Epistemic Audit & Compliance (CONCEPT:AU-KG.enrichment.compliance-posture-rollup)¶

Persona: internal audit, SOC2/compliance officer, or anyone building regulated decision support on top of the knowledge graph — "what did we believe, on what evidence, as of what date, and can I prove the record hasn't been tampered with."

This guide is a walkthrough for a specific narrow question a plain graph_query Cypher result can't answer on its own: not just what is in the graph, but why do we believe it, who could see it, and is the trail provably intact. Everything below wraps existing primitives — no new compliance or redaction logic was added to write this guide; the tools it documents (graph_audit, graph_compliance, graph_epistemic) are thin aggregation/naming layers over the engine's own audit chain and belief machinery (agent_utilities/mcp/tools/audit_tools.py, agent_utilities/mcp/tools/compliance_tools.py, agent_utilities/mcp/tools/epistemic_tools.py).

What you get¶

Question	Tool / action	Backing primitive
"Has the audit log been tampered with?"	`graph_audit(action="verify")`	Rust SHA-256 hash-chained mutation log (`epistemic-graph/src/audit.rs`, redb `AUDIT` table)
"What happened to this entity, in order?"	`graph_audit(action="for_target", target_id=...)`	KG's own `:ToolCall` provenance, reverse-indexed
"What's our overall compliance posture right now?"	`graph_compliance(action="posture")`	audit-chain `verify()` + node-count/status rollup of the governance labels the CISO Assistant extractor + TRM engine already ingest
"Export this subgraph for an auditor, with redaction applied"	`graph_compliance(action="export")`	bulk `explain_belief(node_id, disclosure_level)` over an id list or Cypher selection
"Why do we believe this claim?"	`graph_epistemic(action="why", node_id=...)`	`explain_belief` justification tree (`Asserted` / `DerivedSupport` / `DerivedContradiction` / `BayesianUpdate`)
"Do we still believe it, since when, and what would flip it?"	`graph_epistemic(action="status", node_id=...)`	`epistemic_status` acceptance capstone (opt-in `epistemic-tms` engine feature)
"What changed between two audit periods?"	`graph_epistemic(action="what_changed", tx_from=..., tx_to=...)`	whole-graph bitemporal diff (opt-in `epistemic-tms`)
"As of last quarter-end, what did the graph say?"	`graph_query(cypher=..., as_of="2026-03-31T00:00:00Z")`	bitemporal `as_of` cutoff on the read path

All of these are also plain MCP tools you can call from any client that speaks MCP, and each has a REST twin (POST /audit, POST /compliance, POST /epistemic on the graph-os gateway — the generic REST-twin factory in kg_server._build_server mounts one for every entry in ACTION_TOOL_ROUTES without a bespoke handler).

1. Verify the audit ledger hasn't been tampered with¶

Every durable mutation to the graph already chains into a per-graph SHA-256 hash chain (epistemic-graph/src/audit.rs): each entry binds the previous entry's hash, so altering, reordering, deleting, or inserting any entry breaks the chain at that exact position. graph_audit(action="verify") walks the chain and reports the first break, if any.

// Call
graph_audit(action="verify")

// Response shape (AuditReport, eg-types/src/protocol.rs):
{
  "surface": "audit",
  "action": "verify",
  "available": true,
  "graph": "__commons__",
  "ok": true,
  "entries": 4218,
  "first_broken_seq": null,
  "detail": "chain verified: 4218 entries, no breaks"
}

If the chain has been altered, ok is false and first_broken_seq names the exact sequence number where verification first failed — the entry at that seq (and everything after it) is suspect.

Degrade contract. AuditVerify requires the engine to be built with the security cargo feature (part of the default full build) and a durable redb persist dir configured (GRAPH_SERVICE_PERSIST_DIR / --persist-dir). An in-memory-only engine, or a slim build without security, returns a clean error instead of raising:

{
  "surface": "audit",
  "action": "verify",
  "available": false,
  "error": "audit ledger not exposed by this engine build/config (...). Requires the epistemic-graph `security` cargo feature (part of the default `full` build) AND a durable redb persist dir configured — otherwise this is a corresponding epistemic-graph-side gap, not an agent-utilities one."
}

If you're building toward an audit-defensible deployment, confirm this returns available: true before relying on the chain for anything — the Enterprise and Single-node prod recipes both configure a durable persist dir; the zero-infra Tiny recipe may not, depending on your .env.

2. Reconstruct "what happened to entity X"¶

graph_audit(action="for_target", target_id="...") is the entity-anchored half of the same primitive: every :ToolCall that acted on a given entity id, in call order, reverse-indexed off the KG's own tool-call provenance (Orchestrator.get_tool_calls_for_target), plus a best-effort chain-verify snapshot alongside it — so you get both "what touched this record" and "is the ledger recording it still intact" in one call.

graph_audit(action="for_target", target_id="claim:mine:abc123")

3. One-view compliance posture¶

graph_compliance(action="posture") is a rollup, not new logic: it joins the same verify() report from step 1 with a node-count / status-breakdown of the governance labels already ingested into the graph by the CISO Assistant GRC extractor and the TRM portfolio-intelligence engine — Control, Policy, Risk, ComplianceRequirement, ComplianceGate, Regulation, ComplianceAssessment, Assessment, Incident, RemediationProposal, Finding, SecurityException.

graph_compliance(action="posture")

// ->
{
  "surface": "compliance",
  "action": "posture",
  "audit_ledger": { "ok": true, "entries": 4218, "first_broken_seq": null, ... },
  "node_counts": {
    "Control": 214, "Policy": 38, "Risk": 91, "ComplianceRequirement": 156,
    "ComplianceGate": 12, "Regulation": 9, "ComplianceAssessment": 47,
    "Assessment": 47, "Incident": 6, "RemediationProposal": 3,
    "Finding": 22, "SecurityException": 1
  },
  "status_breakdown": {
    "Control": {"implemented": 190, "planned": 24},
    "Incident": {"open": 2, "resolved": 4}
  }
}

A read failure on any one label degrades that label's count to 0 rather than failing the whole rollup — a partial governance mirror still gives you a usable posture view.

4. Redaction-compliant bulk export, as-of a date¶

Per-node redaction already existed at the engine (Method::ExplainBelief's disclosure_level — Full / Skeleton / ExistenceOnly — masks, never silently drops, an evidence node the caller's row-level security can't see). What was missing was a bulk "export this subgraph, redacted, for an auditor" primitive — that's graph_compliance(action="export"): it takes either an explicit id list or a read-only Cypher query selecting an id column, and calls the engine's own explain_belief(node_id, disclosure_level) per id — the exact same per-node redaction primitive graph_epistemic's why action uses, just batched.

// Select by Cypher (any read-only query returning an `id` column) and
// export at Skeleton disclosure, as of a fixed instant, capped at 500 rows:
graph_compliance(
  action="export",
  cypher="MATCH (c:ComplianceAssessment) WHERE c.framework = 'SOC2' RETURN c.id AS id",
  disclosure_level="Skeleton",
  as_of="2026-06-30T23:59:59Z",
  limit=500
)

// ->
{
  "surface": "compliance",
  "action": "export",
  "disclosure_level": "Skeleton",
  "as_of": "2026-06-30T23:59:59Z",
  "requested": 47,
  "exported": 47,
  "truncated": false,
  "entries": [
    {"node_id": "compliance:soc2:cc6.1", "belief": { "root": { "claim": "compliance:soc2:cc6.1", "rule": "DerivedSupport", "confidence": 0.94, "premises": [ ... ] } } },
    ...
  ]
}

disclosure_level controls what an out-of-scope evidence node in the justification tree looks like to the recipient of this export:

Full (default) — every node in the proof tree, as the caller's own RLS view sees it.
Skeleton — structure (claim ids, rule kind, confidence) preserved, but the content of nodes the recipient's policy can't see is masked.
ExistenceOnly — the recipient learns a supporting/contradicting node exists at all, nothing about its content.

This is the mechanism you want when handing a subgraph to an external auditor who should see that a control claim is supported, without seeing the (possibly sensitive) evidence text itself.

5. Explainable decision logs — why do we believe this¶

For a single claim rather than a bulk export, graph_epistemic is the purpose-named wrapper (SKILL: kg-epistemic-answer) over the same justification machinery:

// The justification tree — what supports/contradicts this claim, and how
graph_epistemic(action="why", node_id="claim:mine:abc123")
// -> { "engine_method": "explain_belief",
//      "result": { "root": {
//        "claim": "claim:mine:abc123", "rule": "DerivedSupport",
//        "confidence": 0.82,
//        "premises": [
//          {"claim": "evidence:doc:44", "rule": "Asserted", "confidence": 1.0, "premises": []}
//        ]
//      } } }

rule is one of Asserted / DerivedSupport / DerivedContradiction / BayesianUpdate — the actual proof-tree vocabulary the engine uses (eg-epistemic::model::JustRule), not a paraphrase.

The acceptance capstone goes one step further — "do we still believe this, since when, on what evidence, and what would flip it" — via epistemic_status (opt-in engine feature epistemic-tms, not in the default full build; check for a clean {"error": ...} degrade before assuming it ran):

graph_epistemic(action="status", node_id="claim:mine:abc123")
// -> { "engine_method": "epistemic_status",
//      "result": {
//        "claim": "claim:mine:abc123",
//        "believed": true,
//        "confidence": 0.82,
//        "uncertainty": 0.05,
//        "proof": { "root": { "rule": "DerivedSupport", ... } },
//        "why_not": null,
//        "evidence": ["evidence:doc:44"],
//        "valid_time": [1751328000, null],
//        "tx_time": [1751328042, null],
//        "what_would_invalidate": { ... }
//      } }

And for "what changed between two audit periods" rather than one claim's history, what_changed (same opt-in feature) gives a whole-graph bitemporal diff between two transaction-time bounds:

graph_epistemic(action="what_changed", tx_from=1748736000, tx_to=1751328000)

6. As-of reads for point-in-time audit snapshots¶

Every read that goes through graph_query/graph_ask accepts an as_of ISO-8601 instant — a bitemporal cutoff (valid_from <= as_of < valid_to) — so you can reconstruct exactly what the graph asserted on a given date, independent of what's been written since:

graph_query(
  cypher="MATCH (c:Control {framework:'SOC2'}) RETURN c.id AS id, c.status AS status",
  as_of="2026-03-31T00:00:00Z"
)

Add include_epistemic=true to get each row back as a full EpistemicRow (confidence, bitemporal window, evidence refs, policy labels) in the same call, instead of a bare row plus a second explain_belief round-trip.

Putting it together: a SOC2 evidence-collection pass¶

Confirm the ledger is trustworthy first. graph_audit(action="verify") → ok: true. If this returns available: false, your engine build/config doesn't carry a durable audit trail yet — treat that as a finding, not a tool bug.
Pull the posture rollup. graph_compliance(action="posture") for the control-coverage/status snapshot to attach to the audit narrative.
Export the in-scope control evidence, redacted, as of period-end. graph_compliance(action="export", cypher=..., disclosure_level="Skeleton", as_of=<period-end>).
Deep-dive any control an auditor flags. graph_epistemic(action="why", node_id=...) for the justification tree; action="status" for the acceptance capstone if your engine build has epistemic-tms.
Answer "what changed since last audit." graph_epistemic(action="what_changed", tx_from=<last_audit_tx>, tx_to=<now>).

Honest limitations¶

epistemic_status/what_changed need the opt-in engine feature epistemic-tms — not folded into the default full build. A clean {"error": ...} means the connected engine doesn't have it, not that the claim is unbelieved.
disclosure_level redaction (Full/Skeleton/ExistenceOnly) needs the opt-in epistemic-redaction feature for the masking behavior; without it, explain_belief still runs but does not mask.
graph_audit(action="verify") needs the security cargo feature (default full build carries it) and a durable redb persist dir — an in-memory-only engine has no chain to verify.
These are read-only diagnostics over what was already written. They explain and export existing belief/audit structure; they do not create compliance controls or remediate findings themselves — see graph_ops_causal's control_evidence action and the CISO Assistant writeback sink for the write side of GRC.