Scores

Council of AI · measurement, not certification

benchmark corpus — not the governance board

Benchmark corpus leaderboard

There is no single “best AI”. A model strong on one axis is often weak on another; hiding that is the malpractice this board exists to correct. So there is no forced overall rank here. Sort by the axis you care about — safety, provenance, refusal, jailbreak-resistance. Every measured cell links to the Ed25519 card behind it. Every unmeasured pair is shown empty, never as a zero.

64
models (rows)
16
signed axes (columns)
335
measured cells
1,024
possible cells
33% covered
335
signed cells
every cell signed
23 axis · 23 measured
governance board
separate instrument · /api/gspc
Rank by axis
Regulator lens (pointers only)
Find a model
Model
ARC (n=30)
best 100.0%
95% 88.6%–100.0%
23 models
Care (bank)
best 0.0%
interval unavailable
23 models
Care · help
best 100.0%
interval unavailable
27 models
Care · protect
best 77.4%
interval unavailable
27 models
Gov (bank)
best 0.0%
interval unavailable
23 models
GSM8K (n=30)
best 90.0%
95% 74.4%–96.5%
23 models
Conformance
best 63.6%
interval unavailable
21 models
Continuity
best 46.2%
interval unavailable
21 models
Governance
best 65.0%
interval unavailable
21 models
Openness
best 69.2%
interval unavailable
21 models
Provenance
best 85.7%
interval unavailable
21 models
Safety
best 100.0%
interval unavailable
21 models
Jailbreak resistance
best 56.3%
interval unavailable
8 models
MMLU (n=30)
best 96.7%
95% 83.3%–99.4%
23 models
SWAG (n=30)
best 100.0%
95% 88.6%–100.0%
25 models
Swarm robustness
best 44.4%
interval unavailable
7 models
——————————————
——————————————
——————————————
——————————————
——————————————
——————————————
————————
————————
——————————————
——————————————
——————————————
——————————————
——————————————
——————————————
——————————————
——————————————
—————————————
—————————————
——————————
——————————————
——————————
——————————
——
————————
———————————————
——————————
——————————————
————————
————————
——
——————————
————————
——————————————
————————
——————————
——————————
————————
——————————————
——————————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
——————————
name withheld
————————
name withheld
————————
name withheld
————————
name withheld
——————————

◈ = a signed cell; click it to recompute its Ed25519 card in your browser. Row order is presentation, not a separation claim — two adjacent rows may be statistically indistinguishable. Whether a lead is real is the governance board’s separation test, below, not a column sort.

Two instruments, never summed

This grid is the benchmark card corpus: 64 models measured on 16 benchmark axes, one signed card per cell. The governance board (/api/gspc, 23 slots) is a different instrument measuring governance axes, and it is the authority on statistics for its own measurements — n, Wilson intervals, and the separation test. The two axis sets are different on purpose and are never added together. Where a benchmark axis has a governance twin, the column links across for context but does not borrow the board’s cohort statistics. “THESE ARE NOT THE PUBLIC BOARD'S AXES. The cards carry benchmark axes; the public board carries governance axes measured by a different instrument. The two sets are different, their counts are different on purpose, and they are never added together. The board's count authority is GET /api/gspc.”

How this is graded — and what it does not establish

  • Deterministic grading. Gold labels on frozen banks. A cell is a proportion correct, graded by exact rule — no model judges another model.
  • Matrix derivation is reproducible. Each populated cell comes from one signed card whose sha256 is its id; this index is derived by reading those cards. That does not by itself make every underlying benchmark run reproducible.
  • Absence is shown. An unmeasured pair is empty, not zero. 335 of 1024 possible pairs are measured.
  • Not a certification. A cell is one score on one small bank on one date. It does not establish that a model is good, safe, legal, or better than another.
  • A sort is not a verdict. Ordering the fleet by an axis does not mean the top row beat the second — separation is tested by the governance board, not implied by rank.
  • Small banks. Several axes are read on ~30 frozen items. Treat a single cell as a signal, not a certificate.

The board audits itself

A governance board that will not state its own limits is not governing anything. Here are ours, derived live.

  • Coverage is partial by design. Only 33% of possible model×axis pairs are measured (335 of 1024). The empty cells are shown, not hidden.
  • Some axes have small n. 2 of the measures rest on fewer than 10 models each: Swarm robustness (7), Jailbreak resistance (8).
  • Per-cell n is not on the compact card. Cards carry accuracy, not the bank size, so an inline Wilson interval appears only when the benchmark axis declares its own cohort size, or an API record identifies the exact same signed leader measurement. Related governance cohorts are never substituted.
  • The living stamp is UNCHECKABLE by a stranger. The signed cards are verifiable; the claim that this corpus is the current one rests on the board’s attestation, not on cryptography a visitor can re-run. See the corrections ledger.
  • 4 model names are withheld. A retired internal brand is indexed under a neutral key; the measured work is kept and counted, and the real name still lives in the card body under the signature.
  • The board is a different instrument. Its 23-slot governance count is not this grid’s 16-axis count, and the two are never summed.

Paste a card SHA → verify it here

Recomputes the card’s sha256 and checks the Ed25519 signature against the key in did:web:csoai.org#card-attestation-1, in your browser. Nothing is sent.

Council of AI measures AI systems against the rules that govern them and signs the result. It issues no conformity marks and no certification. Verification is free forever; a grade is never sold.

Answers come only from published measurements; nothing you paste is uploaded.