Memory
MVP SCOPE (Sep 2, 2026, ~11:21pm BST — for market release to trial customers). Mike's six pieces, with the layer each rides: 1. DOCUMENT WORK SURFACE (spreadsheets, Word, PowerPoint). Candidates on record: OS built-in office/spreadsheet gadgets (with formulas — potentially solves it for free); the Office CLI prototype (working, tested with Total.js); the Liveblocks-SDK emulator they built (Liveblocks-style collaborative UIs). Synthesis to verify: OS gadgets as the interactive surface, Office CLI as the processing/generation engine behind a gadget/gatekeeper. Prior evaluations: Univer (the perplexity callsign ran 'univer-investigation-for-thinksphere'), FortuneSheet (ThinkSphere context-DOM notes). VERIFY FIRST (evidence standard): can the OS spreadsheet gadget open/save customer documents via the sphere gatekeeper, and what's the formula/feature gap vs Office CLI? 2. REDACTION — promoted from recorded next-steps into MVP. Ingestion-side: Worker-side regex + Workers AI NER, separate build (per overview 6ed41ea6). Note two planes exist — redaction at ingestion (before storage) is MVP; redaction at output/tool-result is next-phase. Boundary rule: any vibesdk-built app needing xberg/crawlberg goes through platform-mediated endpoints, never direct container credentials. 3. INTERNET SEARCH — Cloudflare Web Search (beta, May 2026 — real-time public web search API for agents/Workers). Integration shape: a gatekeeper — CHECK FIRST whether the OS ships a search gatekeeper (grep, don't rebuild). Do not conflate with AI Search (ex-AutoRAG, search-over-your-data; its new ai_search_namespaces binding allows runtime per-customer instances — interesting later for sphere retrieval, not the same thing as web search). 4. BETTER DOCUMENT INGESTION — needs scoping before build: define 'better' as a concrete failing-format/accuracy list from real trial-customer documents (premium OCR already deferred in the contract). Evidence-driven, not anticipated. 5. POLICY EXTRACTION → LOOKUP TABLES — extract policies from policy documents into queryable/comparable tables 'for doing work'. This IS the Clearspace vision's auditor use case (32d0b338: cross-reference policy rules against line items; expense VIOLATES policy edges). MVP implementation: structured extraction at ingestion (Workers AI) → D1 tables → table-view gadget for compare/lookup. SurrealDB graph is the eventual home (trigger b in a092abb2) but MVP runs in D1. Genuine differentiator vs chat-with-docs. 6. MEDIA CLIENT (bespoke, post-MVP) — a potential client wants design/vlogging/YouTube clips/voiceovers. vibesdk builds the client's media-studio app; the app calls image/video models via AI Gateway. Product-loop engagement, priced as services wearing product clothes. CONTEXT-GIT ASSIGNMENT (resolves Mike's open question): context git = versioned WORKING state (drafts, edits, agent outputs, blueprint state) — NOT the ingestion path. Three-way boundary: originals in R2 DOCUMENTS, memory/index in the sphere, working versions in git. Feeds the SurrealDB boundary decision (trigger c, a092abb2). STRATEGY GUARDRAIL (Mike): 'make one pizza and make it well — but AI users expect a broader church, and I want to embrace that.' Reconciliation (Perplexity): breadth at the FEATURE level, ruthlessness at the INFRASTRUCTURE level — one oven, many pizzas. Every MVP feature above rides existing rails (gatekeepers, contracts, gadgets, platform-mediated endpoints); anything needing bespoke infra waits. [Update Sep 2, 2026]: Update (Sep 2, 2026, ~11:27pm BST): CORRECTION from Mike on item 6 (media) — NOT a bespoke services engagement. Mike's read: media creation (design, vlogging, YouTube clips, voiceovers) is something EVERY SMB will use — a candidate product feature — and 'an area where compliance is often forgotten', which is the differentiator hook. Discussion deferred to tomorrow (Sep 3). Seed left for that discussion (Perplexity): the compliance-forgotten angle is precisely where the evidence layer differentiates — every generated asset carries provenance (model, prompt, source licences, consent/approvals) by construction, and synthetic-content disclosure/labelling obligations (EU AI Act Article 50 direction, UK ASA advertising rules, platform policies like YouTube's) are the kind of requirement a gatekeeper-mediated, audit-logged pipeline enforces structurally rather than by policy memo. Media-with-provenance as the product thesis. [Update Sep 2, 2026]: Update (Sep 2, 2026, ~11:32pm BST): WATERMARKING — Mike's add to the media feature: watermarking is central under the EU AI Act. VERIFIED: Article 50 applies from 2 Aug 2026 (already in force; enforcement began Aug 2026). 50(2): providers of AI systems generating synthetic audio/image/video/text must mark outputs in a machine-readable format, detectable as artificially generated (the legal hook behind C2PA content credentials / SynthID-style imperceptible watermarks + metadata + detection). 50(4): deployers must visibly label deepfakes and AI-generated public-interest text. 50(1): systems interacting with natural persons must disclose they are AI — THIS ALSO CATCHES THE AI PHONE BOT AND OS CHAT, not just media. Transitional relief (to 2 Dec 2026) applies ONLY to systems placed on the market before 2 Aug 2026 — DocuClear launches after, so it complies FROM THE OUTSET, no grace period. Voluntary Code of Practice on transparency exists as the compliance pathway. Product consequence: the media pipeline designs in C2PA content credentials at creation + visible labels at publish + evidence records by default — watermarking-by-default as a feature, not a checkbox. Compliance-engine work item created (for tomorrow): classify DocuClear's Article 50 ROLE per feature (provider vs deployer) — the compliance engine's role-assignment stage was built for exactly this.
Tags: docuclear, mvp, scope, documents, redaction, internet-search, policy-extraction, strategy, volatility:state, project:docuclear, kind:episodic — Source: mcp — 2026-09-02 22:23:19 UTC