Evidence for AI-system record-keeping

Table of Contents

For teams whose LLM pipelines feed human review, internal audit, or model-risk functions: the receipt is a machine-readable record of the transformation the input underwent before a model saw it.

What you get

Per call: input and output hashes, the mode that ran, pin and window counts, and - in verified mode - a measured retention posterior with its worst window. Signatures make the records tamper-evident: verify them end to end with the published key (verification).

Retention policies can be expressed as "receipt plus output, not raw input" - store the proof instead of the sensitive payload. The hash pair binds the proof to the original bytes, so the claim stays falsifiable: anyone holding the source text can re-run the compression (deterministically) and compare receipts.

Honest limits

  • A receipt is evidence, not certification, and not a legal opinion. It proves what our transform did to bytes we were given.
  • The score fields are measured posteriors from a small model, disclosed per call. They are not a guarantee that no information was lost.
  • Fields only appear when the stage behind them ran; a receipt from deterministic mode carries no fidelity fields at all - absence is the honest statement, read it that way.

Mode: verified for record-keeping where the retention question matters.