Skip to content
ISNADISNAD

Chain-scoped grounding

Chain-scoped grounding answers a question the plain content critic cannot: “is this claim grounded in evidence that actually traveled its own transmission chain?”

The content critic checks whether a claim’s information is present in a merged corpus. In a multi-agent run that merged corpus loses track of which agent fetched which row, so a claim can look “grounded” in a row a different agent on a different branch fetched — context pollution. Chain-scoped grounding makes grounding provenance-aware.

The primitive

Section titled “The primitive”
  • ChainLinkSpec.retrieved_rows — the content of the rows a link retrieved, carried runtime-only (never serialized into to_dict() / the signed record; the signed record commits to document_hashes only, so raw row content never enters the RFC 8785 canonical form or the --redact surface).
  • chain_scoped_corpus(chain) — the deduplicated, order-preserving union of every link’s retrieved_rows — “the evidence that traveled this claim’s own path”.

The scoping is chain-level, not link-level: a synthesis link retrieves nothing and legitimately grounds in the upstream retrieval link’s rows, so a link-local check would wrongly flag every legitimate synthesis.

The policy

Section titled “The policy”

ChainScopedGroundingPolicy (via evaluate_chain_grounding) returns a GroundingResult:

Field Meaning
grounded_off_chain_only True iff the claim is CONSISTENT against off-chain rows and UNVERIFIABLE against its own chain — a grounding gap, not proven contamination
on_chain_verdict the critic verdict against the chain-scoped corpus (None = provenance unknown)
off_chain_verdict the critic verdict against the off-chain rows

The flag is evidence, not an action. It deliberately has no route into the SERVE/REVIEW/QUARANTINE decision matrix — that mapping is a separate decision.

Honest limits

Section titled “Honest limits”
  • Critic capability. The flag only fires when the critic can affirm CONSISTENT. The default EmbeddingCritic is contradiction-only (never returns CONSISTENT), and NLI/LLM critics downgrade CONSISTENT → UNVERIFIABLE when their affirmation gate is on. A CONSISTENT-capable critic requires gate_affirmation=False (or a configured LLM). The serving API surfaces this as grounding.grounding_capable.
  • Persistence. retrieved_rows is runtime-only. A chain reloaded from the database carries retrieved_rows_known=False, and the policy returns on_chain_verdict=None (never flags) rather than falsely flagging a provenance-unknown chain (#241).
  • Spoofable split. off_chain_rows is caller-supplied and unverified — a planted off-chain row can false-flag a rival’s claim; off_chain_rows=[] silently disables the flag. Establishing the split honestly is the caller’s responsibility.
  • Measured, not asserted. The flag’s false-positive rate is measured in experiments/grounding_eval/ (#239) against a real NLI critic — see the “measure, don’t assert” discipline. Small pilot set; treat as a direction, not a tight estimate.

Wiring

Section titled “Wiring”

POST /v1/claims accepts optional retrieved_rows per chain link, computes the chain-scoped corpus, runs the policy, and returns a grounding block (flag, both verdicts, corpus sizes, grounding_capable). The flag does not change the decision.