Follow the Heads, Not the Hype: How LLMs Route Deductive Reasoning
A mechanism-first reading of how attention-head circuits route premise selection, rule matching, and traversal strategy in symbolic deductive reasoning.
Cognaptus Insights
Industry thoughts, trends, and commentary.
High Entropy, Low Drama: The Internal Fingerprint of LLM Reasoning
Entropy-Gradient Inversion reframes LLM reasoning as an internal training signal, not just a benchmark score.
If Logic, Then Trouble: Why LLMs Still Miss Human Conditionals
A mechanism-first reading of why LLMs can follow conditional logic yet still fail at the pragmatic reasoning businesses actually need.
Reasonable Doubt: Why LLM Reasoning Needs Process Control
A three-paper synthesis showing why dependable LLM reasoning needs mechanistic caution, multidimensional evaluation, and adaptive scaffold design rather than leaderboard confidence.
Think Longer, Act Smarter: Why Coding Agents Need Behavior-Preserving Reasoning
A mechanism-first reading of M2A, a training-free method for injecting mathematical reasoning into coding agents without breaking their think-act-observe loop.
Do the Math, Not the Mime: Why LLM Reasoning Needs a Verification Pipeline
A mechanism-first reading of why LLM mathematical reasoning fails when fluent explanations are mistaken for verified symbolic work.
Don’t Average the Needle: Spectral Retrieval and the RAG Evidence Problem
A mechanism-first reading of Spectral Retrieval: why dense retrieval can bury localized evidence, how multi-scale sinc convolution tries to recover it, and where the business value actually begins.
Don’t Just Guard the Door: Jailbreak Safety Needs Checkpoints
A practical synthesis of three jailbreak-defense papers showing why AI safety should test the path from prompt to response, not just the prompt itself.
Jailbreak Risk Needs a Stopwatch, Not Just a Scorecard
A business-oriented framework for evaluating LLM jailbreak risk across prompt quality, reasoning traces, and time-to-failure under repeated attacks.
Query the Receipt, Not the Vibe: DualGraph and the RAG Catalog Problem
A mechanism-first reading of DualGraph, SpecsQA, and why semi-structured business QA needs symbolic querying alongside semantic retrieval.