Evidence · First principles

How to prove an AI decision happened

The question sounds simple until someone disputes your answer. Then every artifact you reach for turns out to have been produced by you.

A customer disputes an automated denial. A regulator asks which model version handled a transaction last March. Opposing counsel argues the record was created after the fact. In each case the question is the same: how do you prove what happened, to someone with no reason to take your word for it?

Why the usual artifacts fail

  • Screenshots prove what a screen displayed to whoever pressed the button.
  • Application logs prove that a process wrote a line, at a time that process asserted.
  • Database exports prove the current state of a table you control.
  • Signed logs with your own key prove you signed something, at a time you chose, with a key you still hold.

Each is useful operationally. None survives an adversary who simply says: you could have made that.

The three properties evidence needs

To settle the question, a record has to carry three things:

  • Integrity. Any change to the record is detectable, by anyone, not just by you.
  • Time. The record demonstrably existed before some moment, witnessed by a party that has no stake in the dispute.
  • Independent verification. The check can be run without the record's owner participating.

The third property is the one people skip, and it is the one that decides disputes. Evidence that requires the accused party's cooperation to verify is not independent evidence.

How anchoring supplies all three

A decision record is hashed and signed at the moment it occurs. The hash is written to a public ledger, where it receives a consensus timestamp from a network the record's owner does not operate. Afterwards:

  • Changing the record changes its hash, which no longer matches the ledger entry.
  • Backdating is impossible, because the ledger's ordering is external.
  • Verification runs against public infrastructure, using the record and the ledger alone.

Check it yourself

Every Rubric attestation resolves publicly, with no account and no API key, and every anchor resolves to a public ledger message you can read without our cooperation.

HCS topic 0.0.10416909 · ML-DSA-65 signatures

Verify an attestation · Read the ledger ↗

What about the inputs?

Sensitive inputs do not have to be published to be provable. Committing to a hash of the input at decision time lets you demonstrate later that a specific input produced a specific output, without ever disclosing the input itself except to those entitled to see it.

Evidentiary framing

In United States federal practice, Rules 902(13) and 902(14) provide for self-authentication of records generated by an electronic process and of data identified by a hash value, subject to certification requirements. Independently anchored records are built to line up with that framing rather than to argue against it.

Different jurisdictions treat electronic evidence differently, and admissibility is a question for counsel. The point of anchoring is narrower and more durable: it makes the factual question answerable.

Related: Tamper-evident AI logs · Attestation vs SIEM logging · Why not build it yourself