Imagine turning a dial that controls the reliability of the platform image. At one end, the fog clears and the outline is sharp. At the other, the same outline is nearly random noise. What should a rational observer do? A strong sensory signal should pull the estimate towards the current image. A weak signal should let prior knowledge carry more weight. Predictive processing names this reliability precision, usually the inverse of variance.

This is more exact than saying that the brain “predicts everything”. A useful predictive model must specify at least three things: the hidden cause it is trying to infer, the observations that cause could generate, and the uncertainty attached to each relation. The model sends a prediction towards incoming data. A residual returns. The residual changes the current estimate only in proportion to its precision. The decisive quantity is not error alone, but precision-weighted error.

A four-number example

Suppose the hidden quantity is the direction of a moving object. Your prior estimate is 20 degrees. A noisy observation suggests 32 degrees. Give the prior a precision of 1 and the observation a precision of 3. Under a simple Gaussian model, the best posterior point estimate is the precision-weighted average:

Reduce sensory precision from 3 to 0.5 while preserving the observation. The estimate becomes 24 degrees. Nothing mystical happened. The model trusted the same observation less. This small calculation contains the core intuition behind precision weighting in much larger hierarchies.

Error is not surprise

A residual such as “observed 32, predicted 20” is a numerical difference. Surprise, or surprisal, is the negative logarithm of the probability assigned to the observation by a model. The same residual can be unsurprising under a broad noisy distribution and highly surprising under a narrow one. Precision supplies that distinction. It converts a bare mismatch into a model-relative quantity.

This matters because loose language can create a false ladder: prediction error becomes surprise, surprise becomes salience, salience becomes attention, and attention becomes consciousness. Each step may be a research hypothesis, but none is a definition. Feldman and Friston developed a precise proposal in which attention corresponds to inferred sensory precision and may be implemented through changes in synaptic gain.[4] That is a mechanistic claim about selection. It neither says that all precision estimation is attention nor that attended information is necessarily experienced.

flowchart TD
  A["The confidence dial"] --> B["What free energy actually buys"]
  B --> C["Where consciousness enters<br/>from outside the equation"]
  C --> D["A bridge test you can run"]
  D -. "re-test" .-> A
  classDef first fill:#0a1424,stroke:#B96D00,stroke-width:2px,color:#e6f3ff
  classDef step fill:#0a1424,stroke:#3f7da0,stroke-width:1px,color:#e6f3ff
  class A first
  class B,C,D step

The platform example had only one hidden variable and two Gaussian sources. Real generative models contain many hidden causes, nonlinear relations and dependencies across time. Their exact posterior distribution may be impossible to calculate. Variational inference offers a practical bargain: choose a tractable family of approximate distributions, then optimise the member of that family which comes closest to the posterior.[6]

Let y be the observed sensory data and z the hidden causes. The generative model p(y,z) says how hidden causes and observations are jointly related. The recognition distribution q(z) is the system’s current approximation to the posterior. Variational free energy is:

The identity is exact when the relevant probability densities and expectations are well-defined. The optimisation still inherits the chosen approximation family. A unimodal Gaussian q cannot faithfully represent a posterior with two separated modes. The algorithm may reach the best member of the wrong family. Free-energy minimisation is therefore model-relative twice: it depends on the generative model and on the permitted approximation.

This is the first place where an impressive equation can outrun its explanatory warrant. Lower free energy means that q fits the chosen joint model better for the observed data. It does not mean the model is true, the implementation is neural, the organism is healthy, the action is ethically desirable, or the inference is conscious. Each additional claim requires a separate bridge and evidence.

Two debts can hide behind one low value

There are two different ways to obtain a reassuringly low objective and still misunderstand the system. The first is model debt. The generative model may omit a real cause, encode the wrong likelihood or assign confidence to a biased sensor. Optimisation can then be exact relative to a world description that is systematically false. A navigation system may converge beautifully while using a map whose bridge has been removed.

The second is approximation debt. The generative model may contain the right alternatives, but the permitted recognition family cannot represent them. Imagine two equally plausible platform figures, one at 10 degrees and one at 34 degrees. A single Gaussian approximation may settle at 22 degrees, a location supported by neither hypothesis. Lower free energy within that family indicates the best available compromise, not a faithful posterior.

These debts matter for consciousness claims because an unexplained residual is often promoted into evidence of unusual experience, while a well-fitted model is promoted into evidence that the relevant experience has been explained. Both moves are unsafe. First test whether the model predicts interventions and whether a richer approximation changes the result. Only then ask whether the remaining structure tracks a declared consciousness target. Optimisation quality, world-model adequacy and consciousness relevance are three separate audits.

Perception changes the estimate; action changes the sample

In the Gaussian example, perception moved the estimate m towards a compromise. An embodied agent has another route. It can act so that future observations better match its predictions. A reaching robot can rotate its camera, move a gripper or touch an object. Each action changes the evidence available to the model.

Active inference develops this perception-action loop through generative models, preferences and policy evaluation.[3][18] The practical idea is powerful: do not separate sensing, estimation and control as if they were independent. Ask how an agent chooses observations and actions that reduce uncertainty while keeping expected outcomes within preferred ranges.

Self-organisation is a stronger claim than good prediction

The free energy principle extends beyond a local inference algorithm. It proposes that systems which persist within a limited set of states can be described, under stated assumptions, as minimising a form of variational free energy through perception and action.[3][17] Markov blankets are used to formalise conditional separations between internal, external, sensory and active states.

This extension is scientifically interesting because it connects estimation to the maintenance of an organised boundary. It is also where hidden assumptions multiply. Technical critiques have shown that blanket definitions, dynamical decompositions and inference interpretations are not automatically equivalent, and that extra conditions are needed for some derivations.[15][16] These critiques do not erase the programme. They tell us to record which theorem, dynamics and boundary are actually being used.

The matched optimiser

Suppose three systems minimise the same Gaussian objective. One is a cortical circuit model. One is a software Kalman filter. One is a mechanical linkage whose equilibrium implements the same weighted average. They can share an input-output function and even a state trajectory at the chosen grain. The equation alone cannot tell us which implementation, if any, has experience. That underdetermination is not a defect in the equation. It is evidence that the equation answers a different question.

A perception model can explain why one interpretation wins. A control model can explain why one action is selected. A self-organisation model can explain how a bounded process remains within viable states. Consciousness introduces a different explanandum: why some states are experienced, why they have a particular phenomenal character, or why information becomes available for flexible report and control. These are not one question.

Hohwy and Seth have argued that predictive processing can serve consciousness science as a systematic basis for relating neural mechanisms to phenomenological structure precisely because it need not be treated as a self-sufficient theory of consciousness.[10][11] On this restrained view, the framework helps organise hypotheses about perceptual content, attention, self-models and uncertainty. Another, stronger proposal now attempts to extract a minimal and testable consciousness theory from commitments shared by active-inference models.[13] That is a legitimate research move. Its success depends on whether the added commitments discriminate consciousness-relevant cases rather than merely redescribe successful inference.[12]

The key question is not whether predictive processing can be used in a consciousness model. It can. The key question is which added property does the discriminating work. Candidates include recurrent stabilisation, counterfactual depth, temporally extended self-models, global availability, precision control or embodied regulation. Once named, each candidate becomes a bridge hypothesis with its own intervention and failure conditions.

Five targets that should not share one score

A bridge audit begins by naming the phenomenon. Perceptual content asks why the shape is seen as a person rather than a poster. State consciousness asks why wakefulness differs from deep anaesthesia. Access asks why information can guide flexible reasoning and report. Phenomenality asks why there is something it is like. Selfhood asks why experience or control is organised around a perspective. One predictive hierarchy may contribute to several of these. A scalar such as prediction error, posterior entropy or free energy cannot be assumed to measure them all.

The strongest alternatives remain live

A functionalist can propose that a sufficiently specific inferential organisation constitutes consciousness. This view must specify the relevant causal roles, rather than only the shared objective. Biological naturalism can accept predictive inference while holding that biological organisation is necessary for experience. Physicalist and emergentist accounts can seek bridge laws linking particular dynamics to phenomenal properties. Idealism can treat predictive processing as a model of how contents are structured within awareness rather than an account of awareness arising. Neutral monism, dual-aspect theory and process philosophy can regard inference and experience as different descriptions or aspects of an underlying process.

None of these positions wins because the formalism is elegant. They change what an experiment must preserve. A substrate replacement matters greatly to biological naturalism. A causal-organisation intervention matters to functionalism. A consciousness-primary view asks whether changes in modelled content and access leave the fact of awareness conceptually untouched. The formal model earns metaphysical relevance only through a bridge whose predictions differ across these alternatives.

The Bridge Burden Test

The practical contribution of this paper is a five-step assurance instrument. It is designed to stop an equation from acquiring consciousness implications merely through repeated relabelling. Call it the Bridge Burden Test.

Target: state exactly what must be explained: content, state, access, phenomenality or selfhood. Bridge: name the extra variable or organisation alleged to connect the predictive model to that target. Matched rival: construct a system that preserves inference quality and free-energy trajectory while lacking or altering the bridge. Intervention: manipulate the bridge independently of task performance as far as possible. Abstention: pre-register which results permit only a claim about inference or report, and which outcome would weaken the bridge.

The central design principle is a matched-objective contrast. If two candidate systems minimise the same objective and reach equally calibrated posteriors, a consciousness-relevant bridge must predict another measurable difference. That difference might concern flexible access, temporal integration, metacognitive sensitivity or a theory-specific neural signature. First-person report can be included, but its production route and contamination risks must be modelled rather than assumed away.

stateDiagram-v2
  [*] --> Claim
  state "The frozen image and the<br/>moving belief" as Claim
  state "A four-number example" as Test
  state "The two witnesses" as Revision
  Claim --> Test: expose mechanism
  Test --> Decision: evidence survives
  Test --> Revision: rival survives
  Revision --> Claim: narrow or rebuild
  Decision --> [*]

The executable lab below isolates the central problem. All three synthetic systems use the same Gaussian generative model, the same precisions and the same gradient descent. They therefore reach the same posterior estimate and reduce the same objective. Only their access and report arrangements differ.

The silent filter computes the posterior but exposes no report. The threshold reporter emits a statement only when confidence crosses a chosen gate. The narrative reporter always emits a fluent sentence. This design is deliberately austere. It makes one result unavoidable: report behaviour can change while variational inference remains fixed. It also makes one limitation unavoidable: no setting tells us which system is conscious.

What the lab tests

The lab tests whether a chosen Gaussian model produces the predicted precision-weighted posterior and whether a simple optimiser approaches it. A positive result permits a narrow conclusion: the implementation is consistent with this variational objective. A negative result can expose an unstable learning rate, a coding error or a mismatch between the assumed and implemented model. Neither result establishes a biological circuit or consciousness.

Its strongest use is counterfactual. Hold the observation, prior and precisions fixed. Move only the report threshold. The posterior, posterior variance and objective reduction stay fixed, while the threshold reporter switches between silence and speech. Then hold the threshold fixed and change sensory precision. Inference changes for every system. These two interventions separate a mechanism of belief updating from a mechanism of report release.

The same experiment in Python

The following program reproduces the core trajectory without a library. Run it with Python 3. It prints the analytic posterior, the iterated estimate and the objective reduction. Change only report_threshold to observe that the inferential result is invariant.

from dataclasses import dataclass

@dataclass(frozen=True)
class Model:
    prior_mean: float = 20.0
    observation: float = 32.0
    prior_precision: float = 1.0
    sensory_precision: float = 3.0

    def objective(self, estimate: float) -> float:
        sensory = 0.5 * self.sensory_precision * (self.observation - estimate) ** 2
        prior = 0.5 * self.prior_precision * (estimate - self.prior_mean) ** 2
        return sensory + prior

    def gradient(self, estimate: float) -> float:
        return (
            self.sensory_precision * (estimate - self.observation)
            + self.prior_precision * (estimate - self.prior_mean)
        )

    def posterior_mean(self) -> float:
        total = self.sensory_precision + self.prior_precision
        return (
            self.sensory_precision * self.observation
            + self.prior_precision * self.prior_mean
        ) / total


def run(model: Model, steps: int = 24) -> tuple[float, list[float]]:
    estimate = model.prior_mean
    learning_rate = 0.72 / (model.sensory_precision + model.prior_precision)
    history = [estimate]
    for _ in range(steps):
        estimate -= learning_rate * model.gradient(estimate)
        history.append(estimate)
    return estimate, history


model = Model()
estimate, history = run(model)
confidence = 1.0 - min(1.0, (1.0 / (model.prior_precision + model.sensory_precision)) ** 0.5 / 3.0)
report_threshold = 0.65

print(f"analytic posterior: {model.posterior_mean():.4f}")
print(f"iterated estimate:  {estimate:.4f}")
print(f"objective reduction: {model.objective(history[0]) - model.objective(history[-1]):.4f}")
print("threshold report:", "emitted" if confidence >= report_threshold else "withheld")
print("consciousness verdict: not licensed by this simulation")

The program tests optimisation under fixed Gaussian assumptions. It deliberately returns no consciousness score.

A preregistered discrimination study

The same logic can guide an empirical study of conscious access. Compare two bridge theories while fitting the same hierarchical generative model. Theory A says access depends primarily on precision-weighted posterior confidence plus broad task availability. Theory B says access depends on recurrent stabilisation even when confidence and accuracy are matched. Orthogonally manipulate cue reliability and recurrent disruption, then staircase stimulus strength to match performance.

Theories must state different predictions. Theory A predicts that access-sensitive measures follow precision after recurrence-independent performance matching. Theory B predicts that disrupting recurrence removes the access signature even when posterior confidence is restored. Both theories must abstain from claims about phenomenality if the measures are reports, confidence judgements or task use alone. A null result that preserves adequate manipulation checks weakens both bridges for this target; it does not show that predictive inference is absent.

Three controls keep the comparison honest

Performance matching is necessary but not sufficient. Two conditions can reach the same average accuracy through different mixtures of easy and hard trials. Match calibration, response time and the full distribution of confidence where possible, not only the mean score. Fit the generative model separately to each condition and verify that its posterior predictions remain comparable. Otherwise a supposed bridge effect may be an unrecognised inference difference.

Manipulation specificity is the second control. A recurrent disruption may also reduce sensory precision, increase fatigue or delay a motor response. A cue-reliability manipulation may change expectation, arousal and strategy together. Use orthogonal manipulations, physiological checks and negative-control tasks. The goal is not a perfectly pure intervention, which may be impossible, but a causal model that exposes the plausible side paths rather than hiding them inside a condition label.

Report contamination is the third control. When the outcome is a button press or confidence judgement, the experiment measures experience only through a chain of memory, decision and action. Add a condition in which the same stimulus guides a later task without an immediate report, or compare report preparation with the target signal in time. A no-report proxy is not automatically closer to phenomenality, but it can reveal whether the candidate bridge is dominated by response construction.

How to read four possible outcomes

If precision changes the access measure while recurrent disruption does not, Theory A gains support for this task, provided the recurrence intervention worked. The result still concerns access, not phenomenality. If recurrence changes the measure after posterior confidence is restored, Theory B gains support and a purely confidence-based bridge weakens. If both manipulations matter independently, the theories may describe complementary conditions rather than rivals; the next study should test their interaction and necessity.

If neither manipulation changes the measure, resist the temptation to declare that the system simply minimised free energy at a deeper level. Check power, manipulation strength and model fit. When those checks pass, the declared bridges have failed for the target. Predictive processing may still model the perceptual estimates, while the consciousness explanation must move elsewhere. A useful theory loses entitlement when its distinguishing prediction fails.

This four-outcome reading prevents a common escape. A framework broad enough to redescribe every result after the event cannot adjudicate among mechanisms. The preregistration should therefore contain a result-to-claim table before data collection. For each cell, state the strongest licensed sentence, the rival that remains and the next intervention. That ledger turns abstention from a rhetorical caveat into part of the experimental instrument.