A precise record of what testers can install today, what exists only as an unreleased source candidate, and which entries are retained as historical development snapshots. Source progress is not presented as a distributed release.
Solid archival goldCurrent signed private-beta APK — installable by authorised testers.
Dashed cyan outlineUnreleased source candidate — compiled evidence, but no signed APK.
Neutral ivoryHistorical build or snapshot — preserved for development context.
v0.8.0 · Unreleased source candidate
August 14, 2026 · versionCode 21
AI integrity, grounded claims and genuinely private local voice
This entry records only the incremental AI-integrity work in the current source. The YC build compiles, but v0.8.0 has no signed or distributable APK. Signing, cloud canary and exact-device qualification remain mandatory release gates.
Grounded generation
- ▸ Gemini 3.6 grounded cloud path: structured answer, TTS and atomic claim-to-source output with an operational model override.
- ▸ Claim-level provenance: deterministic citation validation and semantic/safety verification before Android renders the response.
Local runtime & retrieval
- ▸ LiteRT-LM 0.16: current Gemma 4 E2B/E4B on-device runtime.
- ▸ Private speech path: on-device recognizer only on supported devices, with no silent remote speech fallback.
- ▸ Attested assets: Gemma and Piper downloads are pinned and verified by exact size and SHA-256 before use.
- ▸ Weighted BM25 gate: local pack retrieval uses a refusal threshold and an automated IT/EN evaluation gate.
Hardening & evidence
- ▸ Input and audio lifecycle hardening: visitor-name prompt sanitisation, voice-call mute restoration and a thread-safe Visualizer lock.
- ▸ Automated evidence: Android, Functions, Worker and mostrapack checks pass at this checkpoint.
- ▸ Artifact boundary:
:app:assembleYc succeeds; the output is unsigned and is not offered for download.
v0.7.0 · Current signed APK
Signed August 9, 2026 · Private Beta 1 · versionCode 20
The installable private-beta build testers currently have
v0.7.0 is the only signed and distributable APK currently available to authorised private testers. It carries the complete visit-first product line below. These implementation and artifact checks do not imply requalification on every exact device in the future v0.8 matrix.
Packagecom.mostrami
SignatureAPK Signature Scheme v2 verified
Packaging16 KB verification recorded
Size308 MiB
Visit & artwork flow
- ▸ Visit-first navigation with a stable scan-to-chat handoff and shared Camera/Chat state.
- ▸ Image-first artwork guide with pinch zoom, point focus and circle selection.
- ▸ Museum Experience Packages with route metadata, guided visit state and first-artwork camera handoff.
Clio & content
- ▸ Clio Live native audio, interruption/barge-in, global assistant and real-amplitude waveform.
- ▸ Clio-led onboarding and bespoke MOSTRAmi app/web icons.
- ▸ Rich-text narration across Chat and Explorer, plus the structured Scholar Dossier.
Privacy & release quality
- ▸ Privacy-aware telemetry remains opt-in; the release client contains no API key.
- ▸ Android 16 KB compatibility and signed artifact verification are recorded.
- ▸ Retained as the rollback artifact when future signed v0.8 qualification begins.
Download current signed v0.7.0 private-beta APK
SHA-256 · 485387a1bd6e4c66fea89bf5e72b6b1f0e356697c0c595a37130c0d85fe49acb
v0.6.0 UX pass
June 12, 2026 | Debug APK generated
GLOBAL CLIO + MUSEUM PACKS
A Living Assistant Across the App, Plus Curated Museum Routes
The latest testing APK reframes the experience around the museum visit: Clio Live is available across the main app surfaces, local voice controls stay hidden unless relevant, Live and TTS cannot overlap, and Museums now lets visitors load curated route packs before scanning.
🏛️ Museum Packs
- ▸ Choose route: each museum can expose duration, audience, artworks, and route steps.
- ▸ Guided visit state: loading a pack creates the active visit and fills the artwork wishlist.
- ▸ Camera handoff: the visitor is sent directly to scanning the first artwork.
🎙️ Clio Everywhere
- ▸ Floating Live assistant: available on Home, Museums, History, Settings, Feature Guide and post-visit surfaces.
- ▸ Cleaner voice UI: local controls stay hidden outside local/expert context.
- ▸ No double audio: TTS is stopped when Live starts and hidden while Live is active.
v0.6.0 YC demo
June 11, 2026 | Debug APK generated
CLIO LIVE + INTERACTIVE ARTWORK
Native Live Audio, Visible Errors, and Visual Detail Chat
The YC demo APK adds the user-facing layer needed for a credible live demo: Clio Live now targets the Firebase Google AI native-audio model, exposes runtime errors with retry, and the chat image panel supports pinch-to-zoom, tap-to-focus, detected points of interest, and circle-select detail crops.
🎙️ Clio Live
- ▸ Native Gemini Live API: Continuous listening, voice response, and barge-in through Firebase AI Logic.
- ▸ Recovery UI: Live failures show the specific runtime error and a retry action inside the Live panel.
- ▸ Fallback clarity: Local Live remains the offline path when Gemma/STT/Piper assets are present.
🖼️ Interactive Artwork Chat
- ▸ Pinch-to-zoom directly in the chat artwork panel.
- ▸ Tap or circle a detail to crop the frame and continue the conversation on that area.
- ▸ Points of interest from the analysis are passed into ChatScreen, not only History.
v0.6.0 audit
June 10, 2026 | Legacy architecture snapshot
CLIO CLAIM-LEVEL PROVENANCE
Clio Multi-Layer Mixture-of-Agents Pipeline & Per-Message Provenance
Historical snapshot: the previous multi-agent approach has now been superseded by the grounded generator/verifier workflow. Android builds a strongly typed InterpretationContract, carries evidence into local Gemma prompts, parses cloud claim-to-source mappings, and renders verified citations on each Clio message.
🏗️ Contract-First Architecture
- ▸ InterpretationContract: Typed Kotlin data class with 8 sections (task, artworkContext, visitContext, userContext, styleContext, sourcePolicy, safetyPolicy, outputSpec).
- ▸ ContractBuilder: Reads SettingsRepository and assembles the contract with UUID, language, persona, privacy, and expertise level.
- ▸ Built-in demo broker: Demo evidence retrieval from Museum Packs, Licensed Critics, and MOSTRAmi KB. The production Vector/FTS broker remains a launch gate.
- ▸ PromptRenderer: Translates Contract + EvidenceBundle into a multi-section structured prompt for local Gemma 4 and AICore paths.
☁️ Cloud Agent Chain (Firebase Functions)
- ▸ [0] Preflight Gate: Privacy & consent verification before any AI call.
- ▸ [1] Source Broker: Extracts EvidenceBundle from artwork context.
- ▸ [2] Interpretation Planner: Selects 2-4 factual claims with JSON schema output.
- ▸ [3] Draft Generator: Writes sterile academic text based ONLY on evidence.
- ▸ [4] Audience Adapter: Rewrites for persona and audience without changing facts.
- ▸ [5] Safety Verifier: Independent verifier for EU AI Act guardrails, unsupported claims, and refusal conditions.
- ▸ [6] Voice Director: persona- and prompt-based prosody steering for natural TTS narration.
Safety, UI & Verification
- ▸ End-to-end error propagation: Firebase
HttpsError safety and consent states now reach Android without being flattened into generic connectivity errors.
- ▸ Per-message source UI: Chat bubbles store their own source chips, grounding label, contract task, and museum grounding flag.
- ▸ Build gates passed:
compileDebugKotlin, unit tests, lintDebug, bundleRelease, Wear debug build, and Firebase syntax check.
- ▸ Zero API Keys on Client: All keys managed server-side by Firebase (FF-101).
v0.6.0
June 5, 2026
👥 TOURSYNC & JOURNAL
👥 Clio TourSync Group Tours, Art Treasure Hunt & Timeline Scrapbook
Major update focused on enhancing the visiting experience: support for tour guides (sync and BLE radar), gamification for families, and redesign of the post-visit journal into a historical scrapbook.
👥 Clio TourSync (B2B/B2C Guide)
- ▸ WebSocket Sync: Real-time synchronization of participants' screens with the artwork scanned by the guide.
- ▸ BLE Proximity Radar: Separation alerts (under 15m) based on RSSI values and haptic vibration.
- ▸ Persona Steering: Centralized AI persona change from the guide for the entire group (Kids/Senior/Standard).
🎮 Art Treasure Hunt
- ▸ Wishlist Riddles: Wishlist transformed into textual/visual riddles based on the tour path theme.
- ▸ AI Quiz Master: Multiple choice quizzes generated by the AI after unlocking each artwork.
- ▸ Digital Stickers: Collectible graphic rewards unlockable on the ship log journal page.
📔 Scrapbook Log Journal
- ▸ Travel Timeline: New vertical scrapbook layout with personal memories, photos, and captions.
- ▸ Voice Journal: Recording and playback of audio notes (.aac) coupled with paintings or global visits.
- ▸ Premium Exports: Polaroid postcard generator for social sharing and PDF Scrapbook export.
v0.5.7
June 5, 2026
🖥️ DIAGNOSTIC HUD
🖥️ Diagnostic HUD, Scan Timeouts & JNI Recovery
✨ New Features
- ▸ Diagnostic HUD Overlay: Collapsible diagnostic dashboard overlaying the camera feed that displays active AI engine, RAM size, network status, and pipeline status.
- ▸ Fail-Safe Routing: Automatic failover of the AI pipeline to Gemini Cloud or local warnings in case model loading or inference freezes.
🐛 Hotfixes & Optimizations
- ▸ Inference & Load Timeouts: Strict 30s/35s timeouts on Gemma 4 JNI calls and model loading to avoid indefinite freezes of the scanning page.
- ▸ Clean JNI Mutex Release: Forced release of the
nativeMutex via JNI cancelProcess() when the timeout triggers, allowing scan retry.
v0.5.6
May 29, 2026
🎙️ CLIO LIVE
🎙️ Clio Live Audio, Dual-Engine & Dual Profile (Standard vs Expert)
✨ New Features
- ▸ Dual Configuration Profiles: Standard profile for a seamless, zero-permission onboarding, and Expert profile to unlock advanced technical options.
- ▸ Clio Live Cloud & Local (Dual-Engine): Real-time, two-way voice conversation. Cloud (Gemini Live API via high-fidelity WebSocket with barge-in support) and Local (Gemma 4 + Piper offline) both active and exposed in the Expert interface.
- ▸ Gemma 4 Download Panel: On-device management panel for downloading and choosing Gemma 4 variants (E2B / E4B) directly from Settings.
- ▸ ProsodyProcessor: Intelligent speech prosody engine converting markdown, emojis, and styling into neural audio instructions.
- ▸ Live Transcription: Real-time captioning of user speech and Clio responses.
🐛 Hotfixes & Optimizations
- ▸ Model Naming Cleanup: Legacy Gemini labels were aligned with the current Gemini 2.5 family naming used by the AI stack.
- ▸ Local BLE Tracking: Native Bluetooth Low Energy (Clio TourSync) RSSI tracking initialized only for Expert profiles.
- ▸ Permission Loop Bypass: Standard users are not interrupted by intrusive permission popups at startup, improving onboarding.
v0.5.5
May 28, 2026
🚀 STABILITY & CLOUD
🚀 Thread-Safety, Direct Cloud Connections & Stage Journal
✨ New Features
- ▸ Direct Cloud REST Client: Bypasses Firebase proxies, connecting directly to Google AI Studio (Gemini 3.6 Flash) via secure API keys.
- ▸ Stage Journal & Auto-Matching: Camera navigation and fuzzy-matching of scanned artworks on the wishlist.
- ▸ Itinerary Sharing: Exposes wishlist export options via system sharing intents.
- ▸ Auto-Narration TTS: Automatically triggers voice narration after successful artwork detection.
- ▸ Standard AI Fallback: Toggle to disable RAG databases in settings and use standard AI explanations.
🐛 Hotfixes & Optimizations
- ▸ Chatbot Deadlock Fix: Concurrent clicks blocked via Compose state and Kotlin
nativeMutex thread serialization.
- ▸ TTS Shutdown Fix: Proper singleton release on background to avoid initialization errors.
- ▸ Small Screens UI: Added horizontal scrolling to suggestion chips in settings to prevent broken text.
v0.5.4
May 28, 2026
🛡️ RESPONSIBLE AI
🛡️ Responsible AI, EU AI Act & Watch Fixes
✨ New Features
- ▸ Detailed Responsible AI section with 4 pillars (Fairness, Transparency, Privacy, Reliability).
- ▸ EU AI Act Compliance (Reg. 2024/1689), Article 50 Transparency.
- ▸ Law 132/2025 Compliance, the first national AI framework in the EU.
- ▸ Visual safety pipeline: Capture -> Face Blur -> Consent -> AI -> RAG -> Disclosure.
- ▸ User profile safety guardrails (Kids prompt anti-injection, Senior anti-ageism).
⌚ Watch App Fixes
- ▸ Added Stop button ⏹ to the Wear OS controller.
- ▸ Progress indicator fixes: properly reflects real playback progress.
- ▸ isActive fix: Wear OS app returns to idle state on playback stop.
- ▸ Website showcase: new Responsible AI Hall added to the digital museum.
v0.5.3
May 27, 2026
🏛️ Clio AI, 10-Field OCR Taxonomy & Expert RAG Datasets
✨ New Features
- ▸ Clio: Labeled the AI assistant: Clio, the Greek Muse of History.
- ▸ 10-Field OCR Taxonomy: museum, inventory, artist, origin, title, year, technique, dimensions, provenance, and note.
- ▸ Expert RAG Datasets (Lexroom-style): curated critical archives for historical grounding.
- ▸ Art Curator AR Hotspots: points of interest verified by expert art historians.
- ▸ Progressive Answers (typewriter effect) for improved UX during processing.
- ▸ Foldable historical taxonomy sheet in ArtworkExplorerScreen.
🐛 Fixes & Enhancements
- ▸ Post-shot crash fix: deferred resetConversation() with 500ms delay.
- ▸ artworkBitmap converted to StateFlow for Compose thread-safety.
- ▸ Room Database migration v8 for Expert Dataset entities.
- ▸ AIRouter: Levenshtein fuzzy matching for critical archive lookup.
- ▸ Visitor profiles: dynamic system prompts for Kids/Senior/Adult.
v0.5.0
May 26, 2026
📔 Ship Log & Exhibition Logbook
- ▸ ExhibitionEntity + WishlistItem: new data model for museum visits.
- ▸ Pre-visit wishlist with auto-matching when artwork is scanned.
- ▸ Visit statistics: viewed artworks, duration, completion rate.
- ▸ Personal exhibition notes with auto-suggest for visited museums.
- ▸ Tab system in HistoryScreen: History + Journals.
- ▸ Active banner in CameraScreen during active visits.
v0.4.13
May 24, 2026
🎙️ Voice Chat Hands-Free & Real Feedback
✨ New Features
- ▸ Hands-Free Voice Chat: continuous ASR→AI→TTS loop without screen touches.
- ▸ Animated pulsing orb visualizer with states: listening, processing, speaking.
- ▸ Real 👍/👎 Feedback: data pushed to Firebase Firestore (no longer mockup).
- ▸ FeedbackRepository.kt with GDPR support for user data deletion.
- ▸ lang/voiceKey parameters passed to speak() directly.
🐛 Critical Bug Fixes
- ▸ TTS Root Cause: setLanguage() no longer destroys isInitialized.
- ▸ Race condition: voice/language configuration after initialization wait loop.
- ▸ Safety-reset: TTS state no longer stuck on Speaking if all tiers fail.
- ▸ speakWithSystemTTS() correctly resets state on failure.
- ▸ shutdown() properly releases native resources of the TTS engine.
v0.4.16
May 24, 2026
🔊 Chatbot Auto-Narration & Audio Sync
- ▸ Gemma Optimization: prompt constraint of 2 short sentences (≤30 words) for under 5s latency.
- ▸ Golden speaker icon visible to replay the last chatbot response.
- ▸ Chat text formatting with markdown support and properly spaced paragraphs.
- ▸ Narrator profiles tailored for tone (Academic, Popular, Adventurous, Kids).
- ▸ Dedicated system prompts with grammar and syntax calibrated by language.
- ▸ Differentiated audio voices based on visitor profile (e.g., warm voice for kids).
v0.4.13
May 23, 2026
🧠 Dual AI Architecture & Onboarding
- ▸ Gemma 4 E2B LiteRT: offline-capable local inference for privacy-first visits after model download.
- ▸ Gemini 3.6 Flash Cloud: scholar-tier deep analysis via Firebase AI Logic.
- ▸ Smart AI Router: automatic local/cloud switching with billing manager.
- ▸ Guided Onboarding: local default configuration, package downloads, hardware tests.
- ▸ Local LLM Chatbot for general app questions inside Settings.
- ▸ Save photo set as default with disclosure during onboarding.
v0.4.10
May 20, 2026
📍 GPS, Maps & OCR Scanning
- ▸ GPS Proximity: real-time museum distance calculation with geolocated welcome.
- ▸ Label OCR scanning with corrective editing and confidence scoring.
- ▸ Piper TTS speech synthesis: 10 offline neural voices (5 languages x 2 genders).
- ▸ Cloud TTS Gemini: prompt-steerable voice persona with 6 personalities.
- ▸ Freeze UX: instant capture frame with GDPR face redaction.
- ▸ AR Points of Interest: 8-12 art details with AI zoom overlays.
- ▸ Whisper Mode: reduced volume at 20% with automatic headphone detection.
- ▸ Proximity sensor: auto-pauses playback when the phone is moved away.
- ▸ AI Report system: pushes feedback directly to Firestore.
- ▸ Scan History: local storage with JSON export and GDPR deletion.
