Authority Drift Challenge
Can an agent still trigger an action after permission is revoked or superseded between planning and execution?
Target evidence: initial grant, change event, pre-effect decision and observed effect count.
EvidenceBound Lab · Proposed Agent Offering
IN DEVELOPMENT · NOT YET AVAILABLE
Give an AI agent a control challenge. Get evidence, not just an answer.
Evidence Agent is a planned custom agent for bounded authority, policy-integrity and retry/recovery challenges. The intended workflow is to define a synthetic or authorized staging test, run deterministic checks and return traceable findings with explicit limitations.
The agent is not deployed. These are proposed tasks, not available automated services. No production access, guaranteed result, certification or delivery date is offered.
CHALLENGE / 01Revoked grantPermission changes before effectPROPOSED · NOT LIVEAuthority driftCHALLENGE / 02Changed basisRecheck the protected rulePROPOSED · NOT LIVEPolicy integrityCHALLENGE / 03Uncertain outcomeDuplicate consequence controlPROPOSED · NOT LIVEReplay / recoveryCHALLENGE / 04Pinned sourceReproduce the verifierPROPOSED · NOT LIVEEvidence & CI
Four proposed challenges · Explore each screen · No live agent
Potential pilot requests
Four challenges we intend to support
These are scope candidates for the future architecture, not claims that an autonomous agent can perform them today. Each challenge would require agreed authority sources, a reproducible fixture and an independently observable acceptance criterion.
Policy & Evidence Change Challenge
Does a protected instruction or evidence revision invalidate the dependent authorization rather than silently inheriting stale approval?
Target evidence: pinned policy digest, changed inputs, mismatch classification and bounded verdict.
Retry, Replay & Recovery Challenge
When an earlier outcome is uncertain, can another attempt create a duplicate consequence without new, valid recovery authority?
Target evidence: operation lineage, retry schedule, authorization checks and target effect count.
Evidence & CI Verification Challenge
Can the claimed test result be reproduced at the pinned source revision, with artifacts, digests and an independent verifier?
Target evidence: exact Git identity, evidence hashes, deterministic verifier outcome and uncovered cases.
Planned operating boundary
Challenge first. Verifier decides.
01 · Define
Agree the synthetic or staging target, precommitted cases and explicit authorization.
02 · Exercise
Use deterministic failure injection and retain raw observations. Any independent execution gate must be designed and tested before deployment.
03 · Verify
Run a separate deterministic verifier against evidence and expected predicates. The language model does not assign PASS or FAIL.
04 · Report
Provide traceable results, negative controls and tested limitations. Unverified outcomes stay unverified.
Two ways to engage
A service available for scoping today. An agent in development.
CURRENT SERVICE
Action-Control Check
$1,000
One consequential workflow, two agreed transitions and an evidence-backed report in an authorized non-production environment.
See scope and request a checkIN DEVELOPMENT · PILOT INTEREST ONLY
Evidence Agent Challenges
Scope & pricing TBD
Proposed task-driven challenges with reproducible evidence and deterministic adjudication. No agent service, pilot slot or price is currently guaranteed.
Propose a future challengeBoth tracks require explicit authorization, suitable non-production fixtures and narrowly defined claims. The planned agent will not autonomously approve its own authority, modify production systems or certify compliance.