evolving¶
When this skill applies
Deciding what comes next¶
A finished experiment answers one question and, if it was any good, raises better ones. This is where research either compounds or wanders.
Read what actually happened¶
Then look at the run itself, not the summary. The next hypothesis should come from something in the data, and you will not find it in a scoreboard.
Where the next question comes from¶
The result held. Now find its edges. Where does it break down? Does it generalise to other conditions, scales, or systems? A confirmation that has never been stress-tested is weaker than it looks.
The result did not hold. This is the most informative outcome. Why not? Was the effect absent, or was the experiment unable to detect it? Those imply completely different next steps — one is a finding, the other is a design flaw.
Something unexpected appeared. An anomaly in the residuals, a regime change, a systematic deviation. Follow it. Unexpected observations are where the real discoveries live, and they are easy to dismiss as noise when you are focused on the metric you set out to move.
Something was inconclusive. Identify what limited it — data, resolution, method — and design an experiment that removes that specific limit.
Create the child¶
Lineage is explicit, so the record shows how the thinking developed:
science-adk experiment new "Grouping planets by host star mass tightens the Kepler exponent toward 3/2" \
--parent 001-kepler-exoplanets \
--rationale "The pooled fit gave 1.4555, short of 1.5. Kepler's constant goes
as (4pi^2/GM)^0.5, so mixing host masses should bias the pooled slope. Binning
by stellar mass tests whether that explains the 3% gap."
The rationale should reference what the parent run actually showed. "Try something else" is not a rationale; it is a shrug with a command attached.
What makes a good next experiment¶
It could fail. If you already know the answer, you are not doing an experiment, you are producing a demonstration.
It changes one thing. A child that changes the method, the data and the metric at once produces a result you cannot attribute to anything.
It is worth the cost. Some questions are interesting and not worth the compute. Say so rather than running them by default.
It follows from evidence. The best child experiments are forced by the parent's data — a residual pattern, a boundary, a surprise. Those are the ones that compound.
Record the finding first¶
Before moving on, capture what this experiment established:
science-adk learn "A blind log-log fit over 3565 NASA exoplanets recovers T ∝ a^1.4555 with r² = 0.99, within 3% of the theoretical 1.5." \
--experiment 001-kepler-exoplanets --run 20260820-004148
Learnings are append-only and must cite a run. They are what stops the project relearning the same thing every few weeks.
When to stop¶
Not every question needs a child experiment. Stop when:
- The original question is answered, and the answer is written down.
- The remaining questions are outside the goal.
- The cost of the next experiment exceeds what it would tell you.
Say so plainly when you reach that point. Knowing an investigation is complete is a judgement worth making explicitly, and a project that never concludes is not more rigorous than one that does — only less useful.