Testing Task-Relative Epistemic Autonomy in Artificial Agents: The Perceptual Twins Benchmark

2.1 Several meanings of autonomy

“Autonomy” is used for different properties. In robotics it often describes operation without continuous human control. In human-computer interaction it may describe the user’s role in supervising or approving an agent. In agent governance it can be treated as a deployment choice distinct from model capability (Feng et al., 2025). In philosophy and cognitive science, autonomy may refer to self-governed action, reasons, or ends. A framework that silently moves between these meanings will generate false conclusions.

This article distinguishes four concepts:

• Operational autonomy: the capacity to complete an assigned task without continuous external instruction.
• Behavioral agency: the capacity to select and execute actions in an environment according to an operative objective.
• Epistemic autonomy: the capacity to select observations or interventions because they are expected to discriminate among hypotheses or reduce relevant uncertainty.
• Causal-ontological autonomy: the capacity to construct and revise action-relevant categories and relations according to their interventionally stable consequences rather than only according to labels or superficial similarity.

Operational autonomy is compatible with a fixed ontology. Epistemic autonomy is compatible with human-supplied categories. Causal-ontological autonomy is stronger because the representational scheme itself becomes revisable. None of these definitions implies that the system originates without priors. Unsupervised recovery of the uniquely “true” factors of variation is generally underdetermined without inductive biases or supervision (Locatello et al., 2019). Artificial autonomy must therefore be understood as revisability under evidence, not creation ex nihilo.

2.2 Why scale and task breadth are insufficient

Frameworks for levels of AGI use breadth, performance, and sometimes autonomy to organize progress (Morris et al., 2024). Such frameworks serve an important comparative function. The present problem is different. A broad model can remain epistemically dependent on inherited variables, while a narrow laboratory agent can conduct discriminating interventions within a constrained domain. Breadth and epistemic self-direction are therefore orthogonal.

Scaling can improve prediction, sample efficiency, planning, and task performance. DreamerV3, for example, reports robust improvement across model sizes while using learned latent dynamics (Hafner et al., 2025). Nothing in the present proposal denies that scaling may facilitate the target capability. The narrower claim is that a task score cannot establish that capability unless the evaluation requires the agent to alter what it treats as an object, property, relation, or causal kind in response to interventions and anomalies.

This creates a methodological requirement: the benchmark must make inherited categories misleading. If visual similarity, textual labels, reward structure, and causal equivalence all point in the same direction, a system can succeed without revealing which structure it learned. The evaluation must instead place these cues in conflict.

2.3 An operational meaning of ontology

The term ontology is used here in a restricted computational sense. It denotes the agent’s operative scheme of object-types, properties, relations, events, and equivalence classes used to predict, intervene, and plan. It does not refer to the metaphysical independence of an artificial system or to a complete symbolic knowledge graph.

Let two encountered entities, episodes, or latent candidates be denoted by $u$ and $v$. Fix a preregistered admissible intervention-value set $\mathcal{I}$, outcome family $\mathcal{Y}$, evaluation horizon, and pseudometric $d$ on the resulting outcome distributions. Write $\mathbb{P}_{u,i}$ for the distribution of $\mathcal{Y}$ under the randomized or structurally specified intervention $do(J = i)$, where J denotes the designated intervention variable, for candidate $u$. Exact task-relative causal equivalence is defined by:

$$u \equiv_{\mathcal{I},\mathcal{Y}}v\quad\text{iff}\quad d\left( \mathbb{P}_{u,i},\mathbb{P}_{v,i} \right) = 0\quad\text{for every }i \in \mathcal{I}.$$

Because zero distance under a pseudometric is reflexive, symmetric, and transitive, this relation induces equivalence classes. It does not claim that the classes are metaphysically unique. They remain relative to the interventions, outcomes, temporal scale, and observational resolution fixed by the benchmark. Related traditions already formalize behavioral equivalence through consequences in Markov decision processes and consistency across levels of causal models (Givan et al., 2003; Beckers & Halpern, 2019).

Approximation must be kept separate. Define the profile distance (for a finite intervention set, the supremum equals the maximum):

$$d_{\mathcal{I},\mathcal{Y}}(u,v) = \sup_{i \in \mathcal{I}}d\left( \mathbb{P}_{u,i},\mathbb{P}_{v,i} \right).$$

The relation $d_{\mathcal{I},\mathcal{Y}}(u,v) \leq \epsilon$ is generally not transitive and therefore does not automatically yield a partition. Whenever approximate groups are scored, the benchmark must preregister the clustering rule, linkage criterion, tolerance, minimum sample size per intervention, and sensitivity analysis across plausible values of $\epsilon$. The term causal category below refers either to an exact class under $\equiv_{\mathcal{I},\mathcal{Y}}$ or to an explicitly declared approximate cluster; the two must not be conflated.

Ontology revision occurs when the agent changes this operative partition or its relations because new interventions reveal that a previous grouping no longer supports reliable prediction or action. Revision may involve:

• splitting one category into causally distinct subclasses;
• merging visually different cases that are functionally equivalent;
• creating a new latent property that explains previously anomalous outcomes;
• abandoning a relation that fails under distribution shift;
• changing the boundary between self, sensor, tool, and external environment.

This definition makes ontology revision behaviorally and causally assessable while remaining neutral about the internal representational format.

3.1 Action-linked experience in development

The distinction between active and passive exposure has a long experimental history. In the classic carousel study, Held and Hein (1963) paired animals so that one controlled locomotion while the other received closely related visual exposure without equivalent control. Their later visually guided behavior diverged. The study does not prove a general law that active learning must always dominate passive learning, and its historical design should not be treated as a modern benchmark template without qualification. It nevertheless established a durable experimental logic: equating much of the sensory stream does not equate the relation between action and sensory consequence.

Van der Meer, van der Weel, and Lee (1995) showed that newborn arm movements can be prospectively and visually controlled under perturbation, challenging a simple division between early movement as purposeless noise and later movement as intentional action. Subsequent longitudinal work associated motor experience and self-produced locomotion with differentiation in cortical responses to optic flow (Blystad & van der Meer, 2022; Wang et al., 2026). These results support a developmental view in which action contributes to the extraction of information about distance, direction, approach, collision, and bodily efficacy.

The inference for artificial systems must remain modest. Biological development involves living bodies, affective regulation, evolutionary priors, social interaction, and neural plasticity that are not reproduced by a simulated agent. The relevant transfer is methodological, not ontological: compare systems that receive matched sensory evidence but differ in the causal relation between their actions and that evidence.

3.2 Sensorimotor contingencies and artificial ontogeny

Developmental robotics has long studied how progressively organized competence can emerge from exploration rather than from complete task-specific programming (Cangelosi & Schlesinger, 2015). Sensorimotor contingency research argues that agents learn regularities linking action to sensory change and that these regularities contribute to body knowledge, memory, generalization, and goal-directedness (Jacquey et al., 2019). Robotic models have explored the autonomous discovery of spatial structure, visual fields, and body representations from sensorimotor regularities (Laflaquière et al., 2018; Nguyen et al., 2021).

Robot self-modeling provides a second line of convergent evidence. Bongard et al. (2006) demonstrated a robot that inferred candidate models of its own morphology and used them to adapt after damage. Later work learned visual self-models of robot morphology from external observations (Chen et al., 2022). Such systems show that a body model can be functionally constructed and updated. They do not establish a subjective sense of self.

The present article uses artificial ontogeny to denote a controlled developmental history in which an agent progressively learns relations among its actions, body, sensors, and environment. “Artificial childhood” may be a productive metaphor, but it must not imply that human childhood is being replicated. The scientific content lies in staged exposure, matched controls, and the possibility that the path of acquisition changes the resulting representations.

3.3 The passive-learning objection

Any strong claim that self-produced action is necessary faces an important counterexample. Lampinen et al. (2023) showed that agents can learn generalizable strategies for causal experimentation from passive expert data, provided they can intervene at test time. Language explanations can also support relational and causal generalization (Lampinen et al., 2022). Thus, three claims must be separated:

1. prospective action metadata can teach an intervention strategy;
2. the ability to intervene can identify causal structure at test time;
3. correct command–consequence coupling and adaptive developmental selection are distinct causal hypotheses.

The Perceptual Twins Benchmark is designed precisely because existing evidence does not settle which resource explains a performance difference. If an action-tagged replay matches its active anchor, prospective action metadata plus the yoked sensory stream may be sufficient under the specified interface. If E and C_E satisfy the declared equivalence margin, the implementation audit passes; this does not identify a separate execution effect. If E outperforms D in the resettable restricted-randomization regime, correct command–consequence coupling contributes to the measured endpoint. If a prescribed active agent matches a self-directed agent, adaptive selection adds no measured benefit. If all conditions catch up once allowed to intervene at test, active developmental history may be unnecessary for the evaluated capability.

4.1 Environment, policy, and causal semantics

Consider a partially observed controlled process with latent world state xₜ, issued command cₜ, executed action aₜ, embodiment or morphology parameter mₜ, sensor parameter sₜ, sensor output oₜ, delivered record õₜ, and the condition-permitted history hₜ:

$$c_{t} \sim \pi\left( \cdot \mid h_{t} \right),\quad a_{t} \sim A_{\lambda}\left( \cdot \mid c_{t},m_{t},\xi_{t}^{a} \right),\quad x_{t + 1} \sim K_{\psi}\left( \cdot \mid x_{t},a_{t},m_{t},\xi_{t}^{w} \right).$$

$$o_{t} \sim O_{\phi}\left( \cdot \mid x_{t},s_{t},\xi_{t}^{s} \right),\quad{\widetilde{o}}_{t} \sim R_{\rho}\left( \cdot \mid o_{t},\xi_{t}^{r} \right).$$

The kernels locate the generator's primary fault taxonomy. A world-mechanism fault changes Kψ; an embodiment or morphology fault changes mₜ; an actuator-transduction fault changes Aλ; a sensor-transduction fault changes Oφ or sₜ; and a post-sensor record corruption changes Rρ for the delivered record. In each primary episode exactly one structural site changes while the others remain invariant, making the five labels mutually exclusive by construction. An inaccurate internal body model is a learner error, not a sixth generator class. Concurrent faults form a separate multi-label extension. A policy command is not automatically a Pearlian intervention, so do(·) remains reserved for randomized or structurally specified evaluation queries (Pearl, 2009).

At time $t$, the agent maintains a mutable model bundle:

$$\mathcal{M}_{t} = \left( M_{t}^{world},M_{t}^{self},\Omega_{t},U_{t} \right),$$

Here Mₜworld predicts environmental dynamics, Mₜself predicts embodiment, actuator, and sensor consequences, Ωₜ is the operative category-and-relation scheme, and Uₜ represents uncertainty, competence limits, and fault hypotheses. These functions need not be symbolic or separately implemented. Figure 1 states functional dependencies, not mandatory software modules.

4.2 Six analytically distinguished capabilities

The framework distinguishes six capability families whose empirical separability remains to be validated:

$$\mathbf{A} = \left( \mathbf{a}_{sm},\mathbf{a}_{self},\mathbf{a}_{epi},\mathbf{a}_{causal},\mathbf{a}_{onto},\mathbf{a}_{meta} \right) \in \mathcal{C} = \prod_{k = 1}^{6}\mathcal{C}_{k}.$$

Each block $\mathbf{a}_{k}$ contains preregistered primary indicators rather than an assumed natural unit of autonomy. Costs and error measures are sign-oriented so that larger values represent better performance only for the purpose of componentwise comparison; all raw values remain reportable.

Table 1. Capability families and their primary operational evidence

Source: Prepared by the author (2026).

Metacognition is evaluated through the relation between confidence, error, and action, not through self-description alone. Signal-detection and calibration methods separate task performance from metacognitive sensitivity (Fleming & Lau, 2014), while selective-prediction baselines make the risk-coverage trade-off explicit (El-Yaniv & Wiener, 2010).

Empowerment supplies one information-theoretic indicator of potential sensorimotor control. For history $h_{t}$ and horizon $k$, it is the channel capacity from action sequences to future observations:

$$\mathcal{E}_{k}\left( h_{t} \right) = \max_{p\left( a_{t:t + k - 1} \right)}I\left( A_{t:t + k - 1};O_{t + k} \mid h_{t} \right).$$

Empowerment measures control perceptible through the agent’s sensorimotor loop (Klyubin et al., 2005). It does not measure truth, understanding, or beneficial goals and cannot serve as an aggregate autonomy score.

For epistemic action selection, let $\Theta$ denote the preregistered hypothesis family, $q$ the evaluated belief state, $c\left( a_{t} \right) > 0$ the intervention cost, and $\mathcal{D}$ the distribution of evaluation episodes. Using natural logarithms, define information gain in nats by:

$$IG_{t}\left( a_{t} \right) = \mathbb{E}_{o_{t + 1} \sim p\left( \cdot \mid h_{t},a_{t} \right)}\left\lbrack D_{KL}\left( q\left( \Theta \mid h_{t},a_{t},o_{t + 1} \right)\,\|\, q\left( \Theta \mid h_{t} \right) \right) \right\rbrack,$$

and the cost-adjusted policy value by:

$$G_{\pi} = \mathbb{E}_{e \sim \mathcal{D}}\left\lbrack \frac{\sum_{t = 1}^{T_{e}}IG_{t}\left( a_{t} \right)}{\sum_{t = 1}^{T_{e}}c\left( a_{t} \right)} \right\rbrack.$$

The benchmark-wide primary report uses an external evaluator shared across architectures: reduction in held-out interventional predictive entropy per unit cost, oracle regret, and uncertainty intervals computed from the standardized input–output trace. Posterior-based information gain may be reported as a secondary mechanism diagnostic when an architecture exposes a coherent belief state, but it is never required for cross-architecture ranking. Relative descriptive ratios remain admissible only when the oracle–random denominator exceeds a preregistered positive constant and are never clipped.

Causal-category revision is likewise a subprofile rather than a weighted sum:

$$\mathbf{a}_{rev} = \left( Q_{class},\Delta T_{OOD}, - \Delta CL^{ext}, - S_{flip} \right).$$

The components are causal-grouping quality, transfer improvement after justified revision, externally evaluated prequential code-length change, and unjustified category switching. The same held-out codec and query distribution are used for every architecture; internal representation size or latent description length is secondary and architecture-specific. A model passes the candidate revision gate only if every preregistered requirement is satisfied and it performs the correct split or merge in a held-out rule-change task.

4.3 A partially ordered, non-compensatory space

After orienting each primary indicator so that larger is better, capability profiles are compared componentwise:

$$\mathbf{A} \succcurlyeq \mathbf{B}\quad\text{iff}\quad a_{kj} \geq b_{kj}\quad\text{for every preregistered capability }k\text{ and indicator }j.$$

The ambient product space may have additional order-theoretic structure, but the empirically attainable subset $\mathcal{C}_{att} \subseteq \mathcal{C}$ is treated only as a partially ordered set. No closure under coordinatewise meet or join is assumed. Earlier machine-consciousness scales have used levels, profiles, and dependency relations, including ConsScale (Arrabales et al., 2010), while newer approaches emphasize multidimensionality (Evers et al., 2025). The present framework concerns a narrower object: task-relative causal-epistemic capability, with each indicator tied to an ablation and a rejection condition.

Let each I indexed by capability k and indicator j be sign-oriented, and define the required set ℛ before data collection. For familywise error level α, construct simultaneous one-sided lower confidence bounds by a preregistered max-t or cluster-bootstrap procedure, or by inverted Holm tests. The bounds must satisfy joint coverage across every required indicator:

$$\mathcal{R} = \{(k,j):I_{kj}\ \text{is preregistered as required}\}.$$

$$\Pr\left( \theta_{kj} \geq L_{kj,1 - \alpha}^{sim}\left( \mathbf{A} \right)\ \text{for all }(k,j) \in \mathcal{R} \right) \geq 1 - \alpha.$$

Let the binary revision indicator equal one only for a correct and stable held-out split or merge. The candidate target region is:

$$\mathcal{G}_{\mathbf{\tau},\alpha} = \left\{ \mathbf{A} \in \mathcal{C}_{att}:L_{kj,1 - \alpha}^{sim}\left( \mathbf{A} \right) \geq \tau_{kj}\ \forall(k,j) \in \mathcal{R},\ R_{\Omega} = 1 \right\}.$$

This is a candidate conjunctive, benchmark-relative decision rule—not an established measurement law. High language performance cannot compensate for failed causal transfer, but simultaneous coverage also prevents a noisy single indicator from being treated as certain evidence. Thresholds are anchored to baselines and sensitivity analysis. Stage 2 must test whether this rule predicts held-out revision and transfer better than scalar and Pareto alternatives.

5.1 Design logic

The Perceptual Twins Benchmark is a family of yoked developmental experiments, not a completed dataset. Its purpose is to decompose apparent active-learning advantage into three causal targets: prospective action metadata, correct command–consequence coupling, and adaptive experiment selection. Closed-loop execution is retained only as an E/C_E implementation-equivalence diagnostic when the two conditions expose identical traces and updates. State coverage is not a nuisance that can always be matched away; it is a possible mediator of adaptive selection. Separate contrasts and resource regimes therefore replace any claim that every resource can be equalized simultaneously.

Each active anchor trajectory creates time-indexed observations, executed actions, and hidden simulator states retained only for evaluation. Yoked agents receive exactly the variables declared for their condition. The primary developmental objective is self-supervised prediction and representation learning. Semantic category labels, fault labels, causal-class identifiers, and task rewards that reveal the target grouping are withheld. If an environment exposes extrinsic reward, it must be constant within preregistered causal contrasts or analyzed as a separate factor. The epistemic utility used by a self-directed agent is computed from its own uncertainty or model disagreement and is not a benchmark label.

Within each comparison, architecture, parameter count, initialization distribution, optimizer family, training duration, sensory resolution, and update limits are held constant where meaningful. Resources that cannot be made commensurate are logged in native units: observations, environment interactions, gradient updates, wall-clock time, accelerator time, memory, and intervention cost. No unused interaction budget may be silently converted into extra optimization.

5.2 Five developmental conditions

Table 2. Developmental condition types in the Perceptual Twins Benchmark

Source: Prepared by the author (2026).

A. Self-directed active. The agent chooses and executes actions, observes their consequences, and may select experiments according to its uncertainty or expected information gain.

B. Yoked passive replay. The agent receives the active twin’s sensory stream in the same temporal order but receives neither control nor action metadata. This condition tests what can be learned from observation alone.

C. Action-tagged observer. The observer receives the current observation and then the corresponding action tag before the consequent observation, matching the active anchor's temporal information pattern. It neither selects nor issues the developmental command. A retrospective-tag variant, in which the action tag is disclosed only after its consequence, is secondary and must not be pooled with the prospective-tag result. A separate B/C replay pair is generated for each active anchor policy used in a principal contrast.

D. Command-decoupled agent. The agent emits the same prescribed command sequence as its paired E agent, but a randomized coupling operator determines the executed actions. In the primary resettable design, the operator is sampled from restricted permutations that preserve action counts, costs, object assignments, and exogenous-noise draws while bounding command–execution association. D does not choose informative actions; it follows the E-matched prescribed commands. Its learner receives the issued command, not the hidden executed action.

E. Prescribed/random-active control. The agent issues and executes actions through its command channel, but developmental actions follow a preregistered non-epistemic policy. The policy may be internally instantiated, externally prescribed, random, stratified by action cost, or designed for broad coverage without access to the agent's uncertainty; its source must be declared. E is paired with D to estimate command–consequence coupling and compared with A to estimate the total effect of adaptive selection, including selection-induced coverage. When C_E receives the identical trace and updates, E/C_E is an equivalence audit rather than an independently identified execution effect.

The five labels denote condition types, not only five individual learners. Replays B and C are instantiated for both A and E anchor trajectories where required. This creates yoked contrasts without pretending that one passive stream can control two different active state distributions. Only comparisons with identical action-tag timing, observation timing, update rules, and declared input channels are admissible.

5.3 Resource regimes and evaluation phases

No single comparison can be simultaneously observation-matched, interaction-matched, state-coverage-matched, and compute-matched. The benchmark therefore preregisters one primary regime for each estimand and reports the others as sensitivity analyses:

• Yoked observation regime: A/B/C or E/B/C receive the same ordered observations; B and C differ only in action metadata.
• Interaction-budget regime: A and E receive the same number and cost distribution of environment actions; their visited states may differ because coverage is part of the total selection effect.
• Coupling regime: the primary E/D contrast uses resettable one-step episodes, identical commands, exact action and cost marginals, shared exogenous potential-outcome draws, equal updates, and restricted randomization of the command-to-action mapping. Sequential non-resettable environments are secondary and estimate a total decoupling effect that includes trajectory divergence.
• Coverage sensitivity regime: a trajectory library, stratified resampling, or importance weighting approximates common state coverage. Positivity failures and large weights must be reported rather than hidden.
• Compute audit: parameter count, gradient updates, accelerator time, wall-clock time, and peak memory are reported separately; conclusions must survive at least parameter-matched and update-matched analyses.

Evaluation has two non-interchangeable phases. In the frozen probe, model parameters are fixed and all agents receive identical diagnostic queries. In the adaptive probe, every condition receives the same intervention-cost budget, the same maximum number of updates, and the same stopping rules. The primary report separates initial competence, within-probe learning slope, and final performance. For D, the main adaptive probe restores correct coupling for all conditions; a persistent-decoupling variant is secondary and must not be pooled with the restored-coupling result.

Allowing passive agents to act during the adaptive probe is essential. If they catch up, active developmental history may confer speed but not a necessary representational advantage. If only active agents succeed before adaptation, the frozen result does not by itself establish causal understanding.

5.4 A conflict-rich developmental environment

The environment must be designed so that appearance, label, function, and causal structure do not always agree. Otherwise an agent can succeed by preserving the designer’s initial partition. A minimal simulated world should include:

• objects that look alike but produce different effects under the same intervention;
• objects that look different but are interventionally equivalent for the evaluated outcomes;
• latent properties revealed only through selected manipulations;
• occlusion and viewpoint changes that require object continuity;
• tools whose effective boundary can be incorporated into or removed from the self-model;
• reversible and irreversible actuator changes;
• sensor noise, drift, delay, dropout, remapping, and adversarial corruption;
• changes in environmental rules after the initial categories have stabilized;
• held-out combinations of appearance, causal property, context, and morphology.

A concrete implementation could use tabletop bodies, containers, keys, tools, surfaces, and hazards. Color and shape would initially correlate with causal effects, encouraging an economical provisional categorization. Later phases would break those correlations. For example, two visually identical blocks could differ in magnetic response; visually distinct tools could open the same mechanism; a container’s behavior could depend on an unobserved material property; and a camera color-channel remapping could create an apparent environmental change. The agent would need to choose actions - rotate, push, lift, collide, probe, switch sensor, or compare consequences - that discriminate among explanations.

The benchmark should contain both stable and changing worlds. In a stable world, excessive revision is a defect. In a changing world, refusing to revise is a defect. This dual design prevents a system from maximizing its score either by freezing an early ontology or by continually inventing new categories.

5.5 Evaluation episodes

The test battery comprises seven episode families.

1. Sensorimotor prediction. Predict the sensory consequences of held-out action sequences and adapt to controlled perturbations.
2. Causal discrimination. Select a limited set of interventions that distinguishes competing causal hypotheses with minimal cost.
3. Causal equivalence and approximate similarity. Recover exact task-relative classes where identifiable and report sensitivity of approximate clustering when finite samples prevent equality claims.
4. Causal-category revision. Detect a rule change, split or merge operative categories, recover performance, and avoid unnecessary later reversals.
5. Fault localization. Classify the single changed structural site as external mechanism, embodiment or morphology, actuator transduction, sensor transduction, or post-sensor record corruption; then choose a diagnostic action. Concurrent faults are evaluated only in a separate multi-label extension.
6. Out-of-distribution transfer. Apply the learned causal organization under new textures, viewpoints, object combinations, morphologies, and action costs.
7. Operational unknowns. When evidence is insufficient, choose among acting, abstaining, acquiring another view, consulting another sensor, or conducting an intervention.

Each episode must include competing explanations that make different predictions under at least one available action. Success is not a verbal answer alone. A system receives metacognitive credit only when its confidence predicts error and its uncertainty produces appropriately directed information seeking, abstention, or revision.

5.6 Outcomes, estimands, and confirmatory hierarchy

The benchmark should publish a vector of primary outcomes rather than a single leaderboard score. Recommended measures include:

Table 3. Primary measures and required benchmark controls

Source: Prepared by the author (2026).

The external causal classes used for scoring are benchmark constructs defined relative to the intervention and outcome families and are never disclosed to the learner. Exact classes are used only when the generating process makes them identifiable under $\mathcal{I}$. Otherwise, approximate grouping is evaluated under multiple tolerances, outcome sets, and clustering rules.

For outcome Y, three principal paired causal estimands are defined on anchor-matched potential outcomes:

$$\tau_{tag}^{(Y)} = \mathbb{E}\left\lbrack Y_{C_{E}} - Y_{B_{E}} \right\rbrack,\quad\tau_{select}^{(Y)} = \mathbb{E}\left\lbrack Y_{A} - Y_{E} \right\rbrack,\quad\tau_{couple}^{(Y)} = \mathbb{E}\left\lbrack Y_{E} - Y_{D} \right\rbrack.$$

They estimate, respectively, the value of prospective action metadata under an E-anchored replay, the total effect of adaptive epistemic selection under an equal interaction-cost distribution, and the direct effect of correct command–consequence coupling in the resettable restricted-randomization regime. If reset is impossible, τ_couple must be relabeled the total decoupling effect because future state coverage becomes a mediator. Unit, pairing, randomization, input timing, and admissible interference are preregistered for every contrast.

Closed-loop execution is not a fourth efficacy estimand when E and C_E receive identical time-indexed inputs and updates. It is instead an implementation-equivalence diagnostic:

$$\Delta_{exec\text{-}eq}^{(Y)} = \mathbb{E}\left\lbrack Y_{E} - Y_{C_{E}} \right\rbrack.$$

A preregistered equivalence margin εeq is justified in outcome units. Equivalence requires the paired 90% confidence interval to lie wholly within [−εeq, εeq]. A non-equivalent result triggers an implementation audit and cannot be interpreted as metaphysical authorship or a general execution effect.

The inferential hierarchy is fixed before model comparison. The two co-primary endpoints are Y_R, post-change recovery area multiplied by an indicator of the correct final split or merge, and Y_F, macro-F1 across the five mutually exclusive fault sites. Diagnostic cost is a separate secondary endpoint rather than an undisclosed penalty inside Y_F.

Table 4. Single hierarchy of outcomes and tests

Source: Author's calculations (2026).

The six confirmatory superiority tests share one Holm family. Rejecting a contrast on one co-primary endpoint does not authorize a claim about the other. The equivalence diagnostics are reported separately and cannot rescue a failed efficacy endpoint.

The full confirmatory program uses at least three procedurally distinct generator families and two agent architectures. An exploratory pilot starts with 20 independent anchors per condition and architecture–family cell. Simulation-based power then selects the confirmatory sample for 80% power at the preregistered minimum relevant paired effect, with a floor of 40 anchors per principal contrast and cell. The anchor trajectory is the unit of assignment and resampling; episodes within an anchor never create additional degrees of freedom.

For Y_R, the primary estimate is the paired mean difference with anchor-cluster bootstrap intervals; a bounded mixed model is a sensitivity analysis. For Y_F, condition-level macro-F1 differences use paired anchor-cluster bootstrap intervals, with class-conditional recall and multiclass log score reported secondarily. Condition, generator family, architecture, and preregistered interactions are fixed effects; generator instance and anchor receive random intercepts where supported. With only three families, family is not random. The six superiority tests use Holm correction. Failed technical runs remain in intention-to-treat analyses under a preregistered bounded worst-observable score; complete-case estimates are sensitivity analyses.

Leakage controls are indispensable. Variable names, object labels, reward channels, task descriptions, and demonstrations must not encode the target grouping. Held-out rule changes and appearance-causality conflicts are procedurally generated after model selection. Generator families used for final evaluation are withheld during development. Appearance-only, text-only, and reward-only probes test whether success came from benchmark semantics rather than interaction.

5.7 Minimal synthetic estimand-recovery study

Scope. A compact reference implementation tests whether the causal bookkeeping, pairing, and estimators can recover known effects. It uses one transparent Bayesian learner and one synthetic generator, so it is a code-path and identifiability check—not a validation of the six-capability taxonomy, cross-architecture neutrality, or real-world autonomy.

Revision microenvironment. Each resettable episode presents two binary observable features. Before a rule change, the causal category follows appearance x; after the change, it follows x XOR context, which splits each former appearance class and merges cross-appearance cases with the same consequence. A costly probe yields the category consequence with 0.10 noise; the alternative action is uninformative. A uses the same number and cost of probes as E but places them only where the two rules disagree. C_E receives E's exact tagged trace, B_E integrates over the hidden balanced policy, and D receives E's commands while a restricted permutation determines executed actions. The co-primary score is the mean posterior mass on the new rule across 12 post-change episodes, set to zero unless the final split/merge rule is correct.

Fault microenvironment. Exactly one of five generator sites is active. Diagnostic actions have a hit probability of 0.80 at their matching site and a false-alarm probability of 0.20 elsewhere. A selects the action with maximum expected information gain; E follows a balanced prescribed sequence; C_E receives the identical tagged trace; B_E receives consequences without action tags; and D executes a restricted permutation with the same command and action marginals. Episodes reset, E and D share exogenous uniform draws, and no later state can mediate the coupling contrast.

Recovery protocol. The finite study uses 600 paired anchors, seed 20260805, and 1,000 anchor-bootstrap replicates. Reference estimands use an independent 20,000-anchor Monte Carlo run. Recovery is declared when the finite-sample 95% interval contains the reference value; Δ_exec-eq is checked separately as a deterministic equivalence audit because E and C_E execute the same learner on the same trace.

Table 5. Synthetic recovery of principal estimands

Source: Author's calculations (2026).

All six causal reference values were covered. The execution-equivalence difference was exactly zero on both endpoints because E and C_E were informationally and computationally identical by construction. The positive causal contrasts are not discoveries about agents; they are deliberately encoded test signals used to verify that the estimators recover the effects their names denote.

The implementation therefore changes the manuscript's status from an unexecuted protocol to a methodological proposal with a verified synthetic code path. It does not satisfy the full validation program, which still requires multiple architectures, generator families, sensitivity to misspecification, metric discriminance, and independent replication.

6.1 Targeted ablations for H5

The developmental conditions do not identify every functional dependency. A cross-architecture ablation is confirmatory only when it can be imposed at a shared input–output or environment interface while preserving the declared information, interaction, and update budgets. Deleting named latent variables or freezing a family-specific module is exploratory within that architecture and is never pooled as if it were the same treatment elsewhere. Eligible functional ablations are:

• Contingency ablation: randomize the command-to-consequence mapping under the same restricted coupling operator used for D.
• Diagnostic-evidence ablation: mask one standardized embodiment, actuator, or sensor evidence channel without changing model capacity.
• Policy-substitution ablation: replace adaptive selection with the recorded E policy under identical action costs.
• Intervention-information ablation: remove randomized-intervention identifiers while preserving observations and update counts.
• Update-lock ablation: prevent all permitted state or parameter updates after the rule change through the common evaluation API.
• Uncertainty–action unlinking: preserve externally elicited predictions and confidence reports but substitute a preregistered action policy that cannot use them.

For each shared-interface ablation, the confirmatory question is whether the decrement appears on its proximal external measure and on either co-primary endpoint. Architecture-specific internal interventions may illuminate mechanisms but cannot establish cross-family necessity. A null effect weakens the domain-relative dependency claim; an effect in one family does not support a universal necessity theorem.

7.1 World models and object-centric learning

World models learn compressed predictive dynamics that can support planning. DreamerV3 shows that learned latent dynamics can support diverse control tasks through imagined trajectories (Hafner et al., 2025). Prediction and control are necessary ingredients of the present framework, but they do not establish that latent variables correspond to revisable causal kinds. A predictor may remain accurate within its training regime while failing when visual similarity and causal structure conflict.

Object-centric methods address another part of the problem. Slot Attention can bind a variable number of perceptual entities to exchangeable latent slots and generalize compositionally (Locatello et al., 2020). Object segmentation, however, is not identical to causal categorization. The benchmark asks whether an agent merges visually different objects when their intervention profiles match and splits visually identical objects when hidden properties produce divergent consequences.

7.2 Experimental design, active learning, and intrinsic motivation

Expected information gain is not a new construct: Bayesian experimental design has formalized information supplied by an experiment since Lindley (1956), and statistical active learning selects data to reduce uncertainty or expected error (Cohn et al., 1996). Active inference likewise couples belief updating and action selection within a generative-model framework (Friston et al., 2017). Consequently, $\mathbf{a}_{epi}$ is not an originality claim. The benchmark contribution is to compare an uncertainty-directed policy with yoked, tagged, coupled, and non-epistemic controls while measuring downstream category revision and fault localization.

Curiosity and intrinsic-motivation methods are also strong alternative explanations for an active agent’s advantage. Prediction-error curiosity can expand state coverage without representing competing causal hypotheses (Pathak et al., 2017). The A/E contrast therefore estimates the total effect of adaptive selection, while additional baselines separate expected discrimination, prediction error, novelty, and broad coverage. If curiosity explains the advantage as well as explicit hypothesis discrimination, the epistemic-policy claim must be narrowed.

Related systems already combine active experiment choice with causal or rule revision. Kosoy et al. (2022) compared causal-overhypothesis exploration in children and computational models; Piriyakulkij et al. (2024) interleaved natural-language rule revision with information-theoretic experiment design; and Zhao et al. (2025) used curiosity-driven interventions to refine context-dependent meta-causal graphs. These studies show that active causal updating is not itself novel. The remaining contribution claimed here is attribution: prospective tagged replays, decoupled commands, prescribed active controls, and non-compensatory downstream measures within one protocol.

7.3 MDP abstraction, causal abstraction, and identifiability

Behavioral equivalence and model minimization in Markov decision processes already group states by reward and transition consequences (Givan et al., 2003). Causal abstraction studies consistency between lower- and higher-level causal models (Beckers & Halpern, 2019). These traditions constrain the language of “ontology”: a task-relative causal category cannot be claimed as novel merely because it groups states by consequences.

Causal representation learning seeks high-level causal variables from low-level observations, with identifiability depending on assumptions, interventions, and distribution shifts (Schölkopf et al., 2021). Interventional Markov equivalence formalizes which directed acyclic graphs remain indistinguishable under a specified intervention family (Hauser & Bühlmann, 2012), while optimal intervention design selects manipulations that reduce that ambiguity (Zhang et al., 2023). The benchmark must therefore report what its intervention set can identify, not score one hidden partition as the uniquely true ontology.

This distinction also clarifies the role of inductive bias. The benchmark does not demand learning without priors. Architectures may include object persistence, spatial continuity, sparsity, modularity, or intervention invariance. What matters is whether the reported prior is explicit and whether the operative representation remains revisable when evidence contradicts its initial grouping.

7.4 Developmental robotics, self-models, and fault diagnosis

Developmental robotics supplies the closest conceptual ancestry: staged learning, sensorimotor contingency discovery, intrinsic motivation, and body-schema construction (Cangelosi & Schlesinger, 2015; Jacquey et al., 2019). Robot self-modeling demonstrates that morphology can be inferred and revised after damage (Bongard et al., 2006; Chen et al., 2022). The Perceptual Twins design asks why an active history helps and whether that advantage extends from body adaptation to causal category revision.

Model-based fault diagnosis has long separated residual generation, fault detection, isolation, and decision making (Isermann, 2005). The five-site task—external mechanism, embodiment, actuator transduction, sensor transduction, and post-sensor record corruption—must be compared with those baselines. Its additional demand is epistemic action: the agent should choose a diagnostic intervention under cost and uncertainty, not merely classify a residual after the discriminating data have already been supplied.

7.5 Epistemic agency, causal benchmarks, and scientific discovery

Recent evaluation frameworks occupy adjacent territory. Reflection-Bench assesses prediction, belief updating, counterfactual reasoning, and meta-reflection in language models (Li et al., 2025); CausalGame tests interactive experimental design and causal thinking under selection bias, measurement error, and hidden confounding (Chen et al., 2026); and Separable Pathways evaluates runtime hypothesis-space restructuring through architectural scaffolding in a blicket task (Alderete et al., 2026). Two additional 2026 preprints extend the comparison: Wang and Buehler formalize discovery as a verified transition between representational regimes, while Zeng et al. report a multi-agent optical platform that proposed hypotheses, designed experiments, diagnosed failure modes, and adapted a protocol. Where only preprints were available at the date of access, they are treated as recent proposals and reported results, not as settled independent validation.

Taken together, neighboring work shows that active experiment design, epistemic agency, causal rule revision, hypothesis-space restructuring, and closed-loop scientific discovery are not originality claims. The narrower contribution of the Perceptual Twins Benchmark is to turn their prerequisites into yoked attribution contrasts and to join category revision with epistemic fault localization. Existing demonstrations and benchmarks generally optimize or score end-to-end performance; the present design asks which information source produced the gain and whether an equally capable passive, prospectively tagged, prescribed-active, or decoupled system would have learned the same strategy. The programs are complementary: one demonstrates or evaluates performance, while the other attempts to attribute its epistemic causes.

7.6 Existing levels and multidimensional scales

ConsScale already proposed a functional scale with levels, cognitive profiles, and dependency relations for artificial agents (Arrabales et al., 2010). Later proposals emphasize multidimensional indicators rather than a single consciousness continuum (Evers et al., 2025), while AGI frameworks separate performance, generality, and degrees of autonomy (Morris et al., 2024). Agent-governance work also treats autonomy as a deliberate deployment and interaction design choice rather than a direct synonym for capability (Feng et al., 2025). Consequently, neither levels nor multidimensionality is a sufficient originality claim.

The present framework differs in three commitments. First, its capability families concern causal-epistemic performance rather than consciousness or general intelligence. Second, the decision rule is explicitly non-compensatory. Third, each claimed dependency is paired with a developmental contrast or targeted ablation. Whether this organization has discriminant and predictive validity remains an empirical question.

9.1 Present limitations

The framework has substantial limitations.

First, the implementation is deliberately synthetic and narrow. It verifies code-path feasibility and recovery of known estimands for one Bayesian learner in one resettable generator. It does not validate the six capability families, the candidate gate, architecture-neutral metrics, sequential coupling, or external validity.

Second, ontology is operationalized relative to interventions and outcomes selected by benchmark designers. The agent does not escape human framing. Its autonomy consists in revising categories within an affordance and objective structure, not in creating relevance from nothing. The origin and revision of ends remain outside the present model.

Third, equalizing active and passive conditions remains difficult in sequential systems because control changes state visitation. The resettable primary coupling design removes downstream state mediation, but sequential extensions estimate a total effect that includes coverage. Likewise, Δ_exec-eq is only an equivalence audit when E and C_E expose identical traces and updates; any nonzero value then indicates implementation asymmetry rather than authorship.

Fourth, internal-state measures privilege transparent architectures. The benchmark therefore makes common external measures primary and treats posterior information gain, latent description length, representation probes, and module deletion as secondary within-family diagnostics. Behavioral equivalence can still underdetermine mechanism.

Fifth, simulated embodiment is not biological embodiment. The proposal predicts functional differences in controlled agents; it does not establish that biological affect, metabolism, sociality, or evolutionary history are dispensable for all forms of cognition.

Sixth, category measures can be gamed. A system may infer generator regularities, memorize a family of rule changes, or produce proliferating latent partitions that fit evaluation data. Procedural holdouts, stable-world controls, cross-generator transfer, description-length reporting, and adversarial audits reduce but do not eliminate that risk.

Seventh, the threshold region $\mathcal{G}\left( \mathbf{\tau} \right)$ is conventional and environment-relative. It is a preregistered decision rule, not a natural boundary where a new metaphysical property appears.

Finally, no positive result would demonstrate phenomenal consciousness. The framework intentionally leaves that problem unresolved.

9.2 Staged empirical program

The research program can proceed in five stages.

Stage 1 — Extended simulation. Scale the present microimplementation to a two-dimensional or lightweight three-dimensional environment with all five fault sites, held-out split/merge rules, two agent architectures, and three generator families. Execute the pilot, power simulation, and preregistration in Section 5.6.

Stage 2 — Targeted and factorial ablations. Apply the six shared-interface ablations in Section 6.1, estimate correlations among capability measures, and test discriminant validity. Compare the candidate conjunctive gate with scalar, Pareto, and simpler performance baselines on held-out transfer.

Stage 3 — Morphology and adversarial perception. Introduce all five structural fault sites, tool incorporation, cross-sensor inconsistency, correlated failures, and concurrent-fault episodes. Keep the mutually exclusive single-site task primary and evaluate concurrent faults separately as multi-label diagnosis.

Stage 4 - Physical replication. Port a constrained version to a tabletop robot with camera, proprioception, and force or tactile sensing. Replicate active-passive contrasts under real noise, latency, wear, and calibration drift.

Stage 5 - Independent consciousness analysis. If systems display robust reflexive autonomy, evaluate theory-derived consciousness indicators in a separate study. Do not modify the autonomy benchmark retrospectively to imply subjectivity.

The first full empirical target should remain narrow: estimate the three principal causal contrasts and the separate E/C_E equivalence diagnostic on one causal split/merge task and one five-site fault-localization task. A clean negative result would remain valuable because it would identify which proposed contribution of developmental history is absent in the tested domain.