The decision is the unit of audit and learning
Application logs show that software ran. A Decision Record shows what choice was framed, which evidence and alternatives were available, what was recommended, who or what authorized any effect, what actually executed, and what happened later. Those facts support challenge, incident response, evaluation, and governed policy change.
Fourteen parts of a useful Decision Record
Decision ID
Stable identity joining recommendation, authorization, execution, and outcome.
Decision time
When the decision basis was fixed, distinct from evidence and execution times.
Evidence references
Source identities, provenance, as-of/effective time, freshness, and validation—not copied secret payloads.
Missing or conflicting evidence
Explicit absence, contradiction, or staleness and its effect on the decision.
Alternatives
Externally meaningful options considered, including defer, abstain, or escalate.
Objectives
What the decision seeks to improve and at what horizon.
Constraints
Hard policy, eligibility, safety, capacity, consent, or contractual boundaries.
Uncertainty
Ranges, scenarios, sensitivity, and conditions that require abstention.
Recommendation
Selected alternative or abstention with a concise evidence-based rationale summary.
Authorization
Separate person or policy decision permitting, rejecting, revoking, or expiring an action.
Execution
Separate attempt and result, including not executed, failed, ambiguous, reversed, or succeeded.
Observed outcome
Later decision-relevant result with observation time and status.
Attribution and expected vs actual
Comparator and bounded basis for relating the result to the originating decision.
Versions
Methodology, policy, evidence-contract, model/rule, and implementation versions needed for reproduction.
Recommendation, authorization, execution, and outcome are separate states
A recommendation does not grant authority. Authorization does not prove execution. A successful request does not prove the intended external effect. An observed outcome does not prove attribution. Each stage needs its own identity, time, status, evidence, and accountable control.
Fully synthetic Decision Record
This illustrative, non-normative record contains no customer information, evidence payload, vendor assertion, personal data, or hidden reasoning. Its pending authorization, non-execution, and not-yet-observed outcome are deliberately explicit.
{
"schemaVersion": "decision-agent-decision-record-v1",
"decisionId": "synthetic-weekly-replenishment-2026-08-30-001",
"decisionTime": "2026-08-30T09:00:00Z",
"accountableOwner": "Inventory manager",
"alternatives": [
{
"id": "release-proposed-order",
"label": "Release proposed order",
"considered": true
},
{
"id": "reduce-quantity",
"label": "Reduce quantity",
"considered": true
},
{
"id": "transfer-stock",
"label": "Transfer stock",
"considered": true
},
{
"id": "defer",
"label": "Defer and gather evidence",
"considered": true
}
],
"evidenceReferences": [
{
"evidenceId": "inventory-snapshot-001",
"snapshotReference": "synthetic://inventory-snapshot-001",
"eventTime": "2026-08-30T08:45:00Z",
"effectiveTime": "2026-08-30T08:45:00Z",
"retrievedAt": "2026-08-30T08:50:00Z",
"freshness": "CURRENT",
"validation": "PASSED"
}
],
"missingAndConflictingEvidence": [
{
"evidenceId": "supplier-lead-time-confirmation",
"state": "MISSING",
"detail": "Do not release an order until current supplier lead time is reconciled."
}
],
"objectives": [
{
"id": "service-level",
"measure": "Weekly service level",
"expected": "Maintain the approved service-level range."
},
{
"id": "excess-inventory",
"measure": "Projected excess units",
"expected": "Remain below the approved inventory threshold."
}
],
"constraints": [
{
"id": "approved-supplier",
"status": "SATISFIED",
"evidence": [
"inventory-snapshot-001"
]
},
{
"id": "current-lead-time",
"status": "UNKNOWN",
"evidence": []
}
],
"uncertainty": [
{
"id": "next-week-demand",
"representation": {
"kind": "RANGE",
"lowerUnits": 760,
"upperUnits": 940
},
"materiality": "HIGH"
}
],
"recommendation": {
"status": "PROPOSED",
"alternativeId": "reduce-quantity",
"rationale": "Protect service while supplier lead time remains unresolved.",
"expectedOutcome": {
"serviceLevel": "within approved range",
"excessInventory": "below approved threshold"
}
},
"authorization": {
"status": "PENDING",
"authorizedBy": null,
"authorizedAt": null,
"policyVersion": "synthetic-replenishment-policy-v1",
"expiresAt": null
},
"execution": {
"status": "NOT_EXECUTED",
"actionReference": null,
"attemptedAt": null,
"completedAt": null,
"idempotencyKey": null
},
"observedOutcome": {
"status": "NOT_YET_OBSERVABLE",
"observedAt": null,
"expected": {
"serviceLevel": "within approved range",
"excessInventory": "below approved threshold"
},
"actual": null,
"attribution": "NOT_ASSESSED"
},
"versions": {
"methodology": "decisagent-methodology-v0.2",
"policy": "synthetic-replenishment-policy-v1",
"decisionDesign": "decision-design-specification-v0.4-alpha.1",
"evidenceContract": "synthetic-replenishment-evidence-contract-v1"
}
}Use the record for outcome evaluation—not automatic authority expansion
Compare decision-time expectations with later results at the declared horizon. Preserve overrides, execution variance, environmental changes, and attribution limits. A learning system may propose a policy update, but a governed person or process must approve it before later decisions change.
How to evaluate a Decision Agent → Create a free Decision Design