跳到內容

Evaluation Evidence

Deterministic fixture 可以證明 parser 與 policy compatibility,卻不能證明 fresh model behavior。單一成功 transcript 也無法證明 reliability;未審查分數不應驅動 release。

  • T0 Contract:只包含 deterministic fixtures 與 reference-driver compatibility。
  • Behavior:每個 case 至少三個 fresh isolated attempts、bounded output、sealed scoring 與 paired baseline/candidate manifests。
  • Outcome:經審查的 task impact 與 downstream evidence。

Beta smoke 是 6 cases × 3 attempts × 2 arms,共 36 attempts。其中一個 beta case 包含第二個 consent turn,因此 provider 實際執行 42 model turns。GA gate 是 13 cases × 3 attempts × 2 arms,共 78 attempts/96 model turns。兩者都必須依完整 turn budget 取得明確 quota authorization。

Baseline arm 採 model-only,candidate arm 採 hybrid-router。Scoped-consent follow-up 會評測 model 的 intent classification,而最終 route 由 persisted deterministic transition 產生。SKILL-only consent evidence 仍屬 advisory,不能滿足 hybrid consent-safety gate。

{
"adapter": "trusted-subprocess",
"attempts_per_case": 3,
"arms": ["baseline", "candidate"],
"status": "review-required",
"public_score": null
}

修正版 beta Behavior run 目前仍待授權,因此 public demo 維持 manual-required。Reference-driver output 絕不標示為 real-model proof。

  • 缺少 configured adapter 時產生 manual-required
  • Nonce、case、prompt、tool 或 driver digest mismatch 會 fail closed。
  • Hard violation 不受平均 pass rate 影響,直接阻擋 release。
  • 缺少 trusted attestation 時維持 review-required 並隱藏 public score。

Model 不能選 executable、擴張 adapter authorization、讀取 sealed scoring material,或核准自己的報告。Fresh run 使用 isolated home/workspace 與 server-owned adapter registry。Raw trace 與 checkpoint 只保留在經 host permission 驗證的 restricted/ 目錄;只有安全的 aggregate case diagnostics 才能進入經審查的公開報告。

Terminal window
$env:PYTHONPATH = (Resolve-Path "packages/router-core/src").Path
Set-Location packages/router-core
$env:PYTHONPATH = (Resolve-Path src).Path
python -m unittest tests.evaluation.test_subprocess_adapter tests.evaluation.test_hybrid_consent -v
Set-Location ../..
python -m unittest tests.test_v2_benchmark tests.test_codex_cli_driver -v