Gate 1: the reproducibility gate
The full-chain-walk — before any accuracy claim, the bytes have to match.
Before DEINO makes any claim about accuracy, it makes a claim about reproducibility. Gate 1 is the gate that earns the second word by passing the first: the full-chain-walk, where an output is replayed bit-for-bit and the bytes have to match. No match, no accuracy claim. The gate comes before the forecast.
Reproducibility is the property that, given the same inputs, the same model, the same prompt, and the same graph versions, the system produces the identical bytes. It sounds simple. It is not. Non-determinism hides in hash-map iteration order, floating-point reductions, allocator layout, time-based seeds. A replay that produces a different output because of any of these is a replay that cannot be used in an audit — because the auditor cannot tell whether the difference is a bug or a drift. Gate 1 is the gate that checks, end to end, that the bytes match.
The full-chain-walk
The full-chain-walk is the test. Take an output — a forecast, a cone, a manifest — and replay the entire chain that produced it: the GNL NodeVersions, the personality TOML at its stamped prompt version, the model hash, the agent trace, the normalization in deino-oracle. Walk the full chain and compare the output bytes to the recorded output bytes. If they match, the chain is reproducible. If they do not, the chain has a non-determinism, and non-determinism is a defect, not a feature.
Before any accuracy claim, the bytes have to match. A replay that cannot tell a bug from a drift cannot be used in an audit — and a manifest that attests a non-reproducible computation is a signed lie.
Why this gate comes before accuracy
Accuracy is a measurement against a realized timeline. You cannot measure accuracy if you cannot reproduce the forecast — because you cannot tell whether a difference in score is a difference in the model or a difference in the replay. Gate 1 is the precondition for Gate 1b: before you can claim the shadow run beats a baseline, you have to prove the shadow run is replayable. Reproducibility first, calibration second. The gates are ordered because the properties are ordered: you cannot calibrate a system that does not reproduce.
The manifest attests the walk
The full-chain-walk is what the manifest attests. The manifest (BR-842) records the inputs, the model, the prompt version, the agent trace, and the output, hash-linked. A third party with the manifest and the code can walk the chain and verify the bytes. That is what makes the manifest a record rather than a claim: it is checkable, by anyone, with the same result. A manifest that attests a non-reproducible computation is a signed lie, and Gate 1 is the gate that makes sure the manifest is not lying.
- Full-chain-walk: replay from NodeVersions to output bytes; every link checked; every link must match.
- Non-determinism is a defect, not a tolerance — a mismatch is something to fix, not to accept.
- Gate 1 before Gate 1b: reproducibility is the precondition for calibration. You cannot calibrate what you cannot replay.
- The manifest attests the walk — it is checkable by a third party, with the same bytes.
What the gate does not claim
Gate 1 does not claim the forecast was correct. It claims the computation is reproducible. Those are different statements, and the difference is the one the category rests on: provenance and reproducibility, never truth. A reproducible forecast can be wrong; a non-reproducible forecast cannot even be audited. Gate 1 is the gate that makes the forecast auditable — the accuracy is a separate loop (Gate 1b), and it comes second. We are pre-revenue, building the POC; the full-chain-walk is the gate, and the gate is what lets the rest of the claims be made in order.
Reproducibility is the precondition for calibration. You cannot measure whether a forecast was accurate if you cannot prove the forecast is replayable. Gate 1 is the gate that proves it, and it comes before the gate that measures it.
— gate design note