A Self-Fulfilling Prophecy of Noise: Phantom Percepts as Self-Sealing Model Edits in a Homeostatic Algorithmic Agent

Tinnitus and chronic phantom pain are best understood not as broken sensors but as the brain's rational, self-locking decision to stop paying attention to a channel it cannot fix.

The core idea is simple. When a cochlear injury silences real sound, the brain doesn't get quiet — it gets flooded. Central gain cranks up, neurons fire spontaneously, and the deafferented channel now blasts high-variance noise that is always on, carries no useful information, and doesn't change no matter what you do. The brain faces a persistent, unexplained stream of activity. Its job is homeostasis, and compression is just a tool it uses when it helps. The rational move, under a minimum-description-length logic, is to build a compact model of that stream: "there is a noisy source at this frequency address; its fine detail is not worth tracking." This is the structural edit the paper proposes — a learned nuisance model $M_\phi$ consisting of a latent source, a filter capturing the source's spectral shape, a tolerance model for the residual, and an update gain.

Here is where it gets interesting. Once installed, this model sits early in the brain's hierarchical inference cascade and subtracts its estimated contribution before passing the residual downstream. Downstream models never see the original noisy data again — they only see what's left after $M_\phi$ has explained its piece away. If $M_\phi$ later becomes wrong (say, the peripheral injury heals), downstream stages can't easily tell, because they're adapting to the wrong residual rather than reconsidering the upstream model. Meanwhile, the model's own internal audit — the only thing still looking at the raw data — gets its update gain learned down, because the source is stochastic and action-invariant: no action changes it, so the Planning Engine stops searching it, and the Modeling Engine stops refining it. The model's own content — "this is an uncontrollable, useless noise source" — is literally the instruction to stop running the long-window statistics that could prove it wrong. The self-seal is not logical unfalsifiability; it's a strong hysteresis built from rational resource allocation.

The percept is the inferred cause, not a replay of the noise. What you hear is the model's latent variable $z_\phi$, not the random fluctuations it explains away. This is why tinnitus pitch is vague within a band (the model is coarse about detail) but reliably tracks the frequency of the hearing loss (the model is precise about address). Suffering is a separate phenomenon: the Objective Function mistags this uncontrollable percept as a threat, triggering chronic stress dynamics — learned helplessness, allostatic load — in patients who cannot habituate.

The paper derives nine concrete predictions from this framework. The most distinctive: in-band speech-in-noise deficits should be confined specifically to the tinnitus frequency band; peripheral nerve block should relieve the phantom early after onset but progressively fail as the drive centralizes (weeks in rodents); and restoring structured, controllable input to the deprived band should reduce the phantom more than loudness-matched noise, because only structured input moves the innovation monitor that can reopen the model. The cure, the paper argues, must come from outside the loop — new structured input, restored plasticity, restored control, or reappraised threat — because the loop itself has been designed to ignore internal evidence for revision.