The Decision and Execution Record

The definition

What a record captures

A decision and execution record is the structured account written when an agent works through a governed workflow. It captures what the decision rested on and what happened as a result — enough for a reviewer to reconstruct the decision without having been in the room.

The contents of a record substantiated

Evidence The inputs the decision rested on, with references back to their sources.
Source context Where each input came from and the state it was in when used.
Assumptions The material assumptions the workflow made, stated rather than buried.
Tool activity Which tools ran, on what, and what they returned.
Policy checks The risk class, authority level, and gates the decision passed through.
Confidence The calibrated confidence attached to the output, plus any flagged gaps.
Approvals The human decisions around the workflow — who approved what, and when.
Exceptions Anything that deviated from policy or was escalated rather than executed.
Resulting action What actually happened — and, just as importantly, what did not.

What it is not. A record does not expose private model chain-of-thought. It captures the evidence, checks, and actions a reviewer needs to hold a decision accountable — not the raw internals of a model.

Why it matters

So a decision can be shown, not argued

If a partner, regulator, or your own risk team asks why an agent did something, the answer should be a record you open — not a story you reconstruct.

Three things a record buys you substantiated

Replay The sequence can be reconstructed end to end — what was gathered, checked, gated, and done.
Accountability Every approval and exception is attributed, so responsibility is legible after the fact.
Audit Records are written to hash-chained logs, so the history can be verified rather than trusted.

The mechanics behind this — hash-chained logs, evidence chains, structured traces — are covered in the Trust Center and on the Architecture page.

Annotated sample

One record, read end to end

Below is a single illustrative record from a venture-intelligence workflow: an example company being considered for partner review. Every value is representative sample data, constructed to show the shape of a record.

The record at a glance illustrative

decision_record · venture-intelligence Illustrative example

Request & workflow

+Request: Should the example company advance to partner review?
+Workflow: venture-intelligence · version 0.9-example

Evidence

+Eight evidence items gathered, each referenced to its source
!Two items marked low-confidence
!Material gaps flagged: customer retention cohort · reference call

Policy & evaluation

+Risk class: medium · authority level: recommend
+Human approval required before the recommendation progresses
+Quality gate: passed in this example · confidence: illustrative value omitted

Decision & resulting action

+Recommendation: advance with conditions
+Human decision: approved in this example
+Resulting action: partner-review package prepared — no external message or transaction executed

Illustrative example with representative sample data — not customer output, a live trace, or a performance claim. In a live deployment each evidence item links to its source and the record is written to the hash-chained audit log.

How a reviewer reads it illustrative

Request & workflow Pins the decision to a specific question and a specific workflow version — so you know exactly which logic produced it.
Evidence Shows what the recommendation rests on, and is honest about weakness: low-confidence items and material gaps are flagged, not papered over.
Policy Shows the boundary the agent operated inside — medium risk, recommend-level authority, human approval required. The workflow may prepare an internal artifact; it may not communicate or execute externally.
Evaluation Shows whether the output cleared the quality gate and with what confidence, so the reviewer knows how much weight to put on it.
Decision Separates the agent's recommendation from the human's decision — both are recorded, and neither is confused for the other.
Resulting action States what happened in the world. Here: a package was prepared, and nothing left the boundary.

The raw shape illustrative

The same sample as structured data — an excerpt of the illustrative record above, not customer output.

{
  "record_type": "illustrative_governed_decision_record",
  "status": "illustrative",
  "workflow": { "id": "venture-intelligence", "version": "0.9-example" },
  "request": "Should the example company advance to partner review?",
  "evidence": {
    "items": 8,
    "low_confidence_items": 2,
    "material_gaps": ["customer retention cohort", "reference call"]
  },
  "policy": {
    "risk_class": "medium",
    "authority_level": "recommend",
    "human_approval_required": true
  },
  "decision": {
    "recommendation": "advance with conditions",
    "human_status": "approved in example",
    "resulting_action": "partner-review package prepared"
  },
  "disclaimer": "Synthetic demonstration. Not customer data, a live trace or a performance claim."
}

The four rungs

Where records sit on the ladder

The decision and execution record is how the governed-autonomy ladder — Evidence → Calibrated confidence → Gated authority → Full audit trail — becomes inspectable. Each rung leaves its mark in the record.

Evidence

The record's evidence and source-context fields show what the decision rested on — including the gaps.

Calibrated confidence

The evaluation fields carry the confidence attached to the output, so weak signals read as weak.

Gated authority

The policy fields show the risk class, the authority the agent held, and the gate a human had to pass.

Full audit trail

Records are written to hash-chained, replayable logs — the trail is the accumulation of records like this one.

See a record produced live

In a walkthrough we run a workflow on representative inputs and show the record it leaves behind — every rung of the ladder, end to end.

Browse the demo library → Visit the Trust Center →