Figure 1.1
Files: Fig01_01.png
Caption: The criterion reads a completed state trajectory but does not change the update. The final interpretation is a scientific argument outside the agent, not another internal state.
Alternative text: Flow diagram: observations enter a state update, which stores a trajectory. A separate decision criterion reads that trajectory and produces a report; interpretation follows the report.

Figure 1.2
Files: Fig01_02.png
Caption: Both runs start at 0.5 and use alpha 0.35. Reordering the same observations changes the trajectory and final report at criterion 0.55. This demonstrates order sensitivity of the specified accumulator, not experienced temporal continuity.
Alternative text: Two line traces start at accumulator score 0.5 and undergo eight updates. Reversing the observation order changes the intermediate scores and leaves a lower final score than the given order.

Figure 2.1
Files: Fig02_01.png
Caption: Selected relations from the nine-node conceptual registry. Dashed connections to Experience mark inferences, not entailments. The graph encodes assertions and does not implement the capacities named by its nodes.
Alternative text: Conceptual graph: a program implements attention and access; attention modulates access. Access enables report and guides behavior. Dashed arrows from report and behavior to an experience target denote inferential links.

Figure 3.1
Files: Fig03_01.png
Caption: A theory-to-experiment structure. In this laboratory the generator assigns probabilities; it contains no routing mechanism. Comparing a prediction with its generated observations tests recovery, not a causal broadcasting theory.
Alternative text: A theoretical idea leads to a specified generator. The generator supplies a declared prediction and recorded trials; both enter comparison and interpretation.

Figure 3.2
Files: Fig03_02.png
Caption: Accuracy from 200 generated trials per condition-task cell, seed 202603. Condition names select assigned probabilities, not implemented routes. Bars recover a sampled encoded contrast; no causal broadcast or sham operation is tested.
Alternative text: Grouped horizontal bars show synthetic direct and transfer accuracy under local, broadcast, ablated and sham labels. Direct accuracy is similar across labels; transfer is lower for local and ablated conditions than for broadcast and sham.

Figure 4.1
Files: Fig04_01.png
Caption: The synthetic observer has separate decision and visibility paths from shared evidence. A report-only intervention preserves the decision path; neither path is a direct measurement of experience.
Alternative text: Signal and stimulus onset asynchrony generate noisy evidence. One branch reaches a decision threshold; another adds report noise and reaches visibility categories. Both branches enter evaluator counts.

Figure 4.2
Files: Fig04_02.png
Caption: Seed 202604: accuracy uses 500 trials per SOA; mean visibility uses 250 signal-present trials. Visibility divided by four is a scaled ordinal report mean, not a probability. The four SOAs are assigned simulation conditions and neither series measures experience.
Alternative text: Horizontal bars compare accuracy with mean visibility divided by four at four assigned stimulus onset asynchronies, from 16 to 83 milliseconds. Both series rise across these synthetic conditions.

Figure 5.1
Files: Fig05_01.png
Caption: Two selectors receive the same scored display in parallel. Target labels are used by the evaluator, not by either policy. Equal output capacity does not imply equal computation cost.
Alternative text: Salience and goal relevance combine into priority plus noise. The same scores feed hard top-k selection and softmax sampling in parallel, and both are evaluated for target inclusion.

Figure 5.2
Files: Fig05_02.png
Caption: Target inclusion on 250 paired displays per cell, seed 202605, capacity two, softmax temperature .30. Low and high priority-noise SDs are .15 and .45. The result compares these policies under assigned feature distributions, not all attentional mechanisms.
Alternative text: Grouped horizontal bars compare hard top-k and softmax target inclusion for six or twelve items under low or high noise. Hard selection exceeds softmax in all four cells; more items and higher noise reduce inclusion.

Figure 6.1
Files: Fig06_01.png
Caption: One recurrent update combines current input and prior state, then reads the updated state. Repeating this update is distinct from merely drawing a deep feedforward chain; finite unrolling can nevertheless reproduce its output.
Alternative text: Current input and prior state feed a weighted sum, followed by a tanh update, an updated state and a readout. The readout occurs after the state update.

Figure 6.2
Files: Fig06_02.png
Caption: Accuracy on 250 paired eight-sample sequences per cell, seed 202606. Recurrent and unrolled states agree exactly; both use more history than the memoryless control. Pooling is a distinct rival. Thresholds are shared but not separately calibrated, and window storage remains a resource difference.
Alternative text: Four groups of horizontal bars compare memoryless, recurrent, unrolled and pooled predictors under two signal amplitudes and two noise levels. Recurrent and unrolled bars coincide in every condition.

Figure 7.1
Files: Fig07_01.png
Caption: Candidate information passes through competition and admission before routes are evaluated. The target identifier is evaluator-only. The introductory wiring demonstration models perfect availability along permitted routes, not measured message delivery.
Alternative text: Specialized processors feed competition and workspace admission. Admission leads to permitted recipient routes. The evaluator reads these routes and a separate target identifier.

Figure 7.2
Files: Fig07_02.png
Caption: Paired synthetic routing comparison with 2,400 bases per cue timing. Correctness is measured across all trials, including abstentions. All three protocols converge when the delivery budget is sufficient; late-cue encoding losses limit all three. One credit is one directed-edge packet delivery, not a hardware multicast transaction.
Alternative text: Two line plots compare broadcast, selective routing and a local ring as delivery budget increases. Selective routing reaches its plateau sooner; all three protocols converge with sufficient budget. The late-cue plateau is much lower than the early-cue plateau.

Figure 8.1
Files: Fig08_01.png
Caption: The buffer's identities and strengths are separate from probe evidence and the response criterion. The model can isolate these operations, but its state variables do not identify a biological maintenance mechanism.
Alternative text: Presented items pass through priority admission into a retained buffer. Decay, refresh and interference interact with the buffer. Buffer contents supply probe evidence, followed by a criterion and decision.

Figure 8.2
Files: Fig08_02.png
Caption: The criterion sweep reuses exactly the same Maintained evidence from 900 synthetic bases. Lower criteria admit additional positive responses, including errors, without altering storage. False alarms are mathematically excluded above the lure bound 343/512 under this generator; the curve is not participant data.
Alternative text: Across increasing response criteria, target hit rate declines slightly from about 0.9, while lure false alarms fall from about 0.1 to zero and remain there for the higher criteria.

Figure 9.1
Files: Fig09_01.png
Caption: The monitor reads declared evidence and targets a first-order decision. Its estimate can feed a report and a controller, while correctness is computed separately. These are functional relations, not proof of conscious introspection.
Alternative text: Evidence feeds a first-order decision and a target-linked monitor. The monitor produces a confidence report and informs a selective controller. The controller and decision feed an outcome evaluator.

Figure 9.2
Files: Fig09_02.png
Caption: All monitors share the same 1,200 decisions, but the controller retains 720 trials using each monitor's confidence. Selective accuracy changes while first-order accuracy remains fixed. This batch-selection comparison assumes known synthetic difficulty and does not establish real-time help-seeking or subjective uncertainty.
Alternative text: All four monitors share identical first-order accuracy. At 60 percent coverage, selected-trial accuracy is higher for each monitor; matched and overconfident monitors have equal selected accuracy, while noise and a blind spot reduce it.

Figure 9.3
Files: Fig09_03.png
Caption: Human confidence on held-out observations. A: each confidence bin gives equal weight to its twenty participant means; vertical bars are 95% participant-bootstrap intervals. The dashed line denotes equality between confidence and accuracy. B: each line joins a participant's raw and recalibrated held-out Brier scores. Seventeen of twenty improve; lower is better. Data: Shekhar and Rahnev (2021), Confidence Database. This figure is an original reanalysis.
Alternative text: Two panels show held-out human confidence. A reliability curve plots participant-weighted accuracy against confidence with vertical bootstrap intervals and an equality line. A paired plot joins each of twenty participants’ raw and recalibrated Brier scores; seventeen lines descend.

Figure 10.1
Files: Fig10_01.png
Caption: The exact laboratory separates the intervention ensemble, update rules, and cut mechanism. Its final loss characterizes this declared educational comparison, not an IIT calculation.
Alternative text: State variables and time define uniform interventions and network update rules. These lead to a full next-state distribution and a declared noisy partition, which are compared as whole-versus-cut loss.

Figure 10.2
Files: Fig10_02.png
Caption: Exhaustive uniform interventions give the modular system the most output entropy but zero toy partition loss. Both quantities use bits but answer different questions; the cut loss is not a measure of consciousness.
Alternative text: Horizontal bars compare output entropy and toy partition loss in three three-node systems. Modular: three bits entropy and zero loss. Feedforward: two bits entropy and one bit loss. Recurrent XOR: two bits for both.

Figure 11.1
Files: Fig11_01.png
Caption: The Bayesian cycle stores the forecast before using its outcome. The final arrow advances to the next time step; this is passive filtering with specified parameters, not learned or action-conditioned control.
Alternative text: A state prior and emission model yield a pre-outcome forecast. The observation, prior and emission model inform the posterior. A transition advances that posterior to the next prior, closing the filtering cycle.

Figure 11.2
Files: Fig11_02.png
Caption: On the seed-202611 Python sensor trace, the reference filter has the lowest valid log loss. Last-observation accuracy exceeds the base rate, yet its log loss is worse. All models receive the same sixty observations; this single synthetic schedule is not a population estimate.
Alternative text: Horizontal bars compare valid pre-outcome log loss. The reference latent filter has the lowest loss; rigid mismatch is next, then the base rate, while last-observation prediction has the highest loss.

Figure 12.1
Files: Fig12_01.png
Caption: The schema predicts errors of the primary selector and controls requests for additional evidence. The intervention can misalign risk or disconnect the controller; target labels remain with the evaluator, outside this policy loop.
Alternative text: A primary cue feeds a two-channel selector, a primary-error risk model and final selection. The risk model controls rechecking; a requested diagnostic cue also enters final selection.

Figure 12.2
Files: Fig12_02.png
Caption: On the seed-202612 Python diagnostic fixture, aligned and always-on sensing both reach 58 of 60 correct choices, using 25 and 60 requests respectively. Utility subtracts 0.08 per request; the direct-margin rule exactly matches aligned behavior and the disconnected consumer returns to baseline. These are finite synthetic outcomes, not consciousness measures.
Alternative text: Grouped bars compare accuracy and utility for six policies. Aligned, always-recheck and direct-margin policies share the highest accuracy. Aligned and direct-margin utility exceeds always-recheck utility because they make fewer sensing requests.

Figure 13.1
Files: Fig13_01.png
Caption: Current observations and internal deficits guide action; consequences update body state, while environmental availability follows the fixed schedule. Memory selection occurs after the run and has no return edge to the policy. The quantities are software regulatory states, not measures of felt affect.
Alternative text: Environment and resources supply current sensors. Sensors and body deficits guide action selection. Action consequences update body state; post-run memory selection reads body state and action consequences without feeding the policy.

Figure 13.2
Files: Fig13_02.png
Caption: Mean Euclidean distance from the reference state across the same sixty-step schedule. The selective deficit-channel swap increases distance under this policy. Task tokens and critical violations are reported separately in the text; these are deterministic summaries, not confidence estimates.
Alternative text: Horizontal bars compare mean post-action distance from a reference body state. Aligned control has the lowest distance, followed by fixed, swapped and task-only control.

Figure 14.1
Files: Fig14_01.png
Caption: The policy reads retained content before selecting a response. Ablation clears storage, replacement edits the token, and disconnection blocks the storage-to-policy edge. Public instruction can supply a separate policy route. The final node groups the evaluator, which reads the response, and the later log, which copies the selected rule. Neither feeds the decision; the diagram describes access, not phenomenal privacy.
Alternative text: Public instruction supplies both a policy route and an instruction encoder. Encoded private rule storage feeds the policy before its response. Response and stored content feed a downstream log and evaluator.

Figure 14.2
Files: Fig14_02.png
Caption: Seven implemented controllers on the same deterministic fixture. Retained information supports performance, but an integer code and repeated public cues match the verbal controller. The separate downstream-log intervention changes recorded content without changing responses. The counts follow analytically from the disclosed rules and access schedule.
Alternative text: Bars show correct responses out of sixty for seven controllers. Private verbal storage, integer memory and repeated public cues each achieve sixty. Ablation lowers performance, and a replaced token gives the fewest correct responses.

Figure 15.1
Files: Fig15_01.png
Caption: Architecture, execution, causal intervention and behavior supply different records. Theory and bridge assumptions determine which implemented properties may be relevant to a consciousness indicator. A diagram, normalized output or fluent report does not substitute for this evidence chain.
Alternative text: Architecture and weights lead to executed states and reads. Execution informs observed behavior and controlled interventions. These records, together with theory and bridge assumptions, feed an indicator dossier with alternatives.

Figure 15.2
Files: Fig15_02.png
Caption: Changing the unit of analysis changes the reported resolution rate. The distribution at right shows contrasts within issue and model; the dashed line marks their equally weighted issue mean. Negative values denote fewer recorded AI turns in successful attempts.
Alternative text: Left: resolved-attempt rates are 19.0 percent when weighting runs equally and 11.4 percent when weighting issues equally. Right: a histogram of seventy-eight within-issue and model interaction-length contrasts has a negative mean marked by a dashed line, left of the solid zero line.

Figure 16.1
Files: Fig16_01.png
Caption: Both predictors receive the same public inputs; a separately specified plant generates outcomes. The bundle precedes plant execution in code, but its digest does not establish external preregistration or analyst blinding. The plant has its own modeling assumptions.
Alternative text: A shared task and bridge contract define public trial records. These feed an independent synthetic plant and local and global predictions. Frozen predictions and plant outcomes meet in row-level comparison and amendments.

Figure 16.2
Files: Fig16_02.png
Caption: Both predictors miss two canonical outcomes, but they fail on different rows. Their mismatch counts differ on the complete ninety-case input grid. Each count uses the stated denominator and is conditional on one synthetic plant; the two sets are not equal-sized samples or an empirical ranking of consciousness theories.
Alternative text: Both predictors mismatch two of seven canonical outcomes. On the separate ninety-case grid, Model L has more mismatches than Model G. The two evaluation sets have different denominators.

Figure 17.1
Files: Fig17_01.png
Caption: Both scientific and moral uncertainty inform possible artificial welfare and protection of people. The resulting questions require accountable human review. Neither uncertainty classification nor a software screen supplies research permission.
Alternative text: Scientific and moral uncertainty each inform possible artificial welfare and human and social harms. Both branches lead to accountable human review and authorization.

Figure 17.2
Files: Fig17_02.png
Caption: Screening order is validation, non-compensatory stops, and redesign flags before readiness for human review. Invalid records and triggered conditions halt progression. The three fictional proposals produce one ready-for-review, one redesign and one do-not-run state; all fifteen human checks remain pending and no authorization is granted.
Alternative text: A screening flow proceeds from declared project facts through required-field validation, hard-stop rules and review flags to readiness for human review. Independent human checks remain pending after this software screening.

Figure 18.1
Files: Fig18_01.png
Caption: Six commitments make a project inspectable from question to conclusion. Every bridge remains contestable; a successful run supplies neither a consciousness verdict nor authorization for a research extension.
Alternative text: Six boxes form a project sequence: question and claim boundary; theory and computational bridge; interfaces, state and timing; predictions, controls and rival; preserved execution evidence; bounded conclusion and human review.

Figure 18.2
Files: Fig18_02.png
Caption: Deterministic results over all twelve trials and all five delayed trials; nonreports count as errors. DirectRival matches baseline. Local dependencies do not establish broadcast necessity or consciousness.
Alternative text: Grouped bars show accuracy across all twelve trials and the five delayed trials. Baseline, sham, no-monitor and direct-rival results coincide. Removing sharpening reduces performance; removing memory eliminates delayed accuracy, and removing broadcast eliminates both scores in this fixture.

Figure 18.3
Files: Fig18_03.png
Caption: A frozen initial policy and its selected revision on two untouched final sets of 4,000 synthetic cases each. The revision improves utility on the same-distribution test but performs worse than the initial policy after the feature-label relationship changes. The final results never enter the proposal or acceptance loop.
Alternative text: Grouped bars compare initial and revised policy utility. On the same-distribution final test, utility rises from −0.219 to 0.702. Under the shift challenge, it falls from 0.511 to 0.072. Each final set contains 4,000 synthetic cases.
