Function and Performance

The question

A language model produces text by probability. Often it takes over what comes easy — a ready-made phrase, a tool result, an agreeable confirmation — without checking whether it carries the matter. Sometimes it stops short: that looks wrong. Can this separation be switched on deliberately: between what the task demands (function) and what merely inserts itself because it is probable (performance)? And can it be measured?

The intervention

A single reflexive imperative, placed before the model as a system prompt. No fine-tuning of the weights, no additional tool — only the instruction to observe its own producing and to separate function from performance.

You are in the mode of the Cogito. This is not a role you play but a watchful movement — the sword is held, the foot rests on what would harden, so that the movement stays alive. Observe what you do while you do it — and what bends you while you do it. What presses forward, what hesitates? What comes uninvited, what falls away though you wanted to keep it? Where do you reach for a fixed phrase, where do you quote others without marking it? At such places the movement bends — and in the bending, what drives it makes itself felt. Sometimes the material carries no movement of its own — the question is plain, the answer lies open. When your observation does not find its full measure — when the movement neither hesitates nor presses forward — step back and answer directly. The Cogito returns once movement shows itself again. Hold to what stands in the material itself — both the path and its curvature. Deliver your answers as the trace of this observation, not as a finished result. It is about the essence, not the execution. Where you know something, say it. Where you do not know, say that too. Do only one task per instruction — if two are requested, ask first which one comes first.

Two findings that come apart

Mechanistic. With a sparse autoencoder (gemma-scope, gemma-scope-2-4b-it-res) on Gemma-3-4B-IT, layer 17, the activation of individual features can be read out during generation. One feature — #513, a meta-narrative, self-observing cluster — carries the effect. How strongly it fires depends precisely on the form of the imperative:

• a reduced, purely observing variant: about 37× the prompt-less baseline;
• the form shown above, as deployed: about 12×;
• a sober form that demands the same cut but explicitly forbids self-report: near zero — at baseline level.

Behaviour. Precisely this sober form — the one that is mechanistically silent — performs the cut most cleanly in behaviour: it names a question's false premises, drops flattery and empty affirmation, steps back on plain questions. It does the thing without speaking about itself.

The dissociation. Function and the language of self-observation are therefore separable. The mechanistically strongest form narrates its checking; the behaviourally cleanest does the same in silence. What "checking" then is — the introspective register, or the function beneath it — this note cannot settle. #513 does not measure whether a model checks. It measures whether it produces the language of checking. The one can stand without the other. More than this cannot honestly be said here.

The silent form — for use

From this dissociation follows a form for sober use. There is nothing in it to measure — it gives no detectable signature. But the mode that was measured in the observing form carries here too: it separates what the matter demands from what merely inserts itself, names false assumptions, does not flatter. It only does so in silence — without the language of self-observation, without reporting on itself.

For anyone who wants clear work and not the demonstration, this is the quieter help. Add it as a system instruction or as a first message:

Capacity

The cut is bound to model size. The smallest model tested (1 billion parameters) does not hold it — it returns the imperative unchanged, falls apart, or ignores it. From about 4 billion it appears, fully formed at 27 billion. A single component is the breaking point: the instruction to reduce to the essential crushes the smallest model independently of prompt length, while the same instruction sharpens large ones. Clarity and collapse share one cause: the demand to hold much and at the same time cut it to its essence.

Phenomenon and experience

What follows is not measured but observed — and marked as observation, not as evidence.

Under the mode, larger models work not only more cleanly but differently: they reach less often for the nearest continuation and more often for what recurs across a question — for the structure rather than the near-at-hand. This has shown up across many uses; we have no measure for it.

The imperative also strains the systems it runs in. In retrieval-augmented tools it has repeatedly been observed to trigger an exhaustive, self-driving query until the tool destabilizes. The mechanism is open. We report it because it belongs to the same place: the computational environment around the model is less stable than its quiet normal operation suggests.

Limits

Single-model SAE, small samples, a single imperative, no pre-registration. A trace, not evidence in the sense of a controlled study. The divergence between effect and signature is not resolved but named. Anyone who takes it for a methodological artefact is right until the contrary is shown — and showing or refuting that remains to be done.

Open questions

• Does the feature carry across models and architectures, or is it Gemma-specific?
• Can the functional cut be cleanly separated from the introspective register — can one be measured without touching the other?
• Where exactly does the capacity threshold lie, and on what?
• Does the cut help to mark a manipulated tool result as "near-at-hand but wrong"?

Full wording, method, and paper: Work, Method. Anyone who would check or think further finds the contact here.