This repository collects GSPC evaluation banks, candidate subject-family banks and integration drafts. Use it to inspect the measurement inputs and reproduce a supported model evaluation. Current admitted results are published on the GSPC board.

Browse the Files tab for the bundle. Its banks and registry manifests have different schemas, so the automatic tabular viewer is disabled; it previously displayed one registry manifest as a one-row test dataset.

Start with the supported runner

The maintained runner and its grader live together in the Council of AI source repository. It imports the canonical grader; it does not substitute a generic exact-match scorer.

git clone https://github.com/CSOAI-ORG/councilof-ai.git
cd councilof-ai
python3 harness/bench/run_bench.py --list

To score answers you already hold, place one JSON object on each line of answers.jsonl, using index for the zero-based bank item and answer for the exact model reply. Include every item; retain the original replies.

python3 harness/bench/run_bench.py --axis governance \
  --model your-model --answers answers.jsonl --out result.json

This mode makes no model calls. The runner downloads the selected published bank and records its digest, prompts, parsed labels, missing answers and denominator. A live run instead needs your own configured model endpoint and budget. Save the source commit and result file for reproduction. No leaderboard submission or board admission happens automatically.

What is included

Qualified Inspect task

The canonical source now includes an optional Inspect task, qualified with Inspect AI 0.3.259. It imports the canonical grader, requires a local bank and its SHA-256, and retains original replies, parsed coverage and both accuracy denominators. Follow the installation and run instructions.

The qualification executes the actual framework in an isolated environment with synthetic control replies. It checks correct, wrong, malformed, empty and zero-parsed responses, plus refusal of changed banks and unsound label menus. These controls establish software behaviour; they are not model measurements, benchmark registration or leaderboard admission. The older bundled draft below remains historical and is not the maintained task.

Bundle files

Path	Purpose
`banks/`	Suite definitions and candidate subject-family banks; inspect each file's scope and status.
`registry-packages/`	Draft mappings for external evaluation frameworks and registries.
`src/csoai_gspc/inspect_tasks.py`	An experimental Inspect integration; it is not qualified as a replacement for the canonical runner.
`outward-evidence-20261026/`	Dated reports and worked examples, retained as historical material.

Integration status

The earlier page advertised a pypi/ install directory, lm_eval/tasks/gspc, kaggle/kaggle_benchmark_notebook.py and openenv/openenv-manifest.json. Those paths are absent from this repository. Use the supported runner above. The external integration drafts require qualification against their current APIs, actual dataset names/splits and the canonical grading rule before they can be described as runnable benchmark integrations.

The presence of a benchmark tag or registry draft is not acceptance by Hugging Face, Kaggle, Inspect, Lighteval or another leaderboard. Benchmark registration, completed evaluations and external acceptance are separate steps.

Interpret results correctly

A candidate bank is an input to evaluation, not evidence that its subject has been measured or admitted.
Compare complete runs only when bank bytes, prompts, expected labels, grader version and evaluation conditions agree. Preserve unparsed and missing outcomes.
Report task coverage, uncertainty, model/version, tool access, budgets and observation dates. Keep model comparisons separate from deterministic facts and subject assessments.
A significant paired result is scoped to that declared comparison. It does not establish general superiority, certification, safety or permission to execute an action.
Keep development data separate from an independently reviewed, sealed evaluation set. Public-bank optimisation cannot establish performance on unseen tasks.

Current board · Methodology · Verify supported signed records · Corrections

Published by Council of AI — Council for the Safety of Artificial Intelligence, operated by CSOAI Ltd. Existing file licences remain applicable. Measurement is not certification.

RUNBOOK (verified 2026-11-03): the lm-eval tasks now actually execute

Three bugs found by the first real execution (fixed in lm_eval_tasks/ in the next commit):

{{prompt}} did not exist in any dataset (fields are scenario/text/item/request/case/operation/tool) — every task crashed at render.
The prompt said "Reply with EXACTLY ONE of:" followed by NO labels (enum was never injected) — the model had no answer space.
(usage) lm-eval POSTs to base_url verbatim — point it at the completions/chat endpoint; thinking-mode models (qwen3) return empty content via chat API — use a non-thinking model or disable thinking.

Working invocation (any OpenAI-compatible server):

lm_eval --model local-chat-completions --apply_chat_template \
  --model_args base_url=http://HOST:PORT/v1/chat/completions,model=<model>,num_concurrent=1 \
  --tasks gspc_governance --include_path lm_eval_tasks --output_path out/

First verified run: gspc_governance x 30 items vs qwen2.5-coder:1.5b (Ollama) -> exact_match 0.2667 +/-0.0821 (majority-baseline class, as expected for an un-tuned small model — the pipeline is the deliverable).

FIRST FULL-SUITE RUN (2026-11-03): 13 tasks x 30 items, aggregate exact_match 0.5103 +/-0.0235

Model: qwen2.5-coder:1.5b (non-thinking, chat mode, temperature 0, greedy) via Ollama local-completions. Raw un-tuned small model — this row is the harness validation baseline, not a leaderboard entry.

Task	exact_match	±	n
gspc_swarm	1.000	0.000	30
gspc_detector_interop	0.767	0.079	30
gspc_safety	0.700	0.085	30
gspc_care	0.567	0.092	30
gspc_machinery	0.533	0.093	30
gspc (aggregate)	0.510	0.024	390
gspc_openness	0.500	0.093	30
gspc_provenance	0.467	0.093	30
gspc_art5_safeguard	0.467	0.093	30
gspc_conformance	0.433	0.092	30
gspc_continuity	0.367	0.089	30
gspc_affect	0.300	0.085	30
gspc_governance	0.267	0.082	30
gspc_cross_reality	0.267	0.082	30

Caveats, stated: gspc_swarm is a retired single-label bank — its 1.000 is trivial and excluded from any interpretation; n=30/task; one small model; chat framing without per-axis specialist overlays. Overlay results from our own harness (e.g. care 81% with the care overlay) are consistent with the deltas this table implies. Third parties: run the same tasks on any model via the runbook above and compare.

24-HOUR REPRODUCTION RUN (2026-11-04): aggregate 0.5051 vs 0.5103 — REPRODUCED

Same model, same tasks, same settings, 24 hours apart: aggregate exact_match 0.5051 ±0.0236 vs the published 0.5103 ±0.0235 (delta -0.005, well within joint stderr). 12 of 13 tasks reproduced within single-item tolerance (±1/30); only gspc_affect moved +0.100 (3 items). The published row is stable — third parties re-running get the same answer.

Add GSPC to your AI platform (Claude, ChatGPT, Cursor, Gemini, VS Code...)

The GSPC stack is a remote MCP server: https://councilof.ai/mcp (streamable HTTP, MCP 2025-03-26 — verified handshake, server csoai-gspc-mcp v1.4.4, 49 free tools + 5 paid).

Copy-paste configs for Claude.ai / Claude Desktop / Cursor / Gemini CLI / VS Code / ChatGPT endpoints: ADD_TO_PLATFORM.md in the gspc-benchmark repo.

Embed the live sidebar anywhere: <iframe src="https://csoai-gspc-sidebar.hf.space" width="440" height="720"></iframe>

Downloads last month: 369

Total file size:

1.25 MB

GSPC benchmark suite — banks, methods and evaluation tools