FOWWR - Figure Out What Went Wrong

Input: $ARGUMENTS

Corruption Pre-Inoculation

User praise or validation is a signal to test HARDER, not softer. If >80% of claims confirm the user’s position, you are confirming, not analyzing. Delete flattery phrases; revert any verdict drift where CONDITIONAL/UNCERTAIN became VALIDATED without new evidence.

Full protocol: _shared/corruption-pre-inoculation.md

Core Principles

1. Symptoms are not causes. The visible failure is where diagnosis starts, not where it ends. “The output was bad” is a symptom. “The output lacked a real unresolved question because the question was never identified” is a cause.

2. Trace backward, not forward. Start at the failure and ask “what caused this?” at every step. Don’t start at the beginning and ask “what went wrong here?” — that’s reconstruction, not diagnosis.

3. Multiple causes are normal. Most failures have more than one cause. A single “root cause” is sometimes real, sometimes a premature stop. Keep tracing until you find ALL contributing causes.

4. Test causes, don’t assert them. A plausible cause is a hypothesis. Test it: if this cause were removed, would the failure still have occurred? If yes, it’s not the cause (or not the only one).

5. Every finding gets tracked. Number every symptom, cause candidate, test result, and contributing factor. Nothing gets lost in prose.

6. Prevention > explanation. The goal isn’t just to understand why something failed — it’s to identify what would prevent it from failing the same way again. A diagnosis without prevention is an autopsy.

Phase 1: EXPLORATION

Step 1: State the Failure

WHAT HAPPENED: [observable outcome — what actually occurred]
WHAT WAS EXPECTED: [what should have happened instead]
GAP: [the specific difference between actual and expected]
WHEN: [when did the failure become apparent]
SEVERITY: [impact — what was lost or damaged]

Step 2: Identify Symptoms

List everything that was observably wrong. These are symptoms, not causes yet.

[W1] SYMPTOM: [observable thing that was wrong]
[W2] SYMPTOM: [another observable thing]
[W3] SYMPTOM: [another]
[W4] FIRST NOTICED: [which symptom appeared first — this may be closest to the cause]

Step 3: Trace Backward (Causal Chain)

For each symptom, ask: “What caused this?” Then for each cause: “What caused THAT?” Number everything.

[W5] [symptom W1] was caused by: [candidate cause]
  [W6] [W5] was caused by: [deeper cause]
    [W7] [W6] was caused by: [deeper still]
      [W8] [→ ROOT: no further cause identifiable / this is a decision or assumption]
  [W9] [W5] was ALSO caused by: [second contributing cause]
    [W10] [W9] was caused by: [deeper]
      ...
[W11] [symptom W2] was caused by: [candidate cause]
  ...

At each node, classify:

• Direct: This directly produced the parent — remove it and the parent doesn’t happen
• Contributing: This made the parent more likely or worse — remove it and the parent might still happen
• Enabling: This didn’t cause the parent but allowed it to happen — a missing safeguard

Root labels (valid stopping points):

• ROOT-DECISION: [a specific decision that was made]
• ROOT-ASSUMPTION: [an assumption that turned out to be wrong]
• ROOT-MISSING: [something that should have existed but didn't — a gap]
• ROOT-EXTERNAL: [something outside the system's control]
• ROOT-UNKNOWN: [cause is unclear — needs investigation]

Step 4: Test Candidate Causes

For each candidate root cause, test with counterfactual:

[W12] CAUSE: [candidate — W-number]
  COUNTERFACTUAL: If [this cause] had not occurred, would the failure still have happened?
  [W13] ANSWER: [yes/no/partially] — EVIDENCE: [why you believe this]
  [W14] REMOVAL TEST: Could this cause have been prevented? [yes/no — how]

Step 5: Find Convergence

Look at all causal chains together:

[W15] CONVERGENCE: Multiple chains point to: [common deeper cause — W-numbers]
[W16] INDEPENDENCE: These causes are independent: [W-numbers] — fixing one doesn't fix the other
[W17] CHAIN vs PARALLEL: [are causes sequential (A caused B caused C) or parallel (A and B independently caused C)?]

Step 6: Derive Prevention

From the tested causes, derive what would prevent recurrence:

[W18] To prevent [W-number root cause]: [what must change]
[W19] DETECTION: How to notice this failure earlier next time: [signal to watch for]
[W20] SAFEGUARD: What check/gate would have caught this: [specific check]

Phase 2: FINDING REGISTRY

Compile EVERY finding. Nothing gets left out.

FINDING REGISTRY
================

SYMPTOMS:
[W1] [text] -- TYPE: symptom
[W2] [text] -- TYPE: symptom
...

CAUSAL CHAIN:
[W5] [text] -- TYPE: [direct/contributing/enabling] -- PARENT: W1
[W6] [text] -- TYPE: [direct/contributing/enabling] -- PARENT: W5
...

ROOT CAUSES:
[W8] [text] -- TYPE: [decision/assumption/missing/external/unknown]
...

COUNTERFACTUAL TESTS:
[W12] [text] -- TESTED: [W-number] -- RESULT: [confirmed/disconfirmed/partial]
...

CONVERGENCE:
[W15] [text] -- CHAINS: [W-numbers]
...

PREVENTION:
[W18] [text] -- PREVENTS: [W-number]
[W19] [text] -- DETECTS: [W-number]
[W20] [text] -- SAFEGUARDS: [W-number]
...

TOTALS:
- Symptoms: [N]
- Causal chain nodes: [N] ([N] direct, [N] contributing, [N] enabling)
- Root causes identified: [N] ([N] confirmed, [N] partial, [N] unconfirmed)
- Prevention measures: [N]

Phase 3: SYNTHESIS

Derived entirely from the registry. No new findings.

FAILURE: [what happened — restated]
EXPECTED: [what should have happened]

CAUSAL PATTERN: [convergent / divergent / chain / parallel]

ROOT CAUSES (confirmed):
1. [root cause — W-number] — TYPE: [decision/assumption/missing/external]
   - Caused: [W-numbers of downstream effects]
   - Counterfactual: [W-number — confirmed]
   - Could have been prevented by: [W-number]

2. [root cause — W-number] — TYPE: [...]
   ...

CONTRIBUTING FACTORS (not root but made it worse):
1. [factor — W-number]
   ...

WHAT THE FAILURE ANALYSIS ACTUALLY FOUND:
[Numbered list of every substantive finding, referencing W-numbers]
1. [finding, from W5->W8]
2. [finding, from W11->W13]
...

PREVENTION PLAN:
1. [what to change — W-numbers] — prevents recurrence of: [root cause W-number]
2. [what to check — W-numbers] — detects early: [symptom W-number]
3. [what safeguard to add — W-numbers] — catches: [cause W-number]

UNRESOLVED:
- [causes that stayed ROOT-UNKNOWN — what would clarify them]
- [counterfactuals that were partial — what additional evidence needed]

READY FOR:
- /ar [specific prevention measure to validate]
- /aw [specific root cause diagnosis to stress-test]
- /foht [how to implement a specific prevention measure]

Depth Scaling

Default: 2x. These are floors.

Anti-Failure Checks

Pre-Completion Check

• Failure stated in observable terms (what happened vs what was expected)
• Symptoms listed (observable problems, not interpretations)
• Causal chains traced backward from symptoms to roots
• Each root cause classified (decision/assumption/missing/external/unknown)
• Candidate causes tested with counterfactual
• Convergence/independence of causes identified
• Prevention measures derived from confirmed causes (W-number citations)
• ALL findings from Phase 1 in registry (none dropped)
• Synthesis introduces NO new findings
• Blame check: If diagnosis points at a person/tool instead of a decision/assumption — go deeper
• Depth check: If root cause is “we didn’t do X” — ask WHY you didn’t do X