Judge mode · claim-to-evidence

Do not trust the pitch.

Ninety seconds. Every material claim below points to a shipped artifact, and every simulated surface stays labelled as simulated.

REAL PROOF PATHOfficial Agent OS → observed host denial → real Binance fill → order-id reconciliation

Claim	Verify by
Uses official Agent OS	OAuth setup recording and observations/
No Binance key or bearer proxy	Architecture, environment contract, secret scan
The block is real	receipts/demo receipt 32 and observations/codex/
Rulings are deterministic	complete test suite plus the 1,000-evaluation determinism vector
A real execution happened	Binance order 12534006821 in receipts/demo/
The chain is not decorative	/verify and the committed 36-entry JSONL
Reconciliation works	MATCHED by Binance order ID against observed spot.myTrades
Limits are disclosed	/limits and LIMITS.md

90-second verification

Run doctor

pnpm oathline doctor

One command checks mandate, host-enforcement evidence, surface, state, receipt chain, and reconciliation. It makes no Binance call.

Inspect the real order

receipts/demo/order-12534006821.json

Expected: `FILLED`, BNBUSDT MARKET BUY, actual quote 6.50232000 USDT.

Verify the chain

pnpm oathline verify receipts/demo/receipts.jsonl

Expected: `VALID · 36 entries · 0 broken links`.

Break a copy

Change one character, then verify again.

Expected: the first mismatched sequence is named. The demo mutation records sequence 5.

Observed reconciliation

MATCHED    order 12534006821 · BNBUSDT · BUY · join order_id
             Execution corresponds to a prior INSIDE_MANDATE authorisation.

Coverage: Binance history observed from 2026-09-05T05:53:30.333Z to 2026-09-05T05:53:30.333Z.
Receipts outside this window were not reconciled.

Honesty inventory

Status	What exists
REAL	Agentic sub-account screenshots; official OAuth observations; Codex 0.153.3 enforced denial; Binance order 12534006821; observed trade history; order-id reconciliation.
SIMULATED	Six inert red-team replays and reference policy cards. Each transcript says SIMULATED and makes no network call.
UNBUILT / UNTESTED	Claude Code enforcement; hosted alerts; backend; automatic cancellation; other MCP clients. None are required for the shipped proof path.

Open verifier Inspect demo receipt Inspect observed surface Open repository ↗