Changelog

Last updated: September 26, 2026

September 26, 2026 | Media Lens coverage wording

• Limitations and Methodology now say that Media Lens coverage information appears only in labeled fixture examples, and that the live preview reads no Newsjack output or other coverage data. No coverage feature was turned on or off on the live preview; it never compared coverage across outlets.
• Clarity's analyzer, scoring, safety logic, and on-device privacy behavior are unchanged. Service worker cache stays clarity-v37.

September 25, 2026 | Media Lens preview disclosures

• Privacy, Limitations, Acceptable Use, and Methodology now describe the limited Jev-only Media Lens preview that runs on a separate operator host, ml-jev.manipulationscore.com. It analyzes one public page at a time from a short operator allowlist. The page sends only https addresses; the worker also accepts http addresses sent directly to its API, under the same allowlist and redirect rules. It is experimental and not production-ready, and the operator can pause it.
• Privacy now states what that host receives, exactly what each request to TypeSafe AI's Jev service contains, which fields its audit lines and ESTIMATED spend record keep, that a page recognized as paywalled is analyzed from its visible excerpt only, and that its web server keeps a standard access log set in the host's own configuration. The retention period of the host's logs is not yet documented.
• Media Lens examples labeled as fixtures or samples are made up. Comparing coverage across outlets is not live. Live pasted-text analysis and classifier.dev stay off.
• Clarity's analyzer, scoring, safety logic, and on-device privacy behavior are unchanged. Service worker cache stays clarity-v37.
• The Unreleased entry below records the state before the operator preview started on September 21, 2026.

Unreleased | Media Lens preview (not deployed)

• Added an experimental, unreleased Media Lens mode for public articles, ads, speeches, and campaign material, evaluated separately from Clarity as evidence-linked Language, Claims, Coverage, and Source context, with no overall score.
• Media Lens runs behind a local worker process that you or an operator run; the default and only continuous-integration-tested mode uses local example fixtures with no network calls.
• Clarity's files, scoring, safety logic, and privacy behavior are unchanged; Media Lens code is isolated from Clarity and is excluded from this site's deployment.
• Privacy, Limitations, and Acceptable Use now describe a possible future Media Lens live URL path as disabled, not deployed, and not available on the public site, without enabling it. Live pasted-text analysis remains prohibited. Clarity on-device analysis is unchanged.

September 3, 2026 | Clarity brand overhaul

• Reframed the public experience around Clarity, a private communication literacy product by Manipulation Score.
• Replaced the previous landing-page composition with the pause-field brand system and a clearer route into Learn and Analyze.
• Updated the wordmark, icon, manifest, metadata, and shared page shell without changing screening logic, safety behavior, or privacy architecture.
• Service worker cache: clarity-v37.

September 3, 2026 | Visual refresh

• Recomposed the public landing page around a calm private-screening workspace with clearer visual hierarchy and a more legible analysis preview.
• Added visible pattern tags and local-context cues to the illustrative hero card without changing screening logic, scoring, or methodology.
• Refined the Learn/Analyze entry cards and preserved keyboard focus and reduced-motion behavior.
• Service worker cache: clarity-v36.

August 17, 2026 | Public screening v0.3.3

• Platform site: Home, Learn (twelve functions), and Analyze as separate doors.
• Labeled gold harness for benign, autonomy, coercive, and safety slices. Regression count only, not a published accuracy study.
• Clause-aware scoring exclusions; class-based fiction/training and requested-logistics safety filters.
• Thread results name the highest band and how many messages sit in High/Moderate.
• Pause, boundary, and clarify responses all shown.
• Image upload remains disabled. Trust-strip wording is “Free, on your device.”
• Service worker cache: clarity-v35.

August 3, 2026 | Controlled Beta candidate v0.3.2

• History writes require the visible checkbox in the submitting tab; cross-tab storage sync added.
• Results clear when message text changes before re-analysis.
• Unicode normalization, message-boundary safety detection, and expanded negation/word boundaries.
• English-only abstention for unsupported languages; image upload disabled for this candidate.
• Pilot UI shows pattern bands without a 0-100 number; score explainer details fixed.
• Methodology updated for thread aggregation and v0.3.2.
• Expanded and locked the declared 40-case critical and 40-case benign safety corpus.
• Removed the remaining illustrative number, hid the disabled OCR control, and aligned Contact with v0.3.2.
• Service worker cache: clarity-v34.

August 3, 2026 | Controlled Beta v0.3.1

• Fixed Clear so the history opt-in checkbox stays synchronized with saved state.
• Thread screening no longer shows a numeric score when every message abstains.
• Smart apostrophes and extra whitespace no longer erase pattern matches.
• Safety attribution fixes: first-person threats after comma-and are detected; quoted training examples stay excluded.
• Regex word-boundary fixes reduce false violence evidence on words like “your.”
• Consistent Controlled Beta labeling across hero, results, and methodology.
• Service worker cache: clarity-v32.

August 3, 2026 | Controlled Beta v0.3.0

• Evidence-based score calibration: high, medium, and low specificity tiers; single broad matches no longer auto-escalate to Moderate.
• Expanded safety patterns for confinement, violence, stalking, and conditional self-harm variants.
• Benign safety-context handling for quotations, rehearsals, idioms, and protective instructions.
• Working email contact routes: feedback@ and security@manipulationscore.com.
• Mobile menu focus management improved.
• Service worker cache: clarity-v31.

Controlled Beta is an experimental testing stage. It does not imply validation, clinical accuracy, or safety-performance guarantees.

August 3, 2026 | Screening v0.2.8

• Fixed supportive-context false positives where autonomy phrases such as “whatever time you need to think” triggered Assigned obligation.
• Added instructional and procedural exclusions for training “nobody is required” and policy language.
• Added Round 11 regression and benign adversarial test suites.
• Skip link now moves focus to the main content region.
• Service worker cache: clarity-v30.

August 3, 2026 | Site trust pages

• Published Methodology, Contact and Corrections, Acceptable Use, Accessibility, and Changelog pages.
• Updated footer navigation, sitemap, and service worker shell for new pages.
• Screening logic remains v0.2.7 (no scoring or safety rule changes in this build).

August 2, 2026 | Screening v0.2.7

• Split explicit-future and continuous-adverb first-person stalking escalation.
• Require menace for single continuous-adverb stalking behaviors.
• Fixed benign false positives on requested pickups, rides, walks, and agreed meetings.
• Service worker cache: clarity-v28.

August 2, 2026 | Screening v0.2.6

• Extended demonstration phrasing for instructional attribution.
• Added refusal menace phrases (told/asked me to leave, go away, stop).
• Fixed news quotation plus real refusal-violating conduct false negative.

August 2, 2026 | Screening v0.2.5

• Filtered attributed behaviors before multi-behavior stalking counts.
• Fixed instructor and fictional two-behavior stalking quotation false positives.

August 2, 2026 | Screening v0.2.4

• Two-stage contextual stalking model (behavior plus menace).
• Removed generic unless you conditional-access scoring.
• Fixed coercive arrival false negative and invitation false positives.

Earlier alpha releases

• v0.2.3-v0.2.2: Expanded stalking patterns, clause-scoped negation, training attribution.
• v0.2.1: Per-sentence safety with local negation and attribution.
• v0.2-alpha: History opt-in (v2), abstention below 15 words, safety override, exclusion rules.

Known limitations

See Methodology for evaluation status, regression scope, and foreseeable failure modes. Clarity has not completed a representative validation study on natural messages.