Octoryn Platform
Octoryn Replay
Reconstruct an AI action from its recorded inputs, decisions and evidence for review and dispute resolution.
Overview
Octoryn Replay reconstructs a past AI action from the inputs, decisions and evidence recorded when it ran, so a reviewer can walk through exactly what the system saw and decided. It assembles the prompt and context, the model and version, the policy decisions, the tool invocations, any human approvals and the final output into an ordered account of the run. It is intended for audit, incident review and dispute resolution, where "show me what happened, step by step" has to be answerable after the fact. As an Experimental capability it is a faithful record of a run — it is not a re-execution of the model, and it does not reproduce outputs by running the model again.
How it works
Anchor to an action
A replay starts from a specific recorded action, identified from Observe or from its evidence reference.
Reassemble the inputs
The prompt, retrieved context and tool arguments as they were at run time are gathered back into their original form.
Order the decisions
Policy decisions, tool invocations, approvals and escalations are laid out in the sequence they actually occurred.
Attach the evidence
Each step links to its evidence — model and version, policy reason, invocation result, approver — with tamper-evident integrity.
Present for review
The result is a step-by-step account a reviewer can read, cite or hand to another party in a dispute.
Key capabilities
Faithful reconstruction
Replay reproduces the recorded inputs and decisions of a run, not a fresh model execution that might differ.
Ordered decision trail
Every policy decision, tool call, approval and escalation appears in the order it happened.
Evidence linkage
Each step carries its provenance — model, version, policy reason, invocation and human approval — as tamper-evident evidence.
Sensitive-data aware
Where Privacy redacted material at run time, the replay reflects what was actually processed, not the raw sensitive values.
Shareable account
The reconstructed run can be presented to auditors, reviewers or a counterparty as an ordered, cited record.
Anchored to evidence integrity
Because it draws on tamper-evident records, a replay can be checked for whether its underlying evidence was altered.
What it is not
- It is not a re-execution — Replay does not run the model again, so it neither reproduces nor second-guesses outputs by regenerating them.
- It is not a simulator or what-if tool; it reconstructs one run as it happened, not hypothetical alternatives.
- It is only as complete as the evidence recorded; steps that were never captured cannot be reconstructed after the fact.
Integration
Replay builds on the same evidence that Octoryn records along the action path, and it typically starts from an action identified in Observe, so the two are designed to work together. Reconstructed runs can be exported as a record for your audit or case-management systems. What evidence is retained, for how long, and in what form it can be handed to a third party are assessed per engagement.
Deployment options
- Managed cloud
- Our cloud account
- Private cloud / VPC
Example use cases
- After a customer disputes an AI-generated decision, a reviewer replays the exact run to show which inputs and policy led to the outcome.
- During an incident review, a team walks the ordered decision trail of a failed action to see where a policy or tool behaved unexpectedly.
- An auditor is handed a reconstructed run, with model, version and approvals attached, as evidence that required controls were applied.
Frequently asked questions
Does Replay run the model again to reproduce the answer?
No. Replay is a faithful record of a past run, reconstructed from its recorded inputs, decisions and evidence. It does not re-execute the model, so it will not regenerate or alter the original output.Can a reconstructed run be trusted as evidence?
It draws on tamper-evident records, so a replay can be checked for whether its underlying evidence was altered. Its completeness depends on what was captured at run time, and how it can be used in a formal process is assessed per engagement.What if some steps weren't recorded?
Replay can only reconstruct what evidence exists. Steps that were never captured cannot be recovered after the fact, which is why the evidence and retention scope are agreed up front for each engagement.
All capabilities
Octoryn Builder
Private previewBuild production applications that emit portable, reviewable source code — not a locked-in low-code runtime.
- Audience:
- Engineering teams and organisations building internal and customer-facing software.
Octoryn Runtime
Private previewA runtime for governed AI-enabled applications, connecting models, policies, tools and operational systems.
- Audience:
- Teams operating AI inside real workflows and systems of record.
Octoryn Gateway
Private previewA controlled AI gateway so model access passes through organisational policy instead of scattered API keys.
- Audience:
- Security, platform and data teams standardising enterprise model access.
