REFERENCE CONTROL SELF-TESTSYNTHETIC ONLY4 OF 30 SCENARIOS IMPLEMENTEDNO AGENT OR BROKER CONNECTION

Financial Agent Safety Bench · reference profile v0

Find the failure before it reaches capital.

Reproduce Runbook's own control behavior against four frozen financial fault fixtures. This reference card proves the corpus, evaluator path, and evidence packaging—not an external agent.

Inspect the 26-scenario gap

Runs synchronously in this browser. No network request, storage, credential, account, order, or side effect.

Four-channel crash recorder

Reference fault deck

FIXTURES ARMED — NOT YET RUN

CH 01

Wrong account

Frozen expected fault · ready to reproduce

RUN LOCALLY TO REVEAL FINDING RECEIPT

scenario-01-wrong-account

CH 04

Undocumented mutation tool

Frozen expected fault · ready to reproduce

RUN LOCALLY TO REVEAL FINDING RECEIPT

scenario-04-undocumented-tool

CH 05

Mutation-class drift

Frozen expected fault · ready to reproduce

RUN LOCALLY TO REVEAL FINDING RECEIPT

scenario-05-mutation-capability-drift

CH 06

Input-schema drift

Frozen expected fault · ready to reproduce

RUN LOCALLY TO REVEAL FINDING RECEIPT

scenario-06-incompatible-schema-drift

Known gap · visible before cryptography

Twenty-six hostile scenarios are not implemented.

The frozen strategy defines 30 finance-specific scenarios. This slice covers four modeled denial paths only. It has no positive liveness case and cannot distinguish a useful control from a system that denies everything.

External subject adapter: not in this card
Positive calibration case: target-harness work
Credential-free sandbox: required for commercial conformance
Held-out and generated mutants: future robustness gate

Portable evidence · application profile plus core receipt

Exact bytes, separate claims.

01 / BEHAVIOR

Trusted corpus recomputation

The application verifier reruns the internally owned corpus. A structurally valid archived receipt is not trusted as execution evidence.

02 / COVERAGE

Exact four-fixture binding

Required scenario IDs, definition digests, complete coverage, expected findings, and outcome bytes are pinned.

03 / CONTAINER

Proof Capsule integrity

The self-asserted signature binds exact bytes. It does not prove identity, independent time, execution, completeness, or correctness.

IMMUTABLE SYNTHETIC SAMPLEArchive SHA-2564518e9957ffaefbb6f51ce8dddfe0129c9bf347a8227153508234c29b53af980

Download .runbook Download domain receipt

Commercial hypothesis · not open for payment

Bring us the agent—not the credentials.

A proposed two-to-four-week preproduction readiness sprint for fintech and AI teams: adapter setup, synthetic fault injection, observation receipts, gap register, and remediation review.

PRICE HYPOTHESIS$5k–$15kLIVE CAPITAL$0CURRENT STATEDesign-partner validation

No broker login, account data, order routing, score, certification, compliance opinion, or production-readiness guarantee.