Status: Active researchEvidence, Audit and ReplayResponsible AI
Evidence and Workstate
Store-untrusting, fail-closed design where verdicts are captured as evidence, contracts are pinned, and work state is replayable and tamper-evident.
Abstract
This programme studies how the state of ongoing AI work can be represented as replayable, tamper-evident evidence. We treat the store as untrusted, fail closed when contracts are violated, and capture verdicts as first-class evidence rather than transient logs.
Problem & motivation
When AI systems act over time, their state is often stored in ways that cannot be verified or replayed, making after-the-fact audit unreliable.
Research questions
- What is the minimum evidence needed to replay a unit of AI work faithfully?
- How should a system behave when stored state cannot be trusted?
Methods
- Design a pinned-contract evidence format.
- Implement fail-closed verification on load.
- Test replay determinism against captured evidence.
Limitations
- Findings are internal and not peer reviewed.
- Tamper-evidence is not the same as tamper-proofing; threat model is bounded and stated per experiment.
Disclosures
- Funding
- Infrastructure support provided by Octopus Core Pty Ltd.
- Conflicts of interest
- Octopus Core offers commercial audit/evidence infrastructure; findings are reported independently.