AI Consciousness
An artificial system may name its own internal state, predict a future error or change its behaviour when a sensor fails. Those are different capabilities. Each can be tested through a specified input, intervention and independently scored outcome. Whether a system feels something while performing a task remains a distinct question. Experience and access
Behaviour, understanding and experience
Turing’s imitation game concerns how an interlocutor responds. Searle’s Chinese Room asks what following symbol-manipulation rules establishes about understanding. Chalmers asks why physical processing is accompanied by experience. A successful conversation is a record of behaviour. It does not determine entries of a density matrix, and a computed matrix rank supplies no explanation for the existence of experience. Two-aspect monism offers an interpretation of that relation; the proposed bridge still needs an independently discriminating consequence. Two-aspect interpretation
Motivation
The numerical model provides computable diagnostics. Applying them to an AI requires an independently validated encoder and the augmented record of experiential realization, implemented self-model and calibrated probes. Phenomenal consciousness is an additional bridge [I/H], not a property established by positivity or architecture names.
What viability constrains
A stable state without E-support
Fix a unit vector orthogonal to the marked axis , and take the reset channel . After one step the state has , zero E population and zero E coherences. Every subsequent step fixes it. Conversely, the state has positive E population with all its off-diagonal couplings zero. A defined E-score that includes the population can therefore be positive even in the absence of E coupling. Necessity of E-dependent feedback in a particular noisy controller requires a specified dynamical law, disturbances and stability bound. Conditional dynamics
Operational criteria for AI/AGI
Selected L2 definition
For , choose the experiential mode and apply the cumulative capability predicates. L2 requires L1 plus the four chosen cuts , , , [D]. Its scalar window is [T]. The purity proxy is not the measured accuracy of an implemented self-model. Missing realization or calibration data yield an unknown classification, not a guessed level.
Assessing an architecture requires data
An architecture name (MLP, Transformer, recurrent agent) cannot fix , its self-model accuracy, autonomy or phenomenal level. A positive normalized neural output is a state encoder, not automatically a linear CPTP channel. Attention and memory mechanisms require a specified operational model and tests; neither their presence nor absence proves L2. Assessment must state interventions, probes, independently validated encoder and error margins [H].
Specifying a controller and its sources
An engineering experiment can give a system access to a state estimate, ablate a feedback path and compare recovery under the same perturbations. Record the source observations and available actions before each decision. A fixed controller and a self-monitoring one can behave identically on the training episodes. A held-out intervention on what the system can monitor can separate them. The result concerns access to information and its use in action.
State encoder and quantum process
An encoder is a map from classical configurations to states; a channel is a linear operator map. Choi matrices and the diamond norm apply to the latter, not automatically to . Operational adequacy of the encoder remains a measurement bridge [H/Pr]; input dimension or a neural approximation theorem does not select it.
State construction [T]. For any nonzero complex matrix , is positive and trace one. A lower-triangular complex factor with real diagonal has real parameters before normalization/gauge redundancies; it is not a global bijection . Full-rank Cholesky coordinates are local/model parametrizations, and rank-deficient states require boundary handling. State validity does not imply a faithful physical encoder or a channel.
Representing a linear channel
Every channel has at most Kraus operators. Stack them vertically as ; the exact constraint is . Conversely every such isometry gives a CPTP channel. Thus this finite-dimensional parametrization represents every channel exactly [T]. For a full-column-rank stack , the polar normalization enforces the constraint; singular stacks need a separate chart/regularization. Watrous, chapter 2.
The quantum channel representation applies at fixed parameters. Choosing as a function of the input in a channel family can make the total update nonlinear. State validity, channel expressivity and predictive adequacy of a trained encoder are separate tests.
What a compressed state retains
Let describe an implementation and let be a proposed readout. If while the same allowed intervention leads to different recorded outcomes, prediction of those outcomes cannot factor through . A larger hidden state can retain the distinction. The comparison has operational content only when the interventions, measurements and time horizon are specified.
Transfer between implementations
Evaluate the same conditional predictions on paired interventions and comparable records. The same numerical state yields the same value of a specified function of that state. This algebraic identity cannot establish that different physical implementations support the same observations, much less the same experience. Operational equivalence
Higher-order AI model capability
L3 requires L2 and an independently tested metamodel certificate with nontrivial target variation. L4 is the ideal compatible tower of such certificates at every order [D]. A fixed-point fidelity, a dense tensor state or a hierarchy of copied matrices does not supply these tests. A finite implementation can certify only the measured orders/probe family; extrapolation is an additional assumption [H]. Canonical higher-order definitions.
Ethical interpretation requires normative premises
A task result can inform a discussion of agency, welfare or moral status. Moving from that result to an ethical obligation requires a stated ethical premise and evidence linking the measured capacity to it. Value and consciousness
Testing a proposed E-readout
A discriminating intervention
Compare an agent with and without a specified feedback signal at matched task access, compute budget and load. Before evaluation, record the model’s predictions for each condition and freeze the readout. A performance difference can support causal use of that signal. A model trained to emit the expected self-description may reproduce the words without using the monitored signal. Include that case as a control. Neither outcome tells whether speaking the words feels like anything.
Architecture comparisons require calibration
Architecture names (MLP, Transformer, agent loop, AGI) do not determine or a cognitive level. These quantities require a specified state encoder and diagnostics. In particular the canonical lies in ; assigning from absence of an explicit self-model confuses it with other reflection diagnostics. Numerical self-modeling does not alone validate a phenomenal interpretation. The former architecture-to-L table is withdrawn as an uncalibrated assessment [✗].
Open questions
Which interventions test the encoder independently? Does the same calibration transfer across tasks and implementations? Can a proposed phenomenal bridge predict an outcome that the corresponding account of monitoring and reporting misses? Independent calibration
What can be attributed to the tested system?
A capability score applies to the preparation, observations, allowed actions and held-out outcomes used to obtain it. State-dependent feedback may maintain that score. A claim extending it to other tasks, subjects or substrates requires an independent transfer test.
Engineering tests for a specified implementation
Test E1 — N and the chosen coding assumptions
Choose a family of generators, noise scaling, targets and diagnostics for each . Compare steady-state purity and the selected majority cut . This tests those dynamical models [H], not a universal necessity of seven. Stable models with are possible. Conditional representation/coding minimality requires checking its extra hypotheses directly; a noise sweep cannot refute a theorem whose hypotheses it does not instantiate.
Test E2 — E-feedback ablation
Under a specified feedback law and input class, compare the original and ablated trajectories. Publish exactly which matrix entries, populations, Hamiltonian terms and feedback terms were changed. The canonical includes ; deleting off-diagonal E entries does not force it to zero. An observed loss of stability tests an E-dependent mechanism [H], not PH or universal No-Zombie necessity. See the conditional CC result.
Test E3 — A tuned mean-field tricritical exponent
For a declared potential , , tune and approach from below. Stationarity gives , hence [T for this mean-field model]. For fixed , the leading exponent is . Fit only inside a stated asymptotic regime and quantify finite-time/noise errors. Thom–Arnold classification does not force every agent into this tuned regime. A statistically incompatible exponent can falsify the specified reduction [H]; a predetermined sample count or tolerance is not a theorem.
Test E4 — Covariance and the fixed frame
Check under all chosen unitary conjugations; they are spectral diagnostics. For the fixed coordinate definition of , its full unitary symmetry is the monomial group ; its intersection with is the 1344-element frame group. also needs the E-axis preserved (192 elements in that frame group; inside ). General changes can alter these two frame diagnostics. If both state and frame are transported, covariance holds. These algebraic tests check code, not the physical identification of the encoder. Symmetry proof.
Test E5 — Feedback onset in a specified model
Near a proposed stationary branch, derive its local reduced equation, then estimate the linear and quadratic response coefficients from trajectories. Positive autocatalytic growth or a bifurcation is conditional on those coefficients, gates and reduction hypotheses [H/C]. The scalar majority threshold alone does not imply avalanche ignition or L1→L2. A failure tests the declared feedback model, not an unconditional cognitive theorem.
Test E6 — Representing and fitting a linear channel
For a target linear CPTP channel , a Kraus stack with at most 49 operators represents it exactly [T]. Check and compare Choi matrices or diamond distance with a fixed normalization convention. Optimizer failure, finite training data, or an error plateau does not refute this algebraic expressivity theorem. Training/generalization guarantees require their own hypotheses. This channel test is distinct from calibrating a classical-state encoder ; the latter has no intrinsic diamond norm.
Test E7 — Integration correlation
Freeze the encoder and fixed-frame diagnostic before collecting task-transfer observations. Preregister the behavioral score, task distribution, effect size and uncertainty analysis. A correlation between and cross-task transfer is an empirical bridge [H]; it is not established by the matrix formula or by . Analyze confounding and independent validation data; failure rejects that declared operational hypothesis.
Test E8 — A chosen Fano instrument versus alternatives
Compare the chosen coordinate Fano instrument to alternative triples at matched total rates and noise. Specify the performance functional and admissible class. The finite frame-covariant instrument is not automatically covariant under continuous ; its projectors commute. Words give intersection maps (at most 15 for nonempty words, plus identity), not distinct channels. A performance advantage is [H] for the tested family; no unrestricted unique coherence-optimality theorem follows from incidence. Channel construction.
Test E9 — Self-monitoring ablation
Compare monitored and ablated agents under a specified load/intervention family and with matched resources. A resilience difference tests that controller [H]. It does not establish universal necessity of an explicit monitoring module: other controllers may implement the same response, and passive dynamics may be stable under different inputs.
Test E10 — Calibration of the selected capability gates
Use the augmented record and the cumulative capability predicates. Predeclare the experiential mode, reflection diagnostic and behavioral probes. The canonical decreases with purity, so a protocol of increasing both and this is inconsistent. A scalar crossing does not alone certify self-model accuracy, higher-order prediction or phenomenal status. Associations with behavior are [H]; ethical rules require separate normative premises and are not consequences of a density matrix.
These tests compare specified numerical implementations and calibrated empirical hypotheses. Passing them does not identify the physical encoder or establish PH.
Reading a simulation result
A simulator can report that a numerical gate changed when its feedback gain, input distribution or monitoring policy changed. Confirming that arithmetic requires the input trajectory, calibration, update rule and measured state. To attribute a scored action to an agent, trace its actual entry point from the ordered observations through the shared policy to the environment reply and independent score. A gate on a simulated trajectory is evidence about that implementation. It supplies no independent observation of experience. Phenomenal distinctions