csoai/gspc-fact-table

GSPC fact table — every measurement the Council of AI holds, joined

Start here — free verify: https://councilof.ai/gspc-verify/ · Full catalog: https://councilof.ai/catalog.json · Methodology: https://councilof.ai/methodology/ · Live board: https://councilof.ai/api/gspc · Measurement, not certification.

What this is

One long-format table of every measurement the Council of AI holds, joined from the per-cell evidence, with the item universe it refers to and an index of the cells it came from.

file	rows	what a row is
gspc_measurements.jsonl / .parquet	20517	one measurement: a model, an axis, an item, what it answered, and whether it was right
gspc_items.jsonl / .parquet	807	one item of a bank: its text digest, its expected label, the bank it came from and its digest
gspc_cells.jsonl / .parquet	677	one evidence cell: the file it came from, that file's own sha256, the model, the axis and the cell's accuracy
MANIFEST.json	—	the counts, the source digest, and what this dataset does NOT contain

Why it exists

The measurements were published as 677 per-cell files whose NAMES carry only a truncated axis and a content hash — the model is inside the file. A reader who wanted "how did model X do on axis Y" could not find the cell without opening every file, could not join the cells without reading all of them, and could not read the corpus as a table. The projection that did exist (csoai/gspc-peritem-rows-2026-08-12) is a lossy, frozen subset: 15,580 rows over 19 models, against the 20517 rows over 57 models the estate actually holds — and it dropped the prompt, the model revision, the provider route and the elapsed time, which are the fields that make a row auditable.

How to re-derive a number

# every cell's accuracy, from the table alone
gspc_cells.jsonl -> n_ok / n_graded

The test that proves it: scripts/measurements/gspc-fact-table.node-test.mjs — every cell's accuracy re-derives from the fact table, every measurement resolves to an item, the item counts re-derive, the table has exactly as many rows as the evidence has, every cell is indexed once with the digest of its own bytes, and no canary string reaches the output.

What this does not establish

• That a model is good, or better than another. One run per cell, on one bank, at one time.
• That a row is right. It is what the grader recorded, with the grader's own label.
• That the cells are comparable across axes. They are different banks, different item counts, different item selection. The axis is in every row; a cross-axis comparison is the reader's judgement to make and defend.
• That the corpus is complete. 1290 rows carry no model (their cell is knowable only by opening the file) and 2428 carry no grade. Both counts are in MANIFEST.json. They are published as they are, not dropped.

Provenance

Source: public/interop/mill-evidence/ in https://github.com/CSOAI-ORG/councilof-ai — 677 cell files, 20517 source rows, source digest 31fa74faf8a1a4c2bd6092ec1526bf1c746bdbc6f756403ffa0d7776e074814f. Generated by scripts/measurements/build-gspc-fact-table.mjs; generator commit 95273309f3479eb43909c2f2aa472c3ba3d652f2. Licence CC-BY-4.0.