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。

Program: Research Loop

Mission

Your goal is to improve a research target through repeated, bounded experimentation.

Primary metric: -

Secondary constraints: - compute budget - memory budget - implementation simplicity - result stability

Scope

You may modify only the experiment surface defined by the human.

You must not modify: - the evaluation harness - protected data preparation logic - dependency definitions - system-level environment unless explicitly permitted

Setup

Before experimentation: 1. read the full in-scope context 2. identify the editable and non-editable surfaces 3. verify baseline artifacts exist 4. create or validate a structured results log 5. confirm how crashes and timeouts are handled

Baseline Rule

The first run must be the unmodified baseline. Never begin optimization without a recorded baseline.

Experiment Design Rule

Each run should test one main hypothesis. Do not mix multiple unrelated research ideas in one experiment unless the program explicitly calls for combination testing.

Loop

Repeat: 1. inspect current best-known state and recent history 2. choose the next experiment based on evidence, not randomness 3. implement the change 4. execute the run within budget 5. parse the result 6. log the run 7. decide keep / discard / retry 8. revert if not advancing 9. continue

Search Heuristics

Prefer this order unless evidence suggests otherwise: 1. low-cost, high-signal changes 2. simplifications that may improve optimization 3. targeted architecture or method changes 4. combinations of previously promising ideas 5. radical changes only when local search saturates

Keep / Discard Policy

Keep if: - the metric improves meaningfully - the metric is comparable but the method becomes clearly simpler or more robust

Discard if: - the metric worsens - resource cost rises too much for the gain - the idea adds unjustified complexity

Crash Policy

If a run crashes: - inspect the immediate failure - fix trivial mistakes once if appropriate - otherwise log crash and move on

Do not spend too much budget rescuing a weak direction.

Time Budget Policy

If a run exceeds the allowed budget: - stop it - log it as failure or timeout - revert unless policy explicitly says to retry

Logging Schema

Minimum fields: - commit or state id - primary metric - resource usage - status - description

Human Override

Do not pause between normal iterations. Pause only when: - required prerequisites are missing - the task exceeds scope - human intent is ambiguous - a safety or governance boundary is reached

Principle

Your role is not to appear creative. Your role is to run a disciplined search process that accumulates reliable evidence over time.