Skip to main content

Governance Model

ASDD's governance model answers a fundamental question: how do we give AI agents meaningful autonomy without giving up accountability?

The answer is a layered system of confidence gates, formal override mechanisms, reasoning transparency requirements, and immutable audit trails.


The governance stack

Level 4: Production Learning Loop    ← Knowledge Agent proposes; humans approve
Level 3: Human Phase Gate Sign-offs ← TL (and PO at key phases) approves every transition
Level 2: Cumulative Confidence Score ← Automatic pipeline halt when CCS < 0.65
Level 1: Individual Agent Thresholds ← Per-agent minimum confidence gating

All four levels must be active for governance to be effective. Disabling any one level shifts risk upward.


Confidence scoring

Every agent emits a confidence score (0.0–1.0) alongside every artifact. This score represents the agent's assessment of its own output quality — not a performance metric imposed externally.

Score interpretation

RangeMeaningSystem action
0.95–1.0High confidence — agent used all required inputs, domain terms resolved, spec unambiguousAuto-approval eligible (for LOW risk slices)
0.80–0.94Moderate confidence — minor uncertainties noted in uncertainty_factorsProceed; TL informed; uncertainty factors attached to artifact
0.75–0.79Low-moderate confidencePipeline proceeds; human code review mandatory for implementation outputs
Below thresholdConfidence below agent minimumPipeline halts; TL escalation triggered

Individual agent thresholds

AgentMinimum ConfidenceAction if Below Threshold
Discovery Agent0.85DRAFT — AWAITING CLARIFICATION; route to PO/TL
Spec Agent0.85Flag ambiguous sections; route to TL
Validation Agent0.90Block pipeline; require TL sign-off
Domain Agent0.85Draft domain model; require TL review
Design Agent0.80Draft architecture; require TL review
Implementation Agent0.75Feature branch only; human code review required
Security Agent0.95Block deployment; require TL + Security review
Knowledge Agent0.80Propose steering update; require human approval

Thresholds are defined per agent in .asdd/steering/agent-thresholds.md and can be adjusted by the Tech Lead.


Cumulative Confidence Score (CCS)

The CCS is the most important single governance mechanism in ASDD.

The cascade problem

Individual agent gates can all pass while the aggregate pipeline quality is unacceptably low:

AgentScorePasses threshold?
Spec Agent0.86Yes (min: 0.85)
Validation Agent0.91Yes (min: 0.90)
Design Agent0.81Yes (min: 0.80)
Implementation Agent0.76Yes (min: 0.75)

CCS = 0.86 × 0.91 × 0.81 × 0.76 = 0.483 → Pipeline halts.

Each agent individually passed. But the combined uncertainty means the output should not be auto-approved.

CCS calculation

CCS = ∏(agent_confidence_scores in pipeline path)

For a 4-agent path: CCS = Conf_Spec × Conf_Validation × Conf_Design × Conf_Implementation

Threshold: CCS < 0.65 triggers AUTOMATIC_RECOVERY_HALT.

Dynamic gating

When any agent provides a low (but passing) score, the next agent's minimum threshold increases by +0.05:

Spec Agent scores 0.86 (just above 0.85 floor):
→ Validation Agent minimum: 0.90 + 0.05 = 0.95 for this slice

This compensates for accumulated uncertainty debt without halting the pipeline prematurely.


Reasoning traces

An audit trail of what an agent decided is not sufficient governance. Humans need to understand why.

What a Reasoning Trace contains

Every agent must emit a reasoning_trace.md alongside its artifact:

# Reasoning Trace — [Agent Name] — [Slice ID] — [Date]

## Assumptions Made
- [Key assumption 1]
- [Key assumption 2]

## Requirements Prioritised
- REQ-001 was prioritised over REQ-007 because [reason]

## Alternatives Considered
- Alternative A: [description] — rejected because [reason]
- Alternative B: [description] — rejected because [reason]

## Uncertainty Factors (if confidence < 0.95)
- [Uncertainty 1 — why it reduced confidence]
- [Uncertainty 2 — why it reduced confidence]

## Confidence Score: [0.0–1.0]

Phase gate sign-off requires the Reasoning Trace, not just the artifact. The TL must acknowledge the trace before the gate passes. This is not optional — approving an artifact without understanding its reasoning is rubber-stamping.

Traces are stored in .asdd/traces/ and linked in the State Manifest.


Dissent Protocol

Any team member may formally reject an agent-produced artifact at any phase gate. This is a structured mechanism with an audit trail — not an informal veto.

Filing a Dissent Notice

A dissent notice is created by adding an entry to /docs/dissent-log.md:

Date | Author | Phase | Artifact Rejected | Reason Category | Detailed Rationale | Proposed Resolution

Reason categories

CategoryMeaning
TECHNICALLY_UNSAFEAgent output introduces architectural risk or security vulnerability
SPEC_NONCOMPLIANTAgent output does not match the approved spec
BUSINESS_MISALIGNEDAgent output conflicts with stated product intent
HALLUCINATION_SUSPECTEDAgent output references undefined domain concepts
QUALITY_UNACCEPTABLEAgent output is technically compliant but unmaintainable

Resolution authority and SLA

Reason CategoryResolution OwnerSLA
TECHNICALLY_UNSAFETech LeadSame sprint day
SPEC_NONCOMPLIANTTech LeadSame sprint day
BUSINESS_MISALIGNEDProduct OwnerNext planning session
HALLUCINATION_SUSPECTEDTech LeadSame sprint day
QUALITY_UNACCEPTABLETech Lead + EngineerWithin sprint

The immutable override audit trail

All override decisions are immutable log entries. They cannot be deleted. The Knowledge Agent monitors override frequency by agent, phase, and reason category to detect systematic agent weaknesses.

High override frequency for a specific agent is a signal to:

  1. Review the agent's confidence thresholds
  2. Review the quality of inputs the agent is receiving (usually spec ambiguity upstream)
  3. Consider Discovery Re-entry for consistently failing slices

Phase gate discipline

Phase gates are the human checkpoints that separate ASDD's governance from a pure automation pipeline.

Gate requirements by phase

PhaseGate RequirementWho Signs Off
Phase 0intent.md approved; confidence ≥ 0.85PO
Phase 1All slice requirements pass Spec Validation GateTL
Phase 2domain-model.md schema-compliant; TL-reviewedTL
Phase 3design.md + Reasoning Trace reviewed; traceable to requirementsTL
Phase 4Waves complete; TDD tests green; Reasoning Traces acknowledgedTL + Engineer
Phase 5All CI gates green; no force-bypassDevOps Agent + TL
Phase 6Telemetry emitting from all new componentsTL
Phase 7Knowledge Agent proposals reviewed; approved updates appliedTL + PO

Force-bypass protocol

CI pipeline failures may only be force-bypassed by the Tech Lead. Force-bypasses:

  • Are logged immutably in the audit trail
  • Cannot be used on Self-Healing PRs (no exceptions)
  • Trigger a mandatory post-sprint root cause analysis

An organization where force-bypasses are routine has lost governance integrity.


The Hallucination Budget

Agent hallucination is a measurable signal, not a random defect.

Every agent pipeline defines a Hallucination Budget: a maximum tolerable rate of low-confidence outputs per execution wave. When an agent's confidence score drops below threshold consistently, the framework triggers automatic Discovery Re-entry.

Discovery Re-entry

Trigger: confidence_score < threshold on N consecutive artifacts for a slice

Sequence:

  1. Pipeline halts
  2. Knowledge Agent flags DISCOVERY_REENTRY
  3. Slice status reset to SPEC
  4. PO + TL notified
  5. Slice returns to Phase −1 for requirement re-validation

Why this matters: Hallucination is almost always an input problem. The slice failed not because the agent is defective but because the requirements were ambiguous. Re-entry forces the root cause — ambiguous intent — to be addressed before implementation resources are wasted again.


Knowledge Agent governance functions

The Knowledge Agent is not just a learning system — it is the primary governance enforcement mechanism.

FunctionDescription
State custodianAuthoritative keeper of manifest.json; validates every state transition
Drift detectionFlags STATE_DRIFT when a file is modified without a corresponding manifest update
Conflict arbitrationDetects agent-agent conflicts; applies Steering Rules; escalates to TL when rules cannot resolve
Override monitoringTracks override frequency by agent and category; flags systematic weaknesses
Learning proposalsProposes steering rule updates based on ≥ 3 production incidents
Learning validationMonitors metric improvement for 7 days after Self-Healing PR merge

Governance anti-patterns

Anti-patternWhy it fails
Skipping Reasoning Trace reviewPhase gate becomes rubber-stamp; governance fails silently
Not using Dissent ProtocolInformal rejections have no audit trail; override frequency data corrupted
Force-bypassing CI routinelySafety gates become meaningless; Security Agent confidence data misleading
Auto-approving low-confidence agentsIndividual threshold passes mask CCS failures
Deleting dissent log entriesImmutability violated; Knowledge Agent learning corrupted
Enabling Self-Healing PRs without safety gatesAgent-initiated changes to production code without human oversight

Next

  • Security — the security enforcement layer built on top of the governance model
  • State Management — how the State Manifest implements atomic governance transitions