The words this category needs, defined by the people creating it.

The category

Agreement layer

The layer above an enterprise's systems of record where those systems are compared to one another. Distinct from integration, which moves data between systems, and from a warehouse, which stores it. The agreement layer asks a different question: when these systems describe the same object, do they say the same thing, and if not, which one counts.

System of record

An authoritative source for a class of fact. Each one is internally correct. The problem this category addresses is that being internally correct does not make a set of them collectively consistent.

Disagreement

Two or more sources reporting different values for the same canonical object. Not an error, and not a data-quality defect. A disagreement is a fact about the estate that must be preserved rather than resolved away.

Contested tie

Two sources reporting different values for the same object at the same timestamp, so recency cannot arbitrate. The system must surface the conflict rather than select a winner. This is the single most important behaviour in the category: a recoverable error is survivable, a confident wrong number is not.

Canonical object

The single representation of a real-world thing across every system that describes it, holding each source's latest reported value rather than a merged average or a last-writer-wins overwrite.

Evidence

Evidence pack

The complete trace behind a material figure: claim, source, timestamp, transformation, rule version, calculation, and the record reference. Its defining property is that it reconstructs later, by someone who was not there.

Provenance

The recorded origin and derivation path of a value. Provenance answers where a number came from. Evidence answers whether you should believe it.

Decision ledger

An append-only record of decisions, each carrying the authority that made it, the alternatives that were considered and not chosen, and the timestamp. The alternatives matter as much as the choice: a ledger that records only outcomes cannot explain a decision months later.

Receipt root

A content hash over a shipped tree, re-derivable with no vendor credentials. A release with a receipt root can be checked. One without can only be believed.

Outcome record

What was expected at the point of decision, what was measured afterward, and the gap between them, retained rather than absorbed. Without it a system can be confidently wrong indefinitely.

Governed intelligence

Claim adjudication

Ranking competing answers to the same question by the strength of what supports them rather than by the confidence of whichever system produced them.

Claim ranking order

Source evidence, then deterministic calculation, then verified derivation, then model consensus, then single model output. Model agreement ranks below arithmetic. Two models agreeing is a fact about the models, not about the world.

Truth categories

Observed, derived, inferred, assumed, scenario, recommendation, decision, action, outcome. Every material output carries exactly one. Representing one category as another is the defining failure mode of enterprise AI.

Refusal

A recorded declination to answer, with its reason. A system that has never refused anything has never been tested. Refusals belong in the record alongside answers.

Status labels

SEALED

In the current release, content-hashed, and the harness re-derives it.

CONFORMED

Proven against a published contract by executing the suite. Customer-environment validation is recorded per engagement.

SYNTHETIC

Runs on a synthetic estate. Figures are illustrative and are not customer results.

MODELED

Computed from inputs. Not measured from an outcome.

FOUNDER-GATED

Complete, and waiting on a decision only the principal can take.

ON THE CALENDAR

Committed and scheduled. The window is stated when it is set.

IN THE REGISTER

Specified, not built, and listed publicly so nobody discovers it late.

NOT CLAIMED

Deliberately not asserted. Published as prominently as everything else.