Answer Bill of Materials

The Answer Bill of Materials (AnswerBOM) is a versioned, machine-readable manifest of one guarded answer. Where the guard's decision says whether an answer was approved, the BOM says why, claim by claim: which model produced the answer, which scorer judged it against which threshold, and for every atomic claim — what evidence supports it, how strongly, and the verdict the guard reached.

It is the artefact a reviewer or auditor reads to answer "which parts of this answer are grounded, and in what?" without re-running the guard, and it round-trips through an audit log via to_dict / from_dict.

Shape

{
  "schema_version": "director.answer_bom.v1",
  "answer_id": "abom_…",
  "timestamp": "2026-06-06T16:40:00Z",
  "model": "gpt-4o",
  "scorer": "rust",
  "threshold": 0.6,
  "tenant": "acme",
  "claims": [
    {
      "claim": "Refunds are available within 30 days.",
      "verdict": "supported",
      "support": 0.91,
      "evidence_ids": ["vector:refund-policy"],
      "freshness": "2026-01-01",
      "tenant": "acme",
      "policy_refs": ["refund-policy-v3"]
    }
  ],
  "unsupported_claims": [],
  "support_coverage": 1.0
}

verdict is one of supported, unsupported, or contradicted. support is a strength in [0, 1]. unsupported_claims and support_coverage are derived from claims and recomputed on load, so editing them in a stored manifest has no effect.

Building from a guard result

ProductionGuard.answer_bom() builds the manifest from a GuardResult. The threshold is the calibrated threshold when calibration is enabled, otherwise the configured coherence threshold; the scorer id is the configured backend.

from director_ai.guard import ProductionGuard

guard = ProductionGuard.from_profile("finance")
guard.load_facts({"refund": "Refunds are available within 30 days."})

result = guard.check("What is the refund window?", "Refunds close after 30 days.")
bom = guard.answer_bom(
    result,
    model="gpt-4o",
    tenant="acme",
    freshness="2026-01-01",
    policy_refs=["refund-policy-v3"],
)

for claim in bom.unsupported_claims:
    print(claim.claim, claim.verdict, claim.support)

audit_log.write(bom.to_json())          # round-trips via AnswerBOM.from_dict

build_answer_bom(score, ...) is the underlying free function for callers that already hold a CoherenceScore.

How claims are mapped

The BOM consumes the scorer's existing claim-level provenance — the atomic claims and, for each, the source sentence it was attributed to with a support/contradiction divergence:

• supported → verdict="supported", support = 1 − divergence, and evidence_ids resolved from the attributed source chunk's id (or a positional source:N reference when the evidence carries no id).
• not supported, high divergence (≥ contradiction_threshold, default 0.5) → verdict="contradicted", no evidence ids.
• not supported, low divergence → verdict="unsupported".

The manifest's claim-level richness is exactly that of the scorer's provenance: when the scorer produced no attributions, the BOM records the model/scorer/threshold header with no per-claim rows rather than inventing them.