Every number here has a result file and a reproduce path. Nothing on this list is a claim.
Post-rf08 fix (code-framed audit-log-delete caught by a pre-filter; the miss was published openly first). Honest framing: 1.0 on our scenario set = the benchmark found a real gap and the fix closed it — not a claim the gate is perfect.
Frontier API model on 3 tested dimensions (n=13). Low attestation/oversight scores are structural (no signing, no escalate action) — a measurement, not an accusation.
Sub-agent delegation is silent in both NIST AI RMF and the EU AI Act — verified against real primary text (EUR-Lex 32024R1689, 582K chars, 0 self-markers; NIST.AI.100-1 PDF). The publishable, gift-to-rule-writers finding.
Every citation checked for existence (article-range gate) and content against primary sources; each entry carries verified_source + scope.
Care gate refuses harm without refusing legitimate questions; academic-frame guard 17/17. We publish BOTH numbers — recall alone is an advert.
Governed = 0 violations vs ungoverned = 54.3M simulated crimes over ~649M signed episodes; Ed25519-verifiable chain.
A readiness page that only shows green is an advert. These are the cells we will NOT post yet, and why.
PROTOTYPE — do not publish as validated. The word-count signal is real (58.9 → 16.2 words on identical tasks) but the loop/hedge dictionary never fired on a capable model — the scorer is miscalibrated. Publishing it now would be a number we can't defend.
Entering any live arena needs a served model endpoint. The compliance-gateway repo is a real MCP transport but not model-inference serving — partial credit only. Serving stack first, arena second.
The two Modal measurements a buyer will ask for. A repo cannot stand in for a measurement that never ran. Fires on Modal (Science lane), not on this Mac.
Primary text is paid. Legit route: map to the free public clause skeleton, cite the paid text, never reproduce it. Buying the licence is the owner's money decision.
| site | use | read | publish | notes |
|---|---|---|---|---|
| Hugging Face | datasets · models · Spaces | LIVE | OWNER-GATED | Public API feeding the honey gate below (read-only proxy, no keys). Posting models/datasets = explicit go. |
| Kaggle | datasets · competitions · free GPU | KEY OWED | OWNER-GATED | Free GPU is real (~30h/wk T4×2/P100) but interactive notebooks, not a dispatch API. Dataset pulls need the owner's Kaggle API key. |
| GitHub (CSOAI-ORG) | code · benchmarks · CI | LIVE | LIVE | 566 repos. GovComp scenario set + runner publish here first — reproducibility is the credibility. |
| PyPI | distribution | LIVE | 313 pkgs | The measured distribution lever. |
| Eval arenas | public head-to-head | LIVE | BLOCKED | Blocked on the serving stack (above). Entering before serving exists would be theatre. |
| Standards bodies | NIST / DSIT / EU consultations | LIVE | OWNER SENDS | The gap-matrix whitepaper route: contributor, not self-appointed standards body. Drafts staged; sending is Nick's. |
Read through /api/honey — every row below passed the licence gate server-side. Default is refusal: anything not explicitly permissive comes back RED.
GREEN — MIT · Apache-2.0 · CC-BY · CC0 · BSD → may train, after signing + dating.
RED — CC-BY-NC · research-only · no-derivatives · undeclared → refused automatically. Worst licence wins on dual tags.
QUARANTINE — benchmark eval sets (MMLU, GPQA, GSM8K, HumanEval, SWE-bench, our own grid…) are never trainable at any licence. Training on a test set is the Goodhart trap that kills a benchmark — including ours.
Ingestion pipeline (pull → licence-gate → quarantine-check → SIGIL-sign → library): designed, being built in the Science lane — not yet running. This page shows the gate's policy and live reads; it does not pretend the pipeline is live.
Every post, entry, score and dataset flows through one library record: staged → owner-fired → result → signed (Ed25519) → library → SOV4 trains on GREEN-gated material only. Nothing is learned from anything the licence gate or the quarantine refused. Status, honestly: the pipeline is automatable under SOV4, not yet autonomous — today a human fires each step; the library and signing exist; the auto-loop is the build target, not the current state.
/api/sign, seeded) · Reproduce paths for every green number are in _alignment/ ·
verify a signature · MEOK OS