cwhr13-hex-fractions STAC note states ~96.5 M-ac statewide total — computed from H3's GLOBAL average res-10 cell area; the asset measures 102.8 M-ac (no acres are missing)

[cboettig]

Summary

The cwhr13-hex-fractions STAC note states a statewide mapped total that is wrong by ~6%, and it is reaching partners: the ca-30x30 agent quotes the note verbatim to users, and one model reverse-engineered a bogus per-cell constant from it. The underlying data is fine — this is a metadata/arithmetic defect, not an ingest loss.

Current note:

"Note: the statewide mapped total (~96.5 M-ac) matches the mode build; the residual gap to nominal CA area is coastline/centroid quantization, now explicit as the per-cell 1-SUM(frac) remainder + the whr13num=0 share."

What the data actually says

Measured against the live asset (cirrus MCP, mcp-data-server:main):

quantity	value
SUM(frac × h3_cell_area(h10,'km^2')) × 247.105, all classes	102,834,000 ac
same, excluding whr13num=0	102,742,000 ac
whr13num=0 (nodata)	91,371 ac — not a ~5M-ac gap
distinct h10 cells	25,944,500
mode build (cwhr13/hex), SUM(h3_cell_area) over cells	102,834,000 ac — identical
CA-Nature ecoregion extent (SUM(Shape_Area), ecoregion.parquet, EPSG:3310)	101,498,000 ac

So the cell union is 416,155 km² vs California's 410,749 km² — +1.32% MORE, which is what you expect when boundary cells overhang the state edge. There is no deficit to explain, and no acres are missing.

Where ~96.5 M-ac comes from

It is the cell count multiplied by H3's global average res-10 hexagon area instead of each cell's actual area:

25,944,500 cells × 0.0150475 km²  = 390,400 km² = 96,469,758 ac   ≈ the note's "~96.5 M-ac"
25,944,500 cells × 0.016040  km²  = 416,155 km² = 102,834,000 ac  ← actual, per-cell

California's res-10 cells are ~6.6% larger than the global mean, so the global constant understates by ~5%. Note the direction: this makes it look like ~5M acres are missing, and it produces exactly the "~5% undercount" that ca-30x30#73 was opened about in June.

Three separate consequences make it worth fixing rather than leaving:

1. The note contradicts the asset. "~96.5 M-ac" vs 102.8 M-ac measured.
2. The stated mechanism is false. There is no "residual gap to nominal CA area" — the union is larger than CA. Attributing a nonexistent gap to "coastline/centroid quantization" gives readers a wrong causal story, and 1-SUM(frac) + the whr13num=0 share (91k ac) do not account for 5M acres.
3. It propagates. ca-30x30 agents quote it ("classes sum to ~96.5M acres mapped, not the nominal state area"), and in a 2026-07-31 benchmark run qwen reported "H3 resolution-10 cells × 3.718 acres/cell (rough constant for h10)" — that is 0.0150475 km² × 247.105, i.e. the global average, almost certainly inherited from this note. The true California mean is 3.9636 ac/cell.

Suggested fix

1. Correct the note to the measured total (~102.8 M-ac, both builds identical) and drop the "residual gap"/quantization explanation.
2. If a comparison to California's area is wanted, state the direction honestly: the res-10 cell union covering California is ~1.3% larger than the CA-Nature ecoregion extent because boundary cells overhang the state edge — and note that per-class percentages are unaffected, since a uniform area scale cancels in a ratio (which is why the 12-of-13 class agreement with the 2025 assessment holds regardless).
3. Keep the genuinely useful part: 1-SUM(frac) per cell and the whr13num=0 share are real diagnostics; they are just not a 5M-acre explanation.
4. Check the generator, not just the text. If any ingest/QA step computes a "mapped total" from a nominal/global H3 cell area, the same understatement will appear anywhere that code runs. I scanned cwhr13, cwhr, conserved-areas-terrestrial-2025, and ecoregion — only cwhr13-hex-fractions carries this sentence today, so it may be hand-written rather than generated. Worth confirming which, since a generated one would need a code fix.

Verification

-- expect ~102,834,000 for BOTH, and equality between them
SELECT ROUND(SUM(frac * h3_cell_area(h10,'km^2')) * 247.105)
FROM read_parquet('s3://public-ca30x30/cwhr13/hex-fractions/h0=*/data_0.parquet', hive_partitioning=true);
SELECT ROUND(SUM(h3_cell_area(h10,'km^2')) * 247.105)
FROM read_parquet('s3://public-ca30x30/cwhr13/hex/h0=*/data_0.parquet', hive_partitioning=true) WHERE whr13num <> 0;

Filed from ca-30x30 regression work while checking whether our ingest had lost acres (it has not). Related: ca-30x30#73 (opened on the apparent ~5% undercount — this is its explanation), #413 (the separate, real pinyon-juniper footprint divergence), #417 (provenance).