self-consistency-reasoner

Self-Consistency Reasoner

A structured reasoning technique based on the Self-Consistency method (Wang et al., ICLR 2023).

Core idea: Complex problems often have multiple valid paths to the correct answer. Incorrect reasoning, even when confident-sounding, tends to scatter across different wrong answers. By generating N independent reasoning paths and taking majority vote, we reliably surface the correct answer and get a built-in confidence signal for free.

---

When This Fires

This skill is invoked internally by:

• systematic-debugging — during root cause hypothesis generation (Phase 3)
• verification-before-completion — during evidence evaluation

It fires when:

• The reasoning requires 3+ non-trivial steps
• A single reasoning chain could make a wrong assumption or logical error
• The answer has a fixed answer set (a root cause, yes/no, a specific conclusion)
• Being wrong has real cost (wrong diagnosis wastes edits; false "done" wastes review cycles)

---

How Many Paths to Generate

Scale paths to difficulty:

Problem Type	Paths
Binary verification (does this evidence prove the claim?)	3 paths
Root cause diagnosis with 2-3 candidates	5 paths
Complex multi-factor diagnosis or high-stakes verification	7 paths

Default: 5 paths. Research shows gains plateau quickly — 5 captures most of the benefit of 40.

---

The Process

Step 1: Generate N Independent Reasoning Paths

Produce each path independently — don't let earlier paths contaminate later ones. Vary your approach deliberately:

• Use a different starting point or framing
• Work forward from given info in one path, backward from the goal in another
• Decompose the problem differently across paths
• For debugging: start from different points in the call stack, assume different failure modes
• For verification: evaluate the evidence from different angles (what would prove it true? what would prove it false?)

Each path must end with a clearly parsed final answer.