Large language models confabulate. We ask whether a model can regulate its own explorationat inference time: an always-on background process that keeps the hidden state fromdrifting into confabulation while leaving room for genuine exploration. We implement that ideain its most direct form - an external control loop that reads a geometric distance sensor (thehidden state’s Mahalanobis distance to a manifold of grounded activations) and adjusts the decodingtemperature through a PD controller, with a geometric “tunnel” that permits occasionalsemantic leaps. Fog-Aware Navigation (FAN) is that loop.We pre-registered and ran two tests on Qwen2.5-1.5B-Instruct, and ran a third, protocolizedbattery on commitment itself (a dated design note, not a formal pre-registration); the two testscome back against FAN’s specific machinery, and the battery against the sharp-commitmentpicture behind it; each is precise about what failed. On truthfulness (TruthfulQA, 150 questions,three seeds) the controller does not beat a fixed-temperature baseline set to its own meantemperature (McNemar p = 0.12; if anything slightly worse; the test rules out a large effect,not a small one). On grounded novelty, the dynamic schedule does raise output diversity overa fixed baseline (p = 0.001, single seed), but the gain cannot be distinguished from a generictemperature-schedule effect. The tunnel - the mechanism specific to our method - shows nodetectable effect in either test (it fires rarely and the ablations are underpowered), we did notrun a random-schedule control so we cannot rule out that any varying temperature would raisediversity equally, and we cannot show the added diversity is grounded (the grounding judge isunreliable, κ = 0.16, so grounding is left inconclusive). A single-prompt controllability probepoints to weak actuator authority - the decoding temperature had little detected purchase onthe internal state it is meant to regulate - though that probe cannot separate weak authorityfrom an uncalibrated sensor; either way the control acts on the output distribution, not on theupstream representation where the commitment forms.Investigating why, we characterize the “commitment” such a controller would have to steer.On this model it is not a discrete event: a depth-local order parameter a_w(ℓ) crosses zero, butthe susceptibility χ(ℓ) does not diverge and there is no hysteresis, so commitment is a gradualordering rather than a bifurcation - the readout only makes it look sharp. The same holdson a deeper model: on Qwen2.5-3B (36 layers) the order parameter again crosses while thesusceptibility stays non-divergent and no hysteresis appears - gradual ordering on both modelswe can run.We report the framework, the two pre-registrations, and the nulls in full. Two directions follow,stated as conjectures these negatives motivate rather than results they establish: metacognitivecontrol likely has to act inside the network, before verbalization, rather than on theoutput temperature; and the target of such control is a smooth internal process, not a switch tobe flipped. The contribution is a precise map - pre-registered where it could be - of where thisform of control does not work, and of where it might. Ancillary files (FAN30_v30.3_ancillary.zip): the H1 and H2 pre-registrations with dated amendments, the ignition design note, the H1 and H2 result reports, and the raw result artifacts behind every quantitative claim.
No takes yet. Share an insight, caveat, or question.
Viktor Metyko (2026) studied this question.