Experiment Ledger

An experiment ledger is a structured list of experiments, their hypotheses, setup, outcomes, and next decisions. It prevents research from becoming a pile of disconnected runs.

Each experiment can be represented as:

$$e_i=({\mathcal{Q}}_i,h_i,c_i,r_i,d_i)$$

where ${\mathcal{Q}}_i$ is the question, $h_i$ is the hypothesis, $c_i$ is the configuration, $r_i$ is the result summary, and $d_i$ is the decision.

Minimal Columns

Column	Meaning
ID	Public-safe identifier or short slug.
Question	What uncertainty does this reduce?
Hypothesis	What did you expect before running?
Change	What variable changed from the baseline?
Data/Split	Dataset version and evaluation boundary.
Metric	Primary metric or diagnostic.
Result	Public-safe outcome summary.
Interpretation	Confirmed, not confirmed, surprising, or inconclusive.
Next	What changes because of this result?

Public-Safe Rule

The public ledger should generalize or omit private run identifiers, internal task names, private paths, exact server details, collaborator details, and unpublished sensitive metrics.

Checks

• Does every experiment have a pre-run hypothesis?
• Is the baseline explicit?
• Is the changed variable isolated?
• Are negative and inconclusive runs recorded?
• Is the next decision clear?

Related

• Research log
• Experiment design
• Negative result
• Experiment tracking
• Reproducible run record