Skip to content
AletheIA Docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

Execution Audit Record — Extension View

Purpose

Define the pattern-level audit fields an executed task must leave behind: the pattern that was selected, the vehicle it ran in, the skill and tool used, the gate that verified it, the evidence it produced, what it cost, where its state lives, and the decision that closed it. These fields are declared as an extension view over the AHGE governance record — the per-action record defined by the Agent Harness Governance Extension and structured by schemas/agent-harness-governance-record.schema.json.

The AHGE record remains the record authority: it already captures, per action, the tool name, risk and side-effect class, permission decision, decision authority, budget profile, and stop reason. This view does not create a parallel record. It aggregates and references those per-action entries at the level of the executed pattern, so a reviewer can answer “which topology ran, under which gate, at what cost, and with what outcome” without re-reading every action line. Field names here are chosen not to collide with the schema’s per-action fields (action_id, tool_name, permission_decision, budget_profile, stop_reason, …); where they overlap in meaning, the view field is a reference into the record, never a restatement.

This document is docs-first and normative: it declares a reviewable shape; it does not implement a logger, store, or reporting pipeline.

Non-goals

  • No runtime — no audit logger, collector, or reporting mechanism is implemented here.
  • No parallel record — the AHGE governance record is the single audit authority; this view points into it via governance_record_refs and adds pattern-level fields only.
  • No new verdict vocabulary — outcomes referenced here use the existing policy verdicts, the AHGE decision values, and the selection verdicts of the execution pattern selection.

Declaration shape

execution_audit_extension:
  execution_id:
  task_id:

  pattern_selected:            # from execution_pattern_selection.recommended_pattern.type
  vehicle:                     # from execution_pattern_selection.recommended_vehicle.type
  selection_verdict_ref:       # the selection declaration this execution ran under

  skill_used:
  tool_used:                   # references tool_name entries in the AHGE record

  gate:
    gate_type:                 # objective gate, checker, or human review
    gate_result:

  evidence_refs:
    - <evidence-reference>

  cost:
    tokens_used:
    runtime:
    iterations:

  loop_state_ref:              # loop_state record, when the pattern is a loop
  governance_record_refs:      # the AHGE per-action record entries this view aggregates
    - <record-reference>

  decision:                    # the outcome that closed the execution

Normative rules

  1. The AHGE record is authoritative. Every claim in the extension view must be resolvable to per-action entries in the AHGE governance record via governance_record_refs. A view entry that points at nothing is not auditable and is not admissible.
  2. The selection is part of the audit. pattern_selected, vehicle, and selection_verdict_ref must match the execution pattern selection declaration that approved the topology. An execution whose audited pattern differs from its selected pattern is a governance finding, not a footnote.
  3. Gates leave evidence. gate.gate_result and evidence_refs must be filled whenever the selection declared objective_gate_required, maker_checker_required, or human_review_required — a gate without a recorded result is a claim, not a check (see Objective Gate Policy and Maker-Checker Policy).
  4. Cost is recorded, not estimated. cost reflects what the execution actually consumed; the budget semantics and exhaustion handling belong to AHGE.
  5. Loops point at their state. When the executed pattern is loop_until_done or scheduled_stateful_loop, loop_state_ref must point at the loop state record.
  6. Knowledge Governance restricts what may be audited. Evidence references and view fields must respect the restricted knowledge usage policy; restricted content is referenced, never copied into the audit surface.

Comprehension debt review

Volume-generating patterns — fan-out, generate-and-filter, tournament, and recurring loops — can produce more artifacts than reviewers absorb. The audit surface therefore includes a periodic comprehension debt review per execution or recurring pattern; the concept is defined in concepts/comprehension-debt.md.

comprehension_debt_review:
  execution_id:
  pattern:
  period:
  artifacts_generated:
  artifacts_reviewed_by_human:
  accepted_change_rate:
  unread_artifact_risk: low | medium | high
  reviewer_capacity:
  debt_signals:
    - repeated_confusion
    - unknown_generated_area
    - fragile_gate
    - high_output_volume
  decision: continue | reduce_scope | pause | retire

The review decision is binding on the pattern, not on a single run: reduce_scope, pause, or retire feeds back into the next execution pattern selection — a pattern whose output is not being read has lost the proportionality that justified it.

Was this page helpful?