Changelog
Release Notes
Internal
VERIFIED · Authentication Platform
Current: v23 — March 2026
Design system unified around Poshmark brand. Findings redesigned as evidence presentation. Loader, color, hierarchy, and mobile improvements.
Design System
- Poshmark accent color — primary action color unified to
#7B1F5E across all interactive elements. Apple blue removed entirely.
- Warm color tokens — background shifted from clinical white/blue-black to warm off-white and near-black across light and dark themes.
- 3-signal verdict system — green/amber/red for authentication verdicts only. All other UI elements monochrome.
- Body gradients removed — pure background, no atmospheric effects competing with content.
Findings
- Evidence presentation — findings redesigned as a clean list. No containers, no badges. Severity through typography only.
- Severity-based expansion — Critical and Major findings expanded by default. Minor and Informational collapsed, tap row to expand detail.
- Dot-only severity signal — 5px CSS dot replaces emoji severity indicators. Red for critical, muted for everything else.
Verdict Card
- Confidence label enlarged — 36px weight 300. First thing the eye lands on.
- Confidence bar removed — replaced with inline percentage next to descriptor.
- Photo requests elevated — surfaces above findings when RAVEN needs more photos. Clearest next action.
- Begin analysis → New Submission — analyze button swaps to New Submission after verdict renders, resets on new submission.
- Scroll on analyze — page scrolls to loader immediately when analysis begins.
Loader
- Magnifying glass removed — replaced with text-only loader. No icon, no spinner.
- Blur-to-sharp headline — "Checking [Brand] [Category]" starts blurred, resolves once over 1.6s. No oscillation.
- Single phrase per submission — one brand-specific phrase picked from pool, held for the full analysis. Rotates on new submission.
Mobile
- Sticky analyze button — "Begin analysis" fixed above bottom nav when photos are uploaded.
- Header overlap fix — main content padding corrected on mobile.
- Floating bug button removed — lives in corner menu only.
Intake Form
- Dropdown auto-confirm — AI brand/category detection auto-confirms without manual step.
- Upload zone refined — solid border, no dashed line, no plus icon. Typography carries it.
- Microcopy updated — "Begin analysis", "Drop photos or click to upload", "Select brand and category to view required angles".
Changelog
Release Notes
InternalVERIFIED · Authentication PlatformCurrent: v22 — March 2026
Major UI overhaul, motion system, mobile layout, analytics rebuild, live KV injection, and RAVEN enrichment schema.
Navigation
- Corner menu — Settings, docs links, CSV export, and bug report consolidated into a Claude-style avatar dropdown. Primary nav reduced to three tabs: New Submission, Library, Analytics.
- "New Submission" tab label — renamed from "New" for clarity.
- Sun/moon theme toggle — replaced text button with monoline SVG icons. No emoji.
- New Submission CTA — full-width primary button added to the bottom of the verdict card. No hunting after a verdict.
- CSV export — moved from header to corner menu. No longer occupying primary nav space.
Motion System
- Blur-replace transitions — verdict card, panel switches, and corner menu use
blurReveal animation. Content materializes rather than snapping in.
- Finding cascade — findings blur in with 60ms staggered delay per item when a verdict renders.
- Learning mode detail reveal — finding bullet blocks use blur-reveal with cascaded delay. Each block reveals independently.
- Theme switch blur — brief blur-through effect on light/dark toggle replaces the previous hard snap.
Analytics
- Card-based modules — modular accent cards replace flat stat tiles. Authenticated (green), rejected (red), suspicious (yellow), override (orange) each carry distinct visual weight.
- Submission trends chart — day-bucket bar chart showing submission volume. Period-aware when filters are applied.
- Period, submitter, and brand filters — all cards update simultaneously. Submitter and brand lists auto-populate from history.
Library
- Self-contained history detail — opening a past submission from the Library tab now shows an inline detail view within the Library panel. Previous behavior redirected to the New Submission panel, breaking library context.
- Back navigation — detail view includes a back button returning to the library list without losing scroll position.
RAVEN
- Enrichment schema — RAVEN returns structured enrichment fields alongside the verdict:
material, pattern, hardware_tone, origin, model_family, size_variant, model_description. Populated at confidence-appropriate thresholds only. Displayed as chips beneath the verdict label.
- System prompt field constraints —
brand_detected and category_detected constrained to clean single-value fields. Eliminates analytics pollution from full listing descriptions being stored in brand/category slots.
- Listing title auto-generate removed — RAVEN no longer generates listing titles.
model_description enrichment field handles item description with better accuracy.
- KV injection active — Worker fetches brand and category knowledge from Cloudflare KV and prepends to prompt on every submission. Currently: LV brand core + 6 category entries.
Mobile
- Full responsive layout — single-column stack on small screens. Safe area insets for iOS.
- Bottom tab navigation — nav tabs fix to the bottom of the screen on mobile.
- Camera capture — dedicated camera button opens rear camera directly via
capture="environment". Photos go straight into submission.
Cloudflare infrastructure deployed. KV knowledge base populated and wired. Analytics bug patched.
Infrastructure
- Cloudflare Worker deployed —
bitter-grass-11f7.marquisfa2.workers.dev. API proxy, KV reads, R2 storage, CORS.
- Cloudflare KV namespace —
RAVEN_KV created and bound to Worker. Serves as the knowledge base layer.
- Cloudflare R2 bucket —
verified-storage created for reference photo storage.
- KV injection wired — VERIFIED sends brand and category slugs with each call. Worker fetches matching entries and prepends to RAVEN prompt.
Knowledge Base — Louis Vuitton
- Brand core entry — stamp morphology (L foot, O roundness, R leg, TT spacing), spacing rules, factory variance notes, hardware tone standard, known exceptions, RFID era handling.
- Bags category — canvas types, vachetta patina logic, hardware piece notes, lock/key matching, era notes, placement guidance.
- SLG category — scale-adjusted marker logic, compact label format variants, macro photography requirement for stamp reads.
- Clothing category — tag-first authentication, 2017 era split documented with different tag color/material/font standards for pre- and post-2017 pieces.
- Shoes category — insole/outsole origin consistency as primary check. Bag factory code logic explicitly excluded. Espadrille and seasonal exceptions noted.
- Sneakers category — inherits shoe logic, adds tongue code structure as primary marker. Dust bag as supporting evidence only.
- Accessories category — organized by construction family: plaque chain bracelet, engraved link bracelet, cord lock bracelet, logo brooch/pin. Micro-engraving as primary marker.
Bug Fixes
- Analytics field corruption — localStorage patch corrected entries where full listing descriptions were stored in
brand_detected and category_detected fields.
Photo editor, three-tier finding format, learning mode, voice standard, LV date code logic, Taigarama spec, and 25+ brand knowledge base.
Photo Tools
- Photo editor — in-browser rotation correction, fine-tune slider, zoom before submission. Edited photos flagged as locked. EXIF orientation handled on upload.
Verdict Rendering
- Three-tier finding format — marker header → lead line → bullet block. Replaces flat text/detail structure. Bullets separated by hairline rule, subordinate to the lead.
- Learning mode — toggle to show or hide finding detail bullets. Off by default. Critical and major finding details always visible regardless of mode.
- Marker chip verdict colors — remapped from visibility status to authentication signal: green (consistent), yellow (uncertain), red (failed), grey (not provided).
- Updated JSON schema —
marker, lead, bullets[] replace text/detail. Graceful degradation for legacy history entries.
RAVEN
- Voice and tone standard — forbidden constructions enforced in system prompt. Assessment reads like a senior authenticator talking to a colleague, not a report.
- Findings explanation standard — four-part requirement: specific characteristic observed, what authentic looks like, how this item compares, what confirms or signals counterfeit. "CC logo consistent" is not a finding.
- LV date code decode — interleaved digit method formalized. Positions 1&3 = week, 2&4 = year. Factory code reference table. Week 00 = Critical finding.
- Taigarama model spec — origin Italy, full Monogram canvas strap, clean side panels, 3-line label format. Défilé Homme 4-line label documented as a known counterfeit tell.
- Over-branding hard rule — exterior tabs, stamps, or hardware absent on authentic examples must be flagged Major or Critical. Non-negotiable.
- Chanel footwear spec — season code format, style number structure, insole foil stamp standard (deboss depth, warm matte gold), outsole markers, origin consistency rule.
- 25+ brand knowledge base — Hermès, Chanel, Gucci, Prada, Dior, Saint Laurent, Fendi, Bottega Veneta, Goyard, Loewe, Miu Miu, Rolex, AP, Omega, Cartier, Tiffany, Canada Goose, Moncler, Arc'teryx, Chrome Hearts, Rick Owens, Amiri, Maison Margiela, Off-White, and others.
Three-tier finding schema and marker chip color remapping. Export updated to match.
Schema & Rendering
- Finding schema updated —
text/detail replaced with marker, lead, bullets[]. Backward compatible with graceful degradation.
- Chip color remapping — marker chips now colored by authentication verdict signal rather than visibility status.
- CSS updated —
.finding-marker, .finding-lead, .finding-bullets, .finding-bullet class hierarchy introduced.
- CSV and text exports updated — export format reflects three-tier structure. Bullets concatenated with separator in CSV.
Reference library, override logging, human verdict log, reviewer notes, and bug reporting. Authentication becomes a full record-keeping system.
Reference Library
- Reference library panel — agents save authenticated items as reference entries with brand, model, year, tags, notes, and photo thumbnail. Persisted in localStorage.
- Library search — full-text search across brand, model, category, and tags.
- Recent submissions in Library — last 6 submissions shown as quick-reference cards at the top of the panel.
- Library context injection — matching library entries injected into RAVEN prompt as authenticated reference context.
Verdict & Logging
- Human verdict log — agents log final verdict (authentic / replica / cannot authenticate) separately from RAVEN's recommendation.
- Override detection — disagreement between agent verdict and RAVEN recommendation flagged as an override in analytics.
- Reviewer notes — freeform field for override reasoning, escalation notes, or additional context. Stored with submission record.
- Additional authenticator field — co-authenticators logged per submission for escalation tracking.
- Bug report modal — in-app reporting with submission context capture.
Analytics panel, agreement rate tracking, cost display, and CSV export.
Analytics & Cost
- Analytics panel — summary stats, agreement rate vs. final verdict, override count, verdict distribution, volume by brand and category.
- Agreement rate — tracks alignment between RAVEN recommendation and agent's logged verdict.
- Cost tracking — per-submission API cost on verdict card. Session and all-time totals in Settings.
- CSV export — full history exportable with order, listing, brand, category, verdict, confidence, next action, reviewer, date, notes, assessment, and findings.
History panel with verdict filtering and replay. Photo lightbox with navigation and rotation.
History & Photos
- Authentication history panel — all submissions persisted in localStorage. Filterable by verdict type. Stat tiles for total, authenticated, flagged, and rejected.
- Verdict replay — clicking a history item reloads the full verdict, photos, and notes for review or re-submission.
- Photo lightbox — full-screen viewer with next/previous navigation, rotation, and access to the photo editor from within the lightbox.
Context panel for incomplete submissions. Confidence bar visualization. Physical inspection and NFC scan fields.
Submission Flow
- Context panel — when RAVEN determines key context is missing, a targeted request card surfaces before the full analysis. Agent fills in available context, then proceeds.
- Physical inspection field — freeform field for tactile observations. Passed as context, not used to adjust verdict.
- NFC scan field — URL or scan result from chip reads on post-2021 pieces. Passed as context for RAVEN to validate format.
- Confidence bar — visual indicator beneath confidence label. Fills by band level. Color follows verdict.
- Confidence descriptor — one-sentence explanation of what the confidence band means for this submission, generated by RAVEN.
Dark mode, sticky header, RAVEN loader animation, partial parse recovery.
Interface
- Dark mode — full dark theme with CSS variable system. Persists across sessions.
- Sticky header — locks to top with backdrop blur. Nav always accessible.
- RAVEN loader — animated magnifying glass with rotating brand-specific loading messages during analysis.
- Partial parse recovery — character-level JSON sanitizer handles malformed API responses. Partial parse flag shown when recovery was required. Verdict still renders rather than erroring.
Marker visibility chips, photo requests panel, next action field, and comparison context.
Verdict Output
- Marker visibility chips — each authentication marker displayed as a chip with its visibility status. Five statuses: clearly visible, partially visible, not assessable, not provided, not applicable.
- Photo requests panel — specific photo requests surface below findings when critical markers need better coverage. Includes angle and lighting guidance.
- Next action — structured action from RAVEN: Approve, Hold for review, Request more photos, or Escalate.
- Comparison context — RAVEN declares reference quality: exact match, near match, component family, or no reliable comparison.
Prescan removed. Photo checklist introduced. Agent confirms brand and category before analysis. Authentication philosophy and Chanel footwear knowledge added.
Submission Flow
- Prescan removed — separate pre-analysis API call for brand/category detection eliminated. Agent confirms brand and category before analysis begins. One full API call removed per submission — approximately 30% cost reduction. Latency step eliminated.
- Photo checklist — pre-upload — appears when brand and category are selected. Shows required and recommended markers with photography tips. Checklists for Chanel shoes, Chanel handbags, LV handbags, Hermès handbags, Rolex watches, plus generic defaults per category.
- Photo checklist — post-analysis — updates in place after verdict delivery with marker coverage results from
marker_visibility. Coverage bar shows required markers assessable.
RAVEN
- Authentication philosophy — three questions framework added to system prompt: do all markers agree, is execution quality consistent throughout, does each detail feel made or assembled.
- Chanel footwear knowledge — initial Chanel shoe logic: insole foil stamp standard, CC hardware proportions, outsole markers, origin consistency requirement.
- max_tokens increased to 4000 — extended output limit for longer findings on complex submissions.
- Prompt version tracking — version string added to Settings panel. v1.1 at this release.