hypothesis-generation

Hypothesis Generation

Adapted from K-Dense for EasyResearch's Experiment-owned artifact roots and Research Assistant-owned user interaction.

Scope

Use this Skill before code or trials when the empirical question needs a testable hypothesis package. A hypothesis is a candidate to challenge, not a finding, causal conclusion, novelty proof, or recommendation.

Apply it inside the experiment root selected by the dispatch:

• local: exact-cwd experiments/;
• remote: only the verified exact-cwd experiment_ssh/ mount.

Do not create a second planning root. Record the accepted hypothesis, rivals, predictions, controls, measurements, and analysis plan in <experiment-root>/formal-experiment-plan.md. Optional structured working files may live under <experiment-root>/outputs/planning/; promote only accepted plan evidence to <experiment-root>/results/planning/.

This Skill does not replace autoresearch: autoresearch owns an already authorized metric-bound trial campaign. This Skill may strengthen the campaign contract or a normal Experiment plan but does not select a winner without data.

Readiness And Safety

Inspect the dispatch, ref_papers/source.json, ref_papers/paper-notes.md, readable sources, existing experiment plan/record, and relevant preliminary results. Distinguish:

• observation;
• research question;
• hypothesis and mechanism;
• prediction;
• rival explanation;
• operationalization;
• analysis plan;
• observed evidence.

Never ask the user directly. Return a blocked handoff when existing decisions cannot resolve a consequential objective, sensitive-data authority, ethics or safety gate, target population, metric, intervention, or resource boundary. Never send unpublished/sensitive content to search or another external endpoint without authority carried by the dispatch.

Stop at required human, animal, biosafety, dual-use, privacy, clinical, legal, or regulatory review. Do not provide harmful operational optimization.

Procedure

1. Freeze the observation with source, unit, population/system, preprocessing, uncertainty, missingness, and whether it was expected or selected after looking at outcomes.
2. Frame one answerable research question and declare its claim type: descriptive, associational, predictive, causal, or mechanistic.
3. Record the dated evidence boundary: indexes, queries, filters, included and excluded source types, and known coverage limitations. Search absence never proves global novelty.
4. Generate genuinely different rivals where plausible: proposed mechanism, measurement artifact, confounding/common cause, selection, reverse causation, context/boundary effects, stochastic variation, and competing mechanisms.
5. If no human-independent rival set exists, label the set AI-generated and require later human review; never pretend independent ideation occurred.
6. For causal work, define target population/system, intervention/exposure, comparator, outcome/time horizon, estimand, and identification assumptions.
7. Derive predictions that distinguish at least two rivals. State observable, measurement, expected pattern, incompatible result, boundary conditions, and indeterminate outcomes.
8. Specify positive/procedural/negative controls only when scientifically valid; a negative control must share relevant bias paths but not the target mechanism.
9. Operationalize every construct with units, timing, instrument, calibration, reliability, validity, missingness, transforms, thresholds, and bias risks.
10. Match experimental unit, sampling/allocation, randomization, masking, outcomes, models, uncertainty, multiplicity, missing-data handling, stopping, sensitivity, and replication to the claim.
11. Mark confirmatory versus exploratory decisions and preserve later deviations instead of rewriting them as a priori.
12. Write/update formal-experiment-plan.md, validate structured sidecars when used, and reference the plan in experiment-record.md.