Enterprise Learning Runtime

Judgment registry	Human decisions, rationales, overrides, and accountable judgment
Control fabric	Controls linked to policies, risks, evidence, agents, and decisions
Decision memory	Historical decisions, similar cases, replay, and outcome patterns
Exception intelligence	Frequency, root cause, cost, resolution, and prevention
SME knowledge	Controller and analyst knowledge captured for governed promotion
Outcome intelligence	Recommendation quality, value, delay, reversal, and rework
Learning runtime	Feedback promoted through review, certification, pilot, and activation

Real today	Decision ledger, evidence, replay, context usage, policy/control assets
Partial today	Decision outcomes, rationale capture, exception mining
Missing hardening	Mandatory rationale, control IDs per decision, formal root-cause taxonomy
Next operating layer	Judgment Workbench and governed learning promotion gates
Message	The real asset is scalable enterprise judgment, not just agents

Live Enterprise Judgment & Learning Runtime Audit

Live runtime

Readiness

93%

Decisions

500

Rationales

493

Outcomes

491

Exceptions

494

Learning signals

498

Human Capital -> Token Capital

Human Capital -> Enterprise Learning Runtime -> Token Capital -> Enterprise Advantage

BP can replace GPT, Claude, Gemini, Bedrock, Azure OpenAI, or local models without losing expertise because the durable value is stored in ontology, policy, replay, evidence, outcomes, learning artifacts, and the enterprise memory graph.

Human judgment captured

Decision replay preserved

Outcome measured

Pattern discovered

Learning artifact proposed

Owner approves

Agent, policy, skill, or evidence ranking improves

Strategic Learning Platforms

These are the durable BP-owned capabilities above agents and copilots: learning runtime, memory graph, and enterprise evaluation/reinforcement.

Platform	Readiness	Vision	Core services	Runtime proof	Gap
Enterprise Learning Runtime Learning Intelligence Hub	99%	Convert every decision, outcome, policy override, and exception into governed institutional intelligence.	Decision Capture Service, Outcome Capture Service, Learning Artifact Generator, Pattern Mining Engine, Continuous Agent Improvement	500 decisions available; 491 outcomes captured; 498 learning signals captured	Make outcome capture and learning-artifact generation mandatory after every governed action.
Enterprise Memory Graph Memory Explorer	80%	Represent BP's operating memory as a graph of people, decisions, policies, evidence, outcomes, cases, experts, and agents.	Enterprise Memory Search, Expert Discovery, Decision Similarity, Organizational Memory, Knowledge Coverage Heatmap	500 replayable decisions; 0 knowledge assets; 13 control assets	Persist graph IDs on decisions, policies, evidence, outcomes, experts, and learning artifacts.
Enterprise Evaluation & Reinforcement Platform Agent Evaluation & Reinforcement Dashboard	99%	Evaluate agents against BP outcomes, then promote improvements through governed backtest, pilot, activation, and rollback.	Business Outcome Evaluation, Human Agreement Evaluation, Policy Compliance Evaluation, Financial Impact Evaluation, Reinforcement Queue	491 outcome-bearing decisions; 500 replay records; 10 recurring decision patterns	Add baseline snapshots, backtest windows, promotion approvals, and rollback controls for every agent improvement.

Token Capital KPIs

Token Capital measures how much BP expertise is captured as reusable, governed, model-portable intelligence.

KPI	Value	Meaning	Runtime proof
Judgment Capture Rate Target: >90%	99%	% of human or agent decisions with rationale captured for reuse.	493 of 500 decisions carry rationale-like fields.
Replay Coverage Target: 100%	100%	% of decisions that can be reconstructed from replay records.	500 of 500 decisions have replay linkage.
Policy Learning Coverage Target: >85%	100%	Coverage of policy and control assets available to become governed learning constraints.	465 policy assets and 13 control assets discovered.
Resolution Learning Rate Target: >80%	99%	% of resolved work that can become reusable patterns or lessons.	498 learning signals and 491 outcome records.
Agent Improvement Rate Target: monthly quality improvement	100%	Readiness to convert outcomes, patterns, and exceptions into agent improvements.	10 decision patterns and 10 exception patterns mined.
Token Capital Growth Target: quarterly compounding growth	99%	Composite index for institutional knowledge accumulated from judgment, replay, outcomes, policies, and patterns.	Weighted from judgment capture, replay, outcomes, learning signals, patterns, and knowledge assets.

Governance And Safety Services

This closes the main audit gap: learning cannot pollute memory or change runtime behavior unless qualification, confidence, approval, lifecycle, trust, simulation, and rollback controls pass.

Service	Readiness	Purpose	Controls	Lifecycle	Quality gate	Remaining gap
Learning Governance Service P0	99%	Qualify learning candidates before they enter memory or runtime use.	Learning Qualification Layer, Learning Confidence Engine, Human Approval Workflow, Learning Lifecycle	Draft -> Validated -> Approved -> Active -> Deprecated -> Archived	Only successful, policy-aligned, evidence-backed decisions can become reusable learning without explicit reviewer override.	Persist learning_qualification, confidence, reviewer, lifecycle_state, and expiry on every learning artifact.
Judgment Capture Service P0	99%	Capture why a human acted, not only what action was taken.	Rationale requirement, Alternatives considered, Confidence capture, Judgment provenance	Captured -> Structured -> Reviewed -> Promoted -> Retired	High-risk decisions require rationale, confidence, policy reference, evidence reference, and alternatives considered.	Make judgment_reason, confidence, and alternatives mandatory in high-risk action drawers.
Outcome Intelligence Service P0	99%	Determine whether decisions were correct, valuable, delayed, reversed, or harmful.	Outcome classification, Value attribution, False-positive/false-negative tagging, Resolution quality	Pending -> Observed -> Attested -> Reconciled -> Closed	No learning promotion without outcome state or explicit attestation gap.	Bind outcomes to source-system completion events and controller attestation.
Enterprise Memory Governance Service P0	80%	Govern temporal, contradictory, classified, and trusted memory.	Temporal Memory, Superseded-By Relationships, Memory Trust Score, Memory Sovereignty Classification	Created -> Valid From -> Superseded -> Restricted -> Retained -> Disposed	Every memory node requires valid_from, source, trust score, classification, owner, and access policy.	Add temporal validity, supersession, trust score, classification, retention, and Entra-backed access to memory nodes.
Enterprise Experimentation Service P1	99%	Prevent unsafe reinforcement with causal attribution, regression protection, simulation, and champion/challenger evaluation.	Causal Attribution, Shadow Evaluation, Champion Challenger, Counterfactual Simulation	Candidate -> Backtest -> Shadow -> Pilot -> Promoted -> Rolled Back	No agent, prompt, threshold, policy, or tool-routing improvement may activate without backtest, shadow pass, rollback plan, and owner approval.	Add champion/challenger scorecards, causal-attribution ledger, and simulation runs before production promotion.
Reinforcement Service P2	100%	Promote approved improvements into agents, skills, policies, and evidence ranking without silent model drift.	Promotion queue, Version pinning, Regression monitor, Rollback trigger	Approved Candidate -> Versioned -> Activated -> Monitored -> Rolled Back	Runtime behavior changes only through approved, versioned, observable releases.	Separate learning detection from runtime reinforcement and expose every promotion in the control plane.

What Learning Means

Learning = observed outcomes + decision rationale + control performance + exception patterns + human feedback + business results.

Learning is not model fine-tuning by default. It is a governed enterprise knowledge loop that measures what happened, detects patterns, creates knowledge candidates, and promotes only approved changes.

Learning type	Captured	Runtime proof	LLM required
Outcome learning	Yes	491 of 500 recent decisions carry outcome-like fields	False
Control learning	Yes	13 control assets are available for control-performance analytics	False
Exception learning	Yes	10 exception patterns are mined from decision reason codes and states	False
Judgment learning	Yes	493 decisions carry rationale-like fields	False
SME learning	No	SME capture workflow is defined as a governed candidate path; live approval workflow remains a gap	optional
Process learning	Yes	498 decisions carry learning signals or confidence calibration data	False

Enterprise Memory Graph

The model can change; BP's decision memory, policies, evidence, outcomes, and expertise must remain portable enterprise assets.

Readiness: 90%

Graph nodes

Person / Expert	Analyst and supervisor identity captured in action and decision records where available.
Case	500 decision records can be linked to case-like entities.
Decision	500 historical decisions available.
Policy	465 policy assets discovered.
Control	13 control assets discovered.
Evidence	Evidence hash and evidence pack fields are present on the decision ledger.
Outcome	491 outcome-bearing decisions.
Learning Artifact	498 learning-bearing decisions.
Knowledge Asset	0 knowledge/playbook assets discovered.

Graph edges

Person / Expert	made or approved	Decision
Decision	resolved	Case
Decision	used	Policy
Policy	enforces	Control
Decision	supported by	Evidence
Decision	produced	Outcome
Outcome	generated	Learning Artifact
Learning Artifact	improves	Agent / Policy / Skill

Evaluation & Reinforcement

Measure agents on BP business outcomes, not generic model benchmarks.

Readiness: 99%

Evaluation	Readiness	Measure
Business Outcome Evaluation	98%	leakage prevented, cash improved, cycle time reduced
Human Agreement Evaluation	100%	agent recommendation versus expert decision
Policy Compliance Evaluation	100%	control adherence and authority boundary compliance
Financial Impact Evaluation	98%	value protected, recovered, released, or delayed
Trust Evaluation	100%	override, acceptance, rejection, and escalation rates
Regression Evaluation	100%	before/after quality and false positive movement

Agent Memory Contract

Every agent should consume case memory, outcome memory, expert memory, and organization memory within policy boundaries.

Memory	Runtime use
Case Memory What happened before?	Retrieve similar cases, evidence, policy gates, and decisions.
Outcome Memory What worked?	Rank actions by actual resolution, value, false positives, and rework.
Expert Memory How do senior analysts resolve this?	Surface approved SME guidance and expert-resolution patterns.
Organization Memory What has BP learned globally?	Apply approved lessons across missions, countries, suppliers, and controls.

Discovery Agent Learning Opportunities

Discovery should not only find automation opportunities. It should identify where BP judgment, expertise, and repeated exceptions can become reusable learning.

Question	Learning output
Which decisions repeat?	Reusable decision pattern candidates and automation boundaries.
Which decisions rely on tribal knowledge?	SME capture targets and documentation gaps.
Which decisions are not documented?	Mandatory rationale and evidence-capture backlog.
Which experts are bottlenecks?	Expertise-risk heatmap and delegation opportunities.
Which cases produce the most learning?	Learning value ranking for discovery, controls, and transformation.
Which policy overrides repeat?	Policy clarification or control redesign candidates.

Enterprise Knowledge Taxonomy

Knowledge is not just documents. It includes process, control, judgment, outcome, exception, and SME knowledge.

Knowledge type	Readiness	Meaning	Runtime proof	Gap
Process Knowledge Enterprise Knowledge Graph + process/workflow packs	100%	How finance work happens across processes, activities, tasks, and handoffs.	18 procedure/workflow assets discovered	Promote process/activity/task IDs onto every decision and action record.
Control Knowledge Control Intelligence Fabric + policy/control packs	100%	How risks are prevented, detected, corrected, tested, and attested.	13 control assets discovered	Add control owner, testing frequency, failure impact, and attestation state to each enforced control.
Judgment Knowledge Enterprise Judgment Registry + decision ledger rationale fields	99%	Why humans approve, reject, override, or escalate recommendations.	493 decisions carry rationale-like fields	Make decision rationale and override reason mandatory for governed actions.
Outcome Knowledge Outcome & Learning Runtime + decision ledger outcome fields	98%	What happened after a recommendation or human action, including value and rework.	491 decisions carry outcome-like fields	Bind executed actions to actual SAP/BlackLine/ServiceNow completion and business-value results.
Exception Knowledge Exception Intelligence Engine + reason-code pattern store	100%	What repeatedly goes wrong, root causes, preventability, and prevention opportunities.	10 exception patterns mined	Add formal root-cause taxonomy, preventability flag, and process/supplier owner.
SME Knowledge Knowledge Registry + governed knowledge candidates	54%	Controller, analyst, and process-owner guidance that is not present in formal policy documents.	0 explicit knowledge/playbook assets discovered	Add SME capture, review, approval, effective date, and retirement workflow.

Foundational Runtime Services

These are not five separate features. They are the foundational runtime services between agents, enterprise judgment, controls, outcomes, and learning.

Runtime service	Entities	Services	UI surface	Integrations
Enterprise Judgment Registry Institutional memory for enterprise decisions	Decision, Judgment, Override, Outcome, SME Knowledge	Decision Capture Service, Judgment Extraction Service, Similar Decision Service	Judgment Explorer + analyst drawer similar historical decisions	SAP, Ariba, ServiceNow, Workflow systems, Human comments
Control Intelligence Fabric Executable, observable, measurable enterprise controls	Control, Policy, Risk, Process, Evidence	Control Registry, Control Mapping Engine, Control Monitoring Engine	Control Command Center	Policy Runtime, Evidence Runtime, Replay Runtime, Agent Runtime
Decision Memory Runtime Searchable historical enterprise decision patterns	Case, Context, Evidence, Outcome, Decision Pattern	Decision Embedding Service, Pattern Discovery Engine, Similarity Engine, Memory Ranking Engine	Decision Memory Drawer	Enterprise Judgment Registry, Replay Runtime, Evidence Fabric
Exception Intelligence Engine Convert exception handling into prevention intelligence	Exception, Root Cause, Resolution, Owner, Impact	Exception Registry, Root Cause Engine, Prevention Engine, Trend Engine	Exception Control Tower	Decision Memory Runtime, Control Intelligence Fabric, Outcome & Learning Runtime
Outcome & Learning Runtime Closed-loop measurement of decision quality, value, and improvement	Recommendation, Decision, Outcome, Variance, Value	Outcome Tracking Engine, Decision Quality Engine, Learning Engine, Value Realization Engine	Learning Command Center	Controls, Continuous Close, Autonomous Controls, Finance Command Center

Capability	Readiness	Purpose	Runtime proof	Remaining gap
Enterprise Judgment Registry	99%	Capture human decisions, rationales, overrides, and accountable judgment as enterprise assets.	500 decisions in canonical ledger; 493 decisions carry explicit rationale; 18 SME/procedure knowledge assets available for promotion	Make override reason and SME rationale mandatory in analyst/supervisor action drawers.
Control Intelligence Fabric	100%	Map controls to policies, risks, processes, agents, evidence, decisions, and attestation.	13 control assets discovered in tenant policy packs; 500 decisions have evidence linkage; 500 decisions have replay linkage	Persist control_id, control_owner, test cadence, and attestation status on every policy-bound decision.
Decision Memory Runtime	100%	Store historical decisions, similar cases, approval patterns, override patterns, and replayable outcomes.	500 historical decisions queryable; 10 recurring decision patterns detected; 500 decisions replayable	Add similarity search across the last 500+ decisions and expose approval/override/rework rates per pattern.
Exception Intelligence Engine	99%	Convert exception handling into root-cause, prevention, policy-friction, and supplier/process improvement intelligence.	494 exception-like decisions detected; 10 reason-code patterns mined; Exception signals are derived from status, action, reason codes, and decision payload	Add formal root-cause taxonomy, preventability flag, and prevention owner per exception.
SME Knowledge Capture	67%	Preserve controller, analyst, and process-owner judgment that is not written in formal policies.	0 knowledge assets discovered; 18 procedure/workflow assets discovered; 493 decisions have rationale-like fields	Add a Judgment Workbench for SME capture, owner review, certification, and promotion to runtime rule candidates.
Policy-to-Control Mapping	100%	Connect policy execution to controls, risks, processes, evidence, agents, and audit obligations.	465 policy assets discovered; 13 control assets discovered; 500 decisions evidence-linked	Promote policy-control-risk-process-agent-evidence graph IDs into the decision ledger.
Outcome Intelligence Runtime	98%	Measure whether decisions were correct, valuable, delayed, harmful, reversed, or improved.	491 decisions carry outcome-like fields; 491 decisions carry counterfactuals; 498 decisions carry confidence bands	Persist actual outcome, business value, rework, delay, and false-positive/false-negative status after action completion.
Enterprise Learning Runtime	79%	Close the loop from feedback to judgment, outcome, control, policy, and agent improvement.	498 decisions carry learning signals; 83 context queries marked used in decisions; 17 context cases pass >=85% completeness	Add governed learning promotion gates: propose, review, certify, pilot, parallel-run, activate, monitor.

Enterprise Knowledge Registry

Knowledge	System of record	Stored fields	Governance
Judgment	AgentDecisionLedger + EnterpriseMemoryEntry	decision, rationale, actor, authority, override, replay	Human action writes remain ledger-backed; promoted guidance requires owner approval.
Lessons	EnterpriseMemoryEntry + CrestKnowledgeArtifact + CrestLearningOutcome	lesson, source, confidence, approval_state, version, content_hash	Lessons start as candidates and must be validated before runtime use.
Outcomes	AgentLearningFeedback + AgentDecisionLedger outcome payloads	recommendation, decision, actual outcome, value, error, rework	Outcome facts should come from SOR completion events or controller attestation.
Control Knowledge	Policy/control packs + Control Intelligence Fabric	control, owner, risk, evidence, testing, attestation	Controls are never auto-changed by learning; they generate change candidates.
Exception Knowledge	Decision ledger reason codes + Exception Intelligence Engine	exception, root cause, frequency, impact, resolution, prevention	Exception patterns can be auto-mined; prevention changes require process-owner review.

Knowledge Governance

Learning may propose changes; governance approves changes. No automatic learning path may change policies, controls, authority limits, or production execution rights.

Observed -> Candidate created -> Evidence attached -> Owner reviewed -> Approved / rejected -> Pilot or parallel run -> Activated -> Monitored -> Retired

Add visible approve/reject/retire actions for learning candidates and require owner certification before runtime consumption.

Learning Provenance

Every lesson must trace back to the decisions, evidence, actors, policies, controls, outcomes, and replay records that produced it.

168 similar decisions

Policy: pol:finance:control

Action: recommend

Top exception signal: Pol:Finance:Control

500 evidence-linked decisions

500 replayable decisions

Readiness: 100%

Learning Effectiveness

Can the system prove learning improved business outcomes?

Measured by: decision accuracy before/after, false-positive rate before/after, false-negative rate before/after, override rate before/after, cycle-time improvement, control exception reduction, value protected or recovered

Create baseline snapshots and post-promotion comparison windows for every approved learning item.

SME Knowledge Capture

Capture controller, analyst, process-owner, and SME rationale that currently lives outside formal systems.

Capture: override reason, approval rationale, escalation note, post-action outcome comment, controller guidance, audit finding response

The capture model is defined, but the UI must make rationale capture mandatory for overrides and high-risk approvals.

Learning Boundaries

Type	Auto learn	Boundary
Metrics	Yes	Can update analytics and confidence monitoring automatically.
Patterns	Yes	Can create pattern candidates automatically; promotion requires review.
Exceptions	Yes	Can mine frequency/root-cause signals automatically; remediation changes require owner approval.
SME Guidance	No	Can be captured as candidate knowledge; must be reviewed before runtime use.
Policies	No	Cannot change automatically. Learning may only propose policy candidates.
Controls	No	Cannot change automatically. Controller or control owner certification required.
Approval Limits	No	Cannot change automatically. Requires formal authority and IAM/governance review.
Agent Authority	No	Cannot expand automatically. Requires certification and control-plane approval.

Learning Architecture Layers

Layer	Purpose	Services
Observation Layer	Capture recommendations, decisions, evidence, replay, feedback, and outcomes.	Decision Ledger, Outcome Tracker, Evidence Runtime, Replay Runtime
Analysis Layer	Detect patterns, trends, exceptions, control friction, and value opportunities.	Pattern Engine, Control Analytics, Exception Analytics, Decision Analytics
Knowledge Layer	Store lessons, SME guidance, decision memory, control knowledge, and exception knowledge.	Knowledge Registry, Enterprise Memory, Judgment Registry, Decision Memory Runtime
Governance Layer	Review, approve, reject, retire, version, and certify learning before runtime use.	Learning Promotion Gates, Owner Review, Versioning, Certification
Runtime Layer	Consume approved learning as advisory context, not as silent model drift.	Policy Runtime, Agent Runtime, Control Plane, Ask Sphere

Institutional Memory Readiness

Area	Readiness	Proof
Knowledge	0%	0 explicit knowledge assets
Judgment	99%	493 rationale-bearing decisions
Controls	100%	13 control assets
Outcomes	98%	491 outcome-bearing decisions
Learning	100%	498 learning-bearing decisions
Replay	100%	500 replayable decisions

Deterministic Learning Core

Enterprise learning is primarily outcome measurement, pattern detection, control analytics, decision analytics, and governed knowledge updates. It does not require an LLM.

Layer	Readiness	Purpose	Services	Runtime proof	LLM required
Outcome Learning Deterministic	98%	Capture recommendation, human decision, actual outcome, value, error, and rework.	Outcome Tracker, Decision Quality Engine, Value Realization Engine	491 decisions carry outcome-like fields; 500 decisions available for outcome measurement	No
Pattern Learning Deterministic	100%	Detect recurring decision, supplier, country, amount, policy, and exception patterns.	Pattern Detection, Similar Decision Search, Decision Analytics	10 recurring decision patterns detected; 10 exception reason-code patterns mined	No
Control Learning Deterministic	100%	Measure control alert volume, evidence linkage, replayability, violations, and false-positive pressure.	Control Analytics, KPI Analysis, Exception Analysis	13 control assets discovered; 500 decisions evidence-linked; 500 decisions replayable	No
Judgment Learning LLM-assisted optional	99%	Extract and classify why humans override, approve, reject, or escalate recommendations.	Rationale Extraction, Comment Classification, Override Clustering	493 decisions carry rationale-like fields; LLM may classify free text; ledger remains source of truth	No
Knowledge Extraction LLM-assisted optional	50%	Transform unstructured comments, SME explanations, and post-action notes into governed knowledge candidates.	Comment Summarization, Knowledge Synthesis, Lesson Candidate Generation	0 knowledge assets discovered; 498 decisions carry learning signals	No
Recommendation Improvement ML / statistics / optional LLM	100%	Promote better features, thresholds, policies, and automation candidates through governed review.	Feature Discovery, Threshold Analysis, Learning Promotion Gates	498 learning signals available; 10 patterns available for promotion review	No

No-LLM learning architecture

Observation -> outcome measurement -> pattern detection -> knowledge update -> behavior change.

Decision

Outcome Tracker

Learning Events

Learning Store

Pattern Engine

Control Analytics

Decision Analytics

Knowledge Graph Updates

Optional LLM-assisted layer

LLMs help extract rationale and synthesize knowledge from unstructured comments. They are not the learning system or source of truth.

Learning Store

LLM Knowledge Extractor

Lessons Learned

Knowledge Registry

Decision Memory Patterns

Entity	Policy	Action	Count	Outcome	Exception	Replay
Invoice	pol:finance:control	recommend	168	168	168	168
Invoice	pol:finance:control	decide	155	155	155	155
Invoice	pol:finance:control	execute	82	82	82	82
Invoice	pol:finance:control	no_payment_timing_action	24	24	24	24
Invoice	pol:finance:control	attach_ingestion_evidence	22	22	22	22
Invoice	pol:finance:control	monitor	22	22	22	22
payment	payment_release_policy	DELAY	17	17	17	17
INVOICE	p2p_invoice_controls	REQUEST_EVIDENCE	6	0	0	6

Exception Intelligence

Root-cause signal	Count
Pol:Finance:Control	473
Cash Headroom Tight	17
Ontology Cash Coverage Ratio	17
Decision Surface	6
Amount Above Green Threshold	4
Context Steward Action	2
Owner Approval Required	2
Price Variance	1

Enterprise Learning Loop

The runtime should learn only after judgment, outcome, control impact, and replay are captured. Learning is a governed promotion path, not silent model drift.

Enterprise Judgment Registry

Decision Memory Runtime

Control Intelligence Fabric

Exception Intelligence Engine

Outcome & Learning Runtime

Enterprise Decision & Control Runtime

P2P / O2C / R2R / Treasury / FP&A

Decision

Human / Agent Action

Rationale

Outcome

Exception Pattern

Control Impact

Policy Candidate

Learning Promotion

Replay

The learning system does not depend on generative AI. Core learning is deterministic and audit-backed through outcomes, patterns, controls, decisions, and observed business behavior. LLMs are optional helpers for rationale extraction, comment summarization, and knowledge synthesis.

Enterprise Learning Runtime - IRIS