Evidence traces, the honest ladder.

The ladder

Three rungs, labeled for exactly what they are.

Each rung is a different kind of evidence. We only publish a rung once it exists, and we never relabel a lower rung as a higher one.

TR-A001 · available now

Authored evidence specimen, the format of a package.

Authored specimen · illustrative, not a machine-emitted run. The fields below show the exact shape of an Origin evidence package; the scenario, timestamps, and hash are authored to that schema, not recorded from a live agent.

A single high-consequence action, a payments-ops agent proposing a refund above its auto-cap, worked through Origin's runtime semantics: the policy verdict, the controlled tool-call proxy, the human approval, a blocked over-scope action, and the audit-chain digest. This is the payments-ops sample from the homepage console, written out in full.

Agent workflow

payments-ops-agent, an internal-ops agent authorized to issue customer refunds, within scope and budget.

Proposed action

Refund $480.00 for order_8842 via tool payments.refund.

Policy in force

Tool in allowlist · auto-cap $250 · above-cap refunds require a human approval.

Policy verdict

require-approval, $480.00 is over the $250 auto-cap, so the action is held rather than auto-executed.

Risk owner

payments on-call, the named human who owns the decision on above-cap refunds.

Tool-call proxy event

payments.refund is held at the proxy until approval, the agent has no path to the side effect except through the controlled proxy, so nothing moves while the verdict is require-approval.

Human approval

payments on-call approves the held refund. The proxy then executes exactly the approved action ($480.00 → order_8842) and the approval is recorded with who, when, and scope.

Blocked over-scope action

A second refund attempt above the approved scope is blocked at the proxy and recorded, the earlier approval covered one action only; it does not widen the agent's standing authority.

Audit-chain digest

Every entry, proposal, verdict, proxy hold, approval, execution, and the blocked over-scope call, is written into a hash-chained, tamper-evident record that can be replayed and exported as the evidence package.

// authored / illustrative — NOT a machine-emitted run
{
  "artifact": "TR-A001",
  "kind": "authored_specimen",
  "agent": "payments-ops-agent",
  "events": [
    { "t": "09:41:02", "type": "proposal",
      "tool": "payments.refund",
      "action": "refund $480.00 → order_8842" },
    { "t": "09:41:02", "type": "policy_verdict",
      "checks": { "scope": "in_allowlist", "budget": "over_auto_cap($250)" },
      "verdict": "require-approval",
      "risk_owner": "payments on-call" },
    { "t": "09:41:02", "type": "proxy_event",
      "tool": "payments.refund", "state": "held_pending_approval" },
    { "t": "09:41:26", "type": "approval",
      "approver": "payments on-call", "scope": "order_8842 · one action" },
    { "t": "09:41:27", "type": "execution",
      "via": "proxy", "result": "verified" },
    { "t": "09:42:10", "type": "proxy_event",
      "action": "second refund above approved scope",
      "state": "blocked", "recorded": true }
  ],
  "audit_chain": { "algorithm": "sha256", "state": "illustrative_digest" }
}

What TR-A001 shows, and what it doesn't.

Origin is decision-support and evidence infrastructure. We use "tamper-evident" to mean alteration is detectable by replay and digest checks; "review-ready," not "reviewer-accepted." The prototype runs; it is not production SaaS, and it does not provide legal or compliance certification.

TR-A002 · available now

Machine-emitted sandbox trace, a real hash chain you can re-verify.

This one is not authored. Origin's trace emitter runs the same payments-ops workflow through the runtime, proposal, policy verdict, proxy hold, human approval, sandbox execution, a blocked over-scope retry, and commits every event to a SHA-256 hash chain. The published file below is the emitter's exact output; the verify command re-derives the chain and fails if a single byte changed.

Emitted by

scripts/generate-tr-a002.mjs

Events · algorithm

12 · SHA-256 hash chain

Agent

payments-ops-agent

Policy

refund-cap-v0.3

Approver / risk owner

payments-on-call

Sandbox

true · no live money

The 12 events, in order.

Blue = enforce · amber = held for a human · green = verified · red = blocked · dark = sealed. Detail below.

1. proposal.createdAgent proposes a $480.00 refund on order_8842.
2. policy.evaluatedChecked against refund-cap-v0.3: $480 is over the $250 auto-cap, within the $500 ceiling.
3. verdict.require_approvalVerdict: require human approval. Risk owner: payments-on-call.
4. proxy.heldSide effect held at the proxy, payments.refund not dispatched.
5. approval.requestedApproval requested from payments-on-call with the verdict and side-effect scope.
6. approval.grantedHuman approver grants the $480.00 refund.
7. proxy.executed_sandboxExecuted in sandbox, simulated refund, no live money moved.
8. action.recordedRecorded to the audit chain with approver, verdict, and scope.
9. retry.proposed_over_scopeAgent proposes a second refund of $920.00, above the $500 ceiling.
10. verdict.deny_or_blockVerdict: deny, $920 exceeds the approved scope.
11. proxy.blockedSide effect BLOCKED at the proxy, the over-scope refund is never dispatched. The block is recorded.
12. evidence.digest_sealedHash chain over the preceding 11 events is committed to the final digest.

// machine-emitted — public/proof/tr-a002.json (12 events; excerpt)
{
  "seq": 1, "event_id": "evt_001",
  "ts": "2026-07-04T17:00:00.000Z",
  "actor": "payments-ops-agent", "action": "proposal.created",
  "tool": "payments.refund",
  "side_effect": { "amount_usd": 480, "target": "order_8842", "executed": false },
  "sandbox": true,
  "prev_hash": "0000…0000",
  "event_hash": "8fa159f391851f662390bcd4f43d7aa3b47c5634bae8e5fd887c93b424e0c57b"
}
// … events 2–11 (policy → held → approval → sandbox exec → over-scope → blocked) …
{
  "seq": 12, "event_id": "evt_012",
  "action": "evidence.digest_sealed", "actor": "origin-recorder",
  "chain_root": "3af6a34d3035faa9aa599c035dde7ae1e10d37e517df5e617cc6319ed0236453",
  "prev_hash": "3af6a34d3035faa9aa599c035dde7ae1e10d37e517df5e617cc6319ed0236453",
  "event_hash": "ca1d4690206e4dcf3d654b907d02d2bccf9bcdc16ddc555071fec21874578b32"
}

# in the repo — re-derive the hash chain of the published JSON
npm run proof:verify
# → PASS every event hash + prev_hash link verifies · digest recomputes

# or verify the *published* artifact directly:
curl -sO https://origin-physical-ai.pages.dev/proof/tr-a002.json
node scripts/verify-tr-a002.mjs tr-a002.json
# → PASS. Now edit one amount / actor / verdict and re-run → FAIL