Federated from workspace · PL-004 · zaixos-intelligence-platform/LEARNING_MODEL.md Do not edit canonical truth here — update the source repo, then re-run npm run docs:sync.

PL-004 Learning Model

Platform: ZIP · Principle: Learning is derived — never overwrites product truth

Purpose

Define how ZIP learns from outcomes — improving signal weights, detection thresholds, reasoning confidence, and intelligence patterns without mutating authoritative business records or auto-executing retries.

Learning inputs

Input	Source	Description
`ExecutionOutcomeEvent`	Product	Actual KPIs after handoff execution
`ApprovalDeniedEvent`	ZIP + approver	Rejection reason and context
`ApprovalGrantedEvent`	ZIP	Positive human validation signal
`PredictionError`	ZIP internal	Planned vs actual comparison
`ValidationFailureEvent`	ZIP	Plan blocked by compliance/feasibility
`PL-003 eval signals`	PL-003 hooks	Agent and prompt performance metrics

Learning pipeline

OutcomeEvent (product)
    → OutcomeNormalizer
    → Link to recommendation bundle + plan revision + reasoning artifact
    → PredictionErrorAnalysis
    → FeedbackRouter
         ├── HeuristicWeightUpdate (ZIP signal/detection weights)
         ├── ConfidenceCalibration (ZIP reasoning runtime)
         ├── PromptPerformanceSignal → PL-003 eval registry
         ├── AgentToolSuccessSignal → PL-003 agent registry
         └── ProductLearningPort (optional — product-owned ML)
    → LearningRecord (append-only, tenant-scoped)
    → KnowledgeEvolutionEngine.evaluate()

Forbidden: Auto-modifying product configuration. Auto-retrying rejected recommendations without new approval cycle.

Learning record schema

Field	Description
`learningId`	UUID
`tenantId`	Platform tenant
`bundleId`	Source recommendation
`planRevision`	Plan version executed
`outcomeType`	`success` · `partial` · `failure` · `rejected` · `expired`
`predictedMetrics`	Map from plan estimate
`actualMetrics`	Map from product outcome event
`predictionError`	Computed deltas
`feedbackSource`	`outcome` · `approval` · `validation`
`recordedAt`	Timestamp

Knowledge evolution

Intelligence patterns (not domain ontology) evolve from learning records:

Action	Condition
Promote pattern	Success rate ≥ threshold across N cycles; zero policy violations
Demote pattern	False positive rate ≥ threshold; repeated approver rejection
Retire pattern	Obsolete objective or product manifest removal
Version pattern	Promotion creates new version key — immutable prior versions

Patterns stored in ZIP knowledge store. Playbook content remains product extension; ZIP stores effectiveness metadata only.

Cross-tenant learning

Mode	Default	Requirement
Tenant-isolated	Yes	No cross-tenant data in learning
Anonymized aggregate	Opt-in enterprise	Executive Class A + DPA
Federated model training	Future	Not in scope until Phase 7+

Feedback to PL-003

ZIP emits structured eval signals to PL-003 — never modifies PL-003 runtime directly:

Signal	PL-003 consumer
Prompt outcome score	Eval framework
Agent tool success rate	Agent registry weighting
Generation cost efficiency	Job orchestration hints

Cross references

Version history

Version	Date	Change
1.0	2026-07-01	Initial learning model — bootstrap

PL-004 Learning Model ​

Purpose ​

Learning inputs ​

Learning pipeline ​

Learning record schema ​

Knowledge evolution ​

Cross-tenant learning ​

Feedback to PL-003 ​

Cross references ​

Version history ​