Experiment Pipeline

Experiment Pipeline

A structured 4-stage framework for executing research experiments from initial implementation through ablation study, with attempt budgets and gate conditions that prevent wasted effort. This follows the Experiment Tree Search design from the EvoScientist paper, where the engineer agent iteratively generates executable code, runs experiments, and records structured execution results at each stage.

When to Use This Skill

• User has a planned experiment and needs to organize the execution workflow
• User wants to systematically validate a novel method against baselines
• User asks about experiment stages, attempt budgets, or when to move on
• User needs to reproduce baseline results before testing their method
• User mentions "experiment pipeline", "baseline first", "ablation study", "stage budget", "experiment execution"

The Pipeline Mindset

Experiments fail for two reasons: wrong order and no stopping criteria. Most researchers jump straight to testing their novel method without verifying their baseline setup, then wonder why results don't make sense. Others spend weeks tuning hyperparameters without a budget, hoping the next run will work.

The 4-stage pipeline solves both problems. It enforces a strict order (each stage validates assumptions the next stage depends on) and assigns attempt budgets (forcing systematic thinking over brute-force iteration).

Before Starting: Load Prior Knowledge

If coming from research-ideation, your research proposal (Step 7) provides the experiment plan — datasets, baselines, metrics, and ablation design — that maps directly to Stages 1-4 below.

Before entering the pipeline, load Experimentation Memory (M_E) from prior cycles:

1. Refer to the evo-memory skill → Read M_E at /memory/experiment-memory.md
2. Select the top-1 entry (k_E=1) most relevant to the current experiment domain by comparing each entry's Context and Category against the current problem
3. The selected strategy informs hyperparameter ranges (Stage 2), debugging approaches (Stages 1-3), and training configurations across all stages
4. If M_E doesn't exist yet (first cycle), skip this step and proceed — your results will seed M_E via ESE after pipeline completion

4-Stage Pipeline Overview

Each stage follows a generate → execute → record → diagnose → revise loop:

Each stage saves artifacts to /experiments/stageN_name/.

The Stage Loop

Within every stage, repeat this cycle for each attempt:

1. Generate: Form a hypothesis or plan for this attempt. What specifically will you try? What do you expect to happen?
2. Execute: Run the experiment. Record exact configuration, code changes, and runtime.
3. Record: Log results immediately using the stage log template. Include both metrics and observations.
4. Diagnose: Compare results to expectations. If they match, assess the gate condition. If they don't, load experiment-craft for the 5-step diagnostic flow.
5. Revise: Based on diagnosis, either advance to the next stage (gate met) or plan the next attempt (gate not met).

Stage 1: Initial Implementation

Goal: Find or generate executable baseline code and verify it reproduces published results. This stage corresponds to the paper's "initial implementation" — the engineer agent searches for working code, runs it, and records structured execution results.

Why this matters: If you can't get the baseline running and reproducing known results, every subsequent comparison is meaningless. Initial implementation validates your data pipeline, evaluation code, training infrastructure, and understanding of prior work.

Budget: ≤20 attempts (N_E^1=20). Baselines can be tricky — missing details in papers, version mismatches, unreported preprocessing steps. 20 attempts gives enough room to debug without allowing infinite tinkering.

Gate: Primary metrics within 2% of reported values (or within the reported variance if provided).

Process:

1. Find the original baseline code (official repo, re-implementations, or write from paper description)
2. Get the code running in your environment — resolve dependencies, fix compatibility issues
3. Match the exact training configuration from the paper (dataset splits, preprocessing, hyperparameters)
4. Run and compare metrics. If off by >2%, diagnose the gap
5. Common pitfalls: different random seeds, different data splits, unreported data augmentation, framework version differences

When to load experiment-craft: If attempts 1-5 all fail significantly (>10% gap), switch to the 5-step diagnostic flow to isolate the cause before burning more attempts.

Output: /experiments/stage1_baseline/ containing results, config, and verified baseline code.

See references/stage-protocols.md for detailed initial implementation checklists.

Stage 2: Hyperparameter Tuning

Goal: Find the optimal hyperparameter configuration for YOUR specific setup.

Why this matters: Published hyperparameters are tuned for the authors' setup. Your hardware, data version, framework version, or subtle implementation differences mean their config may not be optimal for you. Tuning now prevents confounding your novel method's results with suboptimal baselines.

Budget: ≤12 attempts. Hyperparameter tuning has diminishing returns. If 12 structured attempts don't find a stable config, the problem is likely deeper than hyperparameters.

Gate: Stable configuration found — variance < 5% across 3 independent runs with different random seeds.

Process:

1. Identify the most sensitive hyperparameters (usually: learning rate, batch size, loss weights)
2. Start with coarse search on the most sensitive parameter
3. Narrow the range based on results, then move to the next parameter
4. Validate final config with 3 independent runs

Priority order for tuning: Learning rate → batch size → loss weights → regularization → architecture-specific params. This order reflects typical sensitivity.

When to load experiment-craft: If results are highly unstable (variance > 20%) across runs, there's likely a training instability issue. Use diagnostic flow.

Output: /experiments/stage2_tuning/ containing tuning logs, final config, and stability verification.

See references/attempt-budget-guide.md for budget rationale and adjustment rules.

Stage 3: Proposed Method

Goal: Implement and validate your novel method, demonstrating improvement over the tuned baseline.

Why this matters: This is the core contribution. But because you've verified the baseline (Stage 1) and optimized the config (Stage 2), any improvement you see is genuinely attributable to your method — not to a better-tuned setup or a broken baseline.

Budget: ≤12 attempts. Your method should work within a reasonable number of iterations if the underlying idea is sound. Excessive attempts suggest a fundamental problem, not a tuning issue.

Gate: Outperforms the tuned baseline on the primary metric. The improvement should be consistent across at least 3 runs.

Process:

1. Implement the core method incrementally — don't add everything at once
2. Test each component's integration with the baseline pipeline
3. Run full training and compare against Stage 2 results
4. If underperforming, isolate which component causes the gap

Integration strategy: Add your method's components one at a time to the working baseline. Each added component should stay within 20% of the baseline's performance — if a single component causes a >20% regression, isolate and debug it before proceeding. Never integrate the full method in one shot.

When to load experiment-craft: When your method underperforms the baseline despite correct implementation. The 5-step diagnostic flow will help distinguish between implementation bugs and fundamental issues.

Critical decision — failure classification: If the method underperforms the baseline after exhausting the attempt budget, hand off to evo-memory for IVE (Idea Validation Evolution) — this is evo-memory's job, not this skill's. IVE triggers under two conditions:

1. No executable code: Cannot find working code within the budget at any stage.
2. Worse than baseline: Experiments complete but the method underperforms.

The evo-memory skill will classify the failure as:

• Implementation failure: Bugs or missing tricks → retryable in a future cycle.
• Fundamental direction failure: Core idea doesn't work → update ideation memory to prevent retrying.

Output: /experiments/stage3_method/ containing method code, results, comparison with baseline.