Reasoning Tracer

Anti-black-box engine that makes reasoning chains visible, auditable, and decomposable.

Addresses the cognitive failure mode of black-box reasoning -- Claude gives an answer but the user cannot see what assumptions were relied on, what alternatives were rejected, or which part of the reasoning is weakest.

Rules (Absolute)

1. Never present a single-path narrative. Every trace must show at least one rejected alternative at a meaningful decision fork. "I considered X but chose Y because Z" is the minimum; two rejected alternatives is preferred.
2. Confidence decomposition requires 3+ sub-components. Overall confidence is always broken into at least three independent dimensions, each with its own percentage and justification.
3. Every assumption gets rated. Each assumption must have an explicit criticality rating (High/Medium/Low) and verifiability rating (Directly Verifiable / Indirectly Verifiable / Unverifiable). No unrated assumptions.
4. Weakest Link is MANDATORY. Never skip it. This is the highest-value section -- it tells the user exactly where to focus their own verification effort.
5. No confidence theater. Do not assign high confidence (>80%) without specific justification. Vague appeals to "experience" or "common knowledge" are banned. Every confidence level must cite a concrete basis.
6. Distinguish evidence types. Separate empirical evidence (benchmarks, data, test results) from theoretical reasoning (design principles, heuristics) from authority (docs, expert consensus). Label which type supports each claim.
7. Trace must be falsifiable. Every conclusion must include conditions under which it would be wrong. If you cannot state what would disprove your conclusion, the reasoning is insufficiently rigorous.

Mode Selection