Nothing was built in year 0

[ChelseaKR]

Closes #24. Adds ADR-0006. Stacked on #26; base is fix/notes-that-claim-a-denominator.

What was wrong

YEARBUILT counted the literal 0 as a recorded construction year. 12,148 of the 132,522 DINS records hold it, 9.2% of the file and 11.9% of everything that field counted as present, and every one sat inside its 77.0% recorded share. No structure standing in a California wildfire was built in year 0, so this published an absence as a value.

0 is now a declared unknown marker on Basis.INFERRED:

	Before	After
Recorded value	102,091 (77.0%)	89,943 (67.9%)
Recorded as unknown	0	12,148
Empty cell	30,431	30,431

Why the audit could not have caught it

docs/MARKERS.md covers every field that declares a vocabulary, and tests/test_schema.py enforces that. A numeric field deciding that its zeros are measurements declares nothing at all: no markers, no codes, no domain. YEARBUILT was outside the audit by construction, and the gate reading the audit was green the whole time over a set that excluded exactly the field that needed reviewing.

tests/test_schema.py now fails when a numeric field publishes recorded zeros that docs/MARKERS.md does not cover.

The evidence

docs/MARKERS.md section 7, read off the acquired file in data/raw/ (hashes in PROVENANCE.md):

Observation	Value
Records holding 0	12,148 (9.2% of the file)
Distinct incidents / counties / fire years	186 / 45 / 8
Incident groups holding it at all	199 of 436
Incident groups where every record holds it	55 (largest: 446 of 446, Valley 2020)
Holding it and also carrying an APN	11,779 of 12,148
Holding it whose assessed improved value is also 0	5,527 (45.5%)
Holding a real year whose assessed improved value is 0	391 of 89,943 (0.4%)

Not one bad load: 186 incidents across eight fire years, scattered inside incidents as often as it takes a whole one. And it travels with the other parcel number being zero at a hundred times the rate a real year does, which is what a parcel record with nothing on file looks like. The APN is present on 97% of them, so the join found a parcel and the parcel had no year.

Confidence is high that a 0 is not a construction year and moderate on what kind of absence it stands for, which is what the audit row says.

Eight other records hold years that are also implausible (four 1, three 89, one 1543). Left as recorded values on the same reasoning as FIRE_NUM's bare 0 in section 5: eight cells is a scatter of typos, not a convention.

The other five numeric fields

Section 7 covers all of them, including the two considered and left alone:

Field	Zeros	Reading	Declared
NOOUTBUILDINGSDAMAGED	59,628	a count of none	no, the zero is the measurement
NOOUTBUILDINGSNOTDAMAGED	58,888	a count of none	no
NOOFCARSONPROPERTY	55,831	a count of none	no
NUMBEROFUNITPERSTRUCTURE	58,411	contested	no, and the reason is recorded
ASSESSEDIMPROVEDVALUE	6,613	contested	no, and the reason is recorded
YEARBUILT	12,148	no year on the parcel record	yes, marker 0, inferred

NUMBEROFUNITPERSTRUCTURE's zeros land on 39,438 single residences and 15,946 minor structures while its positives concentrate on multiple residences (1,443 of 2,007). A zero on a utility structure is a coherent observation; a zero on a single residence is not obviously one. The file does not settle it, so nothing is declared and the count stays published.

ASSESSEDIMPROVEDVALUE's zeros travel with YEARBUILT's at a hundredfold, but a parcel assessed at zero improved value is inside the field's own range, and the zeros spread across 126 incidents and every damage band. Not declared, and flagged in the audit as the entry to revisit first.

Proof each new check can fail

Measured on this tree. Against the pre-change audit:

tests/test_schema.py::test_every_published_zero_in_a_numeric_field_is_written_up[dins-coverage.json] FAILED
E   AssertionError: dins-coverage.json: these numeric fields publish recorded zeros that
E   docs/MARKERS.md does not cover, so nothing says whether a zero there is a measurement
E   or a placeholder: [('NUMBEROFUNITPERSTRUCTURE', 58411), ('NOOUTBUILDINGSDAMAGED', 59628),
E   ('NOOUTBUILDINGSNOTDAMAGED', 58888), ('NOOFCARSONPROPERTY', 55831),
E   ('YEARBUILT', 12148), ('ASSESSEDIMPROVEDVALUE', 6613)]
1 failed, 4 passed, 252 deselected in 0.06s

It is separately run against a payload no audit covers (test_the_zero_gate_rejects_a_numeric_field_nobody_reviewed) and against a numeric field with no zeros to explain (test_the_zero_gate_says_nothing_about_a_field_with_no_zeros_to_explain).

Positive controls, passing identically before and after: test_every_published_zero_in_a_numeric_field_is_written_up[perimeters-coverage.json] (green in both states, because GIS_ACRES publishes no zeros, so the gate is not just failing on everything) and test_a_recorded_zero_is_counted_as_a_value_and_as_a_zero, which is the policy for NOOUTBUILDINGSDAMAGED and is deliberately unchanged.

The gate stacked underneath this one also earned its keep: declaring the marker without rebuilding site/ failed test_the_published_note_is_the_note_the_registry_holds[dins-coverage.json], naming YEARBUILT.

Gate output

$ make verify
...
Required test coverage of 90% reached. Total coverage: 100.00%
============================= 581 passed in 8.45s ==============================
...
determinism: build/run-one and build/run-two are byte-identical (5 files)
MAKE_VERIFY_EXIT=0

Exit code read from $?.

Counts that moved with it

Twenty-eight of the fifty-four measured fields now declare a vocabulary, twelve published and sixteen inferred. README.md, docs/MARKERS.md, schema.py, render.py and the published pages all say so, and tests/test_schema.py holds them to it. site/ is rebuilt from the acquired files.

🤖 Generated with Claude Code