For litigators, e-discovery firms, and expert witnesses
When an AI agent causes harm, who is responsible? When memory was poisoned, when can that be established in court? Litigation needs forensic evidence — and AI agent memory currently leaves no fingerprints. We are exploring what a forensic record for AI decisions could look like.
Honest disclosure: this is on our long-term roadmap, not yet built. No demo calls, no follow-ups until there's something tangible to show. Email above just adds you to a low-volume list (~2 emails/year max).
Authorship attribution from the topological signature of memory state. Could the same agent have produced this output? Could the same model family? A fingerprint-class invariant for AI memory.
Reconstruct what an agent knew and when. Deterministic replay of decision sequences, with counterfactual branching to test alternative timelines ("if memory X had been refreshed at T-3, would the harm at T have occurred?").
Formal lifecycle artifacts for the end of an agent's operational life — signed termination certificates, succession records, evidence-grade audit trail of why the agent could no longer be trusted.
Signed W3C Verifiable Credentials, HMAC safety attestations, and Lyapunov convergence-proof PDFs, packaged for evidentiary use. The standards exist; the productisation for legal workflows does not yet.
The concept items above build on top of these production features.
Automated logical consistency checks confirm our healing and compliance policy rules can't contradict themselves (e.g. no rule both allows and blocks the same case; the healing counter is monotonic). These verify the policy rules — not each individual live decision.
Read the methodology →Bit-perfect replay of any decision using pg_override_disabled + config checksum. The forensic substrate is already production-grade.
Cryptographic attestation of memory provenance per tenant. The signing key infrastructure is already in production.
See passport docs →If AI litigation or forensic discovery is your domain, leave a note. We email back when there's something to show, not before.