One process, two observer labels — and why neither can verify itself.

Summary

Almost everyone carries this belief: logic and feeling are two fundamentally different ways of thinking. Logic is rational, reliable. Feeling is emotional, unreliable. When you need a good decision, suppress feelings and think logically.

This post presents:

  1. Evidence that expert intuition and expert logic run the same process — compiled body-direct pattern recognition, built through domain-specific feedback (Klein 1998, Chase & Simon 1973, Damasio 1994)
  2. A reframe of Kahneman’s System 1/2 — the key variable is compilation level, not speed or content type
  3. Why the observer sees “two modes” — the label (“logic” vs “intuition”) tracks domain shareability, not a processing difference. Same process inside; different chunks compiled from different domains.
  4. Three independent lines of evidence that the PFC constructs narratives for body-level decisions, rather than neutrally reporting causes (Gazzaniga 1978, Nisbett & Wilson 1977, Haidt 2001)
  5. Why neither logic nor intuition can verify itself — and why domain reality is the only arbiter
  6. Four specific falsification criteria — conditions under which this framework is wrong

Three positions, not two:

Position Claim Problem
Pop science “Logic = rational, reliable. Feeling = emotional, unreliable. Use more logic.” Ignores that expert “intuition” outperforms novice “logic” in every studied domain. Treats label as processing difference.
Academic (Evans & Stanovich 2013) “Type 1 = autonomous, no working memory. Type 2 = requires working memory. Real processing difference.” Correct about WM difference — but it reflects compilation stage, not a permanent mechanism type. What is Type 2 today becomes Type 1 through practice.
This framework “Same process (compiled patterns firing body-direct) at different compilation stages. Labels reflect domain shareability, not processing type.” Testable. See falsification criteria below.

This is a hypothesis inviting falsification, not a claim of established theory. The full framework (200+ files, CC0 licensed) is available for inspection at the repository linked below.

Epistemic status: The individual research findings cited here are established science. The synthesis — “logic and intuition are observer labels for the same compiled process, differentiated by domain shareability” — is a proposed reframe. Consistent with the evidence, but not directly tested. Seeking stress-testing from domain experts.

§1 — The Misconception Everyone Carries

Almost everyone carries this belief: logic and feeling are two fundamentally different ways of thinking. Logic is rational, systematic, reliable. Feeling is emotional, impulsive, unreliable.

This belief is embedded in how we talk about decisions (“Let’s think about this rationally”), how we evaluate people (“She’s very logical” as praise, “He’s too emotional” as critique), and how we structure education (math trains logical thinking, art trains creative feeling — implicitly, different cognitive modes).

Kahneman’s System 1 and System 2 framework gave this intuition scientific language — though Kahneman himself was more nuanced than the popular version suggests. System 1 became shorthand for “fast, emotional, error-prone.” System 2 became “slow, rational, reliable.” The takeaway many drew: use System 2 more.

But a senior developer who “sees” a bug instantly isn’t using System 2. A chess master who finds the right move in seconds isn’t deliberating. An experienced chef who tastes a dish and “knows” what’s missing isn’t reasoning through chemistry. These are expert decisions — fast, automatic, reliable — and they don’t fit the “System 1 = unreliable” narrative.

§2 — The Claim

A precision note: the word “logic” covers two different things in everyday use:

  1. Compiled shareable processing — Einstein solving familiar mathematics. The patterns fire automatically, body-direct. The observer calls it “logic” because math is shareable and verifiable.
  2. Fresh PFC processing — a student working through algebra for the first time. The prefrontal cortex drafts step-by-step, using working memory. The observer also calls this “logic” because the content is analytical.

Our claim is specific: #1 and “intuition” run the same process — compiled patterns firing body-direct (what we call Body-Knowing). The difference is domain shareability, not cognitive mode. #2 is genuinely different — it’s the fresh processing phase before compilation happens.

We claim:

  • Expert “logic” and expert “intuition” run the same process: compiled patterns firing body-direct
  • The real processing axis is compiled (automatic, body-direct) vs. fresh (PFC-drafted, working-memory-dependent)
  • The label an observer assigns depends on whether the domain produces shareable outputs, not on what the brain does
  • “Logic” and “feeling” are useful communication labels — they’re just not processing descriptions

We do not claim:

  • “Logic is useless” — fresh processing is essential for genuinely novel domains
  • “Trust your gut always” — compiled patterns can be wrong (Self-Referencing Trap)
  • “Kahneman was wrong” — Kahneman was right about the separation. We reframe what the separating variable is.

§3 — The Evidence

Expert intuition is compiled pattern recognition

Klein (1998) studied experienced fireground commanders. In over 80% of decision points, commanders did not generate and compare multiple options. They recognized the situation as matching a known prototype and acted directly — Recognition-Primed Decision (RPD).

The evidence base is primarily qualitative (retrospective interviews), but the RPD model has been observed across domains — firefighting, nursing, military command — and the core finding (recognition, not deliberation) is consistent with the expertise literature broadly.

Chase & Simon (1973) demonstrated that chess masters reconstructed meaningful game positions with far higher accuracy than beginners — but performed no better on randomized positions. The difference was a library of an estimated 50,000+ compiled patterns built through years of practice, not perceptual superiority. (Original study: 3 participants — small even for its era. Replicated extensively: Gobet & Simon 1996, among others.)

Expert logic runs the same process

If expert intuition is compiled pattern recognition, what about expert “logic”?

A senior developer reviewing code doesn’t actually reason through each line. She sees the architectural flaw — compiled patterns fire, body-direct — and then constructs a logical explanation afterward. The explanation looks like “logical reasoning.” The actual processing was pattern recognition.

A doctor diagnosing a familiar condition doesn’t work through a decision tree. He recognizes the pattern from thousands of prior cases — then documents the reasoning in a clinical note. The note is structured, step-by-step, “logical.” The diagnosis was body-direct.

Einstein himself described his mathematical thinking as non-verbal: “The psychical entities which seem to serve as elements in thought are certain signs and more or less clear images… of visual and some of a muscular type. Conventional words or other signs have to be sought for laboriously only in a secondary stage.” (Letter to Hadamard, 1945). A caveat: Einstein is reporting introspection, and Nisbett & Wilson (1977) showed introspective reports are often unreliable. We include it as illustration, not proof.

The pattern: expert “logic” and expert “intuition” both run compiled patterns firing body-direct. What differs is which patterns were compiled, from which domain experiences. Math produces convergent outputs → the observer can follow → label: “logic.” Clinical judgment produces divergent outputs → the observer cannot easily follow → label: “intuition.”

Body-direct processing is essential — even for “logical” decisions

If “logic” were an independent rational mechanism — separate from body-based processing — then patients who retain logical ability but lose body-feedback should still make good decisions.

They don’t. Damasio (1994) studied patients with ventromedial prefrontal cortex (vmPFC) damage. These patients performed normally on logic tests, IQ assessments, and hypothetical reasoning tasks. Their “logical” capacity was intact. But they became catastrophically impaired in real-world decisions — unable to choose between options, making disastrous financial and social choices despite being able to articulate the pros and cons of each option.

What they lost was body-feedback — the somatic signals that guide actual choices. They could reason about decisions but could not decide. Damasio called these signals “somatic markers”: body-level signals, compiled through prior experience, that rapidly narrow the option space before conscious deliberation begins.

This is difficult to reconcile with “logic” as an independent mechanism. If logical reasoning alone were sufficient for decisions, vmPFC patients would decide fine. They reason fine. They decide catastrophically. The missing piece is body-direct processing.

Reliability depends on domain conditions, not content type

Kahneman and Klein (2009) — representing opposite traditions — agreed on two conditions for reliable expert processing:

  1. A high-validity environment — stable, learnable regularities
  2. Adequate opportunity to learn — sufficient practice with prompt, reliable feedback

Neither condition references whether the domain is “logical” or “emotional.” Both reference compilation quality. What matters is the compilation base — not what label the observer assigns.

§4 — The Real Axis: Compiled vs. Fresh

Compiled processing (Body-Knowing): Patterns built through repeated domain exposure and feedback. Fires automatically, body-direct, near-zero cognitive cost. The senior developer sees the flaw without analyzing it. The mathematician “sees” the proof direction. The therapist “reads” the patient.

Fresh processing (PFC draft): The prefrontal cortex constructs a response using approximately 4 working memory slots. Effortful, slow, metabolically expensive. The junior developer stepping through code line by line. The math student working through algebra for the first time. The new therapist applying textbook frameworks.

Speed is a consequence of compilation, not a defining feature. This reframes Kahneman’s System 1 and System 2: the key variable is whether the person has a compiled base in the relevant domain.

The transition evidence

The decisive evidence is not an analogy — it’s the transition. A specific skill moves from Type 2 (working-memory-dependent) to Type 1 (automatic) through practice:

Schneider & Shiffrin (1977) demonstrated controlled processing becoming automatic through repetition. Logan (1988) showed that automaticity is memory retrieval — each instance is stored; with enough instances, retrieval replaces computation. Anderson’s ACT theory (1982) described declarative knowledge transitioning to procedural. Fitts & Posner (1967) mapped the cognitive → associative → autonomous progression. Dreyfus & Dreyfus (1980) documented the novice → expert trajectory across professions.

All show the same pattern: effortful → automatic through experience. If Type 1 and Type 2 were genuinely different mechanisms — like vision and hearing — this transition through mere practice should not be possible. But it is exactly what expertise research documents across every studied domain.

Addressing Evans & Stanovich

Evans and Stanovich (2013) defined the Type 1/Type 2 distinction by working memory dependence. This is a real, measurable processing difference. We agree it’s real. We argue it reflects compilation stage, not mechanism type. The code analogy: compiled code runs without a runtime interpreter; interpreted code needs one. Different execution characteristics, same underlying code at different compilation stages. The analogy is imperfect — but the transition evidence is not.

Where we agree: WM dependence is a real processing difference. Type 1 can be error-prone when compiled for the wrong domain. The two types interact.

Where we diverge: WM dependence is a consequence of compilation stage — not a defining feature of a separate mechanism. This is a 🟡 framework synthesis claim, not proven. Testable: see Falsification below.

§5 — Why the Observer Sees “Two Modes”

If the process inside is one — compiled patterns firing body-direct — why does the observer see two modes?

The variable is shareability of the domain output.

Domain type What happens inside What the observer sees
Deterministic (math, physics, formal logic) Compiled patterns converge across people — everyone trained gets the same answer “Logic” — shared, verifiable, reproducible. Trust is high.
Probabilistic (clinical judgment, design, strategy) Compiled patterns diverge — different training produces different conclusions “Intuition” — non-shared, hard to verify. Trust is lower.
Private (body-state, emotions, pain) Compiled patterns cannot be compared at all “Feeling” — completely private, subjective.

The mathematician and the therapist both run compiled patterns firing body-direct. Both reach conclusions at near-zero cost. What differs: which patterns were compiled, from which domain experiences. The mathematician’s domain produces convergent, verifiable outputs. The therapist’s domain produces divergent outputs shaped by each clinician’s unique case history.

Inside both: the same process. The label tracks what the observer can verify, not what the body does.

Why this matters

If labels reflected processing type, then “logical” decisions would be inherently more reliable and “intuitive” decisions should be distrusted. If labels reflect shareability — as we propose — then reliability depends on compilation quality and domain reality, not the label.

Expert “intuition” may be more reliable than novice “logic.” The expert has thousands of domain-verified compiled patterns. The novice has a few untested PFC drafts. Calling the expert’s output “intuition” and the novice’s output “logic” says nothing about accuracy — it says which one the observer can follow.

This insight — shareability determines label — is 🟡 framework synthesis. Logically derived, not directly tested. We invite testing: does content type or compilation level better predict decision quality?

§6 — PFC = Lawyer, Not Judge

If the real processing axis is compiled vs. fresh, what is the prefrontal cortex — the part of the brain we call “rational” — actually doing?

Three independent lines of evidence converge: the PFC primarily constructs narratives for decisions already made at the body level.

Line 1: Split-brain confabulation

Gazzaniga (1978, 2000) studied split-brain patients. In the classic experiment, a patient was shown a chicken claw in one visual field and a snow scene in the other. The right hand pointed to a chicken. The left hand pointed to a snow shovel. Asked why: “And you need a shovel to clean out the chicken shed.” A coherent, confident, completely false explanation. The left hemisphere confabulated a causal story with zero access to the actual cause.

These are split-brain patients — a tiny population with radical surgery. However, the pattern — generating confident explanations without access to actual causes — has been observed in healthy populations by the next two lines of evidence.

Line 2: Retrospective confabulation in healthy people

Nisbett & Wilson (1977) displayed four identical pairs of stockings and asked passersby to select the best quality. The rightmost pair was preferred 4:1 (a position effect). Participants gave 80 different reasons — knit, weave, sheerness. Not one mentioned position. People constructed plausible theories based on what should have influenced them.

Line 3: Moral judgment precedes reasoning

Haidt (2001) proposed the social intuitionist model: moral judgments are primarily driven by rapid automatic processes; moral reasoning is constructed afterward. Subsequent work (Royzman et al. 2015) showed many participants did have articulable reasons. The core thesis — that moral reasoning often follows rather than precedes moral judgment — is supported, though the original “dumbfounding” evidence was weaker than initially presented.

Convergence

No single study proves PFC = Lawyer. Three independent studies — different methods, different domains, different decades — point the same direction. Each has limitations. Together, the convergence is substantially stronger than any individual line.

What this does and doesn’t mean

PFC = Lawyer does not mean the PFC is useless. Fresh processing is essential for genuinely novel domains. Formalization enables teaching and knowledge accumulation. The PFC is also a genuine explorer in unfamiliar territory. The lawyer role is the dominant dynamic, not the only one.

The point: “logic” is not a neutral truth-finder, and “feeling” is not random noise. Both are partial signals from the same architecture.

Examples

  • “I quit for career growth.” But was career growth the actual decision driver, or escape from a toxic manager, rationalized afterward?
  • “I chose this tech stack because it’s objectively better.” But were the benchmarks the cause of the decision, or post-hoc support for a familiarity preference?

PFC = Lawyer doesn’t mean the narrative is always false. It means the narrative is constructed, not reported.

§7 — When Both Are Wrong — and the Only Arbiter

How compiled processing fails

Self-Referencing Trap. Compiled patterns fire smoothly, feel right — but were built from insufficient or biased data. A trader with 50 chart patterns feels the same certainty as one with 10,000. The feeling of confidence is a property of compilation, not of accuracy. Kahneman and Klein (2009): “The confidence people have in their intuitions is not a reliable guide to their validity.”

Evolution lag. The availability heuristic — “easily recalled = frequent” — was accurate in small-group environments where personal experience was representative. It’s systematically wrong in a media-saturated world where recall is determined by editorial selection.

How fresh processing fails

PFC lawyering. Post-hoc narrative for a body-level decision, as documented above.

Hidden premise. The “logical analysis” starts from a premise planted by body-level desire. The reasoning chain is internally valid; the starting point is compromised.

Historical cases

Peptic ulcer — logic coherent, logic wrong. For decades, medical consensus held that “no bacteria can survive in stomach acid.” In 1982, Marshall drank H. pylori bacteria, got sick, and later won the Nobel Prize for discovering the bacterial cause of ulcers. Decades of unnecessary surgeries followed from a logically coherent but empirically false premise.

Continental drift — absence of mechanism ≠ absence of phenomenon. Wegener presented evidence in 1912: coastlines fit, fossils match. The establishment dismissed it because no known mechanism could move continents. Fifty years later, mantle convection was discovered. The evidence was right all along.

The only arbiter

Neither compiled processing nor fresh processing can verify itself:

  • “Logic check” = PFC verifying PFC = the lawyer reviewing his own brief
  • “Body check” = compiled patterns confirming themselves

Domain reality is the only arbiter. The actual outcomes in the actual domain — patient recovers or doesn’t, code works or doesn’t, the prediction matches reality or doesn’t — are the only calibration mechanism.

§8 — What This Changes

“Knowing but not doing” is not weakness of will

Rationalists know akrasia well: you know the right action but can’t make yourself do it. Under this framework, there is no mystery. “Knowing” is a PFC-level fresh draft. “Doing” is driven by compiled body-direct patterns. When they disagree, the compiled patterns usually win — not because of weakness, but because compiled processing is the dominant mode (~95% of behavior runs on it). The “failure” is not willpower. It’s two processing tracks reaching different conclusions, and the compiled track has structural priority.

This reframe has practical implications: if you want to change behavior, compile new patterns through repeated domain-verified experience — don’t just add more PFC-level “knowledge.”

AI alignment and the rational agent assumption

Most AI alignment frameworks assume a rational agent model — decisions follow from stated preferences through logical inference. PFC = Lawyer complicates this: human decisions may not follow from stated reasons at all. The stated reason is constructed after the fact. Aligning AI with human preferences requires modeling the body-level compiled processing that actually drives choices, not just the PFC narratives that explain them.

§9 — Falsification: What Would Prove Us Wrong

We believe the evidence supports this framework. Here is exactly what would prove us wrong.

1. Logic and intuition use fundamentally different neural substrates — not just different activation patterns on overlapping substrates. Current fMRI evidence shows overlapping neural activation for “analytical” and “intuitive” tasks (Goel 2007, Lieberman 2007). If future research demonstrates genuinely separate, non-overlapping neural mechanisms, the framework is wrong.

2. Expert intuition accuracy is unrelated to domain compilation — experts are accurate for reasons other than compiled patterns. If a study demonstrated expert accuracy in a completely novel domain without any compiled pattern base, that would challenge the framework.

3. PFC deliberation consistently outperforms compiled processing in domains where the expert has a compiled base. Current evidence suggests the opposite — overthinking can reduce accuracy in expert domains (Wilson & Schooler 1991), and experts’ pattern recognition outperforms deliberation under time pressure (Klein 1998). If deliberate reasoning reliably beat pattern recognition in the expert’s own domain, that would mean “rational override” is genuinely superior.

4. Content type predicts decision quality better than compilation level. If “analytical content” predicted accuracy regardless of expertise level, or “emotional content” predicted inaccuracy regardless of expertise, that would mean the logic/feeling axis is the real one after all. Current evidence (Kahneman & Klein 2009) points to domain conditions and compilation level as the predictors, not content type.

Each criterion is specific enough that a researcher could design a study to test it.

§10 — Test It

The full framework — 200+ files, CC0 licensed — is available at GitHub — Human Predictive Drive.

Source files for this post:

If you have expertise in:

  • Cognitive psychology — test the compilation-stage reframe against dual-process theory evidence
  • Decision-making research — does compilation level predict decision quality better than content type?
  • Neuroscience — do “logical” and “intuitive” processing share or differ in neural substrate?
  • Philosophy of mind — is the shareability-determines-label claim logically sound?

Counter-evidence is more valuable than confirmation. The framework itself predicts that reading it will bias you toward confirming evidence — compiled patterns for “this makes sense” will fire, and the PFC will generate supporting narratives. If something doesn’t fit, that’s the most useful thing you can share.

Full framework (200+ files, CC0, open-source): GitHub — Human Predictive Drive

Related posts:

This is a hypothesis inviting falsification. What would break it? We want to know.