Skip to content

Ontology-Grounded ARB Delivery Plan

Use this plan to sequence the smallest changes that turn the current manifest-driven review into an ontology-grounded, agent-owned closed loop. Each work package has an executable exit condition and keeps authority-bearing changes behind existing FDAI gates.

Starting boundary: The structural checker, generic workflow, Process projection, read-only Console view, exact snapshot graph receipt, and injected production evidence attestation exist. They do not yet prove an autonomous 15-agent ARB path or an immutable decision authority chain.

Rollout boundary: Complete the observation-mode vertical slice before adding an authority-bearing decision or execution path.

Work packageOutcomeDepends on
ARB-1 Ontology truthFresh operating intent, topology, and exact change revision are queryableExisting ontology platform
ARB-2 Agent evidence loopRelevant agents publish and Forseti joins verified evidenceARB-1
ARB-3 Decision authorityImmutable decision receipt binds case, evidence, conditions, approvalsARB-2
ARB-4 Effect closureResulting actions close through independent observation and recoveryARB-3
ARB-5 Learning and rolloutRepeated outcomes produce governed candidates and measured promotion evidenceARB-4

Deliver the operating facts that every later decision consumes:

  • project deployment-supplied ServiceObjective, RecoveryObjective, CostObjective, ArchitectureConstraint, Ownership, and ChangeWindow instances;
  • issue authenticated graph revision, completeness, freshness, purpose, and cutoff receipts;
  • bind one normalized planned Change to its exact desired-state digest and plan receipt;
  • keep config/architecture-review.yaml as a review profile, not mutable runtime decision state;
  • model review status as evidence, evaluation, authority, execution, and learning axes.

Exit: One exact planned change produces a time-consistent context snapshot on a pinned ontology release. The assessment carries a typed graph evidence receipt, and any accepted critical or high blocker keeps a current risk or exception record. Missing, stale, mixed-release, unverified, or truncated context still lowers authority and has a typed reason.

Wire the existing ownership boundaries rather than adding an ARB agent:

  • Huginn publishes the immutable Change; Muninn retains the revision and context;
  • select Njord, Freyr, Loki, and Heimdall from active objectives and dependency relationships;
  • compose OperationalEvidenceBundle and OntologyScenarioBranch before judgment;
  • replace caller-supplied workflow branch outcomes with agent-owned topic evidence;
  • let Forseti own the deadline-bound join and produce DecisionCase plus ImpactEnvelope;
  • send only eligible objective conflicts to Odin;
  • attempt bounded evidence reacquisition and deterministic reevaluation before Var escalation.

Exit: A complete low-risk fixture reaches an automatic conformant recommendation in observation mode. Stale, conflicting, missing, duplicate, reordered, or late evidence reaches an explicit hold without mutation or silent loss.

Replace mutable status assertions with a replayable authority chain:

  • version the Decision contract with case, context, evidence, graph, catalog, and condition digests. The reusable content-addressed receipt and additive Decision 1.1.0 shape are implemented;
  • bind approval receipt identities, approver roles, quorum, no-self-approval, and expiry;
  • require a complete risk or exception contract for every accepted critical or high blocker;
  • reuse the injected provider verification of evidence bodies, digests, scope, revisions, freshness, and approver authorization;
  • derive production readiness and ReviewCase projection from the final receipt;
  • represent an accepted constraint or production transition as a separate governance ActionType.

Exit: Changing one evidence item, approval, condition, target revision, or graph revision invalidates the prior decision identity. No manifest edit or workflow context value can create production authority.

Keep ARB control-only while connecting approved work to the ordinary action pipeline:

  • route any resource change back through ActionType promotion, policy, risk, Var, Thor, and Saga;
  • carry the selected option, impact envelope, stop condition, rollback, lock, idempotency, and expected effects unchanged;
  • let Heimdall close each expected effect from an authoritative observation;
  • leave missing or incomplete observation pending rather than reporting success;
  • let Vidar recover within the approved envelope when a failure or stop condition occurs;
  • reopen or expire the review when a condition or protected objective is violated.

Exit: One accepted observation mode case links Change -> DecisionCase -> ImpactEnvelope -> ActionRun -> ObservedOutcome, and a failure fixture reaches recovery without letting the actor score its own effect.

Use reviewed outcomes to reduce future human touchpoints without relaxing authority:

  • let Muninn retain exact case history and outcome cohorts;
  • let Norns propose inert RuleCandidate records only from bounded, consent-eligible evidence;
  • let Mimir revalidate and promote through reviewed catalog-as-code;
  • measure automatic conformance, conditional closure, evidence reacquisition, human approvals, holds, policy escapes, rollbacks, and settlement completeness;
  • promote only the bounded review capability whose observation cohort meets its gate;
  • demote on regression without changing the underlying ActionType authority.

Exit: A pinned observation cohort supports an independently reviewed promotion decision with zero policy escapes and complete decision/effect lineage.

Implement this path before extending the Console or production gate:

Huginn Change
-> Muninn context snapshot
-> OperationalEvidenceBundle
-> ontology scenario branch and diff
-> Forseti DecisionCase and ImpactEnvelope
-> observation-mode recommendation
-> Saga audit
-> derived ReviewCase projection

The slice is complete when:

  • the same change revision creates one immutable case;
  • a fresh, complete graph can produce an automatic conformance recommendation;
  • stale, truncated, conflicting, or target-mismatched evidence produces a hold;
  • protected-objective violations are removed before Odin arbitration;
  • evidence updates resume the same Process and reevaluate the same logical case;
  • no step can approve, mutate, promote, or execute outside its owner boundary.
BoundaryFocused proof
OntologyExact release, path direction, temporal cutoff, freshness, completeness, conflict, truncation
TopicsSingle writer, duplicate, reorder, retry, backpressure, dead-letter, restart, replay
JudgmentHard constraints before arbitration, no-action baseline, bounded options, deterministic identity
ApprovalRole, quorum, no-self-approval, timeout, receipt replay resistance
ExecutionARB control-only; separate promoted ActionType; seven safeguards
OutcomeIndependent target observation, every expected effect closed or explicitly unscorable
ProjectionRead model derived from authoritative lineage; expired or removed evidence reconciled
LocalizationEnglish/Korean structure, links, translation SHA, terminology, punctuation
  • Do not add a sixteenth ARB agent.
  • Do not let the workflow call agents directly or own judgment.
  • Do not let the ontology graph approve or execute.
  • Do not make Bragi or Odin a universal controller.
  • Do not grant automatic authority to architecture baselines, exceptions, irreversible changes, or production promotion.
  • Do not start with UI polish or enforce routing before the observation-mode vertical slice passes.
To learn aboutRead
ARB entry point and decision boundaryArchitecture Review Board Packet
Target ontology and agent flowOntology-Grounded Agent Loop
Evidence and authority contractEvidence and Authority
Operational planning foundationsOperational Planning
Delivery status and remaining workImplementation ledger