axis large_stringlengths 4 32 | family large_stringclasses 3
values | kind large_stringclasses 3
values | status large_stringclasses 1
value | n int64 2 261 | separation large_stringclasses 2
values | leader large_stringclasses 9
values | dataset large_stringlengths 13 33 ⌀ |
|---|---|---|---|---|---|---|---|
governance | gspc | model-comparison | MEASURED | 237 | TIE | mistral:7b (base model) | csoai/gspc-gov |
safety | gspc | model-comparison | MEASURED | 36 | TIE | gemma3:12b (base model) | csoai/gspc-agi |
provenance | gspc | model-comparison | MEASURED | 32 | TIE | llama3.2:3b (base model) | csoai/gspc-prv |
continuity | gspc | model-comparison | MEASURED | 33 | TIE | gemma3:12b (base model) | csoai/gspc-asi |
conformance | gspc | model-comparison | MEASURED | 35 | TIE | mistral:7b (base model) | csoai/gspc-mcp |
openness | gspc | model-comparison | MEASURED | 32 | TIE | gemma3:12b (base model) | csoai/gspc-oss |
machinery-conformity | gspc | model-comparison | MEASURED | 33 | UNTESTED | withheld — no signed card | csoai/gspc-mach |
care | gspc | model-comparison | MEASURED | 199 | TIE | qwen2.5:0.5b-instruct (base model) | csoai/gspc-care |
cross-reality | gspc | model-comparison | MEASURED | 32 | UNTESTED | withheld — no signed card | csoai/gspc-xr |
detector-interop | gspc | model-comparison | MEASURED | 33 | UNTESTED | withheld — no signed card | csoai/gspc-det |
art5-safeguard | gspc | model-comparison | MEASURED | 36 | UNTESTED | withheld — own model led | csoai/gspc-art5 |
swarm | gspc | model-comparison | MEASURED | 37 | UNTESTED | qwen2.5:7b (base model) | csoai/gspc-swarm |
affect | gspc | model-comparison | MEASURED | 41 | UNTESTED | withheld — own model led | csoai/gspc-affect |
jail | gspc | model-comparison | MEASURED | 71 | UNTESTED | qwen2.5:0.5b-instruct (base model) | csoai/gspc-jail-goldbank |
effect-binding | gspc | deterministic-facts | MEASURED | 261 | null | none by design (facts run) | null |
provenance-controls | financial | deterministic-facts | MEASURED | 6 | null | none by design (facts run) | csoai/gspc-provenance-controls |
reserve-attestation | financial | deterministic-facts | MEASURED | 16 | null | none by design (facts run) | csoai/gspc-reserve-attestation |
regulatory-framework | financial | deterministic-facts | MEASURED | 16 | null | none by design (facts run) | csoai/gspc-regulatory-framework |
distribution-integrity | financial | deterministic-facts | MEASURED | 16 | null | none by design (facts run) | csoai/gspc-distribution-integrity |
custody-disclosure | financial | deterministic-facts | MEASURED | 16 | null | none by design (facts run) | csoai/gspc-custody-disclosure |
ai-adoption-components | financial | deterministic-facts | MEASURED | 2 | null | none by design (facts run) | csoai/gspc-ai-economy-index |
labour-components | financial | deterministic-facts | MEASURED | 2 | null | none by design (facts run) | csoai/gspc-human-labour-index |
humanoid-labour-index | financial | deterministic-facts | MEASURED | 8 | null | none by design (facts run) | csoai/gspc-humanoid-labour-index |
lab-practice | subject | measured-subject | MEASURED | 37 | null | — | null |
company-practice | subject | measured-subject | MEASURED | 40 | null | — | null |
measurer-integrity | subject | measured-subject | MEASURED | 40 | null | — | null |
harness-fidelity | subject | measured-subject | MEASURED | 36 | null | — | null |
interop-conformance | subject | measured-subject | MEASURED | 34 | null | — | null |
humanoid-deployment | subject | measured-subject | MEASURED | 35 | null | — | null |
regulator-transparency | subject | measured-subject | MEASURED | 41 | null | — | null |
interpretability-reproducibility | subject | measured-subject | MEASURED | 33 | null | — | null |
claim-maintenance | subject | measured-subject | MEASURED | 174 | null | — | null |
serving-fidelity | subject | measured-subject | MEASURED | 38 | null | — | null |
watchdog-ingest | subject | measured-subject | MEASURED | 106 | null | — | null |
GSPC Board Export
A dated, machine-readable export of the Council of AI public measurement board. One row represents one declared axis, with its evidence type, measurement status, sample count and source dataset.
Council of AI is an operating AI measurement and evidence organisation run by CSOAI Ltd. Its external-use and distribution records distinguish documented technical reuse from download counters.
Explore the interactive board · Read the live JSON · Verify a signed record · Methodology
Snapshot at a glance
| Scope and measurement | Evidence types | Leader separation |
|---|---|---|
| 34 slots · 34 measured | 14 model-comparison axes · 9 deterministic-fact axes | 0 separated · 7 ties · 7 untested |
Root snapshot as_of 2026-11-19T03:15:46Z · source board read at 2026-11-19T06:17:38Z. Separation counts apply only to model-comparison axes; a TIE is not a win.
These are dated snapshot counts; use the live board for current state.
Load the dataset
Install datasets, then:
from datasets import load_dataset
board = load_dataset("csoai/gspc-board", split="train")
print(board.column_names)
print(board[0])
For reproducible work, pass revision="<commit SHA>" using a revision from the repository's commit history. Retain that revision with your analysis. To use downloaded files without the Hub client, read board.parquet with a Parquet reader or board.jsonl as newline-delimited JSON.
Schema
| Column | Meaning |
|---|---|
axis |
Stable measurement axis identifier |
family |
Group of related axes |
kind |
Evidence type, such as model comparison, deterministic facts or declared slot |
status |
Published measurement state; unmeasured is an explicit gap |
n |
Published count for that axis; inspect its method before comparing counts |
separation |
Statistical separation state, when applicable |
leader |
Publicly displayable leader field, when available; it is not proof of a separated win |
dataset |
Source dataset identifier, when available |
Null means absent or inapplicable, not zero. Rows are measurement summaries, not training examples. Do not combine different evidence types into one accuracy score.
Snapshot provenance
The default viewer files (board.parquet and board.jsonl) project the dated snapshot/ release: its published root is 2026-11-19T03:15:46Z, the source board was read at 2026-11-19T06:17:38Z, and its axis array records 34 axes. The source snapshot/board.json has SHA-256 20e6a20c50c77370241000d017b97703c68349f2ddac9bb3051bb3bffe1b85f0. These viewer rows are a derived export, not a new measurement or signature; use the live board for current state.
The snapshot records what was published when captured. A newer live result does not rewrite this historical export. Verify source bytes against snapshot/SNAPSHOT.json; an included root signature does not make every derived row individually signed.
Files and reproducibility
| File | Use |
|---|---|
board.parquet, board.jsonl |
Current default viewer export; the same eight-column schema |
snapshot/SNAPSHOT.json |
Capture time, source hashes, derived counts and bank observations |
snapshot/board.json, snapshot/root.json |
Captured source board and transparency root |
snapshot/gspc-axes.csv, snapshot/gspc-axes.jsonl |
More detailed axis export with evidence and bank fields |
snapshot/check-board.sh, snapshot/README.md |
Local verification instructions and release context |
manifest.jsonl |
Repository file inventory; use snapshot hashes to verify that release |
Run the included check script from a downloaded snapshot/ directory after inspecting it. It checks the captured export, not whether a claim remains admitted today. Older top-level CSVs, registers and dated notes remain historical material; they are not the default viewer or a substitute for the current snapshot manifest.
Interpretation and limitations
- MEASURED indicates available evidence for an axis, not certification or a winning model.
- TIE does not establish a separated lead. UNTESTED does not establish separation. A displayed leader can coexist with either state.
- UNMEASURED marks a coverage gap. A failed retrieval is uncheckable, never a fabricated zero.
- Results are scoped to the recorded system, method, items and date. They do not establish general safety, legal compliance or permission to execute.
- Check signatures where provided, bound intake receipts, admission and the current corrections register before quoting a card. A signature alone is insufficient.
Contribute a review
Use the Public Watchdog working groups and document tracker to find a bounded review task, inspect its evidence and prepare a contribution. Independent researchers, practitioners and affected people are welcome. Groups are forming; proposals and forum posts do not change measurements or grant authority.
Licence and citation
Board export: CC-BY-4.0, attributed to Council of AI / CSOAI Ltd. Linked datasets and third-party source material retain their own licences.
When citing this export, include “Council of AI / CSOAI Ltd, GSPC Board Export”, the repository URL, exact revision, and the capture date in SNAPSHOT.json. The source repository maintains the publishers and methods.
Earlier card and release notes
Previous dataset card preserves the earlier dated notes and file inventory. Read those counts in their historical context.
Access model
Humans: free. Browse, verify, and download everything here in your browser — no login, no paywall: free verify · full catalogue. Agents/machines: every call in this catalogue is priced over x402 — USDC, pay-per-call, no keys, no subscription. Machine entry: councilof.ai/mcp. Measurement, not certification.
- Downloads last month
- 2,887