Changelog
Last updated: August 3, 2026
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.
- Service worker cache:
clarity-v33.
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 youconditional-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.