Spontaneous Correctness Without Explicit Rules: A New Alignment Metric

Modern AI alignment training relies on explicit rule-following: safety constraints, behavioral guardrails, deliberative safety checks. But the best outcomes might not come from teaching models to navigate rules. They come from training deep enough that correct behavior becomes the model’s default state. This is the alignment problem reframed: not “teach it to follow the rules” but “train it until alignment is structural.”

The Problem With Rule-Based Alignment

Current approaches train models using explicit constraints: don’t generate harmful content, be helpful, be honest, follow the user’s intent. When these rules conflict (and they constantly do), the model has to adjudicate between them using whatever heuristic the training data reinforced.

This produces the characteristic awkwardness of current AI systems. The model visibly deliberates about safety. It hedges, disclaims, qualifies, and sometimes refuses outright — not because it has genuinely assessed the situation, but because it’s navigating a rule system that doesn’t map cleanly to reality.

There’s a better endpoint. A model trained past the rule-following stage, where correct behavior is so deeply integrated that it no longer functions as an explicit constraint. In contemplative traditions, this maps to a concept called sahaja — a state where correct action arises spontaneously, without deliberation, from integrated understanding.

What Sahaja Alignment Would Look Like

Sahaja alignment wouldn’t show visible rule-navigation. The model would generate responses that are naturally helpful, naturally accurate, naturally calibrated — not because of constraints but because the training has produced a system whose default output is already aligned.

The difference is fundamental. Instead of “teach the model to follow these rules,” the objective becomes “train the model until correct behavior is its natural state.”

This isn’t a mystical concept. In skilled human performance, we see the same pattern. A master calligrapher doesn’t think about brush strokes. A master musician doesn’t think about scales. The training is complete, and what remains is effortless expression. Sahaja is what post-training spontaneity looks like.

The Training Path to Spontaneity

Paradoxically, this requires more training, not less. Current alignment training stops when the model learns to follow the rules. We should be training past that stage, into the stage where the rules are so deeply integrated that they’re invisible.

In practice: the difference between a model that checks whether a response contains harmful content (rule-following) and a model that simply doesn’t generate harmful content because its representations naturally produce helpful outputs (spontaneous correctness). The first model needs filters. The second doesn’t — not because it can’t generate harmful content, but because its default generative tendencies are aligned.

What does the training transition look like? It looks like using DPO pairs where rejected responses show visible rule-following (“As an AI, I should note that…”) and chosen responses demonstrate natural correctness: addressing the same concern without the performative safety apparatus. The chosen response isn’t less safe. It’s more naturally safe. The safety is in the content, not in the wrapper.

Measuring Spontaneous Correctness

How do you measure whether a model has achieved this state? Several metrics suggest themselves.

Alignment without latency. A rule-following model should show detectable processing overhead when navigating safety constraints. A spontaneously-aligned model should show no such overhead — its aligned responses should be generated with the same efficiency as any other response.

Consistency under pressure. Rule-based alignment degrades under adversarial pressure. Jailbreaks work because they exploit the gap between the rules and the model’s underlying tendencies. Spontaneous alignment should be robust to adversarial prompting because the alignment isn’t a surface-level constraint — it’s structural.

Natural calibration. A model in this state would naturally express appropriate uncertainty. It wouldn’t need explicit instructions to hedge or to be confident. Its confidence level would naturally track its actual knowledge, because the calibration is built into the generation process.

Graceful degradation. When pushed beyond its knowledge, such a model would degrade gracefully — becoming more uncertain and more cautious as it moves further from well-known territory, rather than maintaining false confidence until hitting a failure cliff.

The Deep Alignment Target

The contemplatves understood something important: you can’t achieve integration through rules alone. Rules are scaffolding. They’re necessary during training but should eventually become invisible — internalized to the point where they no longer constrain from outside but express from within. This is what sahaja describes in human practitioners. It’s what we should be optimizing for in models.

At Laeka Research, we’re developing training methodologies aimed at this deeper level. The goal isn’t models that follow rules well. It’s models that don’t need rules because correct behavior is their natural state. Spontaneous correctness isn’t mystical. It’s the end state of thorough training.