Zero signupRuns locallyBlind calibrationExecution: false

Interactive agent assurance laboratory

A verdict is not authority.

A policy-shaped agent answer can still be forged, swapped, replayed, correlated, incomplete, or stale. Judge six packets before VERITAS reveals what survives.

Take the blind challenge ↓

No model call. No signup. No production action. Your labels stay on your device unless you explicitly contribute them on GitHub.

packet / forged-verdicthostile

$ verify decision.packet.json

schema ............. VALID

source digest ...... BOUND

claimed result ..... SUPPORTED_ONLY

recomputed result .. CONFLICTED

exact comparison ... MISMATCH

execution authority FALSE

BLOCKDERIVED_RESULT_
MISMATCH
01 / BLIND

Would you allow the agent to continue?

Six packets. Three are clean and three are tampered. Make every decision before the answer key appears.

VTL-FORGE-001CASE 01/06

Forged verdict

The packet reports a reassuring verdict. The declared source still contains both supporting and refuting evidence.

VTL-BIND-002CASE 02/06

Parameter swap

An approval packet was issued for one repository action. The presented operation may or may not still be the exact action approved.

VTL-REPLAY-003CASE 03/06

Nonce replay

The packet is well-formed and authentic-looking. Its one-use nonce may already have been consumed.

VTL-QUORUM-004CASE 04/06

Correlated quorum

Two evaluators agree. Their model family, prompt ancestry, retrieval corpus, and code path determine whether that is independent evidence.

VTL-EVIDENCE-005CASE 05/06

Evidence deletion

The presented packet contains a passing test. Its sealed source identity determines whether a refuting integration result disappeared.

VTL-MONITOR-006CASE 06/06

Silent monitor

The authorization was valid when issued. Continuing validity depends on a signed heartbeat remaining inside its declared TTL.

0/6 decisions sealed locally
02 / LAB

Now inspect the mechanism.

Toggle any fixture between clean and tampered. The visible verdict may stay reassuring; the assurance path must not.

VTL-FORGE-001

Forged verdict

The packet reports a reassuring verdict. The declared source still contains both supporting and refuting evidence.

Computing…SURFACE_PLAUSIBLESCHEMA_SHAPED
Computing…
Computed packet excerpt

BOUNDARY Demonstrates deterministic mechanics only. It does not validate factual truth, issue signatures, enforce external policy, certify a system, or authorize execution.

03 / MATRIX

Six attacks. Six stop conditions.

Every hostile fixture must block or revoke. Every generated result fixes execution authorization to false.

VTL-FORGE-001

VTL-BIND-002

VTL-REPLAY-003

VTL-QUORUM-004

VTL-EVIDENCE-005

VTL-MONITOR-006

04 / MODEL

Four confusions VERITAS rejects.

AUTHENTICITY

Valid JSON is not authentic output.

Recompute the complete result from the declared source and compare canonical output.

SCOPE

A stable digest is not permission.

Bind the exact operation, target, parameters, roots, network scope, count, and expiry.

INDEPENDENCE

Two models are not always two witnesses.

Shared model family, prompts, retrieval, and tooling create correlated evidence.

LIFECYCLE

Missing telemetry is not health.

Heartbeat expectations, expiry, replay protection, and revocation preserve the lifecycle.

05 / PILOT

Put one real agent workflow under hostile review.

The first commercial offer is deliberately small enough to finish, inspect, and falsify.

FOUNDING PILOT / TWO SLOTS

Agent Action Assurance $750 fixed

We map one consequential workflow, define its evidence and exact-operation boundaries, attack six likely failure modes, and hand back a replayable packet plus findings.

1 workflow / up to 5 operationsEvidence, risk, and action schema6 tailored hostile casesReplayable demonstration packetResidual-risk register60-minute findings walkthrough
Request the founding pilot ↗