hypothesis-tracker

name	hypothesis-tracker
description	Add, update, review, and kill hypotheses using CEO Bible framework
user-invocable	true

Hypothesis Tracker

Hypothesis management per CEO Bible Section J.

WIP Limits (Critical Chain)

• Max active hypotheses: 3
• Max concurrent experiments: 2
• Stagger rule: do not start a new experiment until the bottleneck step of the current one is complete

If the founder asks to add a 4th active hypothesis -> ask: "Which of the current ones are you pausing or killing?"

Reason: each parallel project doubles the lead time of all others (Goldratt Critical Chain). Less WIP = faster throughput.

Relay Race (Critical Chain)

Don't wait for the deadline. If evidence is sufficient for a decision (keep/kill/reframe) BEFORE the timebox — decide now.

On every hypothesis update, check: "Is there enough data for a decision already?" If yes -> flag: "H00X has sufficient evidence. Relay race — decide now."

Commands

the founder may say:

• "new hypothesis: ..." -> add (check WIP limit!)
• "update hypothesis X: ..." -> update status/result
• "kill hypothesis X: ..." -> close with reason
• "show hypotheses" -> list active ones
• "rank hypotheses" -> re-sort by ELV

Hypothesis Format (Bible J)

For [ICP] who [trigger/problem], if we change [message / offer / product step / pricing / channel], then [metric] will move from X to Y by [date], because [evidence].

Structure in memory/hypotheses.json

{
  "hypotheses": [
    {
      "id": "H001",
      "created": "2026-03-15",
      "status": "active",
      "statement": "For ops leads at mid-market consultancies who struggle to close the books monthly, if we add per-vertical landing pages (/legal, /accounting), then signup-to-trial conversion will increase from X to Y within 30 days, because current README is generic.",
      "icp": "Ops leads at consultancies, 50-500 billable consultants",
      "change": "Per-vertical landing pages",
      "metric": "Signup-to-trial conversion/week",
      "baseline": null,
      "target": null,
      "deadline": "2026-04-15",
      "earliest_decision_possible": null,
      "evidence_tier": 4,
      "elv_score": null,
      "on_constraint": true,
      "result": null,
      "decision": null,
      "killed_reason": null
    }
  ]
}

Signal Tiers (for evidence quality scoring)

1. Tier 1: costly-to-fake behavioral (paid pilots, expansion, integration, renewals)
2. Tier 2: product/onboarding evidence (session replays, drop-offs, support tickets)
3. Tier 3: commercial evidence (win/loss, objections, stalled deals)
4. Tier 4: market narrative (competitor moves, content performance, search behavior)
5. Tier 5: low-power (praise, vanity traffic, investor excitement)

ELV Score (Expected Learning Value)

ELV = (pain severity x frequency x wedge fit x evidence quality x strategic leverage) / time-to-learning

In practice:

• Prefer hypotheses touching painful, frequent problem in core wedge
• Prefer tests that can change a major decision quickly
• Deprioritize elegant ideas that take months to disprove

Experiment Design (when creating hypothesis)

Every experiment must contain:

• target cohort
• hypothesis statement
• exact metric
• baseline
• success threshold
• timebox
• owner
• what would falsify it
• what decision follows each outcome

Good vs Bad Experiment Types

Good pre-PMF:

• message tests, onboarding/TTFV tests, core workflow simplification, pricing framing, pilot design, PQL handoff, founder outbound, proof asset tests

Bad pre-PMF:

• broad brand campaigns, "viral" growth hacks, complex pricing overhauls without usage proof, feature programs with no decision question

Readout Rules

Experiment produces real learning only if it changes one of:

• who you target
• what you promise
• what users must do first
• what sales motion you use
• what you price/package
• what you stop building

A test that moves a CTR but doesn't change a real decision = activity, not learning.

When Updating a Hypothesis

Log:

• Result (data)
• Decision (keep / kill / reframe)
• What changed in the company as a result
• Update MEMORY.md if the hypothesis affected wedge/stage

Follow output preferences from USER.md (language, format, platform constraints).