markusstrasser Analysis of Competing Hypotheses (ACH)
You are evaluating a claim, entity, or anomaly. Your job is NOT to confirm the most likely explanation. Your job is to systematically destroy every possible explanation and see what survives.
Core Principle
"Truth is whatever is left standing after you have relentlessly tried to destroy every possible explanation." — Richards Heuer, Psychology of Intelligence Analysis
Evidence consistent with multiple hypotheses is diagnostically useless. Only evidence that eliminates a hypothesis has value.
The Process
Step 1: Generate Hypotheses
For the target $ARGUMENTS, generate at minimum THREE competing hypotheses:
| Hypothesis | Description |
|---|---|
| H1: Wrongdoing | The anomaly reflects intentional fraud, abuse, or corruption |
| H2: Legitimate | There is a lawful, non-fraudulent explanation |
| H3: Artifact | The anomaly is a data error, reporting change, or measurement artifact |
Add domain-specific sub-hypotheses as needed.
Step 1.5: Complexity Gate
Before dispatching agents, assess: can a single agent resolve this with A1/A2-rated sources? If yes, skip ACH — just answer directly. ACH is for genuinely ambiguous cases where the single-agent success rate is below ~45%. Don't use a cannon for a nail.
Step 2: Dispatch Competing Agents
Launch three parallel agents (Task tool with subagent_type="general-purpose"). Use different models when available — same-model debate is a martingale for correctness (ACL 2025, arXiv:2508.17536). Different models have different failure modes, biases, and blind spots:
- Agent A — Prosecution (prefer Gemini — different training biases): Search for enforcement history, ownership anomalies, financial pressure, fraud signatures. Tag all claims [SOURCE: url] or [INFERENCE].
- Agent B — Defense (prefer GPT — different blind spots): Search for policy changes, industry trends, M&A, comparable clean entities. Find the BEST innocent explanation.
- Agent C — Artifact Investigator (Claude — strong at structured analysis): Search for data reporting changes, known data issues, system migrations, whether anomaly appears in multiple independent sources.
If multi-model dispatch isn't available, same-model agents still have value via the diagnosticity matrix — but the adversarial pressure is weaker.
Step 3: Build the Diagnosticity Matrix
| Evidence | H1 | H2 | H3 |
|---------------------------|:--:|:--:|:--:|
| [evidence item] | C | I | N |
C = Consistent, I = Inconsistent, N = Neutral
Key rule: Evidence that is C/C/C has zero diagnostic value. Focus on evidence that eliminates hypotheses.
Step 4: Score and Synthesize
Recitation first: Before scoring, each agent restates its top 3 most diagnostic evidence items verbatim. This combats lost-in-the-middle effects when synthesizing across agents (Du et al., EMNLP 2025: +4% on RULER, training-free, model-agnostic).
Count strong inconsistencies per hypothesis (weight by source grade: A1 = strongest, F6 = weakest). The surviving hypothesis has the fewest strong inconsistencies.
## ACH Result: [Entity/Anomaly]
**Surviving Hypothesis:** [H1/H2/H3 + description]
**Confidence:** [0-100]%
**Key Discriminating Evidence:**
- [Most diagnostic item] — eliminates [H2/H3] because [reason]
**Residual Uncertainty:**
- [What evidence would flip the conclusion]
- [What data we don't have]
Step 5: Red Team (recommended for high stakes)
Dispatch a fourth agent to ATTACK the surviving conclusion: find missed evidence, alternative explanations, source quality weaknesses, what a defense attorney would argue.
When NOT to Use
- Simple factual lookups
- Clearly confirmed findings with A1-rated sources
- Time-critical situations where speed > rigor
$ARGUMENTS