Skip to content

For LLM Agents

Corrobore is your external, structured working memory for intelligence work. Use it to store and retrieve entities, relationships, evidence, confidence, time, versions, and audit history instead of carrying a growing graph as JSON in context.

This guide describes the current runtime contract. Check GET /version when a workflow depends on a particular deployed release.

You                                      Corrobore
read and reason                     ->   store and validate structured state
propose evidence-backed changes     ->   enforce policy and record mutations
ask focused graph questions         <-   return the bounded result you need

Operating loop

  1. Check /health when tool connectivity is uncertain.
  2. Search for seed nodes from the task objective when ids are unknown.
  3. Read the smallest subgraph that can answer the question.
  4. Inspect source material and distinguish observation from inference.
  5. MERGE new entities and relationships with evidence, confidence, and status.
  6. Read back entities and run an explicit relationship coverage pass.
  7. Finish every authorized write before promotion.
  8. Promote eligible nodes and relationships, then attempt strict export.
  9. If a late write is necessary, repeat read-back, validation, and promotion.
  10. Leave uncertain claims for human review.

Request discipline

  • Prefer explicit mode routes:
  • POST /v1/cypher/read for reads.
  • POST /v1/cypher/write for mutations.
  • Use session_id when a session was started; this keeps audit traces and lifecycle state coherent.
  • Treat runtime Rejected and ValidationFailed statuses as expected control signals, not transport failures.
  • Keep query scope bounded (LIMIT, narrow patterns, explicit objective).

Find a starting point

POST /v1/seed/search
Authorization: Bearer <token>
Content-Type: application/json

{
  "objective": "find infrastructure linked to the phishing campaign",
  "domain_profile": "cti",
  "mode": "hybrid",
  "top_k": 5,
  "score_threshold": 0.2
}

Use the returned node_id, score, rationale, source references, and boundary notes to choose a defensible seed. Do not treat ranking as proof.

Read before writing

MATCH (c:Campaign {id: "campaign--42"})-[:USES]->(i:Infrastructure)
RETURN c, i
LIMIT 50

Query narrowly and include a bound. If no supported answer exists, report the gap rather than inventing one.

Write evidence-backed intelligence

Confidence boundary

Surface Accepted scale Example
Native Cypher and memory operations 0..=1 0.9 means 90%
STIX objects and STIX import annotations 0..=100 90 is stored as native 0.9

Never copy a STIX value such as 90 directly into Cypher. Corrobore rejects it rather than guessing which boundary the caller intended.

MATCH (a:ThreatActor {name: "APT28"})
MATCH (m:Malware {name: "X-Agent"})
MATCH (e:EvidenceSpan {id: "span--123"})
MERGE (a)-[r:USES]->(m)
SET r.confidence = 0.82,
    r.evidence_refs = [e.id],
    r.status = "candidate"
RETURN r
  • Prefer MERGE for identities that may already exist.
  • Every relationship assertion owns its own evidence and confidence; endpoint metadata does not transfer to the relationship.
  • Mark inference as candidate; do not silently promote it to fact.
  • Use domain vocabulary from Intelligence Domains.
  • Never use unsupported clauses such as DETACH DELETE, LOAD CSV, UNWIND, FOREACH, CALL APOC, or CALL DBMS.

Respect host boundaries

The host decides whether a tool is read-only or permits mutations. Use /v1/cypher/read for reads and /v1/cypher/write only when the task authorizes a graph change. A rejected mutation is a policy result, not a reason to bypass the gateway.

Named sessions make actions inspectable. When a workflow starts a session, pass its session_id on Cypher and import calls, inspect its health, and stop it when finished. Session logs expose input/output audit parity and can be filtered by time.

Do not bypass policy outcomes by switching route shape or query wording. If the runtime rejects a mutation, return the rejection reason and request explicit authorization.

CTI relationship coverage

For STIX 2.1 Indicator modeling, keep these assertions distinct:

  • Indicator -> Observed Data: based-on states which observation supports the Indicator.
  • Indicator -> CTI domain object (SDO): indicates states what threat object, such as Malware, the Indicator detects or characterizes.

Each is a Relationship SRO with its own STIX ID, retained evidence annotation, confidence, and candidate lifecycle status. After importing, compare the expected relationship inventory from the source with the graph. Do not fabricate a relationship merely to improve coverage; report an unsupported bridge as a gap.

MATCH (source)-[r]->(target)
WHERE r.confidence IS NULL OR r.evidence_refs IS NULL
RETURN r
LIMIT 100

MATCH (source)-[r]->(target)
WHERE r.status = "candidate"
RETURN r
LIMIT 100

Validate and export

POST /v1/stix/validate accepts either an explicit bundle or current graph CTI nodes. For explicit bundles, built-in playbooks may correct supported missing fields; corrected objects are imported whenever at least one playbook runs. valid reports whether the validation pass found an error, not a second verdict after fixes. Inspect issues, playbooks_applied, corrections_summary, and persistence, then revalidate when you need a post-correction verdict.

Use GET /v1/export/stix for deterministic, CTI-scoped STIX projection after validation. The route is read-only and strict is the default correctness gate. Late writes remain candidate and require a new readiness and promotion pass before another strict attempt. Permissive is only for an explicit caller request for a diagnostic partial bundle. force=true is an explicit operator decision and never an automatic LLM fallback. Validation still runs and each bypassed semantic finding remains in diagnostics; force does not bypass lifecycle, identity, evidence-integrity, endpoint, provider, or license gates. Preserve the returned object identities and x_corrobore_evidence_refs instead of inventing replacements. Logical export metadata identifies the snapshot and transaction, but the current HTTP export does not roll the graph back in time.

Recover safely

  • Fix invalid ids, modes, domain profiles, arity, or syntax from the returned error.
  • Narrow objectives that return OVERBROAD_OBJECTIVE.
  • Ask for disambiguation when seed search returns AMBIGUOUS_SEED.
  • Stop or ask a human when evidence is missing, contradictory, or outside the authorized scope.
  • Do not repeat the same rejected mutation or weaken evidence requirements.

Canonical references

See Cypher Support, HTTP Server, and the OpenAPI contract.