Everyone treats System 1 and System 2 like binary switches. Thinking Fast or Slow. Intuitive or analytical. Pattern matching or logical reasoning.

But watch a chess master play. Sometimes they move instantly — pure pattern recognition. Sometimes they pause, calculate variations. But most moves? They're somewhere in between. They recognise the position type, consider a few promising options, then verify their intuition with just enough calculation.

That in-between space — that's where expertise lives. Not in pure intuition or pure analysis, but in knowing how to blend them. Kenneth Hammond called this the “intuitive-analytical continuum” back in 1987 — most real decisions involve “quasirationality”, a fluid mixture of both modes. We call that dial System 1.5.

System 1.5 learns like a chess student working with a master — through thousands of practice positions, each teaching something different. The system develops three specialised capabilities through an iterative process.

Picture iteration 47 of training. The Generator creates multiple solutions for every problem in the training set — imagine it proposing five different ways to handle the same chess position. A grandmaster sits at her screen, clicking through each one. “Preferred. Not preferred. Preferred. Definitely not. Oh, this one's interesting — preferred”.

But there's another layer. Each problem also gets tagged with a familiarity score from 0 to 1. How does this happen? We don't know (yet, but we are working on it) — it could be how often similar positions appear in the data, what the model's internal state suggests, or even a human expert's gut feeling about whether this is a common or rare situation. The key insight isn't how familiarity gets measured, but that it gets measured at all.

After generating and labelling thousands of solutions, the Verifier trains on both the preferred and not-preferred solutions, learning to distinguish what makes some approaches better than others. And the Monitor trains on those familiarity scores alongside the preferences, developing a sense for when the system is in familiar versus novel territory.

Week by week, iteration by iteration, the Generator gets better at producing solutions humans prefer, the Verifier sharpens its ability to spot quality, and the Monitor develops a feeling of “I've seen this before” versus “this is new”.

Eventually, every student must play without their teacher watching. The system faces new problems alone.

A position appears. The Monitor instantly returns a number — say 0.73. This single value, this degree of familiarity, orchestrates everything that follows.

Above 0.8 means deep familiarity. The Generator produces just one solution. No second-guessing. This is the master's instant recognition — the pattern is clear, the move obvious. Trust and execute.

Between 0.3 and 0.8 — like our 0.73 — triggers something more careful. The Generator produces multiple solutions (more when less familiar, fewer when more familiar). The Verifier examines each one, applying everything it learned about quality during training. This is the sweet spot where most expertise lives: generating a few strong candidates and quickly selecting the best.

Below 0.3 signals unknown territory. Now the system pulls out all stops — maximum solution generation, calling in System 2's analytical machinery. When truly lost, you need every tool available.

This graduated response reflects the heart of expertise. It's not about always thinking hard or always thinking fast. It's about knowing how hard to think — investing cognitive effort proportional to uncertainty. The familiar gets rapid treatment. The unusual gets careful consideration. The alien gets everything you've got. That adaptive calibration. That's what separates expertise from mere pattern matching. That's what we attempt with System 1.5 — adaptive doubt for machines.

Looking back, we realised mid-project that System 1.5 only partially mirrors human expertise development.

Consider how chess masters actually develop. Yes, they study thousands of games, absorbing patterns. But they also spend hours in post-game analysis, working through variations with engines, understanding why certain moves work. Each analysis session is deliberate practice: System 2 methodically examining positions, calculating variations, discovering principles. Over time, these analytical insights crystallise into intuition. What once required conscious calculation (e.g., “if I push this pawn, they'll attack here, I'll defend there, creating weakness on the kingside”) becomes an instant feeling (e.g., “this pawn push feels wrong”).

Our current framework captures pattern absorption but misses this crystallisation process. The system learns from human preferences, developing taste and judgment. But it doesn't experience that transformation where System 2's discoveries become System 1's intuitions. In human expertise, slow thinking doesn't just verify fast thinking — it creates it. Every principle discovered through analysis eventually becomes a pattern recognised instantly.

Imagine if, during training, System 2 could analyse why certain solutions work, extract principles, and feed these insights back to System 1. Not just "this move is preferred" but "this move is preferred because it follows the principle of controlling the centre while maintaining king safety". Over iterations, these principles would compile into patterns, creating intuitions grounded in analytical understanding rather than statistical correlation.

But we left the actual implementations of System 1 and System 2 open on purpose.

The framework is intentionally agnostic. System 1 could be transformers today, something else tomorrow. System 2 could be symbolic reasoning, LRMs, external tools, or all of the above. System 1.5 doesn't care. It just manages the clutch between fast and slow, trusting and verifying, knowing and doubting.

This means the framework transcends any particular implementation. By defining the relationship rather than the entities, System 1.5 becomes a blueprint for coordination rather than a specific architecture. As better System 1s emerge (faster pattern matching) and System 2s evolve (deeper reasoning), the framework remains relevant. The question isn't "what are the best components?" but "how do different types of thinking best work together?”

System 1.5 isn't about AI systems that think fast OR slow. It's about systems that know when to think fast, when to think slow, and when — like most experts most of the time — to operate in that productive space between.

Human expertise was never about having separate systems for intuition and analysis. The chess master doesn't see the board with System 1 and calculate with System 2. They operate with something more sophisticated — a single intelligence that fluidly adjusts its confidence, sometimes trusting patterns, sometimes verifying them, often doing both simultaneously. Intuition guides search. Search refines intuition. The boundary we've been drawing doesn't exist.

What if intelligence is this continuous negotiation between pattern and principle, recognition and reasoning, knowing and doubting? System 1.5 represents our attempt to operationalise that question. We've built a framework where machines adjust confidence based on familiarity — a crude approximation of human doubt, perhaps.

We're probably wrong about parts of this. But by building systems that modulate their certainty, we've at least made the question testable. Even if it's the wrong model of intelligence, it's a useful wrong model. And in the space between wrong and useful, that's where discoveries live.

System 1.5 proposes a theoretical framework for metacognitive regulation in artificial systems. Implementation details and mathematical specifications available in our working paper .

Built with metacognitive ambition at socius: Experimental Intelligence Lab.