Truth maintenance

Truth maintenance evaluates committed claims and maintains the current answer for a subject and predicate. The answer can be one value, several coexisting values, or unresolved uncertainty. The system records the evidence and rationale so a reader can inspect how it reached that answer. Suppose one note says Mara works at Northstar Labs and another says she works at Cedar Works. Those claims might describe a job change, concurrent roles, or a mistaken identity. Their ingest order alone does not decide which interpretation is correct.

​

Corroboration and conflict

Detection first groups claims by exact value equality. Corroborating duplicates merge their evidence, deduplicated by assertion and span, and are soft-deleted as duplicate claim rows. Their validity windows are not closed: another source agreeing with a claim does not mean the claim stopped being true. For relationships, the ordinary slot includes both endpoints. A rel_set slot additionally compares the different objects for one subject and predicate. That cross-object check is used when cardinality is one or unknown. A predicate explicitly allowing many values, or an event-like predicate, does not create that form of cross-object conflict.

​

Truth slots are materialized answers

A TruthSlot contains the current values and the metadata needed to interpret them: A slot’s top-ranked value is not automatically an unambiguous fact. Inspect the value status, conflict state, rationale, and staleness together. Slots support attributes, individual relationships, cross-object relationship sets, and events. They are partitioned by the security envelope. Claims in different envelopes never share a conflict group, even if an administrator can read both. This protects isolation and deliberately leaves cross-envelope contradictions outside truth maintenance’s comparison scope.

​

How a decision is made

The managed runtime wires the adjudicator when the relevant extraction feature and Anthropic credentials are available. It receives rendered claim values, evidence notes, timestamps, confidence, and known predicate semantics. See pipeline truth maintenance for deployment settings and work kinds.

​

Adjudicator outcomes

Equivalence groups can merge duplicates of a selected winner. A coexist verdict currently retains all claims approved rather than merging every equivalence group independently. If semantics are unknown, the adjudicator must include a semantics proposal. That proposal is recorded as llm_proposed and cannot overwrite an operator’s configured semantics. See predicate governance.

​

Deterministic fallback

When there is no adjudicator or the configured provider cannot produce a usable verdict, the stored predicate policy provides the fallback: With no semantics, the general fallback is allow_parallel. Do not expect an unconfigured deployment to infer which employer replaced the other. The current relationship/attribute adapter supplies no context buckets to split_by_context, so those claims share the default bucket. Setting that policy alone does not infer roles or time periods from prose. A deterministic fallback gives a defined result during provider absence. Its rationale should remain visible so callers can distinguish it from an adjudicated decision.

​

Evidence changes, stale decisions, and retries

Truth maintenance runs asynchronously after graph commit:

1. Detection creates or updates the slot and any conflict group.
2. Resolution evaluates the current competing claim set.
3. Commit writes current values and claim statuses if the live claim set still matches the one that was evaluated.
4. Registry or alias changes can mark dependent slots stale and enqueue recomputation.

If the claim set changed during adjudication, the stale commit is skipped and the slot is re-detected. A model response cannot overwrite a newer set of claims merely because it finished later. Adjudication is memoized using a hash of the claim IDs. An unchanged claim set does not call the model again, including after a needs_review result or re-corroboration of an already-superseded value. Re-running the same work is therefore not a way to ask for a different answer. Inspect the claim-set and review state when deciding how to resolve it. Recomputation commits single-value cases directly. Multi-value cases enqueue resolution work rather than making an unbounded model call inside the recompute sweep.

​

Read an answer responsibly

Use list_truth_slots to find an entity’s slots and get_truth_slot_history to inspect competing assertions and conflict information. Both tools must be available in the selected MCP profile and authorized for the caller. For request examples and troubleshooting, continue with the truth slots guide.