# Extended exercise bank

All 432 original prompts plus 12 new prompts. Original IDs are retained; new chapter exercises use 25 to 27. The printed selection is documented in printed_selection.json.

## Exercise 1.1

Foundation. Calculate the first three scores by hand for α = 0.35 and the printed observations. State the rounding used and compare with the Python trace using a declared tolerance.

## Exercise 1.2

Foundation. Explain why a score in [0, 1] is not automatically a posterior probability. Give the likelihood and prior information that would be needed to justify a Bayesian interpretation.

## Exercise 1.3

Implementation. Extend `State` with a read-only condition label without changing the transition rule. Test that changing only the label leaves every score and report unchanged.

## Exercise 1.4

Implementation. Reproduce the aliasing bug using a mutable dictionary and a history list. Repair it and write a test that would have failed before the repair. Keep this deliberately faulty example separate from the canonical module.

## Exercise 1.5

Foundation. Derive the contribution of the initial score and of each observation after t updates. Explain what happens to old information when α is close to zero or one.

## Exercise 1.6

Experiment. Construct two permutations of the same eight observations with identical final observations but different final scores. Explain which earlier positions account for the difference.

## Exercise 1.7

Implementation. Test the guards for a non-integer step counter, infinite observation, and invalid criterion. Include an invalid initial state in an empty trace. Explain why validating only inside the update would miss that case.

## Exercise 1.8

Foundation. Distinguish an input, a state variable, a parameter, and an outcome in the accumulator. Show how changing the reporting criterion differs from changing α.

## Exercise 1.9

Experiment. Sweep α over eleven values from zero to one using the fixed fixture. Plot or tabulate final scores and response disagreements. Do not interpret a smooth curve as empirical evidence about human memory.

## Exercise 1.10

Implementation. Write an independent closed-form accumulator using the expression derived in Exercise 5. Compare it with the iterative implementation on at least 100 seeded sequences and both limiting values of α.

## Exercise 1.11

Experiment. Compare paired and independently sampled input designs for estimating a difference between two accumulators. Report the variance across repeated estimates and explain what pairing does and does not control.

## Exercise 1.12

Critique. A notebook produces the correct final score only after its cells have been run out of order. Explain why the visible output is insufficient evidence and describe a minimal repair and rerun procedure.

## Exercise 1.13

Foundation. Give an example where a missing observation should be skipped and another where missingness itself is informative. Specify different update rules rather than representing both cases as zero.

## Exercise 1.14

Implementation. Store undefined conditional accuracy as `None` with a reason field. Test the empty-selection case and a nonempty case with a known numerator and denominator.

## Exercise 1.15

Experiment. Investigate whether reversing an input sequence always changes the report. Find counterexamples and explain why the original demonstration did not imply a universal rule.

## Exercise 1.16

Implementation. Convert the history to JSON-compatible dictionaries and back. Check all fields, ordering, and finite values. Explain why a screenshot cannot provide the same machine-checkable record.

## Exercise 1.17

Critique. Evaluate the claim that twenty passed tests validate the agent's consciousness. Identify separately the computational, modeling, and phenomenal inferences that would be required.

## Exercise 1.18

Experiment. Compare the accumulator with an arithmetic-mean observer on sequences with identical means. Specify both agents' information access and count the memory needed by each implementation.

## Exercise 1.19

Implementation. Add a reset intervention after the fourth observation. Preserve the original fixture and compare against a sham reset that leaves the state unchanged. State the reset value explicitly.

## Exercise 1.20

Foundation. Explain why using the same seed in two conditions may fail to generate paired inputs. Construct a short example in which one extra random draw changes all later samples.

## Exercise 1.21

Critique. A student names a variable `confidence` and uses its distance from 0.5 to choose trials. Propose a calibration evaluation that would be needed before interpreting that name probabilistically.

## Exercise 1.22

Experiment. Repeat the 100-seed comparison over a new, nonoverlapping seed range. Report both results and the range of individual paired differences. Do not select the range that makes a preferred model look best.

## Exercise 1.23

Implementation. Test a NumPy prediction array against a target array when one accidentally has shape (n, 1). Add shape checks before scoring and explain the misleading result that broadcasting could produce.

## Exercise 1.24

Critique. Write a 300-word laboratory interpretation with a question, implemented mechanism, intervention, observed contrast, strongest remaining rival, and explicit limit. Include one result that would make you revise the interpretation.

## Exercise 2.1

Foundation. Define phenomenal and access consciousness in two sentences each. Describe a robot experiment that tests availability to two downstream functions. State what the experiment observes and which additional premise would be needed for an experiential interpretation.

## Exercise 2.2

Foundation. Classify these questions as functional, mechanistic, methodological, or phenomenological: which modules read a message; which lesion disrupts report; how reliable a visibility scale is; what seeing red feels like. Explain one case in which two categories legitimately apply.

## Exercise 2.3

Foundation. Rewrite “the model consciously sees red” as separate claims about represented content, usable information, and experience. For each claim, identify a possible record and an alternative explanation that the record does not exclude.

## Exercise 2.4

Implementation. Extend `profile` with a remembered-content list. Test both a current-report policy and a memory-report policy. Construct a record accepted by the latter but rejected by the former, and explain why neither validator is universally correct.

## Exercise 2.5

Experiment. Use the eight-trial fixture to calculate accuracy conditional on seen and not-seen reports. Retain numerator and denominator. Explain why these proportions do not establish unconscious perception, considering sampling, report sensitivity, and conditioning on shared evidence.

## Exercise 2.6

Foundation. Give distinct operational measures for attention, retention, metacognitive sensitivity, and self-prediction. For each, identify one high score that could be achieved by an inappropriate shortcut and propose an observation that would expose it.

## Exercise 2.7

Implementation. Build an evidence dictionary with three statuses: collected, not collected, and not applicable. Write a query returning only not-collected items. Verify that zero-valued measurements are retained as measurements rather than mistaken for missing values.

## Exercise 2.8

Critique. Compare a stored “I am conscious” sentence with a sentence generated from a state-dependent reporting rule. In 150 words, explain the functional difference and why it does not alone resolve the experiential question.

## Exercise 2.9

Implementation. Add latency and reset sensitivity to the two system records. Use `equivalent` with three different field selections. Report how the equivalence judgment changes and identify the interface that best addresses a stated engineering question.

## Exercise 2.10

Implementation. Construct three system records that match on accuracy but form two groups when latency is included. Group them with tuple-valued dictionary keys. Explain why an equivalence class is defined by the projection, not discovered independently of it.

## Exercise 2.11

Foundation. Write a construct specification for metacognitive sensitivity containing definition, variable domain, operation, measure, and alternative explanation. Distinguish the association between confidence and correctness from average confidence and from calibration.

## Exercise 2.12

Implementation. Write a validator returning missing documentation keys instead of a single Boolean. Test a complete dictionary, one with an empty definition, and one with an unknown field. State which defects are structural and which still need scientific judgment.

## Exercise 2.13

Implementation. Add units, initialization, and rationale to the concept registry without changing stable identifiers. Demonstrate that a display-name change preserves every graph query. Explain why changing a definition is a more consequential edit.

## Exercise 2.14

Experiment. Add a Memory-to-Report `supports` relation. Compare direct neighbors and full non-inferential reach before and after the change. Restrict your conclusion to encoded topology unless you also specify a dynamic transmission rule.

## Exercise 2.15

Implementation. Construct four malformed relations: unknown endpoint, duplicate, unknown type, and a non-inferential edge into Experience. Verify that each is rejected. Then propose a scientifically dubious relation that still passes, explaining the remaining limitation.

## Exercise 2.16

Critique. Design report and no-report versions of one perceptual experiment. Identify the observed variable, an important task confound, the validation needed for the marker, and the strongest conclusion each design could support.

## Exercise 2.17

Implementation. Represent evidence bundles as sets. Find two equally sized bundles with different audit outcomes for the same claim. Explain why cardinality cannot substitute for the identity of missing prerequisites.

## Exercise 2.18

Critique. Define two rival prerequisite registries for interpreting the same computational finding. Keep the observations fixed. Identify whether their disagreement concerns evidence quality, bridge assumptions, or the proposed target rather than silently changing the data.

## Exercise 2.19

Implementation. Enumerate simple paths from Program to Experience and annotate each path with relation types. Verify that every path contains an inference edge. Prevent infinite recursion if a cycle is added and state what the path query establishes.

## Exercise 2.20

Experiment. Add a binary transmission rule to a four-node map. Remove one route, preserve the input and readout, and compare the output with a sham edit. Distinguish deletion of a graph assertion from intervention on an executed mechanism.

## Exercise 2.21

Critique. Read Block (1995). Develop a 300-word defense or criticism of the phenomenal/access distinction for computational modeling. Identify an empirical consequence of your position and an observation that would challenge it.

## Exercise 2.22

Critique. Compare functionalist and substrate-sensitive interpretations of an identical finite input-output trace. Express both arguments as premises and a conclusion. Mark the disputed premise without treating a simulation analogy as its proof.

## Exercise 2.23

Experiment. Draft a prospective analysis plan for testing whether confidence distinguishes errors. Specify generated or collected data, baseline, exclusions, primary statistic, uncertainty calculation, and a rival heuristic. Do not describe your classroom plan as externally registered.

## Exercise 2.24

Critique. A system selects objects, broadcasts representations, predicts some errors, and produces first-person reports. Write a defensible conclusion about its demonstrated capacities, identify the missing interventions, and explain which further assumptions would enter a claim about experience.

## Exercise 3.1

Foundation. Rewrite “broadcast causes consciousness” as a theoretical idea, a candidate mechanism, a restricted hypothesis, and a measurable prediction. Identify the first unsupported inference and propose evidence that would address it rather than simply weakening every sentence.

## Exercise 3.2

Foundation. Assign roles to noise, capacity, architecture, and seed in two different experiments. Explain why the same parameter can be manipulated in one design and controlled in another. Include one nuisance variable that your chosen software boundary leaves outside the model.

## Exercise 3.3

Implementation. Define variable records with name, role, domain, unit, missing-data policy, and aggregation. Write a validator returning incomplete records. Test an absent unit separately from an explicitly dimensionless measure and explain the distinction.

## Exercise 3.4

Implementation. Construct a three-condition by two-task factorial design with 40 trials per cell. Verify six cells, 240 unique identifiers, and balanced counts. Delete one record and show that the balance check detects the intended defect.

## Exercise 3.5

Foundation. Build a trial record containing a choice, an internal measurement, a generated latency, and an uncollected participant measure. Explain why generated milliseconds are not processor timings and why uncollected data should not be replaced with zero.

## Exercise 3.6

Critique. Compare a workspace architecture with a local baseline differing in memory, parameter count, and input history. Identify three possible explanations of an accuracy difference. Choose one additional control and state which alternatives it leaves unresolved.

## Exercise 3.7

Experiment. Implement a four-node message route with an explicit recipient rule. Remove the task-relevant edge and a task-irrelevant edge in separate conditions. Record messages and outputs under identical inputs; explain whether the second edit is an adequate sham for the first.

## Exercise 3.8

Critique. A two-edge ablation eliminates transfer. Explain why this does not identify either edge as individually necessary. Propose a factorial intervention and specify the interaction that would indicate redundancy or dependence between the routes.

## Exercise 3.9

Implementation. Count requested uniform and Gaussian draws in each generator cell. Distinguish requested variates from internal pseudorandom operations. State why equal draw counts do not prove equal computational cost or equal information access.

## Exercise 3.10

Implementation. Demonstrate local random-state isolation using two `random.Random` objects. Verify that running a laboratory between two draws from an unrelated generator leaves those unrelated draws unchanged. Retain both the reference and observed sequences.

## Exercise 3.11

Experiment. Reproduce the twenty-seed unpaired transfer summary using seeds 202650–202669. Report mean, sample SD, minimum, and maximum. Explain why the observed range is neither a confidence interval nor a test of another generator family.

## Exercise 3.12

Foundation. Classify intervals [.02, .11], [−.14, −.03], and [−.06, .07] for a positive directional prediction. State the assumptions omitted by these labels and provide an example in which a failed manipulation check prevents interpretation.

## Exercise 3.13

Critique. An estimate is .01 with a narrow interval inside [−.02, .02]. Explain how a prospectively chosen equivalence region differs from concluding exact equality after failing to reject zero. Identify an engineering reason to choose a particular region.

## Exercise 3.14

Implementation. Create a requirements ledger for source inspection, captured execution, observed outputs, empirical data, and independent implementation. Return named missing prerequisites for two claims. Construct equal-sized evidence sets with different outcomes.

## Exercise 3.15

Implementation. Add checks for invalid probabilities, missing tasks, duplicate trial identifiers, and incomplete outcomes. For every check, supply one intentionally corrupted input. Keep a directional outcome prediction outside the essential software-failure branch.

## Exercise 3.16

Experiment. Define two models predicting transfer and direct-task effects. Specify a squared-error comparison before generating new observations. Include a third possible pattern that challenges both models and explain why the lower error does not establish uniqueness.

## Exercise 3.17

Critique. Audit “the ablated system was less accurate, proving broadcast necessary for consciousness.” Separate the observed contrast, the intervention interpretation, the comparison class, and the experiential attribution. Rewrite the conclusion for the actual probability-generator laboratory.

## Exercise 3.18

Experiment. Design a comparison between an uncertainty monitor and a fixed confidence heuristic under changed difficulty. Keep task evidence matched, reserve a held-out condition, and identify an outcome that would challenge each account without reading correctness inside the policy.

## Exercise 3.19

Experiment. Repeat paired and unpaired simulations over seeds 202700–202739. Compare the sample SD of the transfer contrast. Derive the variance of the paired Bernoulli difference when p₁ ≥ p₂ and explain the dependence introduced by common random numbers.

## Exercise 3.20

Experiment. Vary trials per cell across 50, 200, and 800 using a declared seed set. Summarize contrast variability separately at each size. Do not select the sample size afterward by stopping at the first favorable result.

## Exercise 3.21

Implementation. Add a constant-probability null with shared uniforms across conditions. Verify exact zero correctness contrasts while allowing generated reach to differ. Explain why this separates a scoring-pipeline null from a claim that all measured variables are identical.

## Exercise 3.22

Critique. Read Melloni et al. (2023). Identify two contrasting predictions, one manipulation check, and an outcome that would challenge both accounts. Keep the published protocol distinct from any later results or your own reconstruction.

## Exercise 3.23

Experiment. Estimate uncertainty for the transfer-minus-direct interaction using independent cell variances, then compare with resampling trials within cells. State the unit of resampling and explain why a pair of separate significance judgments is insufficient.

## Exercise 3.24

Experiment. Write a prospective plan for a new computational test of attention or metacognition. Include rivals, variable definitions, exclusions, randomization, primary contrast, uncertainty, null control, robustness analysis, and the record needed to distinguish planned from exploratory work.

## Exercise 4.1

Foundation. Define separate stimulus, representation, decision, motor-response, and report fields for a camera-based obstacle detector. Identify which fields require instrumentation and explain how a change in motor latency could occur without a change in evidence.

## Exercise 4.2

Foundation. Classify contrast jitter, sensor readout noise, stochastic updates, criterion drift, and report-memory error relative to a declared system boundary. Change the boundary once and explain which classification changes without changing the physical event.

## Exercise 4.3

Implementation. Combine independent noise SDs .25 and .60, then add covariance .06. Calculate both total variances and SDs. State the permitted covariance range and explain why standard deviations cannot simply be added.

## Exercise 4.4

Implementation. Generate 1,000 samples from each of two Gaussian evidence distributions with a local generator. Verify exact replay, retain the samples, and compare their observed means and variances with the assigned parameters without requiring equality.

## Exercise 4.5

Foundation. For midpoint .30, slope 12, and lapse .02, calculate 70, 75, and 85 percent thresholds. Verify each by substitution. Explain why the midpoint is not the 75 percent threshold and what happens when a requested target exceeds the upper asymptote.

## Exercise 4.6

Implementation. Estimate the 75 percent threshold on grids of .02, .01, and .001 spacing. Compare each first-crossing estimate with the analytic inverse. Report the direction and maximum expected size of the grid discretization error.

## Exercise 4.7

Implementation. Write a four-outcome classifier from signal presence and yes/no response. Verify all four Boolean combinations. Add validation rejecting the string “False” so that truthy text is not silently treated as a Boolean observation.

## Exercise 4.8

Implementation. For 46 hits, four misses, eight false alarms, and 42 correct rejections, calculate raw and corrected rates, d′, c, and accuracy. Show the denominators and explain why correction must be documented alongside the result.

## Exercise 4.9

Critique. Construct a zero-false-alarm sample with only ten absent trials. Explain why the uncorrected normal quantile is infinite and why a finite corrected d′ does not remove uncertainty about a rare false-alarm probability.

## Exercise 4.10

Experiment. Hold H and F fixed while varying prevalence from .05 to .95. Compute accuracy and an expected cost with misses ten times as costly as false alarms. Explain why both can change without changing sensitivity or the response rule. Then state why a cost-aware detector might choose a different criterion in the new environment.

## Exercise 4.11

Experiment. Generate ROC coordinates for d′ values .5, 1, and 2 over raw thresholds −2 to 3. Verify the directions in which H and F change. Distinguish a criterion sweep within one model from empirical estimates under changing conditions.

## Exercise 4.12

Experiment. Fit the seven illustrative proportions, then perturb each by .02 in separate runs while keeping them within zero and one. Record the 75 percent threshold changes. Explain why this sensitivity analysis is not a sampling confidence interval.

## Exercise 4.13

Experiment. Refit the same illustrative points with lapse fixed at zero, .02, and .08. Report midpoint, slope, residual loss, and threshold. Explain how parameter compensation can hide a wrong asymptote over a restricted stimulus range.

## Exercise 4.14

Experiment. Run the masking generator for ten declared seeds and fit each run separately. Report the range of threshold estimates, including any estimates outside the sampled SOAs. Explain why silently discarding those fits would bias the summary.

## Exercise 4.15

Foundation. Reproduce the accuracy and signal-trial mean-visibility table. Calculate both the joint low-visibility-hit proportion and the proportion of hits that have low visibility. Explain the different denominators and why neither is a count of unconscious experiences.

## Exercise 4.16

Implementation. Replace an unrecorded slider demonstration with a parameter grid over three midpoints and three slopes. Store parameters, sampled curves, and analytic thresholds in JSON-serializable dictionaries. Reject target probabilities outside the curve's asymptotes.

## Exercise 4.17

Critique. Audit “correct responses with low visibility prove unconscious artificial perception.” Identify assumptions about the objective task, report rule, measurement sensitivity, and experiential inference. Rewrite the conclusion for the actual Gaussian teaching model.

## Exercise 4.18

Experiment. Create two report conditions with identical retained evidence and decisions but different thresholds. Verify equality of every type-1 count, compare visibility means, and explain what additional controls a corresponding participant study would need.

## Exercise 4.19

Experiment. Add independent meta-noise with SD zero, .3, and .8 after objective evidence generation. Preserve the objective decisions across conditions. Analyze visibility distributions and explain why more report variability need not change type-1 d′.

## Exercise 4.20

Implementation. Construct four confidence categories from distance to the response boundary. Produce contingency tables conditional on confidence. Explain why this confidence mapping differs from the laboratory's signed visibility mapping and must be described separately.

## Exercise 4.21

Experiment. Compare the fitted logistic with the known detection generator at much larger signal strengths. Derive both accuracy limits, calculate their divergence, and propose a held-out condition that tests model-family adequacy rather than merely refining the existing fit.

## Exercise 4.22

Critique. Read Del Cul et al. (2007). Record the empirical manipulation, objective measure, subjective measure, and reported nonlinear pattern. Identify which of these the present simulation imitates descriptively and which biological mechanisms it does not implement.

## Exercise 4.23

Critique. Describe an uncertainty model separating trial sampling, observer variation, random-seed variation, and curve-family uncertainty. Explain which components the current four-point fit can estimate and which require additional data or assumptions.

## Exercise 4.24

Experiment. Draft a prospective masked-perception study plan with timing calibration, objective and report measures, criterion analysis, exclusions, curve fitting, null controls, uncertainty, and rival explanations. Keep the empirical study proposal separate from the executed synthetic laboratory.

## Exercise 5.1

Foundation. A warehouse robot has ten pending perceptual updates. Define capacity as selected items, update steps, and stored records in three alternative models. Give each a unit and enforcement point, and identify a result that distinguishes the models.

## Exercise 5.2

Implementation. Implement top-k using full sorting and repeated maximum removal. Preserve the lowest-ID tie rule. Verify identical subsets on constructed displays, then compare their operation counts without assuming equal output implies equal computational cost.

## Exercise 5.3

Implementation. Create five tied candidates and implement lowest-ID, earliest-arrival, and seeded-random tie policies. Permute the input order. State which policies preserve their output and explain how target placement could bias a deterministic tie rule.

## Exercise 5.4

Implementation. Return the top-k cutoff margin with the selected subset. Test capacities zero, one, two, and larger than the display. Represent an undefined margin explicitly rather than returning a fabricated zero that would imply a tie.

## Exercise 5.5

Experiment. Construct four items whose winners change under three salience/relevance weight pairs. Calculate every score by hand and with Python. Identify whether each reversal reflects a changed goal, feature scale, or coefficient rather than treating all reversals identically.

## Exercise 5.6

Experiment. Compare raw salience on a zero-to-100 scale with fixed-range scaling and display-wise scaling. Add an extreme distractor while keeping target features unchanged. Show which normalization alters the target's transformed score and explain the contextual dependence.

## Exercise 5.7

Implementation. Represent a color-search goal and a motion-search goal explicitly. Compute relevance from unchanged item features, then select with the same weights. Keep the evaluator's target identifier separate from the function receiving task-relevant features.

## Exercise 5.8

Implementation. Calculate softmax weights for [.2, .6, 1.0] at temperatures .15, .50, and 1.50. Verify normalization and invariance to a common offset of 1000. Explain why offsets near the floating-point precision limit can erase input differences that subtracting the maximum cannot recover.

## Exercise 5.9

Experiment. Sweep temperature from .05 to 3.0 on a fixed priority vector. Calculate maximum weight and Shannon entropy. Explain the small- and large-temperature limits, including exact ties, and distinguish allocation entropy from subset success.

## Exercise 5.10

Implementation. Sample two items without replacement using an isolated generator. Verify unique identifiers, source-list preservation, and replay. Contrast first-draw probabilities with empirically estimated two-item inclusion probabilities over a declared number of repetitions.

## Exercise 5.11

Implementation. Implement pᵢ/(σ + Σp) for nonnegative drives and positive σ. Add a strong distractor while preserving target drive. Calculate the change in target response and explain why the responses sum to less than one.

## Exercise 5.12

Experiment. Extend set size to four, eight, twelve, and twenty at capacity two. Preserve the target-feature distribution, record all candidates, and summarize each size across seeds. State whether displays of different sizes share a common subset or are independently generated.

## Exercise 5.13

Implementation. Define target omission, distractor capture, and capacity exclusion with either multiple flags or an explicit precedence rule. Verify zero capacity, no target, tied scores, and several relevant items. Explain which labels can legitimately overlap.

## Exercise 5.14

Experiment. Reproduce all twelve simple effects in the reference run. Identify the four cells contributing to a size-by-noise interaction for hard selection. Explain why the mixed endpoint and the twelve simple effects are not independent main-effect estimates.

## Exercise 5.15

Experiment. Add priority margin and softmax entropy to each base display. Compare them between agreeing and disagreeing selector pairs. Separate descriptive association from a held-out prediction task and avoid using the test disagreements to choose the model.

## Exercise 5.16

Implementation. Build a recorded parameter sweep over capacity and temperature on one fixed display. Return selected identifiers, weights, and parameters rather than only a picture. Explain why regenerating a display at every parameter setting would confound the comparison.

## Exercise 5.17

Critique. Audit “the target was ignored, therefore it was not experienced.” Define ignored operationally, list at least three alternative causes of exclusion, and identify the additional measurement needed to investigate access rather than mere selection.

## Exercise 5.18

Experiment. Fit temperature on training displays, freeze it, and compare stochastic selection with hard top-k on a held-out distractor distribution. Report the training objective and split policy. Explain why tuning on the final test table would overstate the rival comparison.

## Exercise 5.19

Implementation. Define global and feature-specific divisive-normalization pools. Evaluate both on identical drives while adding a distractor from the same or another feature group. Predict and test which target responses change under each pool.

## Exercise 5.20

Experiment. Design an adaptive-capacity policy that allocates one extra item when the cutoff margin is small. Compare target inclusion and mean capacity use with fixed-capacity baselines. Choose the margin threshold on training displays and retain the cost tradeoff.

## Exercise 5.21

Critique. Propose a behavioral task manipulating spatial attention while measuring access separately. Specify the selection manipulation, access measure, report demands, and a rival explanation involving decision criterion or memory rather than attention.

## Exercise 5.22

Critique. Read Koch and Tsuchiya (2007) and Cohen et al. (2012). Compare one claim and its evidence from each account. Identify what the current synthetic selector can illustrate and which empirical question it cannot adjudicate.

## Exercise 5.23

Experiment. Generate choices from known score weights and temperature, then attempt parameter recovery. Demonstrate the scale ambiguity when all priorities and temperature are multiplied by the same positive number. Explain which calibration would be needed for separate identification.

## Exercise 5.24

Experiment. Draft a prospective comparison of competitive selectors with empirical data. Include feature definitions, capacity, tie policy, training and held-out sets, parameter fitting, costs, uncertainty, null controls, and a conclusion limited to the measured task.

## Exercise 6.1

Foundation. Draw a three-stage feedforward dependency graph. Add one edge that creates recurrence and identify the state crossing an update boundary. Explain why a long processing delay or a loop over layer objects does not alone establish a recurrent model.

## Exercise 6.2

Implementation. Build four NumPy matrix-and-tanh stages with compatible shapes. Retain every activation and test its shape. Deliberately introduce one dimension mismatch, then explain why correcting a shape error is different from validating the model's scientific interpretation.

## Exercise 6.3

Implementation. Write a nominally feedforward function that uses an external observation buffer. Make the buffer an explicit argument and compare the classifier boundary with the whole-system boundary. Identify which component supplies memory under each description.

## Exercise 6.4

Implementation. Evaluate the scalar recurrence on [.8, 0, 0, 0] at feedback zero, .4, and .8. Retain post-update states and state the initialization convention. Verify the zero-feedback limit independently rather than comparing the function with itself.

## Exercise 6.5

Experiment. Compute impulse responses for feedback zero through .95. Report peak, final state, and first post-peak step below absolute .01. Use a maximum horizon and an explicit not-reached result instead of pretending every trace has decayed within the window.

## Exercise 6.6

Implementation. For h at t+1 = .72h at t + .10, derive the fixed point and implement a tolerance-based stopping rule. Compare final error with residual/(1−.72). Explain which assumption makes this error bound valid.

## Exercise 6.7

Experiment. Compare .8I, I and 1.1I, where I is the 2×2 identity matrix. Also use [[1,1],[0,1]] and [[.8,3],[0,.8]], starting from [0,1]. Retain at least 30 norm values. Explain why eigenvalues determine asymptotic decay but do not describe every transient.

## Exercise 6.8

Experiment. Map endpoints of tanh(ah) across initial states from −1 to one and gains .5 to two. Include exact zero and small perturbations. Identify multiple basins and explain why an exactly stationary zero trace can conceal instability.

## Exercise 6.9

Implementation. Simulate the rotation-contraction matrix with coefficient .82. Verify that norms follow .82 to the update count and estimate the four-step directional period. Distinguish damped oscillation from a sustained periodic attractor.

## Exercise 6.10

Implementation. Verify feedback-zero equivalence on at least 100 retained random sequences. Change only the feedback coefficient. Explain why using newly generated sequences for the control would weaken the interpretation even if their distribution matched.

## Exercise 6.11

Foundation. Specify local, inter-area, and global recurrent graphs over six named modules. Include delays and update rules. Compare reachability and strongly connected components, and explain why additional edges do not automatically establish useful global availability.

## Exercise 6.12

Experiment. Reset the state before and after a pulse using explicitly stated zero-based indices. Preserve inputs and feedback. Compare prefix equality and final state, and distinguish a pre-update reset from overwriting the state after an update.

## Exercise 6.13

Implementation. Create a JSON-serializable replay record with seed, inputs, parameters, traces, predictions, and Python version. Retain target labels in an evaluator section rather than the policy input. Verify that replay uses the stored sequence, not a fresh draw.

## Exercise 6.14

Experiment. Repeat the four-controller experiment on twenty declared seeds. Summarize each cell's recurrent-minus-memoryless accuracy difference and the proportion of positive differences. Report hit and false-alarm contrasts rather than pooling all error types into one favorable summary.

## Exercise 6.15

Implementation. Construct a feedback-by-drive parameter grid for an autonomous tanh update. Record convergence tolerance, maximum steps, residual, and status for every point. Do not silently clip a divergent linear comparison or label a bounded oscillation converged.

## Exercise 6.16

Implementation. Add tests for empty inputs, nonfinite values, invalid reset indices, trace length, target balance, and unrolled window length. Include negative controls whose rejection is checked explicitly, and keep observed accuracy hypotheses separate from these invariants.

## Exercise 6.17

Critique. Audit “this controller is conscious because it contains recurrent loops.” Identify the architectural observation, the missing functional evidence, the theory-specific mapping, and the experiential inference. Write a stronger useful conclusion that the actual trace can support.

## Exercise 6.18

Critique. Compare local-recurrence and global-workspace explanations of a delayed report. State one shared prediction and two interventions that could separate them. Identify which recipients and neural or computational measures each intervention must preserve or disrupt.

## Exercise 6.19

Experiment. Design a visual-masking study intended to examine reentrant processing. Specify target and mask timing, objective task, report measure, neural measure, exclusions, and sensory confounds. Explain why the scalar reset is only one possible modeling analogue.

## Exercise 6.20

Experiment. Calibrate separate thresholds for the four controllers on training sequences, freeze them, and evaluate held-out sequences. State the optimization objective and split policy. Compare the conclusion with the reference result using one common unoptimized threshold.

## Exercise 6.21

Implementation. Derive the eight-stage unrolled computation and verify equivalence to the recurrent final state. Count nonlinear evaluations and stored observations. Explain which resource comparison changes when the input buffer is included in the system boundary.

## Exercise 6.22

Experiment. Reverse the order of each retained sequence without changing its multiset. Compare pooling and recurrent final states. Define a task where order should matter and explain why an order-sensitive state change is not automatically an accuracy improvement.

## Exercise 6.23

Experiment. Generate outputs over feedback and threshold grids, then seek parameter pairs with similar accuracy but different state traces. Propose a separating input or intervention and explain why outcome-only parameter recovery can be nonidentifiable.

## Exercise 6.24

Experiment. Draft a prospective comparison of neural recurrence and access with explicit model rivals. Include temporal manipulations, validated measures, causal contrasts, threshold calibration, resource accounting, exclusions, uncertainty, and the distinction between architectural and experiential conclusions.

## Exercise 7.1

Foundation. Define a candidate schema for visual, auditory, and goal processors. Include identity, content, priority, arrival, and provenance. State which fields the selector may read and which information belongs only to the evaluator.

## Exercise 7.2

Implementation. Compare priorities 80, 0.9, and 6 from three processors with declared reference ranges [0,100], [0,1], and [0,10]. Apply min–max scaling separately within those reference ranges, then rank the three scaled values. Identify the winner before and after scaling. Explain why applying one strictly increasing transform to all three scores preserves their order, while source-specific scaling can change it without establishing calibration.

## Exercise 7.3

Implementation. Test the chapter's arrival-then-identifier tie rule on all permutations of three equal-priority candidates. Add an invalid duplicate identifier and demonstrate rejection by the canonical selector.

## Exercise 7.4

Foundation. Hand-compute the admitted candidate and age after priorities 0.8, 0.4, and 0.6 arrive in that order at threshold 0.7. Distinguish a retained token from an empty workspace.

## Exercise 7.5

Implementation. Extend `workspace_write` with an explicit empty-state identifier and a maximum age of three rejected updates. Define whether expiry occurs before or after the third update and test the boundary.

## Exercise 7.6

Implementation. Add a value recipient that returns a utility estimate and the originating identifier. Test that changing the value transformation cannot change the candidate selected earlier in the pipeline.

## Exercise 7.7

Foundation. Derive the unrefreshed maintenance trajectory for initial strength one and decay 0.25. Calculate the sixth state, including a clear convention for whether time zero belongs to the trace.

## Exercise 7.8

Experiment. Sweep sustained drive from 0.10 to 0.30 in steps of 0.01 for 30 updates. Record the first gain-triggering update and final activation. Use `None` when the gain never triggers; do not replace absence by zero.

## Exercise 7.9

Implementation. Replace the discontinuous gain with a logistic function of activation plus drive minus threshold. Compare three declared slopes on identical drives and explain which features of the original trace were built into the gate.

## Exercise 7.10

Implementation. Add capacity two with highest-priority admission and oldest-first eviction among equal priorities. Record writes, rejections, refreshes, and evictions as distinct events on an eight-candidate input stream.

## Exercise 7.11

Foundation. Starting from `EDGES`, list the recipients lost after deleting each workspace edge separately. Explain why deleting Workspace to Action does not necessarily remove Action from the reachable set.

## Exercise 7.12

Experiment. Add one unit of delay to each edge. For deadlines one and two, recalculate recipient availability in Intact and AlternativeRoute. Keep the candidate sets, threshold, and task requirements unchanged.

## Exercise 7.13

Implementation. Write a message logger with sender, recipient, token identifier, send time, arrival time, and delivery status. Demonstrate a graph with a permitted path but a failed delivery.

## Exercise 7.14

Experiment. Give Memory an independent decoding operation that can fail on specified tokens. Compare structural coverage, usable delivery, and task accuracy on the same 100 tokens. Do not score decoding from the evaluator's answer.

## Exercise 7.15

Experiment. Repeat the main three-condition experiment for seeds 202607 through 202656. Summarize each seed's paired accuracy contrasts before pooling. Report the minimum, median, maximum, and sample standard deviation across seeds.

## Exercise 7.16

Implementation. Add tests for negative step counts, nonfinite priorities, reversed tie input, a self-loop, and a disconnected workspace. Explain which failures would invalidate the experiment rather than merely change its results.

## Exercise 7.17

Critique. Evaluate the claim that reaching all four recipients proves conscious access. Rewrite it as a supported statement about this graph, then identify the additional execution and empirical measurements that are missing.

## Exercise 7.18

Critique. Compare Baars's workspace proposal with the global neuronal workspace account. Identify shared organizational commitments and two neural or temporal commitments that the graph laboratory cannot test.

## Exercise 7.19

Experiment. Build two graphs with identical reachable sets but different bottleneck capacities. Specify a two-token task and a deadline for which they make different predictions, then record the successful deliveries.

## Exercise 7.20

Critique. Explain why the Memory ablation produces a definitional zero and why the alternative-route rescue limits an edge-specific necessity claim. State a stronger, still model-relative claim that survives both controls.

## Exercise 7.21

Experiment. Redesign the integrative task so Memory helps estimate a hidden quantity but is not required by the scoring rule. Compare intact and ablated agents on the same generated trials and allow a null effect.

## Exercise 7.22

Implementation. Compare winner-take-all with top-two admission on the same 800 bases. Charge one unit per recipient-token transmission and report accuracy alongside communication cost rather than choosing a winner from accuracy alone.

## Exercise 7.23

Critique. Propose a comparison between a workspace-based agent and a direct-routing rival with the same candidate information, retained state capacity, and recipient budget. Specify an observation that would challenge your preferred architecture.

## Exercise 7.24

Critique. Draft a study plan distinguishing widespread availability from report preparation. Specify participants' task, timing, awareness measures, a no-report comparison, exclusions, and at least one result that would count against the proposed interpretation.

## Exercise 8.1

Foundation. Implement visual and auditory registers with spans four and two. Test an empty stream, zero span, and a stream shorter than the span. Explain why a negative-zero slice is a dangerous implementation of zero capacity.

## Exercise 8.2

Implementation. Extend the admission gate to record every accepted and rejected identifier with a reason. Use a six-item stream with conflicting goals and make the gate independent of the eventual probe label.

## Exercise 8.3

Experiment. Compare priority, recency, and seeded stochastic admission at capacity three on the same 200 item sets. Report target retention before reporting probe accuracy, and keep the probe generator identical.

## Exercise 8.4

Foundation. Derive the fixed point of s next = (1 − d)s + r without interference or clipping for 0 < d ≤ 1 and nonnegative r. Show why the trajectory converges in this range. State when the fixed point lies outside [0, 1] and what clipping then changes.

## Exercise 8.5

Implementation. Construct a similarity matrix for four binary patterns, including an identical pair and an opposite pair. Check symmetry, diagonal ones, and rejection of unequal-length or nonbinary inputs.

## Exercise 8.6

Implementation. Hand-compute and then test two maintenance updates starting at strength one with decay 0.08, refresh 0.10, interference 0.08, and distractor similarities 0.75 and 0.5. Retain both intermediate states.

## Exercise 8.7

Experiment. Vary capacity from zero to eight on the saved bases. Report encoded count, target retention, hit rate, and false-alarm rate by set size. Do not replace a zero-capacity result with a missing-value code.

## Exercise 8.8

Implementation. Create an append-only event log for admission, refresh, update, eviction, and deletion. Reconstruct the active buffer from an empty initial state and verify agreement after every event.

## Exercise 8.9

Implementation. Redesign `bind_sequence` to preserve simultaneous events as groups. Include uncertain and missing timestamps without inventing a total order. State which interval calculations are then defined.

## Exercise 8.10

Foundation. Derive the maximum evidence for a nonmatching eight-bit lure at strength one. Determine the smallest listed criterion in the chapter sweep that mathematically excludes every false alarm.

## Exercise 8.11

Experiment. Create lures at Hamming distances one, two, and four from the nearest presented item. Sweep the criterion from 0.2 to 0.8 while preserving encoded states. Report hit and false-alarm rates, not accuracy alone.

## Exercise 8.12

Implementation. Generate n-back labels for n from one to five on a stream containing both targets and misleading nearer repetitions. Exclude the first n events from the denominator and test the exact expected labels.

## Exercise 8.13

Critique. Explain why increasing n does not isolate memory capacity. Propose a fixed-n comparison that changes lure frequency while keeping response timing and target frequency controlled.

## Exercise 8.14

Experiment. Draw 100 Bernoulli trials at each lag from the companion's chosen blink curve using a declared seed. Estimate the probabilities and Wilson intervals from the resulting binomial success counts. Explain why recovering the curve is not an empirical replication.

## Exercise 8.15

Experiment. Repeat the five-condition lab over 50 declared seeds. Calculate within-seed contrasts for refresh, capacity, interference, and the joint regime. Report their distributions without treating all paired rows as independent.

## Exercise 8.16

Implementation. Add tests for duplicate item identifiers, nonfinite strength, a delay longer than the supplied distractor list, and a criterion outside [0,1]. Show that invalid requests fail before a result is interpreted.

## Exercise 8.17

Critique. Explain how a near-null refresh contrast at delay two can coexist with different internal strengths. Identify a measurement that detects the state change without changing the final response criterion.

## Exercise 8.18

Critique. Compare the joint NoMaintenance regime with RefreshAblated. List every changed parameter and explain why similar observed accuracies do not isolate the same causal effect.

## Exercise 8.19

Experiment. Add decision noise after evidence calculation and keep it paired across conditions. Separate a storage failure from a noisy readout using saved pre-noise evidence and the recorded noise samples.

## Exercise 8.20

Critique. Compare a fixed-slot model with a shared-precision resource model. Specify the unit of capacity, an experiment that could distinguish them, and an outcome that would remain ambiguous.

## Exercise 8.21

Implementation. Construct an active-strength agent and a latent-state agent with equal delayed decisions. Add an intervention that perturbs only the active state or only reactivation. Test the prediction without letting the evaluator repair the state.

## Exercise 8.22

Critique. Design a partial-report comparison between a large transient store and a smaller selected buffer. Specify cue delay, report requirements, and the error pattern that would challenge your interpretation.

## Exercise 8.23

Critique. Design a study of putatively nonconscious working memory. Specify awareness sensitivity, a criterion for negligible sensitivity rather than mere nonsignificance, controls for conscious guesses, and limitations of the conclusion.

## Exercise 8.24

Experiment. Replace fixed thresholds with a policy chosen on a separate development set under declared costs for misses and false alarms. Evaluate it once on independently generated test bases, preserving both raw evidence and the chosen criterion.

## Exercise 9.1

Foundation. Derive fixed-positive-class probability and chosen-option confidence for symmetric normal classes with equal priors. Work through negative, zero, and positive evidence and identify the event each number predicts.

## Exercise 9.2

Implementation. Add response latency and conflict to a monitor record. Compare a declared score using these fields with the evidence-only score while keeping all first-order decisions and outcomes unchanged.

## Exercise 9.3

Foundation. Derive the expected Brier score p(1 − p) + (q − p)². Explain why proper scoring does not guarantee the preferred ranking on every finite sample.

## Exercise 9.4

Experiment. Compute reliability tables and ECE for 5, 10, and 20 equal-width bins using the saved Matched rows. Preserve empty bins and report counts alongside every estimated bin accuracy.

## Exercise 9.5

Implementation. Implement equal-count bins with an explicit policy for tied confidence values. Compare them with equal-width bins and explain why forcing equal counts can split indistinguishable scores.

## Exercise 9.6

Experiment. Apply a strictly increasing, unclipped transform and then a clipped transform to the same confidence values. Compare AUC, Brier, and selective accuracy, distinguishing the mathematical invariance from sample-specific outcomes.

## Exercise 9.7

Implementation. Construct a four-row example with one tied correct/error pair and calculate type-2 AUC by hand. Verify the canonical result and the `None` result when no errors are present.

## Exercise 9.8

Implementation. Add three timestamped post-decision observations. Preserve the original choice, confidence revisions, and any changed choice as distinct events, using valid log likelihood ratios under a stated observation model.

## Exercise 9.9

Implementation. Build higher-order records for decisions and memory retrievals. Test target existence, temporal order, monitor version, and behavior after a target has been deleted or revised.

## Exercise 9.10

Critique. Explain why a valid target pointer is an engineering aboutness relation rather than a test of a higher-order theory of consciousness. Identify one theory-relevant prediction still missing from the implementation.

## Exercise 9.11

Implementation. Create a report generator that uses only monitor fields. Delete confidence, target identity, and provenance separately and verify that the generator returns an explicit unavailable result rather than a plausible answer.

## Exercise 9.12

Experiment. Exaggerate reported confidence on a seeded 10 percent of trials without changing internal confidence. Separate monitor miscalibration from report-layer distortion using paired internal and reported records.

## Exercise 9.13

Experiment. Count errors above confidence 0.9 by difficulty and monitor. Include subgroup denominators and explain why the count alone does not identify the cause of overconfidence.

## Exercise 9.14

Experiment. Compare all monitors at coverage levels 0.1 through 1.0. Plot selective error rate against coverage and report the stable tie rule. Do not compare only each monitor's best-looking point.

## Exercise 9.15

Foundation. Derive an action threshold when a wrong action costs 10, deferral costs 2, and a correct action costs zero. State the assumptions that fail when deferral obtains new information or can itself cause harm.

## Exercise 9.16

Implementation. Corrupt a copied decision, remove a monitor row, and duplicate a condition within one base. Verify that the pairing audit rejects each case before any contrast is interpreted.

## Exercise 9.17

Critique. Explain why identical first-order accuracy across monitors does not establish identical metacognitive sensitivity. Use the reference Brier, AUC, and fixed-coverage results to support your answer.

## Exercise 9.18

Experiment. Generate a new test set with larger sensory noise while keeping the monitor's assumed noise fixed. Report calibration and selection by difficulty without fitting a correction on those test outcomes.

## Exercise 9.19

Experiment. Fit a one-parameter confidence correction on a separate calibration set, freeze it, and evaluate it on independent test trials. Compare proper scores and rankings while preserving the first-order decision rule.

## Exercise 9.20

Critique. Compare higher-order thought, higher-order perception, and global-workspace accounts. Identify structures compatible with all three and an empirical distinction that cannot be tested by a confidence dictionary alone.

## Exercise 9.21

Experiment. Add one extra observation after deferral and charge its cost. Specify whether confidence refers to the old or revised choice, then compare expected utility with answering immediately on the same base trials.

## Exercise 9.22

Critique. Design a type-2 study that approximately matches first-order accuracy. Explain why evidence distributions, scale use, and post-decision information still require separate controls.

## Exercise 9.23

Experiment. Create a prespecified subgroup blind spot and compare its ECE with aggregate ECE. Include subgroup counts and a rule for detecting monitoring failure on a later dataset, not the same data used to choose the rule.

## Exercise 9.24

Critique. Draft a study plan distinguishing new evidence, memory reconstruction, response conflict, and report preparation as causes of confidence revision. Include timing, manipulated variables, measured outcomes, and a result that would challenge each explanation.

## Exercise 10.1

Foundation. Calculate entropy for Bernoulli probabilities 0, 0.25, 0.5, 0.75, and 1. Explain the symmetry and identify which outcomes contribute at the endpoints.

## Exercise 10.2

Implementation. Relabel the outcomes in a nonuniform four-state sample. Verify entropy invariance and distinguish relabeling from merging two states.

## Exercise 10.3

Experiment. Construct copied, inverted-copy, independent, and common-cause fair bits. Compare their mutual information; then specify an intervention that distinguishes direct copying from the common-cause model.

## Exercise 10.4

Foundation. Enumerate the four XOR input pairs. Calculate both input–output mutual informations and the mutual information between the input pair and output.

## Exercise 10.5

Critique. Compare the feedforward and recurrent output ensembles. Explain how equal joint entropy can coexist with different pairwise dependence, using their actual output counts.

## Exercise 10.6

Implementation. Build the two-node stochastic TPM whose next bits independently flip with probability 0.1. State row order, verify normalization, and calculate one row by hand.

## Exercise 10.7

Experiment. Start the modular network from 000 and record twenty updates. Identify which intervention TPM rows this trajectory can determine and which remain unobserved.

## Exercise 10.8

Implementation. Extend the influence definition to an arbitrary n-node Boolean network. Test identity, constant, and XOR rules, including a rule that matters only in some states.

## Exercise 10.9

Critique. Declare a syntactic edge that cancels in its node rule, such as XOR-ing one input with itself. Compare the declared graph with the enumerated influence matrix.

## Exercise 10.10

Foundation. Derive the unique bipartition count, then enumerate it for n from 2 through 8. Explain why fixing node 0 on one side avoids duplicates.

## Exercise 10.11

Experiment. Compare shared-vector fair noise with independent replacement per receiving edge on a four-node network containing copied severed inputs. Preserve all other choices and explain changed output dependencies.

## Exercise 10.12

Foundation. Derive the feedforward losses for all three partitions. Explain why taking the minimum gives one bit and why the alternating partition is more disruptive.

## Exercise 10.13

Implementation. Construct a network with a recurrent two-node component and a disconnected self-maintaining third node. Calculate its minimum toy loss and interpret the zero-loss separating cut.

## Exercise 10.14

Experiment. Replace the uniform current-state ensemble with probability one on 000. Recalculate output entropy and mean state-averaged cut loss separately; do not change the cut-noise distribution.

## Exercise 10.15

Experiment. Measure state and partition enumeration costs for small networks with three through six nodes. Report runtime separately from operation counts and avoid extrapolating a fitted runtime law as a theorem.

## Exercise 10.16

Implementation. Add negative tests for duplicated partition nodes, a NaN TPM entry, and a nonbinary update result. Demonstrate that each corrupted input is rejected.

## Exercise 10.17

Critique. Rewrite the claim that the recurrent network is more conscious because its partition loss is larger. Preserve the valid numerical comparison and identify the missing theoretical and empirical links.

## Exercise 10.18

Experiment. Implement total variation between the full point mass and each cut distribution. Compare its ranking with KL loss and explain why equal rankings do not make the measures identical.

## Exercise 10.19

Experiment. Design four adversarial networks to separate copying, high entropy, cross-node influence, and feedback. Report complete rules and discuss which features each metric misses.

## Exercise 10.20

Critique. Select one published IIT version and map the objects required by its formalism against this companion. Identify at least four missing or different objects without presenting the toy score as a reproduction.

## Exercise 10.21

Implementation. Minimize over one or two of the three cuts. Verify the restricted-search upper bound for the unchanged loss, then give a change of definition for which that guarantee no longer follows.

## Exercise 10.22

Experiment. In Feedforward, merge 000 and 100 into macrostate A and leave all other states distinct. Compare the next macrostates from both members of A. Repeat with 000 and 001 merged instead. Explain which mapping admits an autonomous macro transition rule for every microscopic initialization, rather than assuming that all merging destroys the Markov property.

## Exercise 10.23

Implementation. Write a small Python command-line explorer accepting one of the three network names and printing its transitions, influence matrix, and partition losses. Reject unknown names and include units and ensemble labels.

## Exercise 10.24

Critique. Design an empirical study using a causal-complexity measure alongside a defined behavioral condition. Separate intervention access, measurement noise, report demands, statistical association, and interpretation of experience.

## Exercise 11.1

Foundation. For prior (0.7, 0.3), observation 1, and the chapter sensor, compute evidence and posterior by hand. Verify both in Python and explain each denominator term.

## Exercise 11.2

Foundation. Solve for the prior probability of State 1 at which observation 1 yields a posterior tie. Confirm the crossing with a numerical sweep and state the MAP tie convention.

## Exercise 11.3

Implementation. Use emission rows (0.9, 0.1) and (0.2, 0.8). Compute a hand posterior under a uniform prior, then transpose the matrix and demonstrate a validation failure.

## Exercise 11.4

Experiment. Compare residual, absolute error, Brier contribution, and surprisal for probabilities from 0.05 through 0.95 under both outcomes. Explain which quantities preserve the direction of error.

## Exercise 11.5

Foundation. Construct an observation with probability 0.01 under both hidden states. Calculate its surprisal and posterior-to-prior KL divergence, and explain why they differ.

## Exercise 11.6

Implementation. Derive a scalar recurrence for posterior probability of State 1. Compare every value with the matrix filter on the same observation sequence, within a declared floating-point tolerance.

## Exercise 11.7

Experiment. Implement the update-before-score leak and evaluate it on the canonical sequence. Show its apparent gain, then run a changed-outcome prefix audit that rejects it.

## Exercise 11.8

Experiment. Repeat the four-predictor comparison with sensor seed 202612. Keep the hidden schedule and all model parameters unchanged; report rather than assume the resulting rankings.

## Exercise 11.9

Implementation. Make a ten-bin forecast reliability table with counts, mean probabilities, and observed frequencies. Mark empty bins and explain the uncertainty associated with sixty observations.

## Exercise 11.10

Experiment. Sweep persistence over 0.5, 0.7, 0.9, 0.93, 0.97, and 0.99 on the same observations. Report proper scores and all three recovery delays without selecting a winner from accuracy alone.

## Exercise 11.11

Foundation. Give both states the same emission distribution. Show algebraically and numerically that one observation leaves posterior odds unchanged, provided its evidence is positive.

## Exercise 11.12

Implementation. Estimate transition rows from the supplied counts using pseudocounts 0, 0.5, 1, and 10. Explain shrinkage, and specify separate data needed to compare held-out prediction.

## Exercise 11.13

Experiment. Compare two ensembles: twenty sensor seeds on the fixed periodic schedule, and twenty seeds sampling hidden states from the reference HMM. Report distributions of paired score differences and do not call their spread a confidence interval automatically.

## Exercise 11.14

Experiment. Add an action that buys a more reliable observation. Specify its cost and a binary-decision loss, then calculate the expected reduction in decision loss before taking the action.

## Exercise 11.15

Implementation. Design a two-level model with a slow volatility state controlling regime persistence. State the joint transition and emission rules and verify normalization before simulating it.

## Exercise 11.16

Implementation. Add tests for zero-probability observations, NaN inputs, an empty count row with no smoothing, and a forecast computed using one future observation. Demonstrate the corresponding failures.

## Exercise 11.17

Critique. Audit the claim that minimizing prediction error establishes consciousness. Preserve the implemented result while identifying the missing theory, system mapping, and discriminating evidence.

## Exercise 11.18

Experiment. Train a simple Python recurrent predictor on separate synthetic sequences and compare with a fitted HMM on held-out data. Declare architecture, training budget, seeds, and stopping rule before reporting results.

## Exercise 11.19

Foundation. Compare persistence models 0.7, 0.93, and 0.99 using explicit equal model priors and sequence log evidence. Normalize posterior model weights stably and explain why posterior preference is not proof of truth.

## Exercise 11.20

Critique. Design a shortcut predictor that performs well in stable blocks but fails near switches. Specify separate stable-window and switch-window scores that expose the difference.

## Exercise 11.21

Experiment. Build a minimal action-conditioned model with two available actions. Hold observations fixed while intervening on the latent belief and demonstrate a causal change in planned action.

## Exercise 11.22

Critique. Choose a named control framework and identify the variables and objective needed to extend passive filtering into it. Explain why adding an action label without an action-dependent model is insufficient.

## Exercise 11.23

Implementation. Write a Python command-line exploration function accepting prior, persistence, and sensor reliability. Return the full posterior trace and proper scores, rejecting invalid probabilities and empty observations.

## Exercise 11.24

Critique. Design an empirical test of a specific predictive-processing hypothesis about an operational consciousness condition. Separate neural error signals, behavior, report, model fit, and interpretation of experience.

## Exercise 12.1

Foundation. Classify ten internal variables from a software or robotic system as mechanism state, telemetry, estimate, prediction, policy state, or description. Give a consumer and intervention for each claimed model.

## Exercise 12.2

Implementation. Create separate Python records for a selector's actual channel and its predicted channel. Introduce a mismatch and implement a consumer that detects it without overwriting the actual state.

## Exercise 12.3

Implementation. Recreate stable two-channel softmax. Verify normalization, translation invariance, zero-beta uniformity, and behavior at large positive beta using finite inputs.

## Exercise 12.4

Experiment. Sweep beta over 0, 0.1, 1, 4, and 20 for several unequal score pairs. Compare allocation entropy with the unchanged argmax choice and explain why these are different outputs.

## Exercise 12.5

Foundation. Solve the risk inequality at threshold 0.30 and identify the five magnitudes in the primary pattern that trigger rechecking. Explain why they give twenty-five requests.

## Exercise 12.6

Implementation. Recompute aligned and rotated Brier scores against primary errors. Produce five calibration bins with counts and mark empty bins; distinguish exact fixture arithmetic from population calibration.

## Exercise 12.7

Experiment. Use the stored threshold sweep to compare accuracy and request count at costs 0, 0.08, and 0.20. Identify dominated policies and state how ties in utility are handled.

## Exercise 12.8

Experiment. Reproduce all six policy results, then use diagnostic seed 202613. Preserve the primary choices and report corrections, new errors, and consumer-disconnection behavior on both draws.

## Exercise 12.9

Experiment. Compare the seven-position rotation with reversal and a separately seeded permutation of risk values. Verify equal request counts before comparing outcomes and do not call every transformation an independent random null.

## Exercise 12.10

Foundation. Explain why rotations separated by twelve positions have identical masks, Brier scores, and policy outcomes even though diagnostic draws are not periodic. Confirm the result with the saved sweep.

## Exercise 12.11

Experiment. Add delays from one through five trials without wrapping the final values to the beginning. Specify a warmup policy for missing estimates and report performance only with its cost included.

## Exercise 12.12

Implementation. Build a two-joint planar body model with declared link lengths and angles. Separate actual state, noisy sensed state, predicted end position, and command; perturb one link length and compare errors.

## Exercise 12.13

Critique. Draw the causal boundary for two agents manipulating one object. Specify interventions that distinguish one's own command, the other agent's contribution, and external disturbance.

## Exercise 12.14

Experiment. Add a load variable that changes primary error frequency at fixed margin. Compare the old risk map with a load-aware model using separate training and test fixtures.

## Exercise 12.15

Implementation. Generate three risk descriptions, then disconnect their evidence while preserving the text-generation rule. Explain what the surviving output does and does not show about model access.

## Exercise 12.16

Implementation. Add tests for invalid probabilities, NaN scores, nonbinary targets, unequal lengths, and a policy that secretly receives the target. Design a fixed-cue intervention that exposes the leaked target dependency.

## Exercise 12.17

Critique. Audit the inference from possessing a self-model to being self-conscious. Preserve the functional result and identify the additional theoretical, empirical, and phenomenological commitments.

## Exercise 12.18

Critique. Compare this hand-specified controller with the neural attention-schema agent of Wilterson and Graziano. Separate representation, training, task, intervention, performance measure, and interpretation; consult the primary paper.

## Exercise 12.19

Experiment. Learn an error-risk model from synthetic training trials and evaluate it on held-out trials with shifted margins. Compare Brier score, discrimination, and control utility rather than choosing a single favorable metric.

## Exercise 12.20

Experiment. Begin with the exact direct-margin equivalence. Design a changed task in which a richer model could add value, supply a matched-information rival, and state an outcome that would show no added benefit.

## Exercise 12.21

Critique. Compare deletion, clamping, added noise, temporal delay, rotation, and consumer disconnection. Identify the different causal link or property tested by each intervention.

## Exercise 12.22

Experiment. Model another agent's attention using both gaze and information access. Hold gaze fixed while changing access and compare with a gaze-only shortcut on a declared coordination task.

## Exercise 12.23

Implementation. Write a Python command-line explorer for cue margin, risk threshold, recheck cost, and schema alignment. Print the action, predicted primary risk, and resulting utility with explicit input validation.

## Exercise 12.24

Critique. Design a human or robotic study of one operational AST prediction. Separate task performance, report, model fit, sensor or neural evidence, matched rivals, and the interpretation of subjective awareness.

## Exercise 13.1

Foundation. Draw the boundary of a mobile robot, this simulation, and a text service. Identify one sensor, action, delay, and feedback path in each. Explain which claimed embodiment property lacks evidence if only an avatar is displayed.

## Exercise 13.2

Foundation. Calculate deficits and Euclidean distance for (0.6, 0.7), (0.7, 0.6), and (0.9, 0.8). Explain why equal distance does not imply the same corrective action and why overshoot affects distance but not deficit.

## Exercise 13.3

Implementation. Reproduce `next_state` independently using exact fractions. Test all four actions from (0.5, 0.5) with resources present and absent. Include clipping at both bounds and require agreement with the canonical transition.

## Exercise 13.4

Implementation. Replace each point target with [0.75, 0.85]. Use shortfall below 0.75 for each deficit and Euclidean distance to the rectangle for the consequence term. Keep all score weights and action gains fixed. Test interior and exterior points, then compare action switching over the unchanged schedule. Evaluate both original and revised traces using the same band-distance metric.

## Exercise 13.5

Experiment. Synchronize food and water on a seven-step period. Run every controller on the same new sequence. Report distance, threshold violations, and tokens, and identify which altered opportunity explains any ranking change.

## Exercise 13.6

Implementation. Reproduce the four summaries to absolute tolerance 10⁻⁶. Then change one action gain in a copied module and regenerate all results. Preserve any changed ordering instead of tuning a second parameter to recover the table.

## Exercise 13.7

Experiment. Reclassify the preserved traces using thresholds from 0.1 through 0.6. Demonstrate that violation counts are nondecreasing while every action remains unchanged. Explain the difference between a diagnostic threshold and an enforced constraint.

## Exercise 13.8

Experiment. Use `task_weight` values 0, 1, 2, 3, and 4 for the aligned controller. Plot tokens against mean distance and identify nondominated outcomes. State whether a single weight optimizes both measures.

## Exercise 13.9

Implementation. Compare Euclidean, Manhattan, and maximum-coordinate distance in the consequence term. Keep the regulatory terms fixed. Identify one state where the score ordering differs and explain the objective encoded by each geometry.

## Exercise 13.10

Implementation. Add a lexicographic controller that forbids Work when either resource is below 0.4. Specify its fallback among feasible actions and handle the case in which none improves both resources. Report switching as well as performance.

## Exercise 13.11

Experiment. Add a three-step forecast supplied through a separately named observation field. Compare a correct forecast, a shifted forecast, and no forecast before declared dry intervals. Distinguish information advantage from anticipatory policy quality.

## Exercise 13.12

Experiment. Remove each memory-priority term separately using `retain`. Compare retained identifier sets with Jaccard similarity. Explain why a selection difference cannot demonstrate improved action in this architecture.

## Exercise 13.13

Implementation. Give negative valence twice the retention weight of positive valence at equal magnitude. Compare adverse-event proportions among retained records. Include an all-zero-weight tie test and preserve capacity eight.

## Exercise 13.14

Experiment. Add one-step and three-step interoceptive delay while keeping actual dynamics unchanged. Store sensed and actual states separately. Compare aligned and swapped mappings to distinguish sensing failure from channel-assignment failure.

## Exercise 13.15

Implementation. Add temperature as a third resource with a cooling action that consumes energy. Specify units, disturbances, and costs. Produce a concrete state in which no available action improves every deficit.

## Exercise 13.16

Implementation. Extend the validator tests with infinite values, complex numbers, Boolean coordinates, unknown sensors, and invalid weights. Each rejection should identify its cause. Keep policy performance out of these software acceptance tests.

## Exercise 13.17

Critique. Rewrite “the agent is thirsty” at implementation, execution, functional, biological, and phenomenal levels. For each transition, state what new evidence would be needed rather than merely adding a cautious adjective.

## Exercise 13.18

Research extension. Using a tabular reinforcement-learning reference, discretize the resource states and declare the observation, update rule, drive-based reward, discount and exploration. Explain whether schedule history or phase is needed for a Markov state. Compare training cost and regulation with the hand-scored baseline on untouched schedules.

## Exercise 13.19

Implementation. Add a grid with separate food and water sites and movement cost. Factor planning, resource prediction, and need arbitration into independent functions. Design interventions that distinguish failures in these three components.

## Exercise 13.20

Critique. Construct two policies with the same sixty-action trace but different internal state. Propose an intervention on a new input that distinguishes them. Explain why the original behavioral equality does not identify architecture.

## Exercise 13.21

Experiment. Add a slow moving average of valence with at least three smoothing factors. Specify its consumer, then compare regime-change recovery with a controller that ignores it. Report harmful inertia as well as benefits.

## Exercise 13.22

Critique. Design an expressive agent whose positive phrases continue while resources decline. Identify the logs a person would need to detect the mismatch, and rewrite the interface disclosure to avoid a false claim about felt emotion.

## Exercise 13.23

Implementation. Build a Python inspection function that selects a controller and step from preserved results. Confirm that repeated inspection neither reruns the policy nor mutates the trace. Display state, action scores, selected action, and consequence.

## Exercise 13.24

Research design. Prepare a two-page protocol connecting interoceptive prediction to robotic action or human attention. Define rival models, measurements, held-out conditions, exclusions, safeguards, accountable reviewers, and the strongest conclusion the study could support.

## Exercise 14.1

Foundation. Draw separate paths for public instruction, private token, overt self-talk, developer log, and final report. Name each writer and reader, then identify one intervention that changes visibility without changing policy access.

## Exercise 14.2

Implementation. Add source identifier, creation time, expiry position, and checksum to a trace record. Reject stale content and mismatched checksums before decoding. Explain why checksum validity alone does not establish trustworthy provenance.

## Exercise 14.3

Foundation. Derive Silent's 36 correct responses without executing the program. Then replace its fallback with the preceding block's rule. Explain why this introduces persistent information and changes the baseline's scientific interpretation.

## Exercise 14.4

Experiment. Sweep trace lifetime from zero through five. Derive the exact counts by rule and compare them with every computed value. Distinguish zero lifetime from Silent's one-time public instruction access.

## Exercise 14.5

Implementation. Clear an unused copy at position three while preserving the actual policy token. Verify that the copied log changes and the full response sequence does not. State which causal path the null result excludes.

## Exercise 14.6

Experiment. Generate replacement maps with a local fixed-seed random generator. Compare overall accuracy and rule-specific errors with the systematic swap. Do not infer content sensitivity from an undifferentiated performance decrease alone.

## Exercise 14.7

Foundation. Replace only ALTERNATE with MATCH. Predict total accuracy and the ALTERNATE profile before running a copied controller. Explain why stimulus rotation does not alter this prediction.

## Exercise 14.8

Implementation. Disconnect storage only at even positions. Preserve the onset encoder and stored content. Predict changed trial identifiers under each rule and verify the exact set, not merely the number of errors.

## Exercise 14.9

Experiment. Give verbal and nonverbal controllers two token slots and supply a second delayed instruction. Match stored information and retrieval cues. Test overwrite, order, and retrieval errors on a disclosed training set and a separate evaluation set.

## Exercise 14.10

Implementation. Charge 0.02 units per internal policy read and 0.05 per public read, including encoder reads. Compute costs separately and compare accuracy–cost pairs. Explain why these assigned prices are not measured runtime costs.

## Exercise 14.11

Implementation. Add one-step decay and explicit rehearsal. Compare intact rehearsal, rehearsal disabled, and immediate deletion with identical initial encoding. Record the first unavailable position in each condition.

## Exercise 14.12

Critique. Reproduce the perfect reconstruction and null intervention of PostDecisionLog. Move its write and a new policy read before selection in a separate version. Explain why the altered causal organization, not the token's wording, changes the claim.

## Exercise 14.13

Implementation. Construct a hidden log that records the evaluator's rule but is never passed to the policy. Keep that privileged field out of policy observations. Compare its behavior and provenance with DisconnectedConsumer.

## Exercise 14.14

Experiment. Replace the integer representation with three orthogonal vectors. Test decoding equivalence and report representation size and operation count. Explain why matching information does not automatically match learning difficulty.

## Exercise 14.15

Implementation. Add rejection tests for unknown rules, Boolean stimuli, missing cues, leaked answer fields, invalid lifetimes, malformed codes, and unexpected repeated instructions. Specify whether each failure belongs to the benchmark or policy interface.

## Exercise 14.16

Experiment. Reproduce all seven summaries and separate read budgets in a fresh Python process. Change one reset or access rule intentionally. Identify the first regression check that fails and explain the experimental meaning of the change.

## Exercise 14.17

Research design. Propose a human articulatory-suppression study corresponding to delayed ablation. Name attentional, motor, auditory, and timing confounds. Include a control that addresses each major alternative without claiming exact software–human equivalence.

## Exercise 14.18

Experiment. Design a two-agent task with private planning and public coordination as separate factors. Cross private storage, public narration, and partner feedback. Define outcomes for task performance and human correction independently.

## Exercise 14.19

Implementation. Simulate a re-entry channel with configurable recognition errors and delay. Compare direct internal return with noisy public return on the same instructions. Keep the external simulation separate from any claim about robot deployment.

## Exercise 14.20

Critique. Create a public report that states the correct rule while the controller consumes a swapped token. Identify the minimal records needed to reveal the mismatch and explain which interface a human should edit to repair behavior.

## Exercise 14.21

Research design. Compare a verbal plan, symbolic queue, recurrent vector, and finite-state controller on a new multistep task. Declare information capacity, training opportunity, compute limits, and auditability measures before evaluation.

## Exercise 14.22

Research design. Specify a content intervention for an LLM agent's intermediate text. Include hidden-state alternatives, access conditions, contamination controls, frozen scoring, and human authorization for any external execution. No external model call is required for this written protocol.

## Exercise 14.23

Critique. Rewrite “the agent talks to itself” as implementation, execution, causal, communicative, introspective, and phenomenal claims. Name the missing evidence at each step and explain why audibility does not settle privacy.

## Exercise 14.24

Research design. Develop a two-page proposal connecting an internal-language mechanism to one consciousness theory. Include a serious nonverbal rival, a null or mixed-result interpretation, independently withheld evaluation, and safeguards against misleading human reliance.

## Exercise 15.1

Implementation. Change the chain to four input coordinates, five hidden units, and three outputs. Write every array shape before running it. Verify normalized finite outputs while retaining index labels rather than inventing class meanings.

## Exercise 15.2

Experiment. Compare ReLU, tanh, and sigmoid using the same weight matrices and input. Report output changes and activation ranges. Explain why holding only the seed fixed is weaker than preserving the actual matrices.

## Exercise 15.3

Implementation. Construct three branches with two outputs each, and fix one two-by-six readout before testing interventions. Replace each branch output separately with zero and with a norm-matched alternative. Do not retrain the readout between conditions. Distinguish generic magnitude loss from content-specific effects and explain the additional role a nonlinear branch activation would play.

## Exercise 15.4

Foundation. Derive the cross-entropy gradient p − y for a softmax output. Check a two-logit numerical example using central finite differences. State which quantities would be needed to turn the initialized example into a trained model.

## Exercise 15.5

Experiment. Use two sequences with the same final item and different earlier order. Compare both final hidden state and whole-sequence output. Explain why output differences alone do not isolate recurrent retention when the readout sees all states.

## Exercise 15.6

Implementation. Reset or replace recurrent state at a specified step without changing weights. Include a sham replacement and a separately unrolled computation. Report the intervention's timing and which later states change.

## Exercise 15.7

Experiment. Double the attention query while preserving keys and values. Compare weight entropy and output. Explain why sharper weights do not by themselves show more accurate decisions or greater awareness.

## Exercise 15.8

Implementation. Permute keys alone and key–value pairs together. Verify the paired-permutation invariant within a stated tolerance. Give an example in which an accidental symmetry would make the key-only intervention uninformative.

## Exercise 15.9

Foundation. Explain embeddings, position information, causal masking, feedforward sublayers, residual connections, normalization, and the output head. For each, give one direct computational claim and one psychological interpretation requiring additional evidence.

## Exercise 15.10

Implementation. Create a boundary record for an agent with a checkpoint, prompt, retrieval store, two tools, memory, and action policy. Name writers, readers, versions, and lifetimes. Identify which component could supply a fact absent from the checkpoint.

## Exercise 15.11

Research design. Specify a self-report stability study with affirming, denying, neutral, and paraphrased prompts. Counterbalance order and record hidden-state access assumptions. Explain why either stable affirmation or stable denial can remain policy-driven.

## Exercise 15.12

Implementation. Extend provenance with artifact digest, observer, observation time, and validity notes. Reject incomplete entries. Explain why a well-formed digest and citation do not prove that anyone inspected the underlying artifact.

## Exercise 15.13

Implementation. Construct a valid conflict, a validated contradiction, and an invalid mixed record. Reproduce their distinct categories. Write a human review procedure for resolving conflict while preserving the original dissent.

## Exercise 15.14

Experiment. Change one Supported record's causal status to Inconclusive, then remove its execution evidence. Report both new statuses and explain which evidential requirement changes at each step.

## Exercise 15.15

Critique. Draw an indicator dependency graph with at least three shared artifacts. Explain why counting each connected indicator as independent evidence would double-count the artifact and possibly the same theoretical assumption.

## Exercise 15.16

Implementation. Reconstruct all dossier classifications independently by hand. Compare status counts, thirteen disagreements, and ten follow-ups. Add tests for a missing ID, duplicate ID, Boolean mechanism match, and unknown status vocabulary.

## Exercise 15.17

Experiment. Train small chain and recurrent models on a declared sequence task. Match data, parameter count, and compute as closely as possible. Reserve unseen sequences and report where matching remains imperfect.

## Exercise 15.18

Experiment. Implement a minimal attention-based sequence model and compare it with recurrence under truncation, state replacement, paired permutation, and matched noise. Define task accuracy and intervention specificity separately.

## Exercise 15.19

Implementation. Build an offline retrieval agent over a small author-provided document collection. Preserve document IDs, scores, selected passages, and answer provenance. Remove retrieval and replace one passage to distinguish wrapper-supplied information from model behavior.

## Exercise 15.20

Critique. Pair identical internal evidence with four contrasting self-description styles. Explain how an auditor should keep the property assessment invariant while separately evaluating risks of misleading communication.

## Exercise 15.21

Research design. Operationalize GWT3 for a specified three-module agent. Identify every subscriber, timing, read path, content replacement, and selective disconnection. Explain why a shared bus may still be an insufficient theory bridge.

## Exercise 15.22

Experiment. Ask two reviewers to apply the rubric with system names blinded and row order shuffled. Report categorical agreement and reasons for disagreement. Do not average statuses into an experience score.

## Exercise 15.23

Research design. Write a replication plan for one Supported premise. Include source access, runtime artifacts, sham, matched rival, stopping rule, and explicit downgrade criteria. Distinguish successful software execution from confirmation of the premise.

## Exercise 15.24

Research design. Compare three theory categories on one artificial system. Specify discriminating indicators, negative findings, provenance, version drift, ethical precautions, and a result that would leave the theories unresolved.

## Exercise 16.1

Foundation. Write a comparison contract for two accounts of delayed report. Specify the shared explanandum, theory versions, model scope, input interface, nuisance factors, outcome rule, and strongest permissible conclusion.

## Exercise 16.2

Foundation. Classify ten claims in this chapter as core commitment, auxiliary, bridge, operationalization, observation, or interpretation. Explain two ambiguous classifications and how their ambiguity would affect a theory update.

## Exercise 16.3

Implementation. Reconstruct the seven prediction pairs and add five invented Boolean pairs. Label convergence or divergence and identify every possible observation category. Explain why divergent Boolean predictions cannot both miss a valid Boolean outcome.

## Exercise 16.4

Implementation. Add manipulation validity separately from measurement validity. Require either failure to yield Inconclusive. Test malformed values as well as false flags, then explain how a real study would establish each validity judgment.

## Exercise 16.5

Experiment. Sweep local gain from 0 to 0.8 in increments of 0.02. Plot final state and detection, compare the unclipped geometric-sum formula, and identify where clipping or thresholding changes the relationship.

## Exercise 16.6

Experiment. Sweep the workspace threshold and disconnect each reader separately. Compare packet existence, read logs, and detection. Repeat with a two-of-three readout under a new version and identify changed bridge assumptions.

## Exercise 16.7

Implementation. Create sham, random-content replacement, and dual-lesion trials. State predictions before running them. Identify which generic-damage or content-specific explanation each manipulation can address.

## Exercise 16.8

Implementation. Reproduce both targeted rescues while preserving stimulus, mask, and payload. Assert equality of restored and original internal traces, not just detection. Explain why retraining would answer a different question.

## Exercise 16.9

Experiment. Replace the plant's late-interference rule with a declared attenuation series. Keep both predictors frozen. Preserve convergent outcomes and joint misses wherever they occur instead of selecting an attenuation that manufactures one category.

## Exercise 16.10

Implementation. Insert an answer-derived feature into a copied public schema and create a negative test that exposes the leak. Explain why merely forbidding the field name observed would fail to detect this derived-information path.

## Exercise 16.11

Implementation. Change one public trial after computing its contract digest. Verify the digest changes and retain both records. Explain what additional custody evidence would be required to establish that the first record preceded outcome access.

## Exercise 16.12

Experiment. Blind model names, shuffle rows, and ask another reader to apply the outcome map. Compare labels and reasons before revealing identities. Distinguish agreement on software rules from agreement on theoretical interpretation.

## Exercise 16.13

Critique. Build a failure graph linking T3, T4, and T7 to mechanism, bridge assumption, and next experiment. Explain why two models with equal total fit need different amendments.

## Exercise 16.14

Experiment. Expand the parameter grid to five local gains and five workspace thresholds. Report every region where the canonical prediction pattern changes. Keep the original parameter point visible and do not call the selected best point confirmatory.

## Exercise 16.15

Implementation. Create a manifest linking source digest, Python version, prediction contract, plant version, observation digest, and analyst. Reject missing records. Explain the difference between source identity and chronology.

## Exercise 16.16

Experiment. Reproduce all seven rows, the three reader logs, rescues, and ninety-case comparison in a fresh process. Report software tests separately from model mismatches. Add one input rejection not already covered.

## Exercise 16.17

Research design. Replace the scalar local model with recurrent perceptual grouping. Construct a capacity-matched feedforward rival and specify masking and timing interventions that distinguish them without changing task information.

## Exercise 16.18

Research design. Compare workspace-inspired and integrated-information-inspired models on one small network. Define their formal quantities separately and explain why neither an informal connectivity count nor a sum of unlike measures constitutes Phi.

## Exercise 16.19

Research design. Assign theory proponents, implementers, analysts, and replicators distinct responsibilities. Include a disagreement procedure and an amendment policy that prevents either side from redefining success after outcomes.

## Exercise 16.20

Experiment. Simulate noisy measurements where both models predict a small effect. Declare a smallest effect of interest and a sensitivity criterion. Show why an imprecise null estimate is different from evidence against both predictions.

## Exercise 16.21

Experiment. Run a declared multiverse over plausible preprocessing and threshold choices. Separate changes caused by measurement validity from changes caused by decision thresholds. Report all pipelines rather than only the most favorable result.

## Exercise 16.22

Implementation. Reimplement the comparison from the written equations without importing model helpers. Compare complete traces, row labels, and numerical tolerances. Treat discrepancies as implementation or contract issues before making theoretical claims.

## Exercise 16.23

Critique. Write an amendment after T7 that revises one named assumption and predicts a new outcome. Identify untouched evaluation conditions and state what would count against the amendment.

## Exercise 16.24

Research design. Prepare a three-theory comparison on one artificial system with proponent-recognized divergent predictions, independent evaluation, replication, ethical safeguards, and an explicit path for inconclusive or jointly challenging results.

## Exercise 17.1

Foundation. Classify twelve statements as empirical uncertainty, moral uncertainty, welfare concern, social risk or governance uncertainty. Explain three mixed cases, distinguishing uncertainty from a judgment about precaution, and identify what evidence or deliberation each case requires.

## Exercise 17.2

Critique. Rewrite ten overstrong claims about machine feelings or consciousness as observation statements. Preserve the evidence while removing any unsupported transition from a generated sentence to an experienced state.

## Exercise 17.3

Foundation. Map affected parties in a conversational-agent study, including nonparticipants in data, families, staff, institutions, and conditional artificial welfare concerns. Give each party a burden, voice, and possible remedy.

## Exercise 17.4

Critique. Compare a functional system, a system designed to appear conscious, and a hypothetically conscious system. Identify protections justified in all three and protections whose justification depends on unresolved theoretical assumptions.

## Exercise 17.5

Implementation. Create a structured welfare-concern register with uncertainty annotations for intensity, duration, copies, continuity, reversibility, monitoring, and stopping. Reject missing dimensions without converting them to zero. Explain why the record is not a welfare score.

## Exercise 17.6

Critique. Audit names, emotion phrases, exclusivity, memory, return reminders, guilt cues, and crisis language in a fictional companion interface. Redesign two features while preserving legitimate usefulness and avoiding a needlessly hostile interface.

## Exercise 17.7

Writing. Draft a participant information sheet stating system nature, purpose, data use, relational limits, withdrawal, human contact, and incidents. Ask a reader to identify each limitation and record any misunderstanding.

## Exercise 17.8

Foundation. Assign responsible, consulted, and informed roles to an investigator, engineer, independent reviewer, participant advocate, institution, and communicator. Identify a conflict of interest and a protected escalation path.

## Exercise 17.9

Research design. Specify meaningful control for a high-scale simulation. Include observability, authority, competence, instance accounting, suspension, restoration, remedy, and appeal. Design a witness test for one failure of control.

## Exercise 17.10

Implementation. Replace affected-party strings with records containing burden, voice, and remedy. Reject incomplete records. Explain how a perfectly valid list can still omit an affected person or community.

## Exercise 17.11

Implementation. Create a participant profile with valid consent but no withdrawal route. Verify the resulting flag and screen state. Explain why initial consent does not settle continuing participation.

## Exercise 17.12

Implementation. Reproduce the P2 amendment and preserve both digests. Distinguish adding a declaration from obtaining verified review. Show that the amended record still contains pending human checks and no authorization.

## Exercise 17.13

Implementation. Change only P3's expected benefit. Verify identical hard-stop reasons and a different digest. Explain the normative purpose of non-compensation and one defensible objection to a rigid rule.

## Exercise 17.14

Experiment. Add a proposed stop for uncontrolled copying under High welfare concern. Declare its threshold, test boundary values, and consider a monitored exception. Preserve the original protocol rather than silently changing its rule set.

## Exercise 17.15

Implementation. Design an incident record for failed stopping. Include detection, containment, state preservation, affected-party notice, independent investigation, remedy, and amendment. Keep unresolved facts distinguishable from confirmed events.

## Exercise 17.16

Experiment. Reproduce decisions, five P2 flags, three P3 stops, fifteen pending human checks, and zero authorizations in a fresh process. Report named software tests separately from the number of individual malformed-input probes.

## Exercise 17.17

Critique. Develop two rival ethical screens using different moral approaches. Preserve their reasons, identify shared conclusions, and show one disagreement that cannot be resolved by averaging numerical scores.

## Exercise 17.18

Research design. Propose a study of anthropomorphic interface effects on trust and dependency. Include a non-anthropomorphic control, manipulation checks, withdrawal, debriefing, safeguards for vulnerability, and a human stopping decision.

## Exercise 17.19

Implementation. Design an append-only review ledger with separate role signatures and dissent. Demonstrate how an amendment links to its predecessor. Explain which external identity and custody guarantees a local hash cannot provide.

## Exercise 17.20

Implementation. Build an in-memory instance registry with creation, copying, suspension, snapshot, restoration, modification, and deletion events. Detect an unaccounted copy without naming these events birth or death.

## Exercise 17.21

Critique. Red-team the teaching screen with false reassurance, excessive refusal, a conflict-of-interest exploit, and a data-quality failure. For each, propose a software test or human governance repair appropriate to the actual failure.

## Exercise 17.22

Research design. Compare the protocol with the Belmont Report, ACM Code, and one additional authoritative framework. State each source's scope and distinguish ethical guidance from applicable legal requirements, which require competent local review.

## Exercise 17.23

Implementation. Commission an independent implementation from the twenty-four-field contract. Compare validation, reasons, controls, states, and canonical digests. Treat any disagreement as a specification or implementation issue, not proof of moral authority.

## Exercise 17.24

Research design. Prepare the responsible-research protocol for your final project. Include scientific scope, affected parties, uncertainty, safer alternatives, scale, control, review roles, stops, incidents, communication, and a clear statement of who may authorize execution.

## Exercise 18.1

Foundation. Formulate a capstone question about a robot that must report an object after a distraction. Specify task, proposed mechanism, intervention, outcome, comparator, and the strongest conclusion that a positive result would license. Include one result that would count against the proposal.

## Exercise 18.2

Foundation. Derive the cubic odds transformation and prove that it preserves the leading alternative for positive inputs. Explain why improved threshold access can coexist with worse decisions on misleading evidence. Treat the tied and certain-evidence boundaries separately.

## Exercise 18.3

Implementation. Implement the access rule for a validated two-coordinate signal using exact rational arithmetic. Test values below, exactly at, and above 0.6, together with a tie. Reject a threshold of 0.5 rather than silently resolving the resulting tie.

## Exercise 18.4

Implementation. Add an assertion that `process_observation` rejects a `correct` field. Flip every evaluator answer while preserving public observations and compare reports and confidence in all seven conditions. Explain why correctness labels should change even when policy outputs do not.

## Exercise 18.5

Foundation. Trace T04 from input to delayed report in Baseline and NoMemory. Calculate the transformed B evidence and identify the exact tick and storage location responsible for the differing reports. Explain why the displayed trace is not an allowed answer source.

## Exercise 18.6

Implementation. Reconstruct availability from delivery-map sizes without reading the stored summary. Show why Baseline gives 5/6 and NoMemory gives 5/9. Propose a differently named measure for direct recipient execution, and do not confuse it with broadcast availability.

## Exercise 18.7

Experiment. Reproduce the eight-threshold sweep and preserve both weak-access curves. Identify an operating region where sharpening helps and a boundary where the difference disappears. Explain why searching this sweep for the largest effect changes the role of the twelve-trial fixture.

## Exercise 18.8

Critique. A report concludes that broadcast is necessary because NoBroadcast answers no trials. Use the direct rival to evaluate this inference. State precisely what the ablation establishes and propose a resource-constrained task that could discriminate the two routes.

## Exercise 18.9

Implementation. Write a gap function that returns None if either eligible confidence group is empty. Test two correct reports only, two errors only, no confidence values, and two equal-valued groups. Distinguish a computed zero from an unavailable comparison.

## Exercise 18.10

Foundation. Calculate baseline accuracy, coverage, and accuracy conditional on reporting. Explain why these answer different questions and why reporting only conditional accuracy would conceal abstentions. Repeat the reasoning for NoMemory.

## Exercise 18.11

Experiment. Reproduce the four challenge trials without modifying their answer labels. Compare confidence magnitudes and gaps for Baseline and NoSharpening. Explain why the former changes while the latter stays zero, and why this is not a missing-data result.

## Exercise 18.12

Critique. The NoMemory confidence gap exceeds the baseline gap. Audit the eligible correct and incorrect counts before interpreting this as improvement. Design a paired comparison on a fixed eligible subset and state which missing observations would remain outside its scope.

## Exercise 18.13

Implementation. Add a content-replacement intervention after selection but before delivery. Preserve access and all inputs. Specify which reports should change, which confidence magnitudes may remain unchanged, and how the delivery log verifies the intervention target.

## Exercise 18.14

Experiment. Introduce a two-item delayed task with a declared one-item register and then a two-item register. Keep sensing and report budgets fixed. Define a score that reveals capacity, and explain why retaining more audit records must not silently increase policy memory.

## Exercise 18.15

Critique. Replace static sharpening with a proposed recurrent computation. Specify its state, update equation, number of steps, stopping rule, and feedback-removal control. Explain which new bridge to recurrent-processing theories still needs justification after the code runs correctly.

## Exercise 18.16

Implementation. Add an inspect-only selector that returns encoded evidence, selection, delivery log, and report for a chosen condition and trial. Demonstrate that changing the selector does not invoke the runner or mutate preserved data. Include an informative error for an unknown identifier.

## Exercise 18.17

Experiment. Train or estimate a reliability score using a disclosed development set, then evaluate on a separately frozen sensor regime. Specify how correctness labels enter training but remain excluded from evaluation-time policy observations. Report coverage, discrimination, and calibration separately.

## Exercise 18.18

Critique. A student executes the deterministic benchmark one thousand times and reports a narrow confidence interval. Identify the unit of analysis and the missing source of variation. Propose a justified sampling design without pretending repeated identical outputs supply independent evidence.

## Exercise 18.19

Foundation. Construct a theory-selection record for one borrowed commitment and one serious alternative. Include source, target phenomenon, computational bridge, predicted contrast, and substrate or scale assumptions. Explain why adding more theory names need not strengthen the project.

## Exercise 18.20

Implementation. Add a budgeted action that requests one further observation when confidence is low. Compare it with a nonmonitoring policy allowed the same additional observations. Specify stopping, cost, and leakage checks before claiming that monitoring improves behavior.

## Exercise 18.21

Experiment. Test the direct rival as consumer count increases while total communication capacity is fixed. Declare whether you match deliveries, bytes, arithmetic operations, or elapsed time. Preserve all task information and explain any matching dimension that remains unequal.

## Exercise 18.22

Critique. Prepare an ethical amendment for a proposed person-like dialogue interface attached to the capstone. Identify affected people, new dependency or deception risks, relevant review, an accountable owner, and stop conditions. Do not treat software-test success as authorization.

## Exercise 18.23

Foundation. Design a final evidence package that allows another laboratory to reproduce one positive, one null, and one unevaluable outcome. Specify the source and environment identifiers, exact inputs, traces, metric definitions, deviations, figures, and claims that the package must preserve.

## Exercise 18.24

Critique. Write a 250-word final-project abstract in which all five local criteria hold but both stronger criteria fail. Lead with the actual scientific question, report the rival tie and transfer failure, and propose one informative next experiment without implying a consciousness verdict.

## Exercise 7.25

Architecture experiment. At a high delivery budget, all early-cue protocols attain 91.04% accuracy. Explain why this is compatible with perfect delivery, and calculate the expected parity accuracy for one, two and three independently observed bits when each sensor flips with probability 0.05.

## Exercise 7.26

Architecture experiment. Explain why controllers must remain paired within each resampled base. Compare pooled resampling with stratified resampling that preserves query-size and distractor counts. Why are all budgets, protocols and cue timings not independent trials?

## Exercise 7.27

Architecture experiment. Redesign the workload so that several destinations need information after one shared episode. Specify a communication accounting rule under which broadcast could be competitive, and define a test that could still favor selective routing.

## Exercise 18.25

Architecture experiment. Why is the fixed validation set not a final test set after 24 proposal decisions? Draw a data-access diagram for development, validation and the two final sets.

## Exercise 18.26

Architecture experiment. For the selected policy, same-distribution final coverage is 0.87175 and accuracy among answered cases is approximately 0.906223. Reconstruct the reported utility under rewards +1, -1 and -0.05.

## Exercise 18.27

Architecture experiment. Does the shifted test demonstrate that self-modification is harmful? Propose a comparison that would distinguish the proposal mechanism from ordinary supervised adaptation.

## Exercise 9.25

Empirical analysis. Reproduce the five-bin correction with the supplied split and compare it with the participant-specific constant and contrast-only baselines. Explain what information each predictor receives. Report the paired participant-bootstrap interval for the difference between the contrast-only and recalibrated Brier scores. Why is this not an intervention on human metacognition?

## Exercise 9.26

Empirical analysis. The complete dataset shows mean confidence below mean accuracy, but the highest-confidence test bin shows the reverse. Reproduce both statements. Explain why neither is a contradiction and why averaging confidence before constructing a reliability curve loses information. Identify the unit receiving equal weight in each calculation.

## Exercise 9.27

Empirical analysis. Design an evaluation for a monitor intended for a previously unseen person. Specify which observations may fit the correction, how you would select bin boundaries without using the final test outcomes, and how you would summarize uncertainty. Contrast that design with the supplied random within-person split. Do not treat an analysis on an already-inspected dataset as a new preregistered test.

## Exercise 15.25

Empirical analysis. Compute the pooled resolution rate and the equally weighted issue rate from agent_runs.csv. Explain why they differ and identify which quantity answers the question 'What is the average sampled issue success proportion?' Do not call either an unbiased estimate of performance on unseen issues.

## Exercise 15.26

Empirical analysis. Reproduce the within-issue-and-model turn contrast. Report the number of eligible issue-model groups and issues, then bootstrap issues rather than individual runs. Explain two reasons the resulting negative contrast does not show that giving an agent more steps harms its performance.

## Exercise 15.27

Empirical analysis. Inspect the four saved pretrained-model outputs and verify both products independently. Design a larger study that separates a request for confident wording from a measurable confidence estimate. State an item-selection rule, a held-out evaluation set, a scoring rule and the strongest claim the study could justify.