Audited cases
8
Context Layer Audit Console - IRIS
Completeness, lineage, and usage proof for enterprise context
AuditContext usageDecision gating
Context Layer Audit Console
This console answers the credibility question enterprise reviewers will ask: is enterprise context complete, and is it actually used by agents, policies, evidence, decisions, replay, and the UI? It scores context completeness deterministically and shows the signal-to-outcome trace.
Avg completeness
68%
Context usage
42%
Agent bypasses
15
Critical gaps
23
Replayable
1
What this audits
Enterprise context is treated as a governed runtime asset: entities, relationships, policies, evidence, decisions, outcomes, and learning. The page checks whether those assets are present and whether they influence the decision.
| Concept coverage | Required entities are present for the case type |
|---|---|
| Relationship coverage | Supplier -> PO -> Invoice -> Receipt -> Payment -> Exception trace |
| Policy linkage | Policy ID/version linked to decision ledger |
| Evidence linkage | Evidence hash/source refs available for audit |
| Replay linkage | Replay hash available with context version |
| Agent usage | Agents show the context queries and returned objects |
| Bypass detection | Missing context usage is flagged |
| Negative simulation | Mask context and prove the recommendation changes |
Why this matters
Without this proof, an Enterprise Context page can look like selected metadata. With it, bp can see that context is complete, auditable, used in decisions, and able to block or degrade actions when key context is missing.
| Completeness score | Required context present / required context expected |
|---|---|
| Decision gate | Missing policy or evidence can block execution |
| Usage proof | Agent context queries are visible |
| Lineage trace | Signal -> Context -> Policy -> Evidence -> Recommendation -> Outcome |
| Replay with context | Decision can be reconstructed with exact context hashes |
| Safe degradation | Missing context forces warning, review, or block |
Live Context Completeness + Usage Audit
Live runtime
Audited cases
8
Avg completeness
68%
Context usage rate
42%
Agent bypasses
15
Missing critical context
23
Replayable decisions
1
Scoring model
Deterministic
Persistence
persisted
Case completeness map
| Case | Type | Score | Status | Assets | Policy |
|---|---|---|---|---|---|
| INV-EKF-L5A-FINAL-20260712 | invoice | 83% | degraded | 10/12 | invoice_triage_policy |
| CREDIT-CP-ATLANTIC-ENERGY | counterparty_credit | 68% | blocked | 7/11 | policy missing |
| MAINT-ACCRUAL-WO-7781 | maintenance_accrual | 71% | blocked | 8/12 | policy missing |
| TRD-BRENT-FX-EXPOSURE-202606 | trade_exposure | 68% | blocked | 7/11 | policy missing |
| TREAS-LIQ-HEADROOM-202606 | liquidity_position | 68% | blocked | 7/11 | policy missing |
| FPA-LNG-THROUGHPUT-JUN26 | forecast_variance | 55% | blocked | 5/10 | policy missing |
| JE-R2R-ACCRUAL-8421 | journal | 65% | blocked | 6/10 | policy missing |
| INV-LIVE-DUP-1778594510 | invoice | 65% | blocked | 7/12 | policy missing |
Lineage trace · INV-EKF-L5A-FINAL-20260712
| Business Signal | invoice_submitted |
|---|---|
| Context Objects Loaded | invoice INV-EKF-L5A-FINAL-20260712; 3 source refs |
| Relationships Loaded | 2 relationship links captured |
| Policies Applied | invoice_triage_policy |
| Evidence Retrieved | d94cb4334297f45d94f72e94a5b37efa5a20194820f5da9f16187a6f121a9e31 |
| Agent Reasoning | Tier AMBER on invoice exception triage: matched invoice_triage_policy v2026.04.15, action REVIEW routed to ap_specialist. Variance 7.14%. Estimated value preserved: $900.00. |
| Recommendation | REVIEW |
| Outcome | ACTIVE |
| Replay | 283380d0033dd303474bf510048f910c66bf7adff4c03468382607e1bbf2be0f |
| Learning | Learning signal pending |
Agent context usage
| Agent | Query | Objects | Used | Impact |
|---|---|---|---|---|
| Duplicate Invoice Detection Agent | invoice + supplier + amount + prior payment | 5 | Yes | Duplicate risk increased by prior payment and supplier/PO linkage. |
| Policy Agent | duplicate policy + approval rules | 1 | Yes | invoice_triage_policy evaluated for hold threshold. |
| Evidence Agent | invoice, PO, receipt, prior payment source refs | 3 | Yes | Evidence pack/source lineage used for auditability. |
| iaf-invoice-decision-agent | resolved context snapshot | 1 | Yes | Context hash linked to decision replay. |
Skills + Policy Framework Audit
Skills represent enterprise capabilities. Policies represent enterprise guardrails. The runtime proves that an agent can act only when a skill is authorized by policy, supported by evidence, and governed through replay.
Skill
Hold Duplicate Payment
Authority
recommend + controlled execute
Policy
invoice_triage_policy
Authorization
authorized
Context
Skill Activated: Hold Duplicate Payment (SKILL-P2P-DUP-HOLD)
Policy Evaluation: invoice_triage_policy
Evidence Collection
Decision Recommendation: REVIEW
Human / Agent Action
Replay & Audit
Learning
Skill audit questions
| Skill ID | SKILL-P2P-DUP-HOLD |
|---|---|
| Version | None |
| Owner | P2P Control Owner |
| Risk Class | High financial control |
| Purpose | Prevent duplicate supplier payment before release. |
| Inputs | Invoice, Supplier master, PO, Prior payment, Evidence pack |
| Outputs | Hold recommendation, Supplier validation request, Payment release control |
| Policies | invoice_triage_policy |
| Authority | recommend + controlled execute |
| Execution Mode | governed action |
| Evidence | Invoice, Purchase order, Receipt, Prior payment, Supplier history |
| Explainability | Yes |
| Replay | Yes |
| Learning | Pending outcome feedback |
| Human Override | Analyst can hold; supervisor required for release override. |
Policy audit questions
| Source | Policy Runtime / authority matrix |
|---|---|
| Owner | Finance Governance |
| Scope | invoice |
| Enforcement | Mandatory |
| Version | 2026.04.15 |
| Versioning | Replayable |
| Exception Handling | Analyst can hold; supervisor required for release override. |
| Evidence Requirement | Invoice, Purchase order, Receipt, Prior payment |
| Monitoring | Compliance measured through ledger, usage proof, and replay |
| Auditability | Traceable to replay |
Negative context dependency simulation
Simulation masks context in the audit evaluator only. It does not delete source records, evidence, or snapshots.
Run a simulation to prove the recommendation changes when context is missing.
Enterprise positioning
The runtime does not treat context as documents or embeddings. It treats context as governed enterprise assets that every agent decision must prove it used.
Implementation boundary
This first version derives completeness and usage from resolved snapshots and the decision ledger. Persistent agent-context usage records should be added next for production-grade ContextOps.