Experiment Design Checklist
Prevent the "I ran experiments for 3 months and they're meaningless" disaster through rigorous upfront design.
The Core Principle
Before running ANY experiment, you should be able to answer:
- What specific claim will this experiment support or refute?
- What would convince a skeptical reviewer?
- What could go wrong that would invalidate the results?
Process
Step 1: State the Hypothesis Precisely
Convert your research question into falsifiable predictions:
Template:
If [intervention/method], then [measurable outcome], because [mechanism].
Examples:
- "If we add auxiliary contrastive loss, then downstream task accuracy increases by >2%, because representations become more separable."
- "If we use learned positional encodings, then performance on sequences >4096 tokens improves, because the model can extrapolate beyond training length."
Null hypothesis: What does "no effect" look like? This is what you're trying to reject.
Step 2: Identify Variables
Independent Variables (what you manipulate): | Variable | Levels | Rationale | |----------|--------|-----------| | [Var 1] | [Level A, B, C] | [Why these levels] |
Dependent Variables (what you measure): | Metric | How Measured | Why This Metric | |--------|--------------|-----------------| | [Metric 1] | [Procedure] | [Justification] |
Control Variables (what you hold constant): | Variable | Fixed Value | Why Fixed | |----------|-------------|-----------| | [Var 1] | [Value] | [Prevents confound X] |
Step 3: Choose Baselines
Every experiment needs comparisons. No result is meaningful in isolation.
Baseline Hierarchy:
-
Random/Trivial Baseline
- What does random chance achieve?
- Sanity check that the task isn't trivial
-
Simple Baseline
- Simplest reasonable approach
- Often embarrassingly effective
-
Standard Baseline
- Well-known method from literature
- Apples-to-apples comparison
-
State-of-the-Art Baseline
- Current best published result
- Only if you're claiming SOTA
-
Ablated Self
- Your method minus key components
- Shows each component contributes
For each baseline, document:
- Source (paper, implementation)
- Hyperparameters used
- Whether you re-ran or used reported numbers
- Any modifications made
Step 4: Design Ablations
Ablations answer: "Is each component necessary?"
Ablation Template: | Variant | What's Removed/Changed | Expected Effect | If No Effect... | |---------|----------------------|-----------------|-----------------| | Full Model | Nothing | Best performance | - | | w/o Component A | Remove A | Performance drops X% | A isn't helping | | w/o Component B | Remove B | Performance drops Y% | B isn't helping | | Component A only | Only A, no B | Shows A's isolated contribution | - |
Good ablations are:
- Surgical (one change at a time)
- Interpretable (clear what was changed)
- Informative (result tells you something)
Step 5: Address Confounds
Things that could explain your results OTHER than your hypothesis:
Common Confounds:
| Confound | How to Check | How to Control | |----------|--------------|----------------| | Hyperparameter tuning advantage | Same tuning budget for all | Report tuning procedure | | Compute advantage | Matched FLOPs/params | Report compute used | | Data leakage | Check train/test overlap | Strict separation | | Random seed luck | Multiple seeds | Report variance | | Implementation bugs (baseline) | Verify baseline numbers | Use official implementations | | Cherry-picked examples | Random or systematic selection | Pre-register selection criteria |
Step 6: Statistical Rigor
Sample Size:
- How many random seeds? (Minimum: 3, better: 5+)
- How many data splits? (If applicable)
- Power analysis: Can you detect expected effect size?
What to Report:
- Mean ± standard deviation (or standard error)
- Confidence intervals where appropriate
- Statistical significance tests if claiming "better"
Appropriate Tests: | Comparison | Test | Assumptions | |------------|------|-------------| | Two methods, normal data | t-test | Normality, equal variance | | Two methods, unknown dist | Mann-Whitney U | Ordinal data | | Multiple methods | ANOVA + post-hoc | Normality | | Multiple methods, unknown | Kruskal-Wallis | Ordinal data | | Paired comparisons | Wilcoxon signed-rank | Same test instances |
Avoid:
- p-hacking (running until significant)
- Multiple comparison problems (Bonferroni correct)
- Reporting only favorable metrics
Step 7: Compute Budget
Before running, estimate:
| Component | Estimate | Notes | |-----------|----------|-------| | Single training run | X GPU-hours | [Details] | | Hyperparameter search | Y runs × X hours | [Search strategy] | | Baselines | Z runs × W hours | [Which baselines] | | Ablations | N variants × X hours | [Which ablations] | | Seeds | M seeds × above | [How many seeds] | | Total | T GPU-hours | Buffer: 1.5-2x |
Go/No-Go Decision: Is this feasible with available resources?
Step 8: Pre-Registration (Optional but Recommended)
Write down BEFORE running:
- Exact hypotheses
- Primary metrics (not chosen post-hoc)
- Analysis plan
- What would constitute "success"
This prevents unconscious goal-post moving.
Output: Experiment Design Document
# Experiment Design: [Title]
## Hypothesis
[Precise statement]
## Variables
### Independent
[Table]
### Dependent
[Table]
### Controls
[Table]
## Baselines
1. [Baseline 1]: [Source, details]
2. [Baseline 2]: [Source, details]
## Ablations
[Table]
## Confound Mitigation
[Table]
## Statistical Plan
- Seeds: [N]
- Tests: [Which tests for which comparisons]
- Significance threshold: [α level]
## Compute Budget
[Table with total estimate]
## Success Criteria
- Primary: [What must be true]
- Secondary: [Nice to have]
## Timeline
- Phase 1: [What, when]
- Phase 2: [What, when]
## Known Risks
1. [Risk 1]: [Mitigation]
2. [Risk 2]: [Mitigation]
Red Flags in Experiment Design
🚩 "We'll figure out the metrics later" 🚩 "One run should be enough" 🚩 "We don't need baselines, it's obviously better" 🚩 "Let's just see what happens" 🚩 "We can always run more if it's not significant" 🚩 No compute estimate before starting 🚩 Vague success criteria