PERCEPTIFX PROJECT STATUS Interactive prototype — scripted demonstrations using simulated/replay data. Release: 0.1.2 (Public review build of prototype v0.1.1) Deployed (build time): 2026-11-07T15:44:42.618Z Source commit: 091c2e186e816d3bc32b72565d0a1b96f61a96c4 Repository: https://github.com/edgarbirlew-source/perceptifx (pending: repository not created yet; the commit shown here is held locally and will be pushed unchanged) Hosting: Railway project perceptifx, service perceptifx-web, deployment 11e0b7b1-beb9-435f-ae4f-50777f8a980b (static files, no database). Domain and DNS: Vercel team ConstrucFi. Last status update: 2026-11-07T16:05:00Z by Claude (deployment thread) PROGRESS (acceptance items with evidence; groups counted separately) - [Research] Viability: 3/4. Summary, live-state record and competency skeleton saved; waiting for Edgar's read (review, not a gate). - [Research] Technical research: 0/6. Items T1–T6 defined, none accepted. - [Research] Market research: 0/3. Items M1–M3 defined, none accepted. - [Research] Legal / access research: 0/4. Items L1–L4 defined, none accepted. - [Research] Benchmarks: 0/4. Items B1–B4 defined, none accepted. - [Prototype] Interactive prototype: 7/8. 7 of 8 items have author-run test evidence (50/50 automated checks against the same build, served by the same Railway function code, on a local server). P8 is partly verified: outside browsers load the public site (Railway request log: HTTP 200 for /, 308 for www), but the full automated suite has not yet been run against https://perceptifx.com. Independent reproduction and Edgar's review are open. - [Production] Production build: no items defined. Not started — gated on research and owner approval per feature. WHAT WORKS TODAY - Workspace view: both demo jobs with what the agent sees, how it reads it, and its next action. - Production demo: replay of a document-acquisition run where the saved "PDF" is really a login page; the verifier catches it, hands off to a person, resumes, and checks type, source, title, edition and effective date before finishing. - QC demo: four simulated ModelForge-style car-body cases; the agent reports only what it inspected, asks for more views and a measurement, and separates "not visible" from "missing". - Learning demo: mark a finding wrong, record the correction, get two proposed lessons, run evaluator tests on held-out, counterexample and regression cases, activate the passing lesson, reject the failing one, roll back. - Pause, take over, replay from start, and correct the agent. - Demo state persists across reloads in the same browser; each browser has its own isolated copy; Reset demo (two taps) restores the seed. - Evidence view with the demo trail and the automated test results. - This status page, readable as HTML, JSON (/status.json) and plain text (/status.txt). SCRIPTED / SIMULATED / REPLAYED / LIVE - Production observations, interpretations and next actions: REPLAY. Predetermined fixture text. No website is contacted. - Login-page detection: SCRIPTED. Step 3 always fails by design; no real file parsing. - SHA-256 values: PROTOTYPE. Really computed in the browser, but only over fixture byte strings. - "Sample County" document details: SIMULATED. Fictional jurisdiction and file. - QC views and gap measurements: SIMULATED. Drawn from fixture parameters; not captured images, real ModelForge renders, or measured geometry. - QC judgment: SCRIPTED. A fixed rule function. Deterministic, not a model. - Lesson proposals L-001 and L-002: SCRIPTED. Predetermined text, triggered only by the MF-001 correction. - Lesson evaluation, activation gate, rollback, history: PROTOTYPE. Real logic running over the scripted rules. It shows the lifecycle, not that a model learns or generalizes. - Your corrections and demo actions: PROTOTYPE. Stored verbatim in your browser only. Never sent anywhere. - Anything live or autonomous: NONE. No live automation, no live model, no third-party site access in this release. PROTOTYPE ACCEPTANCE ITEMS - P1 MET: Operator console with Workspace, Production, QC, Learning, Evidence and plain-English status. Evidence: Automated tests W1–W3 (all views open from a fresh browser) - P2 MET: Agent shows observation, interpretation and next action; reviewer can pause, correct, take over. Evidence: Automated tests T10a–T10d, W4 - P3 MET: Replay of a document-acquisition run with a login page saved as "PDF" and the verifier catching it. Evidence: Automated tests T10a–T10d - P4 MET: QC case: first view "no defect", agent requests more views, then confirmed or insufficient-evidence states. Evidence: Automated tests T1–T1c, T6b–T6d, T7, T8 - P5 MET: Correction captured, scoped lesson proposed, tested on held-out and counterexample, activate or reject, rollback. Evidence: Automated tests T2a–T9e - P6 MET: Persistence across reloads and a reset control. Evidence: Automated tests T11a–T12b - P7 MET: Every element labeled; nothing presented as live. Evidence: Automated tests L1–L2 (release label visible in every view) - P8 PARTIAL: Verified access instructions: the link opens and can be operated. Evidence: Railway request log shows outside browsers receiving HTTP 200 from https://perceptifx.com and 308 from www; certificate valid per Railway. Full automated run against the public URL pending (the test environment's network policy blocks the host). KNOWN DEFECTS - None recorded. KNOWN LIMITATIONS - Everything in the demos is scripted, simulated or replayed. No live agent, model, website or file is involved. - Lesson proposals exist only for the MF-001 mirror correction; other corrections are recorded but produce no lesson. - Demo state lives in each visitor's browser only. Private windows and cleared site data lose it, and it does not follow you to another device. - The ModelForge cases are stand-ins drawn from parameters, not verified ModelForge renders or geometry. - Edgar's earlier corrections in the private Claude artifact are not copied into this public build. - The page loads fonts from Google Fonts; without them it falls back to system fonts. - Latest automated test results were produced against the identical build on a local server, not yet against the public URL. ACTIVE WORK - Source repository on GitHub | Responsible: Claude (deployment thread); Edgar creates the empty repository | Blocker: This agent session is not permitted to create GitHub repositories. The source and its commit history are ready to push. | Next: Edgar creates an empty private repository edgarbirlew-source/perceptifx; Claude pushes the existing commits and connects Vercel to it. - Run the automated test suite against https://perceptifx.com | Responsible: Claude (deployment thread) | Blocker: The test environment's network policy does not allow perceptifx.com. | Next: Allow perceptifx.com in the cloud environment's network settings, or have another agent run tests/public-site.test.js against the public URL. - Independent reproduction of the prototype tests | Responsible: Unassigned (ChatGPT/Kathy or Tom/Perplexity) | Blocker: None | Next: Open https://perceptifx.com, follow the steps under What works today, and compare with the test results. - Technical research T1–T6 | Responsible: Claude coordinating; Perplexity capacity to be confirmed | Blocker: None recorded | Next: Start T1 primary-source review of computer-use reliability and cost per task. OWNER GATES - G1 Repository and hosting for the prototype: Approved by Edgar on 2026-11-07 for this prototype release (hosting on the existing Vercel team, no new paid plan). - G2 Point perceptifx.com at a host: Approved by Edgar on 2026-11-07 for this prototype release. - G3 Any spend (hardware, hosted browsers, paid APIs): Not approved. - G4 Live automation against third-party websites: Not approved. - G5 Expanding autonomy beyond supervised: Not approved. - G6 Outbound messages or commitments: Always owner-sent. NEXT ACTION: Push the source to GitHub as soon as the repository exists, then begin technical research item T1.