The reproducible verdict — byte-identical, years later, without the original model.
A reproducible verdict is a compliance decision that re-derives byte-identically later — the same signed artifact, produced again from its pinned record, by the platform alone. The mechanism is deterministic replay: not re-asking a model, but re-deriving the exact bytes that were signed at write time.
One mechanism across SOC 2, HIPAA, GDPR, EU AI Act & WCAG 2.1 AA.
Deterministic Precision. Experiential Intuition. Autonomous Agents.
Re-explaining is not reproducing.
Three things get called “replay.” Only one of them is evidence a regulator can lean on five years later.
A fresh answer
Ask the model again and it reasons again — possibly differently, and possibly on a model that has since been upgraded or retired. You get a new opinion, not the original record.
A saved picture
A stored screenshot or export proves you once saw an answer. It does not prove the answer re-derives, and it is editable after the fact.
The same bytes
Deterministic replay re-derives the original signed verdict from its pinned record — byte-for-byte, by the platform alone. That is the bar, and it is the one most AI compliance tools cannot clear.
What “reproducible” has to mean.
For a verdict to hold up in a year-five review, replay has to be all four of these at once — not three.
Same inputs, same bytes
Given the pinned record of a decision, replay returns the identical signed artifact every time — no variance, no “close enough.”
Verifiable without trusting us
Anyone can check the signature and the hash chain with standard cryptographic tools, offline — no proprietary viewer, and no requirement to trust acipta as the operator.
Survives a model swap
The model’s reasoning happened once, at write time. Replay re-derives the signed verdict from its record — so swapping, upgrading, or deprecating the underlying model does not change what already shipped.
Still provable years later
The verdict re-derives from a sealed, pinned bundle long after the original engineer has moved on (target: five-year replay). Turnover does not break the chain.
Seal it at the source. Re-derive it on demand.
acipta binds each verdict to a record that re-derives byte-identically later — disclosed here as an outcome, not an assembly recipe.
Signed at write time
Every verdict is Ed25519-signed at the moment it is produced — not batched overnight, not reconstructed before an exam.
Timestamped & chained
An independent RFC 3161 timestamp and a hash-linked ledger make order and integrity provable. This is the cryptographic evidence chain underneath every replay.
Pinned, then re-derivable
Everything the verdict depended on is sealed into the record, so replay reproduces it without re-prompting any model. The flight recorder is where that record lives.
In 2031, a regulator asks you to prove a 2026 decision. Which kind of replay do you have?
Re-running a model gives you a 2031 answer to a 2026 question. A reproducible verdict gives you the 2026 artifact again — the same bytes, signed at the time, verifiable by anyone. That is the difference between “we can re-explain it” and “we can re-prove it.”
An auditor can verify the signature and timestamp independently, and the verdict re-derives by the platform alone — no original engineer, no original model required.
A note on scope. Defensibility here is a property of the evidence machinery — reproducible, attributable, tamper-evident, replayable, signed at write time. It is not a guarantee that a decision was substantively correct, and it is not a substitute for legal judgment; your counsel stays in the loop.
Deterministic replay — questions
What is a reproducible verdict?
A compliance decision that re-derives byte-identically later: the same signed artifact, produced again from the pinned record, by the platform alone. Not a re-explanation — the same bytes.
Isn’t re-running the AI the same as replaying it?
No. Re-running a model produces a fresh answer that can differ, and the model may be upgraded or retired by then. Deterministic replay re-derives the original signed verdict from its pinned record — same bytes, not a new opinion.
Does deterministic replay mean the AI is deterministic?
No. The model’s reasoning happens once, at write time. What is deterministic is the re-derivation of the signed record afterward — the verdict is pinned at the moment it is made and replays without re-prompting any model.
What happens when the model is deprecated in 2031?
The verdict re-derives from the signed, pinned bundle by the platform alone (target: five-year replay). Model deprecation and staff turnover do not break it, because replay does not depend on re-running the original model.
A verdict you can still prove in 2031.
Reproducible, signed, replayable. acipta is the agent-based defensibility platform — workflow-grounded. Full GA August 23, 2026.