Verification & Integrity

The two tools behind cks-mcp's anti-hallucination guarantees. See Security Model for the full trust model these implement.

verify_source

Performs a real HTTP request against an external URL and builds a cryptographically signed VerificationRecord from the actual result. This is the only sanctioned way to produce a VerificationRecord — every other tool treats a record without a valid signature as fabricated (see below).

Parameters: url (string, required), subject_id (string, required — the id of the Knowledge Object this verification is about).

Response

{
  "objects": [
    {"identity": {"id": "vr-...", "type": "VerificationRecord", "name": "verification"},
     "structure": {"checked_at": "2026-07-30T12:00:00Z", "checked_via": "automated_http_check",
                   "http_status": 200, "signature": "<HMAC>"}},
    {"identity": {"id": "rel-...", "type": "Relation", "name": "r"},
     "structure": {"participants": ["<subject_id>", "vr-..."], "relation_type": "verified_by"}}
  ]
}

Merge this straight into your structure with evolve_knowledge (add_object for both entries), or include it directly in a json_data payload. The request itself is hardened against SSRF and DNS rebinding: only http/https are allowed, the resolved IP is validated against public-address ranges (blocking private, loopback, link-local, and cloud-metadata ranges, plus CGNAT-style shared address space), and the connection is pinned to the validated IP for the actual request — redirects are followed manually, one hop at a time, re-validating each new target.

detect_contradictions

Flags logical contradictions in a structure using two constraint types. Read-only.

• mutual_exclusion — the same source–target pair has both of two declared relation types (e.g. both supports and refutes).
• functional_relation — a single source has more than one target via a relation type declared single-valued (e.g. a planet orbits two stars).

To use it, your structure must contain MutualExclusionRule and/or FunctionalRelationRule objects declaring which relation types are subject to which check:

{"identity": {"id": "rule-1", "type": "MutualExclusionRule", "name": "no-support-and-refute"},
 "structure": {"relation_type_a": "supports", "relation_type_b": "refutes"}}

{"identity": {"id": "rule-2", "type": "FunctionalRelationRule", "name": "single-orbit"},
 "structure": {"relation_type": "orbits"}}

Parameters: session_id or json_data (one of the two).

Response

{
  "session_id": "sess-abc123",
  "contradiction_count": 1,
  "contradictions": [
    {"code": "CKS-EXT-FUNCTIONAL-RELATION", "severity": "warning", "source": "extension",
     "message": "...", "metadata": {}}
  ]
}

Why provenance is checked unconditionally

A VerificationRecord looks the same to a downstream reader whether it was genuinely produced by verify_source or hand-crafted by an LLM trying to look convincing. Because of that, validate_knowledge, evolve_knowledge, merge_knowledge, merge_branch, and the json_data fallback paths of serialize_knowledge/explain_knowledge/detect_contradictions all check every VerificationRecord's HMAC signature regardless of whether you requested the verification_record extension. A forged record can never become a persisted, readable version — see Security Model.