Factual fix: lead with free verify + catalog; 19 MCP tools (14 free); methodology link replaces unavailable Zenodo DOI
Browse files
README.md
CHANGED
|
@@ -29,6 +29,8 @@ configs:
|
|
| 29 |
|
| 30 |
# CompBench — per-category capability sweeps
|
| 31 |
|
|
|
|
|
|
|
| 32 |
Per-model capability sweeps in a neutral prompt mode. Each `compbench_*.json` records the
|
| 33 |
`timestamp`, `model`, `backend` and `prompt_mode`, then `correct` / `total` / `pct` for every category
|
| 34 |
(agentic, coding, general, governance, instruction, math) and a composite. Local Ollama lane; small n per
|
|
@@ -39,7 +41,7 @@ category, stated in the file.
|
|
| 39 |
|
| 40 |
[](https://councilof.ai/api/gspc)
|
| 41 |
[](https://councilof.ai/gspc-verify)
|
| 42 |
-
[ — `npx -y csoai-gspc-mcp`. No version is pinned here: ask the registry for the current one rather than trusting a number written on a card. |
|
| 97 |
| Python reader + card verifier | `pip install "csoai-gspc[verify]"` then `csoai-gspc check` — re-derives the board totals from the axis array and exits non-zero if they disagree |
|
| 98 |
-
| This board
|
| 99 |
|
| 100 |
## What is in this repository
|
| 101 |
|
|
@@ -117,7 +119,6 @@ size, its blob hash and a direct URL, so an agent can enumerate the repo without
|
|
| 117 |
title = {CompBench — per-category capability sweeps},
|
| 118 |
author = {{CSOAI Ltd}},
|
| 119 |
year = {2026},
|
| 120 |
-
doi = {10.5281/zenodo.21991104},
|
| 121 |
note = {DOI record unavailable since 29 Sep 2026 (Zenodo account blocked; appeal pending)},
|
| 122 |
howpublished = {Hugging Face Hub, \url{https://huggingface.co/datasets/csoai/compbench}},
|
| 123 |
note = {Live board: https://councilof.ai/api/gspc. Measurement, not certification.}
|
|
@@ -229,7 +230,7 @@ Source: https://councilof.ai/signed/card_index.json (schema csoai.gspc-card-inde
|
|
| 229 |
|
| 230 |
**Verify (free):** https://councilof.ai/gspc-verify/ · method https://councilof.ai/signed/HOW-TO-VERIFY.md
|
| 231 |
|
| 232 |
-
|
| 233 |
|
| 234 |
Caveat: do not fuse card_index axes (16) with board axes (23) or Hub cells (384); they are three different counts. Per-card values are point-in-time; never compare a card's accuracy to the live board as one number. RunPod pod outputs are UNSIGNED and are not counted or cited.
|
| 235 |
<!-- csoai-archive-by-axis:end 2026-10-06 -->
|
|
|
|
| 29 |
|
| 30 |
# CompBench — per-category capability sweeps
|
| 31 |
|
| 32 |
+
> **Start here — free verify:** <https://councilof.ai/gspc-verify/> (or the free MCP tool `verify_card` at `https://councilof.ai/mcp`) · **Full catalog:** <https://councilof.ai/catalog.json> · Methodology: <https://councilof.ai/methodology/> · Measurement, not certification.
|
| 33 |
+
|
| 34 |
Per-model capability sweeps in a neutral prompt mode. Each `compbench_*.json` records the
|
| 35 |
`timestamp`, `model`, `backend` and `prompt_mode`, then `correct` / `total` / `pct` for every category
|
| 36 |
(agentic, coding, general, governance, instruction, math) and a composite. Local Ollama lane; small n per
|
|
|
|
| 41 |
|
| 42 |
[](https://councilof.ai/api/gspc)
|
| 43 |
[](https://councilof.ai/gspc-verify)
|
| 44 |
+
[](https://councilof.ai/methodology/)
|
| 45 |
|
| 46 |
`GET https://councilof.ai/api/gspc` — quote `totals.public_count`. This Hub card is a **printer** of that GET, never a second
|
| 47 |
engine. If the fetch fails the honest answer is `UNCHECKABLE` — never a fabricated `0.000`.
|
|
|
|
| 92 |
| How to verify, by hand | <https://councilof.ai/signed/HOW-TO-VERIFY.md> |
|
| 93 |
| Transparency root (Merkle) | <https://councilof.ai/root.json> |
|
| 94 |
| Every CSOAI repo on the Hub | <https://huggingface.co/csoai> |
|
| 95 |
+
| Methodology | <https://councilof.ai/methodology/> |
|
| 96 |
| MCP endpoint — read `tools/list` at viewing time (never typed on this card) | `POST https://councilof.ai/mcp` |
|
| 97 |
| MCP Registry | `io.github.CSOAI-ORG/gspc` — version not pinned here; the registry is the authority |
|
| 98 |
| npm — MCP server | [`csoai-gspc-mcp`](https://www.npmjs.com/package/csoai-gspc-mcp) — `npx -y csoai-gspc-mcp`. No version is pinned here: ask the registry for the current one rather than trusting a number written on a card. |
|
| 99 |
| Python reader + card verifier | `pip install "csoai-gspc[verify]"` then `csoai-gspc check` — re-derives the board totals from the axis array and exits non-zero if they disagree |
|
| 100 |
+
| This board, live | <https://huggingface.co/datasets/csoai/gspc-board> |
|
| 101 |
|
| 102 |
## What is in this repository
|
| 103 |
|
|
|
|
| 119 |
title = {CompBench — per-category capability sweeps},
|
| 120 |
author = {{CSOAI Ltd}},
|
| 121 |
year = {2026},
|
|
|
|
| 122 |
note = {DOI record unavailable since 29 Sep 2026 (Zenodo account blocked; appeal pending)},
|
| 123 |
howpublished = {Hugging Face Hub, \url{https://huggingface.co/datasets/csoai/compbench}},
|
| 124 |
note = {Live board: https://councilof.ai/api/gspc. Measurement, not certification.}
|
|
|
|
| 230 |
|
| 231 |
**Verify (free):** https://councilof.ai/gspc-verify/ · method https://councilof.ai/signed/HOW-TO-VERIFY.md
|
| 232 |
|
| 233 |
+
Methodology: <https://councilof.ai/methodology/> · verify by hand: <https://councilof.ai/signed/HOW-TO-VERIFY.md>
|
| 234 |
|
| 235 |
Caveat: do not fuse card_index axes (16) with board axes (23) or Hub cells (384); they are three different counts. Per-card values are point-in-time; never compare a card's accuracy to the live board as one number. RunPod pod outputs are UNSIGNED and are not counted or cited.
|
| 236 |
<!-- csoai-archive-by-axis:end 2026-10-06 -->
|