0018 — A measured conversion has to show what it measured

Status: accepted Date: 2026-08-03 Supersedes: nothing Relates to: 0010 (conversion registry), 0011 (conversion verification), 0017 (dataset)

Context

registry.unit_conversion.basis is a ConversionBasis — measured, published_standard, estimated, assumed_default. Decision 0011 made that column the thing a lender reads when it wants to know how much of a farmer's tonnage rests on a real weighing.

The column does not carry enough to answer that. measured means somebody put containers on a scale; it says nothing about how many, in what state, or of which container. Twelve bags in one store on one morning and four hundred bags across a district over a season produce the same value here. The §13 field exercise will produce the first of those, and the lender view in work order D7 prints "conversion trust" as though the two were interchangeable.

The second problem is identity, not strength. The field question is not only "how heavy is a bag" but "which bag". A kaveera packed tight and dried is not the container the trader means when he says the same word. Coop C in the adversarial seed exists precisely to model this: a bag there is a genuinely different container, and the mismatch has to be legible to a later reader rather than showing up as unexplained variance.

Decision

Six nullable columns on registry.unit_conversion:

Column	Purpose
sample_size	How many containers were weighed
sample_min, sample_max	The observed range
sample_stddev	Dispersion, so a wide sample is distinguishable from a tight one
condition	The state the commodity was in, e.g. dried,tight
local_label	What the container is called where it was weighed, e.g. kaveera

And one constraint:

check (basis <> 'measured' or sample_size is not null) not valid

Nullable columns plus one constraint, rather than a NOT NULL sample_size, because the other three bases legitimately have no sample. A published standard is a citation, not a weighing.

Why NOT VALID

Three rows already carry basis = 'measured': the SI identities kg→kg, tonne→kg, gram→kg. They measured nothing. There is no sample size that is true of them, and inventing one would put a false claim in the registry that append-only would never let us take back.

NOT VALID enforces the check on every subsequent INSERT and UPDATE while leaving those three unexamined. It is not a weaker constraint for new rows; it is the same constraint with an explicit, auditable grandfather list. That the list is exactly three rows and exactly the SI identities is asserted in test/registry/metrics.test.ts — "the SI definitions stay exempt" fails if a fourth grandfathered row ever appears.

FINDING: ConversionBasis conflates definition with measurement

kg→kg is not a measurement that happened to come out at 1.0. It is a definition. Putting it under measured means the enum has no way to say "this value is true by construction" and forces the grandfathering above.

A fifth value — definitional — would separate them cleanly and remove the need for NOT VALID entirely. ConversionBasis lives in packages/schema/src/enums.ts, which is vendored read-only. This is a candidate for schema v0.3. Until then the constraint and its comment carry the finding.

Metrics

/metrics gains kernel_thin_sample_kg_total: tonnage whose conversion is measured but rests on fewer than ten containers.

Ten is not a statistical threshold and is not defended as one. It is the point below which the word "measured" is doing more work than the evidence supports. The adversarial seed's coop A weighs twelve bags, so it sits just above the line — which is the intended reading: twelve is defensible, and the gauge should not punish it. The number is a constant in operations.controller.ts so that moving it is a code change with a diff rather than configuration drift.

Tonnage on a thin sample is reported separately from kernel_assumed_conversion_share rather than folded into it. They are different failures: one is "we did not weigh", the other is "we barely weighed", and an operator who conflates them cannot tell which to fix.

Consequences

• New measured factors cannot be written without a sample size. This is a rejection, not a quality flag, and it is a departure from P6 — but P6 governs a farmer's account of what happened, and a conversion factor is our own administrative record. Refusing to record our own unevidenced claim is not the same as refusing a farmer's.
• condition and local_label are free text and unvalidated. Constraining them would require a vocabulary that the field visits have not produced yet.
• The registry's read path selects the new columns, so anything already reading UnitConversionRow gets them without a further migration.