Experiment

• No hypotheses? Seeds them from CLAUDE.md.
• Hypotheses exist? Picks the most uncertain, designs an experiment, runs it.
• Evidence accumulating? Graduates proven rules, kills disproven ones.
• Everything tested? Reports and stops.

Rules that survive become convictions. Rules that fail become learnings.

```bash
agentdb hypothesis list 2>/dev/null
```

Decision tree (no human input needed):
1. No hypotheses table or empty? → go to SEED phase.
2. Hypotheses exist but all are unproven? → go to PICK phase.
3. Mix of tested/untested? → go to PICK phase (prioritize untested).
4. All have >= 3 experiments? → go to JUDGE phase.
5. Graduation/kill candidates exist? → go to EVOLVE phase.

Parse rule-like patterns:
- Imperative: "Always X", "Never Y", "Prefer Z", "Must W"
- Anti-patterns: block actions, "Don't", "Forbidden"
- Assertions: "X before Y", "X is better than Y"
- Quantitative claims: "reduces by X%", "takes N minutes"
- Conditional: "If X then Y", "When X, do Y"

For each rule:
```bash
agentdb hypothesis add "<statement>" --domain <auto-classify> --source "<file:line>"
```

Domain auto-classification by keyword:
- research, anti-pattern, prior work → methodology
- parallel, agent, spawn, tier → coordination
- test, coverage, edge case, mock → testing
- commit, branch, merge, PR → git
- secret, validation, auth, injection → security
- measure, optimize, latency, profile → performance
- Big 5, review, quality → quality
- module, interface, coupling → architecture

Deduplicate: skip if near-identical statement already exists.
Log count, then immediately proceed to PICK. No pause.

```bash
agentdb emit command "experiment-seed" "" '{"seeded":N}'
```

Priority order:
1. **Most uncertain**: confidence closest to 0.5 (maximum ignorance — any experiment is maximally informative)
2. **Least tested**: fewest total experiments (break ties)
3. **Highest impact domain**: methodology > coordination > security > testing > quality > git > architecture > performance

```sql
SELECT id, statement, domain, confidence, evidence_for + evidence_against as total_evidence
FROM hypotheses
WHERE status NOT IN ('graduated', 'refuted')
ORDER BY ABS(confidence - 0.5) ASC, total_evidence ASC
LIMIT 1;
```

**Falsifiability gate: if no possible outcome could refute the hypothesis, redesign.**
Every experiment defines BEFORE running: method, quantitative measurement, control
condition (what happens WITHOUT the rule), pass_criteria, fail_criteria.

Choose the LIGHTEST experiment type that produces signal:

1. **HISTORICAL** (cheapest — query existing data):
   Query agentdb learnings, session outcomes, error patterns for evidence.
   Use when: agentdb has >= 10 sessions or >= 20 learnings in the domain.
2. **COMPARATIVE** (medium — run a real task two ways):
   Execute WITH the rule applied, then WITHOUT (or find prior without-cases).
   Measure: time, error count, rework, quality.
3. **ABLATION** (medium — remove the rule, observe):
   Temporarily ignore the rule during a real task. Record what breaks.
4. **OBSERVATIONAL** (passive — tag next N tasks):
   Flag the hypothesis; future relevant tasks collect evidence passively.
   Use when: active experimentation would be disruptive.

Minimum sample sizes: methodology/coordination/git/quality >= 3 comparisons;
testing >= 5 tasks per condition; security >= 50 fuzz inputs.

```bash
agentdb experiment add <H_ID> "<method>" "<measurement>" --pass-criteria "<criteria>"
```

**Gate: the control condition was actually tested, not just assumed.**

- HISTORICAL: query agentdb with specific SQL; evidence = query result + interpretation.
- COMPARATIVE: execute the task (spawn agents if needed); evidence = measured delta.
- ABLATION: execute with the rule explicitly ignored; evidence = observed difference.
- OBSERVATIONAL: record the flag; skip to next hypothesis (no blocking).

```bash
agentdb experiment verdict <EXP_ID> <supports|refutes|inconclusive> "<evidence summary>"
```

Confidence update (Bayesian, applied automatically by CLI):
- supports:     confidence += (1 - confidence) * 0.25
- refutes:      confidence -= confidence * 0.3
- inconclusive: no change

Evidence strings must be specific and measurable, never narrative.

Lifecycle transitions:
- unproven → testing: first experiment registered
- testing → supported: confidence >= 0.8 AND evidence_for >= 3 AND ratio >= 3:1
- testing → refuted: confidence < 0.2 AND evidence_against >= 2
- supported → graduated: human approval after sustained confidence
- refuted → killed: human approval to remove from rules
- any → unproven: rule is modified (resets all evidence)

```bash
agentdb learn pattern|failure "<what we learned>" "<evidence>"
agentdb emit command "experiment-conclude" "" '{"H":"ID","EXP":"ID","verdict":"X","confidence":0.XX}'
```

Loop back to PICK for next hypothesis.

```bash
agentdb hypothesis export
```
Write detailed report to _meta/research/experiment-report.md. Proceed to EVOLVE.

**Graduate** (confidence >= 0.8, evidence_for >= 3, ratio >= 3:1):
promote via the artifact ladder with human approval — hook if enforceable, agent if
a role, skill if methodology; CLAUDE.md prose only as last resort.

**Kill** (confidence < 0.2, evidence_against >= 2):
propose rule removal from CLAUDE.md (present to human).

**Mutate** (inconclusive after 5+ experiments):
the hypothesis may be poorly stated; propose a refined version as a NEW hypothesis,
linked to the original (evolution chain).

<ask_user>
  Use AskUserQuestion ONCE at the end of the evolve phase:
  Ask: "{graduated} rules proven, {killed} rules disproven, {mutated} rules refined. Apply changes?"
  Options: apply all, review individually, skip for now
</ask_user>

```bash
agentdb emit command "experiment-evolve" "" '{"graduated":N,"killed":N,"mutated":N}'
```