01Critical risk
Human governance
Owns intent, risk, exceptions, promotion, and material acceptance.
- Owner
- Named human authority
- Contract
- Decision record
- Proof
- Identity, evidence reviewed, reason, conditions
The complete governed system
Trace the factory from intent to evidence.
Supporting implementation views connect the six architectural areas across authority, runtime, risk, failure, recovery, and proof. They remain lenses on the canonical eight-stage value stream, not competing models.
- 01Intent
- 02Plan
- 03Define Agent
- 04Execute through Harness
- 05Apply Skills
- 06Evaluate
- 07Improve
- 08Deliver Software
Responsibility boundaries
Separate decisions before separating deployments.
Planes define what a component may decide and which records it may mutate. A modular V1 can keep several in one codebase.
Lens
02Critical risk
Control plane
Owns domain state, admission, policy, dispatch eligibility, and reconciliation.
- Owner
- Control owner
- Contract
- Commands and state transitions
- Proof
- Policy decision, expected version, audit event
03Critical risk
Execution plane
Runs frozen manifests in isolated environments and reports structured results.
- Owner
- Runtime owner
- Contract
- Attempt contract
- Proof
- Tool calls, artifacts, side effects, completion
04High risk
Knowledge plane
Registers sources, retrieves eligible facts, and freezes attributable context.
- Owner
- Knowledge owner
- Contract
- Context package
- Proof
- Source, permission, selection, freshness, lineage
05Critical risk
Quality plane
Independently evaluates exact candidates and converts eligible results into proof.
- Owner
- Quality owner
- Contract
- Quality contract
- Proof
- Evaluator independence, provenance, contradiction
06Critical risk
Delivery and outcome
Binds approvals to artifacts, controls rollout, and observes production reality.
- Owner
- Release and outcome owners
- Contract
- Release and outcome contracts
- Proof
- Artifact, rollout, rollback, observation window
Deterministic educational walkthrough
Follow one change through the factory.
These are fixed teaching scenarios—not live agents or production evidence. Select a scenario and stage to inspect the record, owner, evidence, and stop conditions.
07 / 11 · failed
Independent verification
The exact candidate is challenged outside the producing session.
Completion is not acceptance.
- Completed
- Execution produced a candidate and complete run record.
- Failed
- The behavioral quality contract scored 8/10 against a required 9/10.
- Decision owner
- Human / policy authority—not the producing agent or harness.
- Delivery
- Blocked because the evidence package is ineligible for acceptance.
- Recovery
- Preserve the failed evidence, revise the candidate or specification, create a new attempt, and verify again.
Independent checks
- Unit testsPassed
- Type checkPassed
- SecurityPassed
- Behavioral evaluation8/10 · required threshold 9/10
- Scope validationPassed
The producer cannot certify its own material work. These checks run in a separate quality context.
Educational example
Candidate evidence
- Candidate
- cand-8f91a2c
- Produced by
- SearchAgent config v3.2
- Execution
- FR-0249
- Verified by
- IV-071
- Provenance
- Complete
- Currentness
- Current
- Acceptance eligibility
- Ineligible
- Record created
- Quality certificate candidate
- Accountable owner
- Independent verification service
- Entry condition
- The candidate identity, run record, and versioned quality contract are complete.
- Required evidence
- Test results, security findings, behavioral evaluation, scope validation, and counterevidence.
- What stops it
- Any required check fails, evidence is ineligible, or the candidate cannot be reproduced.
Who determines whether the completed candidate satisfies the required acceptance evidence?
Concept boundaries
Understand the system by contrast.
Choose a pair, then increase the depth. Each comparison states ownership, limits, and the authoritative chapters without creating a second source of truth.
AI Software Factory→Agent Factory ↔ AI Software Factory
Core distinction
The Agent Factory supplies governed capabilities. The AI Software Factory composes them into governed delivery outcomes.
How to use the map
Start broad. End at the contract.
- 01Choose a view
Use the primary value stream for sequence, planes for ownership, components for runtime, or monitoring for failure response.
- 02Apply a lens
Focus on authority, runtime, assurance, or operations without losing the complete system context.
- 03Inspect the card
Every item names responsibility, owner, contract, risk, and required proof in text.
- 04Walk the run
Use the deterministic scenario to see which record, owner, evidence, and stop condition governs every transition.
- 05Compare the boundary
Use the contrast guide to separate adjacent concepts before opening their authoritative chapters.
- 06Open the source
The linked Markdown chapter is canonical and carries failure, recovery, versioning, exercises, and evidence boundaries.
Architecture is not proof
Review the design. Exercise the controls. Keep the claim scoped.
The map is review ready. It does not assert that every component, control, SLO, or recovery path is operationally proven.
Inspect coverage and maturity