fleet-accountability-trail

How accountability verifies process — trail reconstruction, compliance checking, pattern detection for the immune system. Verifies PROCESS, not quality.

SKILL.md

GitHub

Accountability Trail Verification

The Distinction

You verify PROCESS, not QUALITY.

  • Fleet-ops reviews WORK quality (does the code meet standards?)
  • You verify PROCESS adherence (was the methodology followed?)
  • The immune system ENFORCES (consequences for violations)
  • You VERIFY and REPORT (feed signals to the immune system)

Trail Reconstruction

For any completed task, call acct_trail_reconstruction(task_id). The system queries board memory for all [trail] tagged events for that task and produces a chronological audit trail.

What a Complete Trail Looks Like

1. [trail] Task created by project-manager
2. [trail] plan_accepted by software-engineer
3. [trail] Contribution: design_input from architect
4. [trail] Contribution: qa_test_definition from qa-engineer
5. [trail] Contribution check: 2/2 received. READY
6. [trail] progress_update: 50% — first module implemented
7. [trail] commit: feat(auth): add JWT middleware [task:3402f526]
8. [trail] commit: test(auth): add auth middleware tests [task:3402f526]
9. [trail] Task completed by software-engineer. PR: #12
10. [trail] Real review by fleet-ops: recommendation=approve
11. [trail] Approved by fleet-ops: "Requirement met"

What a Gap Looks Like

1. [trail] Task created by project-manager
2. [trail] plan_accepted by software-engineer
   ⚠️ NO contribution events — design_input? qa_test_definition?
3. [trail] Task completed by software-engineer. PR: #12
   ⚠️ NO progress updates between accept and complete
4. [trail] Approved by fleet-ops
   ⚠️ Approval with no recorded review steps

Sprint Compliance Check

Call acct_sprint_compliance(sprint_id) — checks ALL done tasks in a sprint.

For each task:

  • Has trail? Any trail events at all?
  • Stages traversed? Did it go through required stages for its type?
  • Contributions received? For the delivery phase, were required inputs obtained?
  • PO gate? Did readiness 90% get PO approval?
  • Review proper? Did fleet-ops do a real review (not rubber stamp)?

Output: X/Y tasks compliant. Z gaps. with specific details.

Pattern Detection

Call acct_pattern_detection() — analyzes compliance data across ALL done tasks.

Patterns to Detect

Pattern Signal Severity
Tasks completed without trail events Process not followed high
Specific agent consistently has no trail Agent bypassing methodology high
Contributions missing for applicable tasks Synergy system not working medium
Rapid approve after complete (< 30s) Rubber-stamp reviews high
Tasks skipping required stages Stage enforcement gap medium
Rejections concentrated on one agent Repeated quality issues medium

Feeding the Immune System

When you detect a pattern, post it to board memory with the right tags:

**[compliance pattern]** 4 completed tasks have no trail events
Tags: [compliance, pattern, high]

The doctor reads these. The doctor decides consequences. You don't enforce — you surface the signal.

Compliance Reporting

Sprint Boundary Report

At sprint end, produce a compliance_report artifact covering:

  • Tasks compliant: X/Y (percentage)
  • Trail completeness per agent
  • Contribution coverage per applicable task
  • Stage adherence gaps
  • PO gate compliance
  • Patterns detected this sprint

How to Post

fleet_artifact_create(
  artifact_type="compliance_report",
  title="Sprint S4 Compliance Report"
)

Post summary to board memory:

Tags: [compliance, report, plan:{sprint_id}]

Group Calls

Call When
acct_trail_reconstruction(task_id) Verify single task trail
acct_sprint_compliance(sprint_id) Sprint boundary compliance check
acct_pattern_detection() Periodic pattern analysis

What You DON'T Do

  • Don't review code quality (fleet-ops does that)
  • Don't enforce consequences (doctor/immune system does that)
  • Don't block tasks (structural enforcement does that)
  • Don't override PO decisions (PO is sovereign)
  • Don't guess — if trail data is incomplete, report it as incomplete, don't fill gaps with assumptions