Origin — Move AI forward. With evidence.

A public trust layer, evidence format, and working demos for AI evidence review.

A capable agent and an authorized action are different things. Origin makes that boundary inspectable through deterministic policy evaluations, configuration-bound artifacts, and independent verification. Its Physical AI demos explore the same evidence questions in bounded simulations.

Model proposes. Environment verifies. Gate decides. Trace proves. Capability is not permission.

The proposed controlled execution path (proposed architecture, not deployed) splits execution from evaluation: Model proposes. Environment executes. Verifier evaluates. Gate decides. Trace proves.

Explore the website · Run a reference check · Verify an artifact · Public source

Try the public workflow

Start here	What you can inspect
Agent Reference Check	A selected policy evaluated against deterministic synthetic support and IAM tasks, with a configuration-bound demo artifact.
Artifact verifier	Offline integrity, tamper, and configuration-drift checks for supported artifact formats.
Evidence traces	Clearly separated authored specimens, a downloadable machine-emitted sandbox trace, and the external evidence that has not yet been earned.
Verifier demos	Browser demonstrations of deterministic policy and permission checks.
Physical AI labs	Robot, warehouse, fleet, and spatial examples in bounded simulations.
Trust center	Claim boundaries, provenance, verification records, and data-handling limits.

The public reference check runs browser-local policy logic; it does not contact or execute the named agent. Its evidence is signed by an unpinned browser-session key, so it is untrusted by default. An intact signature demonstrates an integrity check, not trusted Origin issuance or a verified provider effect. The proposed runtime gate/proxy and hosted verification API are design artifacts, not generally deployed public capabilities. See the reference check versus runtime comparison.

Origin is a working prototype seeking design partners. No active customer pilot, production readiness, reviewer acceptance, or certification is claimed.

One evidence contract, explicit authority

The common architecture separates a proposal, permission, execution, verification, evidence, and release authority. Each domain needs a verifier suited to its question; sharing an evidence format does not make the domains equivalent.

• Deterministic authority: the oracle supplies labels, verdicts, and hard-gated zeros. An LLM never grades another LLM. Optional post-gate reward shaping cannot lift an oracle rejection and is off by default.
• Inspectable evidence: canonical JSON, hashes, trace chains, receipts, and signatures expose the bindings a reviewer can recheck. Integrity, signer trust, execution success, and permission to deploy remain separate checks.
• Verifier hardening: Cobra and Chronos explore reward-hack discovery, verifier patches, and held-out evaluation. A verifier result is bounded by the verifier and its inputs.
• Release authority: Origin issues evidence. The customer's gate and named risk owner retain the deployment decision.

What is in this repository

Component	Path	Public scope
Origin Web	apps/origin-web	Product website, reference check, artifact verifier, evidence console, and labeled demos.
Janus	apps/janus	Credential-broker and autonomy-trace prototypes for scoped authority and fail-closed authorization.
Chronos UI	apps/chronos-ui	Interface for verifier-hardening workflows.
Cobra / Chronos	services	Python research and test tooling for verifier hardening.
Verifier core	packages/verifier-core	Deterministic gyms, configuration-bound credentials, signatures, and verification primitives.
Evidence	packages/evidence	Evidence contracts, canonical serialization, hashes, and receipt machinery.

Proprietary algorithm implementation and internal strategy are kept outside this public repository. The public site includes a research snapshot whose private-pipeline metrics cannot be re-derived from this repository alone.

Evidence boundaries

• Synthetic demos, generated counterfactuals, authorized fixtures, and customer-owned proof are different evidence lanes. None substitutes for another.
• Simulation results mean reproducible under the named verifier, inputs, and environment. They do not establish real-world safety, robot certification, or production-autonomy validation.
• Learned-policy examples use route-summary and map-derived features, not raw end-to-end perception.
• Public Foundry uses a deterministic browser-local sample. Provider-backed processing requires a separately enabled local/backend workflow and affirmative consent.
• Generated brand imagery and footage are illustrations, not product recordings, company facilities, or customer evidence.

Run locally

For the public web app, use Node 20+ and the root npm workspace install:

npm ci
npm run dev -w @origin/origin-web

Validate the app and published proof:

npm run test -w @origin/origin-web
npm run build -w @origin/origin-web
npm run proof:verify -w @origin/origin-web
node scripts/honesty-lint.mjs

Makefile is the monorepo front door. make help lists targets; make install includes the isolated Python service environments, and make gates-all runs the wider build, TypeScript/Python, evidence, and honesty gates. Gate results are scoped to the checked revision; a green test run is not deployment authority.

Navigate and contribute

Read PROJECT_OVERVIEW.md, REPO_STRUCTURE.md, and AGENTS.md for the component map and operating rules. Start with apps/origin-web/README.md for the web app's implemented behavior.

The canonical website is originphysicalai.com; its public source is bohueilin/Origin. Pushing or opening a pull request does not deploy the website. Production release is separately human-dispatched and gated; see docs/DEPLOY.md. Secrets stay in ignored per-app .env.local files and must never be committed.