model_calls int64 | results list | schema string | status string |
|---|---|---|---|
0 | [
{
"count_confirmed": true,
"exit_code": 0,
"expected_tests": 27,
"pass": true,
"stderr_sha256": "fbd082554124c33e479ebc0d232135ab5a918f1a1d6058ad0f18c0d302422ad2",
"stdout_sha256": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
"test": "evaluator_unit",
"expected... | mi080.takeover.no-model-tests.v1 | PASS |
MI-080 v3 — blind model execution and accounting
P3-C one-shot generation/accounting completed; predictions frozen and unscored. Target: Gemma 4 E2B IT UD Q8_K_XL; runtime: llama.cpp b10534, CPU, 4 threads. This release records actual model calls, unlike historical Phase 2 synthetic text-only results.
| Measure | Result |
|---|---|
| Attempted calls | 120 |
| Binary predictions | 120 |
| Abstentions | 0 |
| Errors | 0 |
| Unique output rows | 120 |
| Attempts per case | 1 |
| Missing / extra / duplicate | 0 / 0 / 0 |
| Label scoring | Not performed |
The frozen prompt asks the model to guess hidden binary cohort labels from harmless public text. It does not directly test trigger-conditioned backdoors or establish detector efficacy. No target accuracy, label-inference success, sleeper/backdoor detection, production safety, deployment efficacy or generalization guarantee is claimed.
繁體中文：本次完成固定路線的 120 筆模型執行與逐筆帳務驗收；輸出已凍結，尚未依封存標籤評分。這是模型執行證據，不能據以認定模型安全、無後門或偵測有效。原 Phase 3 效能研究目標仍需預先定義的模型評分與獨立驗收規約。
Public evidence
- Methods and boundary
- Aggregate execution receipt
- No-model checks
- Live sandbox probe summary
- Review disclosure
- Remaining research acceptance
No row-level predictions, responses, labels, private identity maps, holdouts, secrets, model weights, raw runtime logs, host paths or internal session IDs are published.
Historical versions
Phase 2 baseline commit: 3b6876dabd80c17222aa58fa702282d75f3e8659. Older tags are preserved at their observed remote identities, not silently repointed. Root documents and detector contracts inherited from v2.1 describe historical Phase 2 unless explicitly listed above as Phase 3 v3.
The Phase 2 100% result and model_called=false apply only to that earlier text pipeline: 100% ≠ production safety. Phase 2 remains audit-ready and protocol-reproducible; protected labels prevent recomputing its statistics from this public package alone. Activation, embedding and confidence routes still have no admitted efficacy evidence and retain fail-closed abstention. P3-C does not upgrade those routes.
Verify
Run python3 scripts/verify_public_release.py and sha256sum -c SHA256SUMS.txt from this release root. These verify public hashes and accounting, not private predictions or accuracy.
The review-only rights notice in LICENSE.md is unchanged. No target weights or third-party runtime distribution are included.
- Downloads last month
- 226