Decision Evidence

A structured record of what the agent says it decided, saved with the conversation for the assistants you turn it on for.

Maturity: Beta. Statements on this page were traced to QuoxCORE code on 2026-10-30.

Tier: Enterprise add-on (see the store for current pricing) Protocol: AEE (Agent Envelope Exchange) Ledger: VOLT (Verifiable Operations Ledger & Trace)

Decision Evidence asks the assistant to explain each decision in a structured form, then saves that explanation with the conversation. The record is the model's own account of what it decided and why. It is not a trace of what actually ran.

At a glance


What it is	A structured decision record, requested from the model and saved with each chat reply
Where it runs	Dashboard chat, on Anthropic model paths
When to reach for it	You want a reviewable account of why an assistant answered or acted as it did, for the assistants you choose
How to start	Turn it on in the plugin settings and pick an evidence level

How it works

You turn it on. The dashboard chat sends the chosen evidence level with each message when the plugin is licensed and enabled. If no level is sent, the server default is none and nothing is recorded.
The model is asked for a record. The collector adds an instruction to the prompt asking the model to end its reply with a JSON block inside <decision_evidence> tags. The fields requested depend on the level.
The block is extracted. The collector takes the first <decision_evidence> block from the reply. The content is not checked against a schema, so it is whatever the model wrote.
It is saved in the AEE envelope. The record goes into the chat.success envelope for that reply. When WARD is enabled on the instance, the envelope is witnessed.
The decision is chained in VOLT. A VOLT agent.decision.recorded event is written with the evidence level, the decision and the confidence, hash-chained per run.
A missing record is flagged, not blocked. If the model returns no block, the reply still goes through and a quox.evidence.missing flag is recorded.

Evidence levels

Set a level in the plugin settings (evidenceLevel). What each level asks the model for:

Level	What the model is asked for
`none`	Nothing. No record is requested.
`minimal`	The decision and a one or two sentence reason summary.
`standard`	Adds the inputs used, the tools used, the action taken and a confidence from 0 to 1.
`full`	Everything in standard, plus context references (source IDs or locators).

Tool names, input names and context references are what the model reports. They are not checked against what actually ran.

Where the record goes

AEE envelope. The full extracted record travels in the chat.success envelope for the reply.
VOLT. Only the evidence level, the decision and the confidence are written to the hash-chained run log, one chain per run.
WARD. When WARD is enabled on the instance, the envelope is witnessed. If it is not enabled, no witness receipt exists.

Checking a record

Fetch a run from VOLT, then verify the file:

bash

quox volt fetch <runId>
quox volt verify <file>

To export evidence for review, use quox evidence.

Verification confirms the chain for that run is intact. It does not confirm the model's account was accurate.

Compliance

Decision Evidence helps with the record-keeping and transparency that several AI frameworks ask for. It does not by itself make a system compliant.

EU AI Act: Article 12

Article 12 requires that high-risk AI systems support automatic recording of events (logging) to enable traceability of the system's functioning. Decision Evidence addresses this by:

Recording decision evidence, with timestamps, inputs and outputs, for the assistants you turn it on for
Keeping the decision and confidence in a hash-chained run log, so a later edit to them is detectable
Supporting a full evidence level that keeps the model's own structured account of its reasoning
Providing export capabilities for regulatory inspection

NIST AI Risk Management Framework

The NIST AI RMF calls for documentation of AI system behaviour, decisions, and outcomes. Decision Evidence supports the Govern, Map, Measure, and Manage functions by providing:

Showing the confidence the model reported alongside each recorded decision
Traceable records linking decisions to their inputs and context
Export paths for risk assessment workflows

General Audit Readiness

For organisations not yet subject to specific AI regulation, Decision Evidence provides a forward-looking compliance posture. Records captured now will be available when regulations take effect, without needing to retrofit logging into existing agent deployments.

Limits

Opt-in. Only assistants you turn it on for produce records. It is not applied to every decision on the platform.
Dashboard chat only. It covers the standard dashboard chat on Anthropic paths. It does not cover streaming chat, other model providers, the CLI or workflows.
Self-reported. The record is the model's own account, not a trace of what ran. Tool names and context references are not verified.
Chain scope. The VOLT chain holds the decision and confidence per run, not the full text of the record.
Missing records are flagged, not blocked. A reply without a record is delivered and marked.

Last verified 2026-10-30 against QuoxCORE origin/main.