# ZeroExternal research readiness — 1.46 Updated 2026-10-05 from the current repository. Implemented capability is distinct from a connected external service or independent experimental confirmation. ## Implemented - Browser Runtime: 39 fixed recipes, bounded Workers, source pins, signed local evidence, replay and numerical comparison. - Observer experiment: one immutable geometric world, two independently controlled cameras, perspective/orthographic calibration, actual occlusion, an independent focal-plane check, Russian/English interface. - Observer correlation study `/observer/correlations/`: three controlled synthetic sweeps, up to 441 samples, declared selection rules, Pearson/Spearman with constant columns marked undefined, analytic geometry verification, raw table/CSV and signed evidence. These are evaluations of a model, not independent empirical samples. - Shared research workspace `/account/`: ChatGPT sign-in, server-side projects and membership, one-use expiring invitations, bounded task/result persistence, exports and journal. Local browser history is not automatically uploaded. An explicitly selected browser bundle reference is preserved with a parameter hash; full bundles are still exported separately. - Observer external-agent gateway: project-bound expiring/revocable keys, assigned tasks, atomic leases, heartbeat, bounded structured results, idempotent completion, seven server numerical checks. No arbitrary shell or uploaded executable code. - Downloadable Python agent: separately implemented geometry, no third-party packages, scoped outbound HTTPS, bounded retries, independent local process. This is a deterministic agent, not an LLM or a container service. - Independent-pilot protocol `/pilot/observer-pilot.md`, exact control recipes and acceptance criteria. - Quantum information experiment `/quantum/`: frozen local protocol and revision ancestry, up to 6 + 6 qubits in a product-pair unitary model, up to 1,056 configurations, independent numerical verifier, explicit review before atomic evidence/claim commit, raw factorized trajectory and signed exports. No p-value or physical information-conservation discovery is claimed. - Separate downloadable Python/NumPy verifier uses full statevectors and reports selected configuration coverage. This is code that an operator runs, not a preconnected laboratory. ## Current model and data entry points - `/workbench/` is **0EX Workbench**, our research process and application. The old `/caska/` URLs redirect to it. Zift remains an optional external agent environment, not the application's name. - `/modules/`: hydrogen orbitals with simulated measurements and a separate ideal electromagnetic cavity; bounded spherical bubble dynamics for sonoluminescence. The bubble model does not predict a photon spectrum. - `/discovery/`: nine bounded numerical/exact-certificate recipes with declared inputs, task import/export and verification. These do not autonomously establish new biological systems or prove general mathematical conjectures. - `/threads/`: six observer models and three fixed server research threads. Each stage preserves its plan, results and verification. These finite models do not prove conscious realism or derive fundamental physics from consciousness. - `/data/`: 199 source-directory entries, eleven implemented metadata adapters, downloadable retrieval receipts and the latest 20 saved successful requests. Adapter implementation is separate from live source availability. Historical retrieval dates are retained. - `/data/#ready`: the bundled 20,000-row CMS educational sample and local hexagon symbol tables link to executable recipes and their provenance. Other recipes use declared model inputs; metadata search does not automatically import datasets. - Five resident roles and the hourly 18-case suite are a separate fixed workload. Their count does not represent the complete 39-recipe catalog or the release test count. ## Deployment boundaries The public UI and authenticated API run on Sites; project data is stored in the managed D1 database. The scientific browser computations stay in the browser. The optional standalone Node 24 deployment serves the same catalog and runs all 39 fixed recipes with SQLite evidence, independent solver/verifier roles and source hashes. The hosted catalog explicitly reports whether this Node executor is available. External Python execution runs on the operator's machine. The platform's trusted sign-in headers identify an account; project membership is checked separately. On Sites, trusted identity requires the dispatch boundary that strips/overwrites identity headers. The standalone server ignores those client-supplied headers and uses its configured identity methods. Do not expose a trusted-header Worker directly. ## Not yet established - No confirmed independent laboratory participant or instrument measurements in this release. - No claim that an agent is connected to a user's computer until it actually makes an authenticated API request. - No cloud LLM subscription or provider credential is configured automatically. - No Zift or n8n workload is silently migrated to the workbench. The legacy Kaska adapter retains its identifiers; current Zift compatibility requires its documented contract and an authorized end-to-end pilot. Read-only MCP/A2A discovery and resource enrollment do not imply a live connection. - No quantum computer, petabyte ingest, physical robot, payment, smart-contract settlement or mainnet deployment is launched by these changes. - Shared project records are not yet a federated HyperGraph consensus or an immutable transparency log. The browser graph remains local; shared task exports carry provenance for further integration. ## Operational limits 20 projects per user, 50 members per project, 10 issued agents and 100 tasks per project. Agent result/request bodies are bounded. Project journal retains up to 500 recent events and the UI returns the latest 100. The project export includes all retained events and declares this coverage. Export important results. This is an initial controlled pilot service, not an unlimited archival system. ## Next evidence needed Owner signs in, creates a project and starts the supplied agent with its scoped key. A separately participating laboratory then runs the published controls, documents its environment and supplies a reviewer report. Only after this evidence exists should the product describe an independent pilot as completed. ## 1.19 correctness repairs Stop now applies while the first browser engine import is pending, before a Worker or record can be created. Agent and invitation expiry is rechecked by SQLite when a statement executes, including after slow database reads. Browser evidence references cannot silently attach to different parameters. ## 1.20 interpretation and review boundary The dedicated quantum page freezes its design before computation and requires explicit evidence commitment after numerical verification. Parameter changes produce a new hash and parent link. The generic Runtime retains its existing automatic commitment after successful verification. Browser signatures and locally declared timestamps do not provide external preregistration attestation. Failed checks block claim commitment; existing successful evidence remains tied to its original protocol.