A grounded field says whether it was measured or extracted

FieldOrigin recorded three origins — GROUNDED, SYNTHESIZED, CALIBRATED — and GROUNDED covered two claims that are not the same claim. Reading the real shards showed how far apart they are: a gss row's age_bracket is a survey answer, carrying assignment_type: direct with empty evidence because the instrument is the source; an amazon row's cuis_vegan is also direct, at confidence 0.8, and its evidence is the review sentence "Not vegan but still good". One is a person answering a question. The other is a model reading a sentence. FieldOrigin gains a fourth value so it can say which: MEASURED, EXTRACTED, SYNTHESIZED, CALIBRATED.

The tier cannot be a property of the source. gss and amazon both report direct, so assignment_type alone does not separate them; and within a single source it varies — wiki grounding is 49% unsupported with a median confidence of 1.00, amazon is 52% summary_inference. A source-level label would be false for exactly the two sources where the distinction matters most. The adapter derives the tier per field from the source and the assignment type together, which is the adapter's knowledge in the same way the packed encoding is.

Considered options

A parallel evidence map beside origins was rejected: two maps per persona must be kept in sync, and the invariants that protect the engine — Population already enforces that gates run only on attributes it declares grounded — become statements about two fields at once, which is where they get written wrongly. Leaving GROUNDED undivided was rejected because every claim the engine makes rests on the distinction; without it, a distribution gate cannot say whether it compared measured values or a model's readings. SCHEMA_VERSION is 1.0.0 and unreleased, so widening the enum costs a mechanical sweep now and a migration later.

Consequences

The tier grades a claim; it does not gate one. Requiring gates to run on MEASURED attributes would refuse every study built on SyntheticCoresetSource, which the quickstart and the whole suite run on and which FINAL_ARCH.md §4 keeps deliberately. Instead the GateReport carries the weakest tier among its gated attributes, following the rule GateReport.reference already sets: a pass is never read as a stronger claim than its weakest gate supports.

One refusal is kept. A gate may not claim GateReference.CATEGORY_TARGETS — "this population matches the measured category" — on attributes that are not MEASURED. The strong claim requires the strong evidence; the weak claim is permitted and labelled.

Audience filters are graded the same way rather than restricted, so one rule covers gates, audiences and sources instead of a rule with exceptions. KNOWN_PERSONA_SOURCES gains the tier mapping, and a study mixing sources takes the weakest tier present.