baseline, freeze preregistration.json, and 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, a provenance manifest and matched SVG, PNG, PDF and native PowerPoint controls. These synthetic controls test localisation and exclusions; they are not independent examples of authorship. 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 are in a local, gitignored folder. None of these workflows installs a calibration artifact.