Every run stamps its prompt version
The ensemble that produced an output is reconstructable — the prompt is part of provenance.
An LLM output without its prompt is a number without a question. You can record the output, the model, and the inputs — but if the prompt that framed the question is gone, the run is irreproducible. The prompt is part of provenance, and in DEINO it is stamped on every run.
Every personality TOML carries a prompt version, and that version is written into the audit manifest for every output the sister produces. The manifest does not say the analyst sister ran; it says the analyst sister, prompt version 7, ran against these NodeVersions with this model. Replaying the run means replaying the exact prompt, not an approximation of it. The ensemble that produced an output is reconstructable down to the wording of the question.
Provenance includes the question
Most provenance records stop at the inputs and the model. Both matter — the GNL NodeVersions and the model hash are in the manifest — but they are not enough. Two runs of the same model against the same inputs with different prompts are two different forecasts, and a manifest that cannot tell them apart is a manifest that cannot tell you why an output changed. The prompt version closes that gap. It records not just what the model saw, but what it was asked.
Version, not snapshot
We stamp a version, not a full prompt snapshot. The personality TOML is the prompt template; the version is a hash of it. Stamping the version means the manifest carries a short, comparable identifier, and anyone with the personality file at that version can reconstruct the exact prompt. Stamping the full snapshot would be redundant — the file is the source of truth — and would bloat the manifest. The version is the pointer; the file is the content.
Two runs of the same model against the same inputs with different prompts are two different forecasts. A manifest that cannot tell them apart cannot tell you why an output changed.
What this means for replay
Replay is the test. When a third party — an auditor, a regulator, a future you — replays a decision, they need the bytes to match. The prompt version is what makes the bytes match. Without it, replay produces a different output because the prompt drifted, and the manifest that attested reproducibility attests a lie. With it, replay reproduces the exact run, and the manifest’s claim survives. The stamp is small; the property it buys is the one the whole category rests on.
- Prompt version (a hash of the personality TOML) is in every manifest — the question is part of provenance.
- Version, not snapshot — the manifest points to the file; the file reproduces the prompt.
- Replay requires the exact prompt — without the stamp, the bytes do not match and reproducibility is a claim, not a property.
- The stamp survives a model swap — pin the prompt version and the model hash; the chain verifies across both.
The honesty of a versioned question
Stamping the prompt version is an act of honesty. It says: this output was produced by a model asked a specific question in a specific way, and if you do not like the output, the question is where to look. A system that hides its prompt is a system that cannot be audited for bias in the question itself. A system that stamps it is a system that hands the examiner the framing along with the answer. We are pre-revenue, building the POC; the stamps run, and the calibration that would make the answers trustworthy is gated (Gate 1b) until it is measured.
The prompt is part of the question, and the question is part of provenance. A manifest that records the answer without the question attests to half a run.
— provenance design note