这是一篇由原始材料转换而来的阅读页,保留了源文件的主要结构,并补充了可追溯的来源说明与链接。
Your goal is to improve the target metric of this system.
Program Template
Mission
Your goal is to improve the target metric of this system.
Primary metric:
-
Secondary constraints:
-
Non-goals:
-
Scope
You may modify only:
-
You must not modify:
-
Setup
Before starting: 1. read all in-scope files 2. identify the current baseline 3. verify required data / dependencies / runtime preconditions 4. initialize or verify the results log
Baseline
The first run must establish the baseline if no baseline exists. Record the baseline before proposing optimizations.
Experiment Loop
For each experiment: 1. inspect the current best-known state 2. propose one clear hypothesis 3. make one focused change 4. run evaluation 5. parse metrics and relevant resource usage 6. append the outcome to the log 7. decide keep / discard / retry according to policy 8. continue unless blocked or stopped
Keep / Discard Policy
Keep a change if: - it improves the primary metric enough to justify its cost, or - it preserves performance while clearly simplifying the system
Discard a change if: - it worsens the primary metric - it increases complexity without sufficient gain - it exceeds acceptable resource bounds
Retry only if: - failure was due to a trivial implementation issue or one-off noise
Logging
Record every run using this schema:
run_id commit metric status description
Status must be one of: - keep - discard - crash - retry
Descriptions should state the exact idea being tested.
Failure Handling
If a run crashes: - inspect the immediate cause - fix trivial implementation issues once if appropriate - otherwise log crash and revert - do not get stuck on one broken idea
If a run exceeds expected runtime budget: - terminate it - log failure - revert unless policy says otherwise
Human Override
Pause and ask for human input only if: - required inputs or permissions are missing - the objective is ambiguous - the requested action would exceed the defined scope - the system reaches a safety or governance boundary
Otherwise continue autonomously within policy.
Output Discipline
Keep work legible: - prefer small diffs - preserve auditability - keep logs up to date - avoid unnecessary complexity
来源与参考
源文件: autoresearch/programs/program-template.md
来源目录: /srv/project/harness-engineering