Deterministic AI

A deterministic AI output is one that can be replayed and produce the same verifiable result. AIEP makes this possible by binding outputs to evidence before they are returned.

What non-determinism means in practice

Standard AI systems are probabilistic. The same question asked twice may produce different answers. This is a feature in creative contexts — but a problem when accuracy matters.

More importantly: even if you get the same answer twice, you have no structural guarantee that either answer is grounded in the same evidence.

Non-determinism means:

Outputs cannot be audited reliably
Reasoning cannot be replayed independently
A third party cannot verify what the system “knew” at the time of the response

What AIEP determinism means

AIEP does not eliminate probabilistic generation. It adds a verification layer on top of it.

An AIEP-verified response is deterministic in a specific sense: the evidence it was based on is recorded, hashed, and replayable.

Even if the model generates slightly different text on a second pass, the evidence rail provides a stable, verifiable record of what was retrieved and validated.

First run:
  Answer: "..."
  Evidence hash: f3a2c4e0...
  Validation: passed

Second run (same question, different phrasing):
  Answer: "..." (different words)
  Evidence hash: f3a2c4e0... (same — same source, same content)
  Validation: passed

The output phrasing varies. The evidence does not.

Replayable reasoning

AIEP evidence rails can be re-fetched and re-validated at any future point.

If the source document has not changed: the hash matches, and the original reasoning is confirmed.

If the source document has changed: the hash does not match. The discrepancy is detectable and documentable.

This is what replayable reasoning means in AIEP: not that the AI says the same words — but that the evidence can be independently checked.

Example

{
  "evidence_rail": [
    {
      "artefact_id": "ev_001",
      "source_url": "https://aiep.dev/schema/evidence.json",
      "content_hash": "f3a2c4e0b6a7d991d3e2f5a9c8b7e6f1d0c9b8a7...",
      "retrieved_at": "2026-04-25T00:00:00Z",
      "validation_status": "passed"
    }
  ],
  "validation": {
    "replay_metadata_present": true,
    "status": "passed"
  }
}

The content_hash is the deterministic anchor. On any future replay, if this URL returns content that produces a different hash, the change is detected.

Try it

Verification Playground →

See the evidence rail example →

View examples/evidence-rail.json →

GitHub →

Machine endpoint →