synapseradio/audit-chain-of-thought

Follow these steps:

1. Extract the reasoning chain

Identify each distinct inferential step from premises to conclusion. Make implicit steps explicit — "A therefore C" often hides "A therefore B, B therefore C."

2. Tag each step by type

• Deductive: conclusion necessarily follows (if premises true, conclusion must be true). E.g. "All users need authentication. This is a user. Therefore this needs authentication."
• Inductive: conclusion generalizes from instances (if premises true, conclusion probably true). E.g. "The last 40 deploys succeeded. Therefore this deploy will probably succeed."
• Abductive: conclusion is the best explanation for observations. E.g. "Performance degraded after the cache change. Therefore the cache change probably caused it."

3. Assess soundness at each step

• Deductive: are the premises true? Is the form valid?
• Inductive: how many instances? How representative? What counterexamples?
• Abductive: what alternative explanations exist? Which is actually best?

4. Identify the weakest links

Where does the chain depend on probabilistic leaps, unverified assumptions, or competing explanations? These are where the conclusion could fail even if earlier steps hold.

5. Surface the overall structure

How much of the chain is deductive vs inductive vs abductive? A mostly-deductive chain is stronger than one resting on multiple abductive leaps.