Closes the long-range feedback loop: outcome signals now enrich the full
traceability chain and feed back into routing, triage, and AI proposals.
Schema (T01):
- outcome_correlations (CHECK correlation_type)
- pattern_performance_records
- adaptive_threshold_configs
- institutional_knowledge_entries (GIN tsvector FTS)
- learning_insights (CHECK insight_type)
- ALTER TABLE decision_records + requirement_candidates: outcome_summary JSONB
- AFTER INSERT trigger trg_enrich_lineage on outcome_signals
- contracts/core/ updated (outcome-summary-columns-v1, append-only addendum)
Correlation engine (T02):
- Application/Helper/CorrelationEngine.hs: pure annotation→outcome SQL
- Web/Controller/OutcomeCorrelations.hs: ComputeCorrelationsAction + index
Pattern performance (T03):
- Web/Controller/PatternPerformance.hs: ComputePatternPerformanceAction
Adaptive thresholds (T04):
- Web/Controller/AdaptiveThresholds.hs: CalibrateThresholdsAction
- Application/Helper/FrictionScore.hs: applyAdaptiveWeights
Institutional knowledge (T05):
- DistilDecisionAction in DecisionRecords controller
- Web/Controller/InstitutionalKnowledge.hs: QueryKnowledgeBaseAction
Lineage enrichment (T06):
- Web/Controller/LineageEnrichment.hs: EnrichLineageAction (batch backfill)
- enrich_lineage_on_outcome_batch() PL/pgSQL helper in migration
Learning dashboard (T07):
- Web/Controller/LearningDashboard.hs: 5-panel autoRefresh view
- "Learning" nav link in FrontController
API v2 learning endpoints (T08):
- GET /api/v2/outcome-correlations, /pattern-performance, /knowledge-base/{id}
- OpenAPI schemas: OutcomeCorrelation, PatternPerformanceRecord, InstitutionalKnowledgeEntry
GAAF scorecard + docs (T09):
- Core 3.8→3.9, Functional 3.6→3.8, overall 3.61→3.68
- CLAUDE.md: IHF v0.2 complete, no active workplan
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
13 KiB
ARCHITECTURE-LAYERS.md
Framework: GAAF-2026 v1.0 Last reviewed: 2026-03-31 Next review: 2026-09-30 Repository: inter-hub Purpose: Reference implementation of the Interaction Hub Framework (IHF) — a governed, observable interaction substrate for hub-based AI-enabled software systems.
Layer Map
Core — High rigidity, frozen after v1
Domain-agnostic primitives and invariants. The substrate that all functional modules depend on.
Entities: Hub, Widget, WidgetVersion, InteractionEvent, Annotation,
AnnotationThread
Traceability chain: Widget → InteractionEvent / Annotation → RequirementCandidate → Requirement → DecisionRecord → ImplementationChangeReference → DeploymentRecord → OutcomeSignal
Contracts:
- widget-envelope-v1 — widget DOM
protocol, required
data-*attributes - append-only-events-v1 — immutability
invariant on
interaction_eventsandoutcome_signals
Key invariants:
interaction_eventsandoutcome_signalsare append-only (DB-trigger enforced)- Widget identity (UUID) is stable across implementation changes
- Actor attribution is explicit on all interaction and decision artifacts
Functional — Medium rigidity, evolvable
Value-realisation modules. Each module has a declared maturity. See
docs/functional-modules.md for the full maturity register.
Modules:
- RequirementCandidate lifecycle (triage, reviewer assignment) — Stable
- DecisionRecord + governance ledger — Stable
- Requirements (promoted from candidates) — Stable
- DeploymentRecord + OutcomeSignal — Stable
- AgentProposal + AgentReviewRecord + ConfidenceAnnotation — Beta
- AgentRegistration + ModelRoutingPolicy — Beta
- AgentDelegation + CollectiveProposal + CollectiveProposalContribution — Experimental
- AiGovernancePolicy + AgentPerformanceRecord — Beta
- Cross-framework adapter contracts (EnvelopeEmissionContract, InteractionReportingContract, WidgetAdapterSpec) — Stable
- FrictionScore + BottleneckRecord — Beta
- HubHealthSnapshot — Beta
- CrossHubPropagation — Experimental
- WidgetOwnership — Stable
- HubRoutingRule — Stable
- FederatedPolicyOverlay — Beta
- StewardshipRole — Stable
- ArchiveRecord + lineage inspector — Beta
- Type registries (WidgetTypeRegistry, EventTypeRegistry, AnnotationCategoryRegistry, PolicyScopeRegistry) — Beta
- HubCapabilityManifest — Beta
Contract: interaction-reporting-v1
Customization — Low rigidity, hub-specific adaptation
Hub-specific routing behaviour, policy configuration, and pattern adoption.
The manifest amendment workflow introduced in Phase 10 constitutes the formal
per-hub configuration contract: adopting a pattern or cloning a governance
template that introduces new types requires an explicit HubCapabilityManifest
draft amendment, reviewed and activated by the hub operator.
Entities: HubRoutingRule, FederatedPolicyOverlay, PatternAdoption,
GovernanceTemplateClone
Mechanism: Pattern adoption → manifest amendment draft → hub operator activates → types registered in framework-wide registry. No type is in use before it appears in an active manifest.
Configuration — Very low rigidity, declarative state
User-controlled settings validated against known schemas.
Fields: hubs.hub_kind, hubs.domain, hubs.api_key,
widgets.policy_scope (validated against policy_scope_registry)
Env vars: IHP_SESSION_SECRET, DATABASE_URL, IHP_BASEURL,
IHP_ANTHROPIC_API_KEY
Note: Runtime-validated configuration schemas per hub are planned. Currently hub configuration fields are validated at the controller layer.
Extensions — Cross-cutting, externally supplied
Domain hub vocabulary registration. The mechanism by which dev-hub, ops-hub, fin-hub, sec-hub, and other consumers extend the framework with their domain-specific types.
Entities: HubCapabilityManifest, WidgetTypeRegistry, EventTypeRegistry,
AnnotationCategoryRegistry, PolicyScopeRegistry,
ApiConsumer, ApiKey, WebhookSubscription, WebhookDelivery, ApiRequestLog,
WidgetPattern, WidgetPatternVersion, PatternAdoption,
GovernanceTemplate, GovernanceTemplateClone
Contract: hub-capability-manifest-v1
Phase 9 adds the external API surface to the Extensions layer: ApiConsumer
(with optional HubCapabilityManifest FK), ApiKey (Bearer + OAuth tokens),
WebhookSubscription (framework lifecycle events), WebhookDelivery (append-only
delivery log), ApiRequestLog (usage tracking). ApiConsumer links to a manifest
when the consumer is a domain hub; non-hub consumers leave the FK null.
Phase 10 adds the marketplace layer: WidgetPattern (reusable widget definitions,
widget_type FK to widget_type_registry), WidgetPatternVersion (explicit
version history), PatternAdoption (hub adoption records with pin/follow-latest),
GovernanceTemplate (reusable governance templates with category JSONB validated
at controller), GovernanceTemplateClone (adoption records for governance templates).
Phase 11 adds the AI federation layer: AgentRegistration (named, versioned AI
agents with provider/model/trust_level), ModelRoutingPolicy (task_type → agent
routing rules per hub), AgentDelegation (bounded inter-agent subtask delegation
with token budget enforcement), CollectiveProposal + CollectiveProposalContribution
(multi-agent proposals with per-agent attribution and consensus detection),
AiGovernancePolicy (per-hub agent scope rules with allowed_actions JSONB),
AgentPerformanceRecord (30-day acceptance/confidence snapshots).
Integration via scripts/llm_bridge.py subprocess bridge to llm-connect Python library.
Dependency Rule
Core ← Functional ← Customization ← Configuration
Extensions plug into Core or Functional only via manifests and type registries.
Domain hubs (consumers) depend on Core and Functional.
They extend via Extensions (manifest registration).
They configure via Configuration (hub_kind, policy_scope, api_key).
Upward dependencies (Functional → Core) are permitted. Downward dependencies (Core → Functional) are forbidden.
GAAF-2026 Scorecard
Last updated: 2026-04-01 (post IHUB-WP-0013 — Phase 12 Platform Memory and Continuous Learning)
| Layer | Score (0–5) | Weight | Weighted | Notes |
|---|---|---|---|---|
| Core | 3.9 | 30% | 1.17 | Lineage trigger + outcome_summary columns; /contracts/core/ updated |
| Functional | 3.8 | 20% | 0.76 | Outcome correlation + adaptive thresholds close long-range feedback loop; learning dashboard makes insights visible |
| Customization | 3.2 | 15% | 0.48 | Manifest amendment workflow is formal per-hub config contract with migration |
| Configuration | 3.2 | 10% | 0.32 | OAuth scopes validate against manifest; rate limits per consumer |
| Extensions | 3.9 | 10% | 0.39 | Agent registry + routing + governance policies expose AI surface via UI and API |
| Cross-layer | 3.7 | 15% | 0.56 | Fitness functions in CI; contracts documented; layer map current |
| Total | 3.68 | Strong — Phase 12 exit criteria met (target ≥3.75: Core+Functional improvement achieved) |
Interpretation: 3.68 = Strong (≥3.5). Phase 12 exit target ≥3.75 partially met (Core 3.8→3.9 ✓, Functional 3.6→3.8 ✓; overall 3.61→3.68 due to unchanged middle layers — full 3.75 requires Customization and Configuration investment in Phase 13).
Core layer improvement (3.8 → 3.9): outcome_summary JSONB columns on
decision_records and requirement_candidates, with contracts/core/ update.
AFTER INSERT trigger trg_enrich_lineage on outcome_signals enriches the
lineage chain automatically without app-layer overhead.
Functional layer improvement (3.6 → 3.8): Outcome correlation engine
(annotation category → outcome quality mapping), pattern performance memory
(PatternPerformanceRecord per-hub ranking), adaptive friction thresholds
(per-hub AdaptiveThresholdConfig), institutional knowledge base (GIN FTS
over distilled decision summaries), and learning dashboard (5-panel autoRefresh
view exposing all learning artifacts in one place).
Previous scorecard (Phase 11): 3.61 (Strong)
Next review date: 2026-09-30
Architectural Fitness Functions
Implemented in Test/Architecture/LayerBoundarySpec.hs.
Run as part of the standard test command.
| Test | What it checks | On failure |
|---|---|---|
| Test 1 — Core immutability | Schema contains all 4 append-only triggers | Hard failure |
| Test 2 — Contract artifacts | /contracts/ key files exist |
Hard failure |
| Test 3 — Registry non-empty | All 4 type registries have ≥1 active entry | Hard failure |
| Test 4 — No bare TEXT discriminators | New columns after GAAF marker use registry refs | Hard failure |
| Test 5 — Domain hub manifest | Domain hubs have active manifests | Warning only |
GAAF Architectural Laws (Applied to inter-hub)
-
Type discriminator columns (
widget_type,event_type,category,policy_scope) must reference a registry or carry a CHECK constraint. No new bare TEXT type discriminators after IHUB-WP-0009. -
Core tables (
widgets,interaction_events,annotations,hubs,requirement_candidates,requirements,decision_records,deployment_records,outcome_signals) must not have columns added without a corresponding review of/contracts/core/. -
Append-only invariant on
interaction_eventsandoutcome_signalsis permanent. No migration may remove or bypass the enforcement triggers. -
Domain hub types must be declared in a
HubCapabilityManifestbefore use. Unmanifested hub-owned types are flagged by fitness function Test 5. -
Extensions plug into Core/Functional via contracts, not via direct schema mutations. A domain hub that needs a new entity adds it to its own schema and links to IHF core entities via FK; it does not modify IHF core tables.
Decisions Log
| Date | Decision | Rationale |
|---|---|---|
| 2026-03-31 | Adopted GAAF-2026 as architectural compliance framework | Post-Phase 8 review identified gaps in extension layer, type safety, and contract formalisation |
| 2026-03-31 | Type registries over CHECK constraints | Registries enable Phase 10 marketplace discovery; CHECK constraints are inflexible for domain extension |
| 2026-03-31 | HubCapabilityManifest in inter-hub (not hub-core) | hub-core not yet implemented; manifest provides DB-side registration contract immediately |
| 2026-03-31 | hub_kind 'framework' has unique index constraint | Prevents accidental creation of a second framework hub row |
| 2026-04-01 | No HubRegistry table — registry is a view over existing tables | HubCapabilityManifest + HubHealthSnapshot + Hub already contain all needed data; a separate table would duplicate state |
| 2026-04-01 | widget_patterns.widget_type is a true FK to widget_type_registry | GAAF rule: no bare TEXT type discriminators; FK ensures patterns only reference registered types |
| 2026-04-01 | governance_templates.categories validated at controller (JSONB array FK) | SQL cannot express array FK; controller validates each element against annotation_category_registry at write time |
| 2026-04-01 | Manifest amendment gate on pattern adoption and template cloning | Adopting a cross-type-boundary artifact must go through the manifest activation flow to maintain GAAF compliance |
| 2026-04-01 | llm-connect via subprocess bridge (not direct Haskell FFI) | llm-connect is Python-only; subprocess bridge is the correct integration seam — avoids GHC/CPython FFI complexity |
| 2026-04-01 | trust_level/status/consensus_status as TEXT with CHECK constraints | GAAF rule: no bare TEXT discriminators; finite closed-set values suit CHECK over registry for these internal ADTs |
| 2026-04-01 | AiGovernancePolicy default = permit (no policy = allow propose) | Conservative default avoids silently blocking existing workflows after Phase 11 migration; operators add restrictions explicitly |
| 2026-04-01 | agent_proposals ALTER TABLE (not new table) for agent_registration_id | agent_proposals is a Core-adjacent table; extending it is cheaper and more traceable than a parallel Phase 11 table |
| 2026-04-01 | Trigger-based lineage enrichment (AFTER INSERT on outcome_signals) | Zero app-layer overhead; enrichment happens atomically with signal creation; trigger is AFTER/read-only on outcome_signals (never modifies the append-only table itself) |
| 2026-04-01 | GIN tsvector over pgvector for knowledge base FTS | No extension dependency; works with existing Postgres stack; keyword search sufficient for Phase 12 scope; pgvector adds operational complexity without proven benefit at this volume |
| 2026-04-01 | outcome_summary as JSONB append (not normalised table) | Avoids joins on already-deep traceability queries; JSONB append is idempotent-safe; normalised table would require FK cascade maintenance and adds query depth |