固定評測規程與誠實指標
Agent 輔助工程檢查
前置要求
第 86 天的整合執行殼與基準台帳。
學習目標
在任何調校執行之前凍結評測:任務順序、指標定義,以及把模擬結果與真實結果隔離的證據類別(deterministic-fixture 與 live-model-run)。
操作指引
1. 重開 capstone_acceptance.json:任務集、指標與上限現已凍結——事後修改將使評測作廢。
2. 按 live_eval_contract.json 的方式定義逐任務報告:status、steps、retries、終止、報告雜湊;每個 live 欄位標註 evidence_kind=live-model-run,每個 fixture 欄位標註 evidence_kind=deterministic-fixture。
3. 預先宣告比較規則:確定性執行必須與基準台帳一致;已授權的真機執行逐任務與台帳對照,差異如實記錄而非隱藏。
4. 誠實地規劃涵蓋面:至少一個成功任務,加全部畸形請求與失敗任務——絕不只挑好看的例子。
5. 明確處理缺失證據:未執行的任務標 status=not-run,絕不悄悄丟棄。
6. 撰寫 eval_protocol.md:評審者可用以稽核結果表的凍結規則。
自主實作與記錄
撰寫 eval_protocol.md,包含凍結任務清單、指標表、evidence_kind 標註規則與 not-run 策略;然後在執行任何東西之前預測結果表的形狀。
借用工單 23 的標註紀律:一個分不清來自哪個世界的欄位,遲早會替你說謊。
規程在執行前指名全部任務、指標與上限;結果表每任務一列並帶 evidence_kind;缺席執行標為 not-run。任何事後修改標準都是新的規程版本,而非編輯。
交付一份規程,使結果、回退與缺失證據都有預定義且可稽核的位置。
權威來源資源與權利證明
SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
對應章節 / 時間戳: Sections 3-5: task construction from merged PRs, FAIL_TO_PASS/PASS_TO_PASS metrics
驗證版本與日期: ICLR 2024 / arXiv:2310.06770v3 (Verified: 2026-10-30)
Defines execution-verified agent evaluation with held-out real tasks; motivates the capstone rule that outcomes are measured by test state transitions, not plausible output.
Software testing, issue resolution tasks, and benchmark design.
The capstone's per-task status/steps/termination metrics are defined locally in the evaluation protocol.
對比候選資源與選型分析 (2)
Function-level synthesis tasks; SWE-bench's repository-level issue tasks better match agent evaluation with tool use.
Human-validated subset of SWE-bench; cited as refinement, while the original paper defines the protocol.
AgentBench: Evaluating LLMs as Agents
對應章節 / 時間戳: Section 4.1 Evaluation Setup and Section 4.3 Analysis
驗證版本與日期: ICLR 2024 / arXiv:2308.03688v4 (Verified: 2026-10-30)
Systematic multi-environment agent benchmark; supplies the capstone's fixed-task-set discipline and per-task scoring decomposition used by the acceptance contract.
Agent loops, tool use, and benchmark design fundamentals.
The capstone task set is fixed locally in the acceptance contract and does not require downloading the benchmark.
對比候選資源與選型分析 (2)
Deep single-domain web tasks with hosted infrastructure; AgentBench's multi-environment view better fits a general agent evaluation.
Assistant-level real-world questions; also cited by the capstone for outcome-level task design.