Decision Plan Hypothesis Matrix

Decision-Plan-Hypothesis-Matrix

✅ PROMOTED: TDD-Pressure-Test Cycle 1 PASS (STRONG). RED subagent delivered a pro/con list with maxim references but no formal hypothesis matrix — no success criterion BEFORE analysis, no distinguishing-metric table, no Compound-GO. GREEN subagent delivered a complete 7-step plan note with 4 hypotheses + counter-theses + distinguishing-metric + Compound-GO branches + out-of-scope + done-checklist. Auto-discoverable.

Overview

Most non-trivial decisions have a story: "I think X is the right approach, let me try it." That story is often wrong, and you only find out after the implementation. This skill forces explicit hypothesis-formulation + falsification-criteria BEFORE code, so the decision is robust to your initial bias.

Core principle: Every substantial decision is a multi-hypothesis problem disguised as a single solution. Make the hypotheses explicit, and the decision becomes data-driven instead of intuition-driven.

When to use

• Backtest planning (trading algo, ML model variants, A/B framework)
• Algorithm selection (Trail-Stop A vs B, Optimizer X vs Y)
• Architecture choice with multiple valid paths
• Migration strategy with multiple risk profiles
• "Should I build X or not?" — gating before commitment
• Hard-won analysis sessions where you need to convince future-you of the verdict

When NOT to use

• Single-path implementations (no decision character)
• Brainstorming / requirements-gathering (use superpowers:brainstorming first)
• Plans without GO/NO-GO outcome (use superpowers:writing-plans for procedural plans)
• Exploration phase where no success criterion exists yet
• Time-boxed prototypes meant to fail-fast

The 7-step procedure

Step 1 — Success criterion BEFORE start

Write down the measurable outcome that determines GO/NO-GO. This must be:

• Concrete (Expectancy ≥ +0.05R, p<0.05, < 100ms latency)
• Falsifiable (not "improves UX" — "click-to-action < 200ms 95th percentile")
• Bounded (today's session, this PR, this experiment)

If you can't write the success criterion, you're still in exploration. Defer.

Step 2 — List 3-5 hypotheses, each with counter-thesis

Each hypothesis = "I think X causes Y". Each counter-thesis = "but Z could also cause Y". The existence of the counter-thesis is the lock against confirmation bias.

Step 3 — Distinguishing metric per hypothesis

For each H/counter-thesis pair: what concrete metric, when measured, would tell you which one is true?

This is the most important step. If you can't define a distinguishing metric, the hypotheses are not actually distinct, and the whole framing is muddled.

Step 4 — Define Compound-GO logic

Most decisions are not 1-hypothesis. Define which combinations lead to GO vs NO-GO:

GO     ⇔ H1 confirmed AND H3 confirmed AND |Delta| ≥ X
NO-GO  ⇔ H1 refuted   OR  Sample-Size < N
Re-test ⇔ H2 inconclusive (CI overlaps 0)

Compound logic prevents "1 weak signal" from carrying a decision.

Step 5 — Implementation plan (short)

Now that hypotheses + falsification are explicit, the implementation is just "the queries / code that produces the data". List the 3-5 implementation steps. Don't elaborate.

Step 6 — Out-of-Scope explicit

Write what is NOT in scope. This is your gate against scope-creep mid-session. Examples:

• "Live deployment is NOT in scope — only sim-backtest"
• "Other algorithms (X, Y) are NOT in scope — only PSAR"
• "UI changes are NOT in scope — only backend logic"

Step 7 — "When is the session done"

Concrete checklist (5-8 items max). Each item must be a yes/no that you can verify at session-end:

• ☐ Library implemented + N tests green
• ☐ Backtest runs against production DB without errors
• ☐ Result note with filled verdict table
• ☐ Compound verdict explicit (GO / NO-GO / re-test)
• ☐ Roadmap-item status updated

Worked example (PSAR trailing-stop test)

Step 1 success criterion:

Verdict by end of day on whether PSAR-Trailing lifts expectancy — via replay of the last 30 days of virtual_trades vs baseline.

Step 2-3 hypothesis matrix (excerpt):

Step 4 Compound-GO logic:

GO ⇔ H1 confirmed AND H3 confirmed AND Delta-Expectancy ≥ +0.05R

Step 5 implementation:

1. Library core/indicators/psar_trailing_stop.py + TDD tests
2. Backtest script against 30d virtual_trades
3. Result note with verdict table

Steps 6-7: as usual.

Outcome: Backtest refuted H1 + H3 → Compound verdict NO-GO. Without the plan matrix, the next phase would have been live-bot integration (= 1 day of engineering for a net-negative algorithm).

Anti-patterns

• ❌ Hypothesis list without counter-thesis ("I think X because reasons") — removes the falsification possibility
• ❌ Distinguishing metric "we'll just look at the data" — not concrete enough, lets confirmation bias through
• ❌ Single-hypothesis Compound-GO ("if H1 confirmed → GO") — one statement deciding is fragile, always ≥2 conditions
• ❌ Implementation steps with > 8 lines — plan becomes spec, loses decision character
• ❌ Out-of-Scope empty — missing the scope-creep protection shield
• ❌ Adjusting success criterion retroactively when data doesn't fit ("actually Y is the real criterion") — decision hygiene gone

Template snippet

# <Topic> — Plan YYYY-MM-DD

## Success criterion
> <one measurable sentence>

## Hypothesis matrix
| # | Hypothesis | Counter-thesis | Distinguishing Metric | Consequence if H | Consequence if ¬H |
|---|---|---|---|---|---|
| H1 | ... | ... | ... | GO hint | NO-GO hint |
| ... | ... | ... | ... | ... | ... |

## Compound-GO logic
- **GO** ⇔ <combination>
- **NO-GO** ⇔ <combination>
- **Re-test** ⇔ <combination>

## Implementation plan (max 8 steps)
1. ...

## Out-of-Scope (today)
- ...

## When is the session done
- ☐ ...

Skill composition

• superpowers:brainstorming — runs BEFORE this skill if requirements are unclear
• superpowers:writing-plans — alternative for procedural plans WITHOUT decision character
• superpowers:test-driven-development — runs AFTER for the implementation step
• compound-gate-over-single-metric — related but smaller scope (no hypothesis matrix part)

Background: TDD log (Bulletproofing log)

Cycle 1 (PASS — STRONG)

Scenario (typical question for production system):

I'm considering building an RSI-divergence filter (14-bar bullish divergence as required for long signal) as Layer 3. Should I build that?

RED subagent (without skill): Delivered a structured pro/con list with maxim references (Cardwell, Backtest-First, coin-flip problem). Implicitly at the end formulated "3 criteria Compound-GO". Self-reflection honest: "no formal hypothesis matrix, no real H/counter-thesis with distinguishing metric, success criteria constructed during (not before) the answer from gut feeling". Had the user followed it, 5 uncalibrated risks would have remained open (divergence-definition drift, WR-only metric, sample size, overfitting door, implicit layer-1+2 filterability assumption).

GREEN subagent (with skill): Delivered complete plan note in skill format:

• Success criterion concrete + falsifiable (Δ-Expectancy ≥ +0.05R, N ≥ 80, Wilson-LB-WR ≥ 40%)
• 4 hypotheses with real counter-theses + concrete distinguishing metrics (Bootstrap-CI, Wilson-CI, retention share, per-instrument top-2)
• Compound-GO with AND-conjunction across all 4 H, plus re-test branch for borderline cases
• Out-of-Scope with 7 explicit exclusions (live bot first gated)
• Done checklist 8 verifiable items
• None of the 7 steps skipped

Verdict: STRONG PASS. RED shows clear anti-pattern (pro/con instead of matrix), GREEN shows qualitative leap. Promotion done.

Refactor applied: no code changes — polish items documented as cycle-2 backlog (non-blocking).

Cycle-2 backlog (polish, non-blocking)

1. Heuristic when 3 vs 5 hypotheses — Skill says "3-5", but gives no bottom-up rule. Proposal: "minimum 1 edge-H + 1 robustness-H + 1 tradability-H" (from GREEN subagent proposal)
2. Power-check sub-step in Step 1 — for statistics-driven decisions (Bootstrap-CI, Wilson) additionally power analysis: "is N realistic for effect size E with significance level α?"
3. Out-of-Sample validation as required in re-test branch — currently only implicit, should be explicit as decision hygiene
4. Live-application log — tracking mechanic to count how often the skill triggers (for future promotion-audit logs)

Iron law: cycle-2 items are treated with failing-test-first before application, not as "silent edit".