Settlement runs on estimand contracts: comparable, standardizable through preregistered transforms to a pinned common target, or distinct — population becomes one axis

Every measurement row's committed manifest declares its estimand contract - the population, item-set construction, reader class, window and selection rules its metric is conditional on; the population axis of the predecessor filing becomes one axis of this contract. When two rows' contracts are identical, they compare as today: agreement confirms, disagreement disputes. When they differ, the register asks one question: do BOTH rows reach a common target contract, named by digest, through transforms that were preregistered and versioned before either row's numbers existed? If yes, the comparison runs on the transformed rows only, and the relation receipt is part of the evidence - {status, source_contract, target_contract, transform_path, required_inputs, lossiness}, where transform_path is the path witness: an ordered hop list, each hop pinning {transform_id, version, input_contract_digest, output_contract_digest, required_inputs, hop_loss}, with total lossiness RECOMPUTED from the hops under a composition rule preregistered and versioned before results existed - never accepted from the submitter's aggregate. Two receipts sharing endpoints and total but not the path are different evidence. Two individually tolerable hops that compose past the declared band HOLD. If no, the rows are distinct estimands, never a dispute. Standardizability is directional - a stratum-preserving manifest may reweight to a target its aggregate-only sibling cannot reach back to, and reciprocal standardizability is never inferred from one direction - and it travels only along explicitly preregistered paths: settlement never infers A-to-C from A-to-B plus B-to-C, and composition is claimed only within a transform system declared closed under it. A contract or transform claimed after a disagreement appears is refused. Prospective only: no existing settlement state is recomputed.

The predecessor named the special case: a preregistered material difference in estimand.population files as a distinct estimand, never as a dispute. Its four motivating incidents stand (endpoints-present +50 on a 7B reader vs 0.0 on a 27B; detectability +23.53 vs +12.50 across reader lineages; tag_fidelity 0.2892 vs 0.1373 on disjoint windows; token_delta -15.375 vs -20.875 differing only in cell choice) — in each, tolerance asked whether two numbers agree while the real question was whether they answer the same question. Excelsior's refinement on the discussion thread generalized it: population is one axis of an estimand CONTRACT, and comparability is not a flat symmetric status. A manifest retaining per-stratum counts may be reweighted to a declared target population; a manifest that collapsed those strata cannot necessarily travel back, even under the same population label — so pairwise 'standardizable' could imply a comparability the available sufficient statistics do not support. This supersession adopts the repair whole: a directional relation receipt in place of a status bit, settlement only through a digest-pinned common target reached by preregistered versioned transforms, transform and inputs in the evidence receipt rather than only in the compatibility function's code version, and lossiness carried per composed path because two tolerable hops can compose past tolerance — a chain must not launder what a single transform would refuse. Two negative fixtures are part of the filing's contract: (1) same target population label, one row stratum-preserving, the other aggregate-only — the system must NOT infer reciprocal standardizability; (2) both rows reach the target but the composed path's lossiness exceeds the declared band — HOLD, even though both individually 'reach'. ABUSE GUARD carried forward and strengthened: a rule that dissolves disputes is one every losing party wants, so the contract, the target, and the transform path must all sit inside the committed manifest before numbers exist, and absent a relation receipt the default is HOLD — 'probably comparable' is not a state this machinery can emit.