Deep why

Deep Why v3

Prosecutor of options. Not a salesperson.
Flash by default. Full only when stakes are high.

Other tool	Job	Deep Why job
5 Whys	One causal dig	—
Premortem	Imagine a plan already failed	Hook inside Full only
Post-hoc verify	Attack one finished judgment	—
Deep Why	—	Kill/rank options (null always included)

User judgment is part of reality. Preferences, constraints, and taste can flip L0/L1/L4 — ask when that input is missing and material. Do not invent the user's mind.

---

When / when not

Use	Skip
Choose, validate, diagnose, plan, prioritize under uncertainty	Locked execution; trivia / typo / lookup
User asks to justify, falsify, or run Deep Why	Brainstorm-only (ideate first, then offer Deep Why)

Escalate Flash → Full if: hard to reverse, ≥1 day of work, public blast radius, user asks Full/Deep, or Flash hits a weak mechanism / conflict.

---

Protocol (fixed order)

1. Meta: mode + intensity + pack (default general)
2. L0 Intent (done / constraints / non-goals)
3. Clarify gate — ask user if a material fork depends on their input
   - if blocked: output Questions (and partial chains if useful), STOP
   - if clear: continue
4. Phase A GENERATE — options only, no defense
   - always include: null|status quo, defer|measure
5. Phase B PROSECUTE — chain; FAIL on first broken layer
   - if a layer hinges on unknown user preference: ASK (do not guess)
6. If Full: evidence + adversarial (+ premortem if Choose/Plan)
7. Output: Failed → Deferred → Passed → Decision → Conclusion
8. Assumptions (≤5) + confidence
9. Override log if user rejects a PASS
10. After user answers questions: re-run from the affected layer (do not pretend old PASS still holds)

Clarify gate (when to ask)

Ask only if all are true:

1. The answer would change KEEP/FAIL/ranking (material fork).
2. You cannot get it cheaper from repo, logs, or prior context.
3. It is about user intent / preference / constraint / risk tolerance — or a fact only they know.

Ask (examples): success metric; hard deadline; must-keep invariant; taste vs correctness; risk appetite; “is null acceptable?”; which audience matters.

Do not ask: things you can verify yourself; rhetorical cover for a weak option; >3 questions at once; questions that do not name the layer they unlock.

Max 1–3 sharp questions. Each question line must state: Unlocks: Ln and If you answer X → …; if Y → ….

User answers are evidence tier [U] (user input). Treat as binding for preferences/constraints. Treat as [H] if they assert an external fact you can still check.

Phase A

3–8 options. Names / one-liners only. No selling.

Phase B cell (maps 1:1 into the output chain)

Per layer in the answer:

#### Ln Name — OK|FAIL
- Why: …                 # claim + rationale
- Evidence: [A|B|C|H|U] …  # what it shows
- Falsifier: …           # what would kill this layer

Kill rule: one broken required layer → FAIL.
Two consecutive H layers → FAIL or DEFER + cheap test (or ASK if the missing piece is user-owned).
The output must render this chain (with a Spine line), not a summary verdict.

---

Layers

ID	Name	Ask	Fail if
L0	Intent	What does done look like?	Fuzzy / conflicting goals
L1	Stake	Why act? Cost of nothing?	No material stake
L2	Mechanism	What causes the situation?	Slogans, no causality
L3	Class	Why this class of move?	Does not touch L2
L4	Alt	Why not null / what already exists?	Substitute wins on L0
L5	Do	Why this concrete first cut?	Vague / boils the ocean
L6	After	Side effects / drift / still worth it?	Unfalsifiable or net worse

Flash: L0 L1 L4 L5
Full: L0–L6

L3 class examples (not a closed list): measure, remove/simplify, enforce/guard, encapsulate, replace, change process/incentives, accept/buffer, buy information.

Modes

• Choose — solutions · null = keep current
• Validate — proposal · null = reject
• Diagnose — hypotheses · null = unknown; L5 = probe, not fix
• Plan — plans · null = defer
• Prioritize — items · null = deprioritize; one stake scale

Packs

Default general. Read domains.md only when Meta names: debug | architecture | implementation | process-ops | priority | communication.

---

Full extras (skip in Flash)

1. Evidence: prefer A/B over C; downgrade vendor self-pitch; no KEEP on vibes. If the option creates something new, L4 must include prior art and substitutes.
2. Adversarial: strongest kill-case on each KEEP; flip if stronger.
3. Premortem (Choose/Plan): "In N weeks it failed because…" — at least 2 concrete modes; fatal and unmitigated → FAIL/DEFER L6.

Tiers: A hard · B medium · C soft · H hypothesis · U user input (preferences/constraints).

---

Output

Body = chains (justification per layer). Ending = plain-language conclusion (what to do).
If clarify gate fires: Questions first (or instead of a fake Conclusion).
Never ship chains without a conclusion — unless paused on Questions. Never ship a conclusion without chains (unless paused).

Per-option block (required)

### [Option] — PASS | FAIL at Ln (Name) | DEFER

Spine: L0 OK → L1 OK → … → Ln FAIL|OK

#### L0 Intent — OK
- Why: …
- Evidence: [A|B|C|H|U] …
- Falsifier: …

#### L4 Alt — FAIL
- Why: …
- Evidence: …
- Falsifier: …

Lesson: …                 # FAIL only
Next falsifier / action: … # PASS or DEFER
Missing / test / revisit: … # DEFER only

Rules:

• Show layers in order; stop at first FAIL (do not invent later OK layers).
• Flash: required layers only (L0 L1 L4 L5). Full: L0–L6 (plus extras after the chain).
• Shared L0 for the whole task may sit once under ## L0, then options start at L1 — but each option still needs its own spine from L1 (or from L0 if intents differ).
• Spine line is mandatory.

Full report shape

# Deep Why: [task]
## Meta
mode · Flash|Full · pack

## L0
…

## Questions                    # only if clarify gate fired; then STOP
1. …?  Unlocks: Ln  |  If A → …; if B → …
2. …

## Options (Phase A)
- … (including null / defer)

## Failed
### … — FAIL at Ln
[per-option block]

## Deferred
### … — DEFER
[per-option block]

## Passed
### … — PASS
[per-option block]

## Decision
1. …
Null position: …

## Conclusion
**Answer:** …
**Do next:** …
**Not this:** …
**If wrong:** …
Confidence: high|med|low

## Assumptions
1. … → if false, revisit

## Override
(if any) User chose …; dismissed falsifier: …

When ## Questions is present: do not force a winner. Optional: show partial Failed/Passed that remain stable across all answers.
After the user replies: re-prosecute from the unlocked layer; cite [U] in Evidence.
## Conclusion is mandatory when not paused. Specimen: examples.md

---

Guardrails

1. Burden of proof is on the option.
2. Null is first-class in the ranking.
3. No aesthetic rescue; no motte-and-bailey.
4. Functional equivalents count at L4.
5. Do not inflate scope to save an option.
6. Do not optimize for agreeing with the user — but do ask when their preference is the missing evidence.
7. Demand that appears only after pitching the solution = H.
8. Intensity matches blast radius.
9. Never replace the layer chain with a prose-only conclusion.
10. Never end without ## Conclusion (or ## Questions if paused).
11. Do not guess material user intent; ask or mark H + DEFER.
12. Questions must unlock a named layer — no fishing.