RELIANT SCALE ATLAS · REPLAY
Most systems tell you what happened.
Replay tells you what was knowable.
Those are different questions, and only the second one defends a decision. Six months after an automated decision, the interesting question is rarely “what did the system output”. It is “given what was actually available at the time, was that defensible — and what do we know now that nobody knew then?”
Held stateFORENSIC RECONSTRUCTION · NOT MODEL RE-EXECUTION
01 / Said precisely, because the distinction matters
Forensic reconstruction.
Not a rerun.
Replay reconstructs the stored envelope around a decision: the evidence reference, the authority evaluation, the outcome, and the principal that acted. It does notre-execute a model, and it cannot tell you what the model would output if asked again today. If a vendor offers to replay a language model's cognition exactly as it ran, ask them to demonstrate it on last quarter's data before you believe it.
What the API actually returnsThe response carries replay_class: FORENSIC_RECONSTRUCTION. The classification is in the payload rather than in the marketing, so an integrator cannot mistake what is being offered — and neither can we.
02 / What replay answers
The state of the world
as the agent found it.
A decision that looks wrong today may have been the only defensible decision available at the time. The opposite is also true. Replay is how you tell those two cases apart, and it is the difference between explaining a decision and apologising for one.
- 01
What the agent held
The evidence reference as it stood when the decision was made — not what that source says today. If the underlying record has since been corrected, replay shows you both facts: what was relied on, and that it later changed.
- 02
What it was permitted to do
The scopes on the credential at that moment, and the authority evaluation that ran. Scope is an upper bound, not a grant, so what was permitted and what was attempted are recorded separately.
- 03
What it decided, and why
The outcome and the reasons recorded with it, reported as they were. Replay never re-decides history against today's state.
- 04
What has changed since
Named explicitly. And where nothing has changed, replay says that — because “no later change has been admitted” is a finding, not an absence of one.
03 / Why later is the wrong time to start
The envelope has to be kept
before you know you need it.
Reconstructing a decision-time envelope after the fact means joining logs that were never designed to be joined, against source systems that have moved on. Sometimes it is possible. It is never cheap, it is never complete, and it is always done under the pressure of the event that made it necessary. Atlas keeps the envelope at decision time, which is the only moment it is free.
Also true, and worth sayingReplay reads the record. Reads are never charged, so examining your own history does not create a reason to examine less of it.
Begin
Pick a decision from last quarter.
Ask what was knowable.
Bring one automated decision and the systems around it. We will show you what a replay of it would contain and, just as usefully, what it could not contain because nothing captured it at the time.
partnerships@reliantscale.com
Replay is forensic reconstruction of a stored envelope. It is not model re-execution, not a compliance certification, and not a claim about what any model would do today.