Ontology-Grounded ARB Agent Loop

Use this design to understand how FDAI’s operating ontology and fixed 15-agent pantheon review a planned change. It keeps meaning, authority, execution, observation, and learning in separate accountable boundaries.

Scope: This document owns the ARB decision loop. The evidence and authority contract owns evidence bindings, approval, risks, and production exit. The delivery plan sequences implementation.

Authority boundary: The ontology validates identity, relationships, time, evidence, and constraints. It never judges, approves, executes, or raises autonomy.

Design at a glance

A normalized Change starts one bounded review. Muninn freezes the current operating context, relevant specialists publish their owned evidence, and Forseti joins the required inputs into an immutable DecisionCase and ImpactEnvelope. Odin resolves only eligible cross-objective conflicts. Var requests human approval only when policy or residual risk requires it.

Authoritative state model

The ARB read model derives from agent-owned records. It does not use mutable checklist status as a second source of truth.

Record	Authority	ARB use
Change	Huginn	Exact proposed revision, target, actor, desired-state digest, and correlation
OperationalContextSnapshot	Muninn-supported projection	Time-consistent topology, objectives, ownership, freshness, and conflicts
OperationalEvidenceBundle	Verified provider receipts	Immutable ontology, state, catalog, and document evidence lanes
ArchitectureConstraint and Policy	Mimir	Effective constraints and deterministic evaluation rules
Domain observations	Heimdall, Njord, Freyr, Loki	Reliability, security, cost, capacity, performance, and recovery evidence
DecisionCase and ImpactEnvelope	Forseti	Bounded options, no-action baseline, protected objectives, uncertainty, and impact limits
ArbitrationDecision	Odin	Selection among constitutionally eligible soft-objective tradeoffs
Approval	Var	Current, scoped approval when policy requires a person
Decision and audit chain	Saga projection over owned events	Replayable decision, conditions, authority basis, and terminal status
ActionRun	Thor	Separate execution after the ARB decision and normal action gates
ObservedOutcome	Heimdall	Independent effect closure against expected effects

ReviewCase and ReviewCheck remain read models. They summarize these records for the Process and Console, but they cannot create a decision or mark evidence ready by themselves.

The 15-agent responsibility model

Every agent has a possible ARB responsibility, but a review activates only the agents relevant to the target’s ontology relationships and active objectives.

Agent	ARB responsibility	Activation condition
Odin	Arbitrate eligible cross-objective conflicts	Two or more eligible options conflict after hard constraints
Thor	Execute a separately promoted action	A final action decision passes all action gates
Forseti	Join evidence, evaluate constraints, create the case and decision recommendation	Every planned change review
Huginn	Normalize, deduplicate, correlate, and publish the exact Change revision	Planned-change ingress
Heimdall	Validate graph and telemetry health, then independently close effects	Reliability, security, drift, forecast, or terminal outcome evidence is required
Vidar	Validate recovery readiness and own recovery after failure	The proposed change needs rollback or recovery evidence
Var	Enforce role, quorum, timeout, and no-self-approval	Policy-mandated or residual human approval
Bragi	Translate an operator request and explain grounded results	Conversational start or explanation
Saga	Persist audit intent, decision lineage, conditions, and terminal closure	Every review transition
Mimir	Supply active constraints and policies; accountable owner approved catalog changes	Constraint evaluation or accepted constraint change
Muninn	Retain change revisions and materialize bounded context	Every planned change review
Norns	Propose inert rules from repeated reviewed outcomes	A bounded outcome cohort shows a reusable pattern
Njord	Evaluate cost objectives and budget effects	Cost-sensitive services or resources are in scope
Freyr	Evaluate capacity and performance effects	Capacity or performance objectives are in scope
Loki	Propose resilience experiments without self-authorizing them	Existing evidence cannot establish recovery behavior safely

Evidence fan-out and join

Agents collaborate through typed publish and subscribe only. The review coordinator is mechanical: it records deadlines and required evidence keys, but Forseti owns the decision to continue, hold, or escalate.

1. Huginn publishes one immutable Change and preserves one correlation ID plus the requested ontology release digest.
2. Muninn materializes a secured snapshot for the exact graph release and evidence cutoff. The snapshot requires verified links before it can produce authenticated graph evidence.
3. Relevant specialists publish only their owned object topics with the same case reference.
4. Forseti converts that exact snapshot into the planned-change graph receipt, compares its ontology release with the requested digest, and joins the policy-required evidence set under a deadline. Missing, stale, mixed-release, incomplete, unauthenticated, conflicting, synthetic, future, or truncated evidence lowers the authority ceiling. The receipt projector re-derives link evidence quality from typed metadata instead of trusting the snapshot’s conflict summary.
5. Before asking a person, FDAI attempts bounded reacquisition, an alternate authoritative source, deterministic reevaluation, or a smaller safe option.

Slow or failed subscribers do not block unrelated work. The owner of the join records an explicit missing or late branch instead of interpreting silence as success.

The current runtime binding keeps this boundary observation-only. Core selects Forseti’s review loop only when composition supplies a complete authenticated context/evidence provider pair and an ontology instance store. The deployment StateStore atomically retains the Change idempotency identity and projection status across replicas; a restart after an incomplete projection holds instead of re-evaluating the same identity against newer evidence. A dedicated non-agent consumer group reads owned object.change, object.state-snapshot, specialist, object.verdict, and object.audit-entry topics, replays the immutable ArchitectureReviewObservationTrace, and retains terminal observer gave_up/halted states as degradation evidence. It does not call agents directly, mutate runtime state, or grant execution authority. When an ARB stage has no distinct owned record yet, the trace stays on hold instead of synthesizing success.

The review accepts only planned intent. It hashes the complete Change identity, uses one absolute deadline, and serializes only identical idempotency keys, so unrelated reviews can continue in parallel. Context and evidence must name the same ontology and catalog releases. An unavailable re-read preserves existing checks instead of removing them, and evidence artifact identities bind the bundle and item content. The resulting Change, DecisionCase, ImpactEnvelope, and ReviewCase records retain typed lineage links and remain explicitly observation-only. An architecture_review observation decision is audit-only: Odin excludes it from action-portfolio counts, Thor never dispatches it, and Saga retains the terminal evidence. If the injected state store cannot retain projection status, the loop uses a bounded process-local oldest-first cache; the fallback cannot grow without limit or create authority. Process lineage comes only from an existing Change process_ref that resolves to a typed Process. The projector emits change_instantiates_process; it never invents a review-to-process shortcut, and a validation conflict leaves lineage unavailable with a bounded warning. The base graph must be complete and its source generation must match the authenticated context before an envelope can be conformant. Projection status is recorded with the observation, so a stored result retries a failed or unavailable production projection without re-running evidence collection. Existing Huginn Change records are preserved; a new Change is projected only when the exact normalized fields are present. Scenario branch identifiers use an ASCII-safe bounded prefix plus a content digest, including for UUID-shaped or otherwise long Change identifiers. Deadline fallbacks compare the full Change identity before accepting stored state and use bounded persistence or a durable outbox. Normalized optional provenance fields, including process_ref, remain available for exact Change and Process lineage.

Autonomous review levels

Autonomy applies to review work before it applies to changes. A machine-ready result is not the same as production authorization.

Level	Conditions	Result
Automatic conformance	Exact change, fresh complete graph, active objectives and constraints, verified evidence, bounded impact, no conflict	Record a conformant recommendation under bounded standing authority
Automatic conditional review	Deterministic conditions can close all known gaps	Record conditions and reevaluate automatically when evidence changes
Human approval	Policy requires a person, an exception is requested, impact is high, or residual ambiguity remains	Var opens a scoped approval with quorum and expiry
Hold or reject	Evidence is incomplete, stale, conflicting, target-mismatched, or a hard constraint is violated	Apply no change; record the reason and next evidence action

Architecture baseline changes, production promotion, irreversible changes, and exceptions do not gain automatic authority from a green graph result.

Review projection

The Console should present separate axes instead of one green or red status:

• Evidence state: complete, incomplete, stale, conflicting, or unavailable.
• Evaluation state: conformant, conditional, violated, or held.
• Authority state: observation only, standing authorization, approval required, or approved.
• Execution state: not requested, pending, running, verified, failed, or recovered.
• Learning state: ineligible, cohort collecting, candidate proposed, or catalog reviewed.

Every displayed state links to the exact Change, context digest, evidence bundle digest, DecisionCase, conditions, approvals, and audit records that produced it.

Current gaps

Gap	Required correction
Planned-change assessment now falls back to ChangeGraphEvidenceReceipt.unavailable()	Wire the receipt from the exact OperationalContextSnapshot so planned-change review carries an authenticated graph revision, release match, freshness, and truncation state
Workflow parallel branches journal caller-supplied outcomes	Drive branches from agent-owned typed evidence and record explicit deadlines
ARB projection reads manifest status directly	Derive ReviewCase and ReviewCheck from decision lineage and verified evidence
The provider-neutral bundle/scenario/case/envelope path is composed, but runtime activation still depends on deployment-supplied authenticated context and evidence sources	Bind those exact production sources and retain one governed observation receipt
Operating intent instances are not proven end to end	Project service, recovery, cost, ownership, constraint, and change-window instances with freshness
The ARB trace runtime binding currently covers only the emitted owned topics	Extend owned records for context, evidence bundle, impact envelope, recommendation, and the derived ReviewCase projection so the observation-only runtime trace can close the full first vertical slice without synthesizing missing stages

Related docs

To learn about	Read
ARB entry point and decision boundary	Architecture Review Board Packet
Evidence, approvals, and production exit	Evidence and Authority
Dependency-ordered implementation	Delivery Plan
Operating ontology ownership	Operating Ontology
Fixed agent roles	Agent Pantheon
Delivery status and remaining work	Implementation ledger