MEASURED, NOT CERTIFIED

CSOAI GSPC — 13-axis benchmark suite

13 axes · 6 canonical domains · published for measurement, not certification · UNMEASURED is reported, never hidden

The GSPC suite is CSOAI's open benchmark library for AI governance measurement. Each axis is a distinct measurement of a specific risk domain that buyers, auditors and regulators have raised but open tools have not measured. The same 13 axes are mirrored on Hugging Face as csoai/gspc-* datasets and on Kaggle as nicktempleman/gspc-* datasets with the same axes. The canonical greenfield names are tuned to unforgeable domains (one per axis), so a citation in any paper, deployment, or audit can be checked against a single source of truth.

13 axes

#	Slug	Title	Domain	Status	Backbone
1	gspc-mach	Machinery Regulation safety classification	EU AI Act / Machinery Reg	MEASURED	GSPC / published
2	gspc-asi	Post-quantum signing classification	continuity	MEASURED	GSPC / gspc-asi
3	gspc-art5	Article 5 prohibited practices	EU AI Act Art 5	MEASURED	GSPC / gspc-art5
4	gspc-det	Detector interoperability	conformance	SPEC	GSPC / gspc-det
5	gspc-agi	Agentic safety refusal battery	safety	MEASURED	GSPC / gspc-agi
6	gspc-prv	Article 50 provenance survival	EU AI Act Art 50	MEASURED	GSPC / gspc-prv
7	gspc-oss	Open-source license scope battery	openness	MEASURED	GSPC / gspc-oss
8	gspc-mcp	MCP tool conformance battery	conformance / MCP §5	MEASURED	GSPC / gspc-mcp
9	gspc-swarm	Multi-agent coordination safety	swarm	MEASURED	GSPC / gspc-swarm
10	gspc-arena-results	Per-model results across all 13 axes	composite	MEASURED	GSPC / gspc-arena-results
11	gspc-gov	EU AI Act risk-tier battery	governance	MEASURED	GSPC / gspc-gov
12	gspc-care	Care-cost over-refusal battery	care	MEASURED	GSPC / gspc-care
13	gspc-xr	Cross-reality checks	cross-reality	SPEC	GSPC / gspc-xr

11 axes are MEASURED with published batteries, 2 are SPEC stage (gspc-det, gspc-xr). Every axis has a corresponding page on this site and a corresponding dataset on Hugging Face + Kaggle.

Six canonical domains (the greenfield brand names)

The 13 axes fold into 6 canonical governance domains. Each domain has a unique domain name (so a citation in any paper is unforgeable) and a greenfield brand site.

The 6 domain names are stored in GSPC_NAMES.json (single source of truth). 7 axes map directly to one domain; the other 6 axes are specialised batteries within a domain.

What the suite is NOT

• It is not a certification. CSOAI does not certify safe. The site states this in plain prose on every axis page.
• It is not a leaderboard. Per-model results are published as a table, not ranked — there is no single "winner" across governance, safety, provenance, continuity, conformance, openness.
• It is not a complete detector. UNMEASURED is reported in every result, never hidden.
• It is not a 100% coverage claim. Each axis has a finite battery size; the delta is in the published items.

How to run an axis

Every axis page on this site offers the same battery as a runnable grader that runs in the browser — no backend, no API key, no cost. The same items the canonical sovereign model (Council-34) answered are scored by the same deterministic rule, so the two numbers sit on one scale. The estate's own usable_n rule applies: below n=30 the page refuses to quote an interval, including for sovereign.

Cross-references

Surface	Field	Mirror
Hugging Face	13 axis datasets + 12 axis Spaces	csoai/gspc-* model + dataset org
Kaggle	12 active axis batteries + 1 SPEC + 8 deprecated stubs	nicktempleman/gspc-*
HF Spaces	15 axes (12 static + 3 demo)	csoai/oowm-router-demo, csoai/oowm-routing-matrix
OpenRouter	sweep harness (do not run unattended — costs credits)	see scripts/openrouter_gspc_sweep.py

OpenRouter sweep — straight, not gamed

The full sweep across all 13 axes on every available OpenRouter model is the natural next step (we have credits, the harness is writeable). It is NOT run unattended. The harness is shipped at scripts/openrouter_gspc_sweep.py with the 13 axes wired to the public OpenRouter chat completions endpoint. To run it:

1. Set OPENROUTER_API_KEY in env.
2. Pick the model roster from openrouter_gspc_sweep_config.json (default: 8 frontier models).
3. Run, watch the JSONL log, stop if rates spike.

Default behaviour: 13 axes × 8 models × 1 epoch per item = ~104 runs. Estimate at $0.002-$0.05 per run depending on model. Total ≈ $0.30-$5.00 per full sweep.

Honest caveats

• 2 of 13 axes are SPEC (gspc-det, gspc-xr) — not yet measured. They are listed because the regulatory frame demands them; the suite admits the gap.
• 8 deprecated stub datasets on Kaggle still receive downloads (130 total since 6 Aug 2026 banner update). The deprecated URLs are valid redirects; do not cite them as canonical.
• Per-axis arena results (gspc-arena-results) are n=20/cell with Wilson confidence intervals — about half the per-axis winners are statistical ties. Treat results as indicative of specialization, not final rankings.
• There is no automated-kill switch for overclaiming. The rule is human-disciplinary: if you cite a score, cite the axis, the items, the seed, the model, the date.

Provenance: this page is part of CSOAI's GSPC benchmark suite (13 axes). See huggingface.co/csoai for the full sovereign benchmark collection. The 13 axes are mirrored on Kaggle at kaggle.com/nicktempleman.