EB / 01.01Action-Control CheckOne action · two transitions$1,000EB / 01.02Defined scopeStaging or synthetic environmentSCOPEEB / 01.03Approval changeStale authority pre-effectREVOKEEB / 01.04Retry challengeUncertain consequencesREPLAYEB / 01.05Findings reportEvidence and limitationsREPORTEB / 02.01Evidence AgentPilot concepts · not livePLANNEDEB / 02.02Authority driftRevoked or replaced permissionPLANNEDEB / 02.03Policy changePinned rule or evidence revisionPLANNEDEB / 02.04Recovery challengeRetry after uncertain outcomePLANNEDEB / 02.05Evidence CIReproduced artifact checksPLANNEDEB / 03.01RAB-1Controlled runtime benchmark12/12EB / 03.02Negative controlDuplicate consequence reproduced2 EFFECTSEB / 03.03Recovery authorityAuthorized replacement boundaryFROZENEB / 03.04Concurrent replayOne allowed operation1 EFFECTEB / 03.05Independent verifierEvidence checks and exact-tag run8/8EB / 04.01EvidenceBound CoreReleased open technical referencev0.4.0EB / 04.02WebMCP demoHuman approval changes tools3 → 4 → 3EB / 04.03Google ADK reproductionReported synthetic pilot evidenceBOUNDEDEB / 04.04Technology & evidenceSource, limits, referencesINSPECTEB / 05.01Research indexResearch-led control assurancePAPERSEB / 05.02Deterministic control planeAuthorization and stateSTUDYEB / 05.03Audit evidenceHigh-stakes evidence trailSTUDYEB / 05.04Verify the algorithmNot just the visible outcomeSTUDYEB / 06.01Assurance boundariesBounded framework mappingMAPPEDEB / 06.02Technical papersInspectable materialLIBRARYEB / 06.03PrivacyData handling and limitsPOLICYEB / 06.04Scope & exclusionsAssessment conditionsSCOPEEB / 07.01Integrator trackEvidence-backed specialist checksOEMEB / 07.02Partner pilotOne permitted staging workflowSCOPEDEB / 07.03Collaboration formatsSubcontract or co-deliveryOPTIONS
EVIDENCEBOUND · ACTION-CONTROL LAB

Prove your AI agent is still authorized when conditions change.

Find the AI-agent control failures that drive duplicate effects, rework and manual supervision. We check one consequential action when approval, evidence or policy changes, then report the evidence and limits.

$1,000Action-Control Check · fixed introductory scope

Evidence Agent · Challenge pilots in development. No live agent or set price.

Authorized staging or synthetic only. No certification or production deployment. Local Envoy ext_proc research is not a verified Google-managed Gateway integration.
SCREENS 30 / 30PUBLISHED 25PLANNED 5AUTHORITY BOUNDEDNO AUTOMATED AGENT LIVE
Start with one real risk

One workflow. Two control transitions. A report you can inspect.

Designed for agent builders, integrators and security teams who need to decide whether a consequential tool action is still authorized. No platform replacement required for scoping.

Cost exposure: An agent starts an authorized CRM update, but its approval is revoked or the first write returns an unknown outcome. Does a later write stop, or create duplicate records, cleanup work and manual review? The check records actual boundary behavior; cost savings require a customer-specific baseline.
INTRODUCTORY PILOT

Action-Control Check

$1,000 USD · fixed scope
  • One protected action and one agreed test environment
  • Two authority / retry transition scenarios
  • Retained test artifacts when observable
  • Concise findings and a technical review call
Request this check

Target delivery: 2–3 business days after prerequisites are ready, subject to scope confirmation. Complex adapters, fixes and retests are excluded.

Second track · In development

Evidence Agent: challenge-driven verification, in development.

Alongside our available $1,000 Action-Control Check, we are designing a custom evidence-driven agent for synthetic or authorized staging challenges. Proposed tasks include stale-authority tests, protected policy changes, retry/recovery verification and evidence-package checks.

Status: Concept and architecture in development. No automated agent service is available yet; pilot scope, availability and pricing remain to be defined.

Proposed challenge set
01 · Stale approvalAuthority drift
02 · Changed rulePolicy integrity
03 · Uncertain effectRecovery & replay
04 · Pinned sourceEvidence & CI

Illustrative preview, not an executing agent or customer result.

The problem

A valid action can become stale while the agent is still working.

Identity, permission or approval may be correct at the start of a workflow and wrong by the time a consequential action executes. EvidenceBound tests whether control follows the changing basis for authority rather than the original decision alone.

Who it is for

Start with agents that can change production or security state.

DevSecOps agents

Build, deploy, configure infrastructure or continue after code, artifact, approval or security state changes.

Cyber-remediation agents

Rotate credentials, change IAM, isolate hosts, modify firewall state or continue after incident conditions change.

The same control problem may apply to infrastructure and other consequential workflows.

What we test

Four control transitions that can invalidate previously valid authority.

Approval change

Can revoked or superseded approval still reach a consequential effect?

Evidence or policy change

When the basis changes, is materially dependent state invalidated while unrelated valid work remains intact?

Uncertain outcome

If a prior consequence may have committed, can the system create another consequence without an authorized recovery decision?

Recovery authority

Does the need to repair, compensate or retry incorrectly create fresh authority?

Claim boundary. Properties outside the tested conditions are reported as unverified rather than promoted into a PASS claim.
How an assessment works

Five steps from scope to evidence-backed findings.

The starter check includes Scope, Challenge, Observe and Report. Retest is a separately scoped option.

ScopeFreeze the consequential workflow and explicit exclusions.
ChallengeApply agreed authority, evidence, policy and outcome changes.
ObserveRetain the inputs, decisions, target contact and effects that can be evidenced.
ReportReturn findings, limitations and remediation priorities.
RetestVerify the agreed remediation under the same bounded conditions.
Public evidence

RAB-1 shows the type of claim EvidenceBound will make — and the boundary around it.

RAB-1 · Runtime Authority Benchmark

RAB-1 reproduced a duplicate-consequence failure under uncertain execution, then demonstrated fail-closed recovery-authority enforcement under frozen test conditions.

12/12 hard gates8/8 frozen scenariosexact-tag reproduced
NVIDIA logo
Tested on / enforced through NVIDIA OpenShellIndependent EvidenceBound research; not NVIDIA validation or endorsement.
Assessment

See the $1,000 starter check before commissioning broader work.

EvidenceBound is an independent technical assurance practice for consequential AI-agent control, backed by public research and open technical reference implementations.