RESEARCH MONOGRAPH / REPRODUCIBLE IMPLEMENTATION IMAGINARIUM Probabilistic Neural Scene Inference, Temporal Mental Imagery, and Evidence-Grounded Video Synthesis A forward-model-first technical architecture for prospective intracortical interfaces, with explicit observation models, Bayesian tracking, temporal reconstruction, calibration, provenance and falsifying experiments. 120 numbered pages | 110 technical topic papers 49-source bibliographic index | experimental algorithms Reproducible synthetic evidence; human efficacy not demonstrated Scientific evidence assessed through November 21, 2026 Original scientific and engineering research prospectus IMAGINARIUM / RESEARCH ATLAS / 2026 FRONT MATTER / METHODS 002 / 120 Abstract and central contribution Research problem Reconstructing freely imagined dynamic scenes is an inverse problem with incomplete sampling, unknown participant-specific encoding, and weak external reference labels. Visual experience is not stored as a literal film, and a high-resolution generator can add plausible detail absent from the recording. Evidence base Human studies demonstrate shared category-related neural activity during perception and imagination [C20], perception-to-imagery transfer using fMRI [C04], and video reconstruction from viewed stimuli using fMRI [C19, C25, C28]. Telepathy publicly targets motor-intention computer control [C01, C08]. These demonstrations do not jointly imply present Neuralink imagery-video capability. Contribution The proposed architecture defines explicit observation and transition distributions, a scene-graph posterior, a generative video interface, confidence calibration, missing-channel handling, abstention, signal-substitution controls, and a strict separation of neural-derived versus model-generated details. Quantitative status Executable software was evaluated on entirely simulated spike-count-like signals. Posterior filtering achieved normalized position MAE 0.0505 against a linear baseline 0.0914 and average-scene baseline 0.3847. These measurements are properties of the toy task, not performance in brains. Core conclusion The correct next milestone is a prospective, consented human study demonstrating decoded imagery attributes from recorded visual-related neural activity against matched-cue and wrong-recording controls. A claim of perfect imagined-video reconstruction is scientifically unsupported without such direct evidence. IMAGINARIUM / RESEARCH ATLAS / 2026 FRONT MATTER / METHODS 003 / 120 Formal targets and identifiability Variables Let S_t denote an internal scene state containing object class, position, motion, relations and event timing. Let N_t denote measured neural events at available electrodes; A_t marks task instructions and Z_t denotes hidden network dynamics. The proposed observation model is a distribution p(N_t | S_t, Z_t, A_t, site, session). Sequential inference The desired posterior has the form p(S_1:T | N_1:T, A_1:T). State predictions are updated using a neural observation likelihood. Filtering is causal and supports real-time output; smoothing uses future samples retrospectively and must be identified as such. Fundamental ambiguity If two scenes induce indistinguishable observations at the sampled sites, their relative neural support is unidentifiable. Increasing model size or producing more detailed videos cannot remedy this. A useful experiment deliberately changes a single scene feature and tests the corresponding neural likelihood change. Decoder versus renderer The decoder produces probability distributions over discrete and continuous scene variables; the renderer samples conditional video realizations. Surface appearance that varies without a corresponding neural likelihood change is labeled generative completion. Clinical meaning Access to an existing movement-control BCI is sufficient for intentional video editing through a cursor, but not evidence of direct image-content recording. Human neural recording and a safe protocol remain separate requirements. IMAGINARIUM / RESEARCH ATLAS / 2026 FRONT MATTER / METHODS 004 / 120 Algorithms, results, and software evidence Baseline experiment The package synthesizes 170 labeled training episodes and 28 withheld episodes using 48 simulated channels and 36 time bins per episode. A Poisson log-link encoding approximation is learned; a 640-particle filter jointly infers position, velocity and category. Controls On synthetic data, neural-aware filtering achieved position MAE 0.05047; linear baseline 0.09142; average-scene baseline 0.38466. Replacing each recording with a different episode raised the error to 0.49806, showing that this software actually uses the artificial input. Calibration defect and repair Initial nominal 80% coordinate intervals achieved only 27.58% measured marginal joint-position coverage on synthetic episodes. Fourteen dedicated calibration episodes were used for a whole-episode split-conformal enlargement. On 14 untouched final episodes the empirical whole-episode joint-coverage was 92.86%; this tiny sample does not establish clinical reliability. Exact algorithm interface The implementation includes a learned observation model, posterior propagation, resampling, explicit missing-channel semantics, normalized per-object probabilities, scene-state output and JSON metadata. Tests cover data shapes, numerical posterior normalization, resampling, and missing electrodes. Scientific limitation The artificial data generator directly encodes the scene variables by design. Thus high performance on the synthetic data supports implementation correctness only. No human neural data, access to Neuralink decoder internals, implant interface, electrode-map validation or clinical approval is claimed. IMAGINARIUM / RESEARCH ATLAS / 2026 FRONT MATTER / METHODS 005 / 120 Research program and structural contents Research design The 110 topical research sheets each specify a causal basis, mathematical or mechanistic interpretation, implementation requirement, disconfirming experiment and decision rule. They are divided into eleven 10-page parts, followed by five bibliographic sheets. Parts I–IV 6–15: scope and information limits. 16–25: scientific evidence and transfer. 26–35: biological populations and spatial codes. 36–45: intracortical recording and proposed Neuralink data exchange. Parts V–VIII 46–55: participants, paired tasks, annotations and leakage controls. 56–65: statistical observation and posterior models. 66–75: filtering, smoothing and dynamic scene inference. 76–85: temporal rendering and provenance. Parts IX–XI 86–95: software implementation, schemas and verification. 96–105: quantitative comparisons, ablation, calibration and adverse cases. 106–115: clinical translation, consent, neuroprivacy, oversight and final decisions. Bibliography 116–120: the indexed source register with direct links to scientific literature and official Neuralink documentation. These sources represent differing measurement modalities; their findings must be interpreted within the specific experiments actually performed. IMAGINARIUM / RESEARCH ATLAS / 2026 I. SCIENTIFIC OBJECT AND EVIDENTIARY SCOPE 006 / 120 PART 01 | TECHNICAL STUDY 001 | 01/10 Definition of a neural-video claim The task is an inverse problem involving private visual content, incomplete observations and a generative output channel. A useful device must recover discriminable content rather than only produce compelling imagery. DECISION STANDARD Only attribute to the brain information that survives a blinded, matched signal-substitution test. Scientific and experimental basis An imagined scene is a changing internal representation, not a literal sequence of encoded RGB frames. The observable is time-stamped extracellular activity sampled near particular neuronal populations; the intended output is a report about hidden scene structure. Proposed explanatory and mathematical mechanism Represent a trial by a latent sequence of entity identities, positions, relations and changes. State what independent observation would distinguish each proposed attribute from a plausible rival. A video renderer is a downstream visualization, not the original measurement. Exact engineering operation Register every output element as neural-supported, task-cue-supported, model-generated or user-edited. Require a typed claim record linking its attribute to signal bins, model version, uncertainty and the external validation task. Experiment that could reject the claim Place two same-category scenes with opposite motion into an evaluation set. If replacing the recording preserves the predicted direction, the output lacks the claimed evidentiary dependence. FORMAL RELATION / Information target I(S;N) = H(S) - H(S|N) latent scene > electrode response > decoder test Measurement > inference > independently tested result SOURCE LINKS C01, C02, C07 IMAGINARIUM / RESEARCH ATLAS / 2026 I. SCIENTIFIC OBJECT AND EVIDENTIARY SCOPE 007 / 120 PART 01 | TECHNICAL STUDY 002 | 02/10 Perception, recall and new imagery DECISION STANDARD A success claim must identify which of the three tasks it has actually passed. Scientific and experimental basis Watching a known clip supplies a physical reference; reconstructing a recalled clip supplies an imperfect memory anchor; composing a novel scene has no independent pixel target. These tasks cannot be pooled as though every trial has equivalent labels. Proposed explanatory and mathematical mechanism Build task-specific observation and target variables. Treat stimulus frames as labels during viewing, retrospective descriptions as noisy measurements during recall, and deliberate discriminative judgments as partial labels for novel imagination. Exact engineering operation Assign immutable task-type metadata to each episode. Estimate and report separate outcomes for viewed, remembered, and independently created sequences. Require a held-out transfer test before claiming shared mechanisms. Experiment that could reject the claim Train on viewing and test on recall without retraining; then reverse direction. Record category, geometry and timing changes separately, including their error intervals. FORMAL RELATION / Identifiability p(N|S_a) = p(N|S_b) means the two states are observationally equivalent. latent scene > electrode response > decoder test Measurement > inference > independently tested result SOURCE LINKS C03, C04, C05 IMAGINARIUM / RESEARCH ATLAS / 2026 I. SCIENTIFIC OBJECT AND EVIDENTIARY SCOPE 008 / 120 PART 01 | TECHNICAL STUDY 003 | 03/10 Ground truth and subjectivity DECISION STANDARD Report recoverable attributes, not a supposed literal correspondence to unknowable private pixels. Scientific and experimental basis A remembered clip and the subject's current recollection can differ; forcing the displayed source video to be the absolute mental target confounds memory error with decoder error. Free imagination may omit lighting, texture or exact durations entirely. Proposed explanatory and mathematical mechanism Treat the subjective scene as a latent construct and collect multiple imperfect indicators: delayed categorical probes, relative-order judgments, blinded pairwise choice and optional immediate visual sketches. These indicators offer convergent but incomplete evidence. Exact engineering operation Freeze a first-pass model reconstruction before collecting any correction. Store corrections as a separate later event stream; never silently use them as the independent accuracy target for the same trial. Experiment that could reject the claim Compare decoder predictions with withheld participant decisions rather than annotations supplied as model input. Blind raters to the trained category distribution. FORMAL RELATION / Decoder evidence Delta = score(N_correct) - score(N_shuffled) latent scene > electrode response > decoder test Measurement > inference > independently tested result SOURCE LINKS C04, C05, C07 IMAGINARIUM / RESEARCH ATLAS / 2026 I. SCIENTIFIC OBJECT AND EVIDENTIARY SCOPE 009 / 120 PART 01 | TECHNICAL STUDY 004 | 04/10 Information cannot appear ex nihilo DECISION STANDARD A vivid prior-driven image is a generative demonstration rather than a validated neural reconstruction. Scientific and experimental basis The generator may be powerful enough to depict an entire city from a weak label such as building. That richness belongs primarily to a learned natural-image prior unless the recording independently distinguishes alternative buildings. Proposed explanatory and mathematical mechanism Apply the data-processing principle: transformations of recorded counts cannot create additional information about hidden scene content beyond what their inputs contain. A posterior can sharpen because of a strong prior, but prior sharpness is not evidence from neurons. Exact engineering operation Save full evidence and prior-only predictions using identical prompts and rendering seeds. Calculate incremental likelihood or retrieval accuracy when neural observations are supplied rather than removed. Experiment that could reject the claim Rerun each trial with same-category wrong-episode signals, zeroed signals, shuffled timestamps and prior-only conditioning. Keep video realism ratings separate from neural-dependence metrics. FORMAL RELATION / Information target I(S;N) = H(S) - H(S|N) latent scene > electrode response > decoder test Measurement > inference > independently tested result SOURCE LINKS C06, C07, C08 IMAGINARIUM / RESEARCH ATLAS / 2026 I. SCIENTIFIC OBJECT AND EVIDENTIARY SCOPE 010 / 120 PART 01 | TECHNICAL STUDY 005 | 05/10 Minimum successful demonstration DECISION STANDARD Deployable complexity follows demonstrated incremental information, not video resolution. Scientific and experimental basis A program can be valuable without reproducing unrestricted experiences. A first honest advance is recovery of a limited feature such as direction of imagined motion on trials with matched external cues. Proposed explanatory and mathematical mechanism Formalize a prediction target before choosing model complexity. Examples include ordered transitions, object count, screen quadrant and direction. Evaluate the error distribution on new episodes and compare with the same model given incorrect signals. Exact engineering operation Construct a feature table and a simple linear baseline first. Add probabilistic state tracking only if its gains survive episode-level testing with independently randomized target combinations. Experiment that could reject the claim Test participant-specific and across-day held-out performance. Require neural signal improvement relative to cue and label-only baselines for each claimed feature. FORMAL RELATION / Identifiability p(N|S_a) = p(N|S_b) means the two states are observationally equivalent. latent scene > electrode response > decoder test Measurement > inference > independently tested result SOURCE LINKS C03, C05, C10 IMAGINARIUM / RESEARCH ATLAS / 2026 I. SCIENTIFIC OBJECT AND EVIDENTIARY SCOPE 011 / 120 PART 01 | TECHNICAL STUDY 006 | 06/10 Model attribution levels DECISION STANDARD If any required dimension is unmeasured, render it as optional illustration rather than neural fact. Scientific and experimental basis Object class, geometry, viewpoint and visual texture vary in observability. Assigning one confidence number to the entire generated clip hides which elements were inferred and which were completed from population statistics. Proposed explanatory and mathematical mechanism Attach a posterior distribution to each scene variable and a separate generative provenance label to unobserved appearance. The conditional entropy of a variable may remain large while another variable is tightly estimated. Exact engineering operation Expose object confidence, position intervals, direction confidence, temporal-order probability and renderer source in a structured manifest. Use a visible uncertainty layer when presenting a clip as scientific output. Experiment that could reject the claim Measure calibration separately for category, coordinates, timing and relations; a good global score cannot substitute for a failing dimension. FORMAL RELATION / Decoder evidence Delta = score(N_correct) - score(N_shuffled) latent scene > electrode response > decoder test Measurement > inference > independently tested result SOURCE LINKS C06, C07, C11 IMAGINARIUM / RESEARCH ATLAS / 2026 I. SCIENTIFIC OBJECT AND EVIDENTIARY SCOPE 012 / 120 PART 01 | TECHNICAL STUDY 007 | 07/10 Structural versus practical recoverability DECISION STANDARD Choose new measurements only when they break a specific degeneracy. Scientific and experimental basis More training data cannot identify two private states that generate indistinguishable activity at available electrodes. Lower noise, longer sampling or additional brain areas may help only when the observation mapping actually changes. Proposed explanatory and mathematical mechanism Write the forward map from latent scene variables to electrode measurements. Identify symmetries and unobserved axes that yield the same likelihood. Compare the Jacobian rank or posterior overlap under deliberately distinct scene states. Exact engineering operation In simulation, set selected feature-to-channel coefficients to zero and verify that the corresponding decoder posterior stays broad. Do not modify the renderer to conceal this loss of observability. Experiment that could reject the claim Use a paired reversal such as right versus left, with identical count statistics under the null encoding. The decoder should admit nonidentification instead of fabricating direction. FORMAL RELATION / Information target I(S;N) = H(S) - H(S|N) latent scene > electrode response > decoder test Measurement > inference > independently tested result SOURCE LINKS C07, C12, C13 IMAGINARIUM / RESEARCH ATLAS / 2026 I. SCIENTIFIC OBJECT AND EVIDENTIARY SCOPE 013 / 120 PART 01 | TECHNICAL STUDY 008 | 08/10 Research decision hierarchy DECISION STANDARD Advance to more invasive work only after task-relevant evidence and independent oversight are established. Scientific and experimental basis Scientific development fails when each new algorithm is judged by its plausibility rather than by an independently falsifiable consequence. The necessary sequence is detect, identify, infer, render and verify. Proposed explanatory and mathematical mechanism Separate physical availability, statistical identifiability, numerical accuracy, calibration, transfer and user usefulness. Each stage inherits the failure status of the prerequisite on which it relies. Exact engineering operation Maintain an explicit milestone record showing data source, split policy, model provenance, confidence coverage, null controls and human authorization. Stop scientific claims at the earliest unsupported link. Experiment that could reject the claim For every candidate improvement, compare against a frozen version on the same untouched episode families. Publish regressions and failure conditions as well as improvements. FORMAL RELATION / Identifiability p(N|S_a) = p(N|S_b) means the two states are observationally equivalent. latent scene > electrode response > decoder test Measurement > inference > independently tested result SOURCE LINKS C01, C08, C14 IMAGINARIUM / RESEARCH ATLAS / 2026 I. SCIENTIFIC OBJECT AND EVIDENTIARY SCOPE 014 / 120 PART 01 | TECHNICAL STUDY 009 | 09/10 Why certainty must be calibrated DECISION STANDARD Never promise a perfect probability algorithm independent of sensor information and population coverage. Scientific and experimental basis A probability of 0.99 from a misspecified observation model measures the model's internal commitment, not the actual proportion of correct predictions. Neural data have drift, correlation, unmeasured inputs and sometimes missing channels. Proposed explanatory and mathematical mechanism Judge numerical probabilities by frequency in independently sampled episodes. Model-check the count distribution, dispersion, conditional independence and out-of-distribution behavior before trusting sharp posteriors. Exact engineering operation Use reliability diagrams, proper scores, credible-region coverage and abstention operating curves. Reserve an episode-level calibration split; change thresholds before the final evaluation, not after. Experiment that could reject the claim Deliberately halve the informative channels and change signal gain. Verify whether the system appropriately widens its uncertainty or refuses to label the episode. FORMAL RELATION / Decoder evidence Delta = score(N_correct) - score(N_shuffled) latent scene > electrode response > decoder test Measurement > inference > independently tested result SOURCE LINKS C07, C15, C16 IMAGINARIUM / RESEARCH ATLAS / 2026 I. SCIENTIFIC OBJECT AND EVIDENTIARY SCOPE 015 / 120 PART 01 | TECHNICAL STUDY 010 | 10/10 Novel contribution and falsification DECISION STANDARD The finished research product is a reproducible model and a bounded evidence claim, not an assertion of universal mind reading. Scientific and experimental basis The proposed advance is an evidence-attributed dynamic scene probability model placed between neural measurements and generative video. Its innovation lies in making each visually realized attribute accountable to a testable neural observation. Proposed explanatory and mathematical mechanism Compare direct latent-video decoding, scene-graph-first inference and an explicit intent-controlled scene editor. The hypotheses differ in neural signal requirements, auditability, and reliance on generative priors. Exact engineering operation Evaluate all candidates on a common dataset, identical stimuli, equal task labels, held-out content combinations, uncertainty calibration and blinded user checks. Publish architecture and source-code revisions. Experiment that could reject the claim Construct ambiguous trials with identical object classes but different order, occlusion or location. A candidate passes only the dimensions it demonstrably discriminates. FORMAL RELATION / Information target I(S;N) = H(S) - H(S|N) latent scene > electrode response > decoder test Measurement > inference > independently tested result SOURCE LINKS C03, C05, C07 IMAGINARIUM / RESEARCH ATLAS / 2026 II. EVIDENCE BASE AND LITERATURE INTEGRATION 016 / 120 PART 02 | TECHNICAL STUDY 011 | 01/10 Classical Bayesian natural-image decoding Current research ranges from classical Bayesian fMRI image recovery to 2026 neural imagery and movie decoding. Results differ sharply by sensory modality, task and whether the model evaluates genuine input dependence. DECISION STANDARD Bayesian decomposition is useful only when both its likelihood and prior are measurable and independently audited. Scientific and experimental basis Early methods estimated image likelihoods from receptive-field measurements and combined them with natural-image priors. This is the formal origin of separating observation evidence from plausible stimulus statistics. Proposed explanatory and mathematical mechanism The posterior is proportional to the recording likelihood under a candidate image multiplied by its prior. An image may rank highly because its neural prediction matches the observation or merely because it is common in the image library. Exact engineering operation Reproduce the principle with a low-dimensional candidate-scene bank and a Poisson neural observation model. Expose both likelihood rank and prior-only rank rather than collapsing their contributions. Experiment that could reject the claim Swap equal-prior candidate images that differ in the single visual feature under study and measure posterior odds attributable to data. FORMAL RELATION / Conditional recovery p(S|N,task) need not equal p(S|N,viewing) viewed images > shared features > imagined test Measurement > inference > independently tested result SOURCE LINKS C06, C17, C18 IMAGINARIUM / RESEARCH ATLAS / 2026 II. EVIDENCE BASE AND LITERATURE INTEGRATION 017 / 120 PART 02 | TECHNICAL STUDY 012 | 02/10 Natural-movie fMRI foundations DECISION STANDARD Temporal decoding during perception is prerequisite evidence, not automatic validation for imagery. Scientific and experimental basis Natural-movie work demonstrated that sluggish hemodynamic responses can still carry information about dynamic sensory stimulation when a motion-energy encoding model is paired with a hemodynamic model. This does not translate directly into imagined movies. Proposed explanatory and mathematical mechanism Distinguish visual content timing from the delayed blood-oxygen response. The forward operator smooths and mixes rapid events; inverting it requires priors and introduces temporal ambiguity. Exact engineering operation Use viewed-video datasets to benchmark a temporal model, but retain task labels and report whether a decoder is selecting from candidates or generating independent motion state estimates. Experiment that could reject the claim Reverse event order while preserving categories and source duration. Establish whether predictions follow actual presented order rather than cue duration. FORMAL RELATION / Source dependence Ablation gain = correct-signal score minus prior-only score viewed images > shared features > imagined test Measurement > inference > independently tested result SOURCE LINKS C18, C19, C03 IMAGINARIUM / RESEARCH ATLAS / 2026 II. EVIDENCE BASE AND LITERATURE INTEGRATION 018 / 120 PART 02 | TECHNICAL STUDY 013 | 03/10 Feature-hierarchy generalization DECISION STANDARD Select representations by measured imagery transfer, not by viewed-image leaderboard rank. Scientific and experimental basis Hierarchical decoding studies show that visual-object features learned from perception can predict some imagined category information, but success varies with visual area and feature level. A high-level semantic code is not a low-level retinotopic map. Proposed explanatory and mathematical mechanism Represent category, outline, location and texture as separately estimated latent channels. Use partial pooling across feature levels and avoid assuming one image embedding captures all physical attributes. Exact engineering operation Fit independent feature heads and a multi-task variant. Measure whether joint representation shares beneficial information or introduces cross-task interference. Experiment that could reject the claim Hold out unseen combinations of familiar categories and backgrounds to test whether the decoder learned compositional structure. FORMAL RELATION / Representation transfer T_view-to-imagine = E[accuracy_imagine | fitted_on_view] viewed images > shared features > imagined test Measurement > inference > independently tested result SOURCE LINKS C03, C20, C21 IMAGINARIUM / RESEARCH ATLAS / 2026 II. EVIDENCE BASE AND LITERATURE INTEGRATION 019 / 120 PART 02 | TECHNICAL STUDY 014 | 04/10 The NSD and NSD-Imagery distinction DECISION STANDARD An imagery decoder needs imagery-specific validation even when most of its pretraining uses perception. Scientific and experimental basis Large visual fMRI datasets contain many viewed images; their extensions to imagery are much smaller and have different noise and label properties. Training on seen content can produce deceptively strong perception metrics without imagery generalization. Proposed explanatory and mathematical mechanism Treat dataset source, presentation type, repeat count, participant and scan session as separate stratification variables. The imagery task is a domain shift, not just a smaller sample of the same distribution. Exact engineering operation Load the official NSD-Imagery material using its published data terms. Record train/validation/test design and benchmark human-label provenance before adding any model-specific feature extraction. Experiment that could reject the claim Evaluate the identical trained decoder on perception and imagery with no target-driven retuning; quantify the transfer gap by participant and feature. FORMAL RELATION / Conditional recovery p(S|N,task) need not equal p(S|N,viewing) viewed images > shared features > imagined test Measurement > inference > independently tested result SOURCE LINKS C04, C05, C22 IMAGINARIUM / RESEARCH ATLAS / 2026 II. EVIDENCE BASE AND LITERATURE INTEGRATION 020 / 120 PART 02 | TECHNICAL STUDY 015 | 05/10 MIRAGE and multimodal features DECISION STANDARD Adopt transfer ideas, not performance figures as interchangeable across imagery modality or duration. Scientific and experimental basis MIRAGE uses a comparatively simple backbone with several complementary semantic and visual guidance features to transfer from perception to static imagery. Reported results concern fMRI image reconstruction, not native intracortical movie capture. Proposed explanatory and mathematical mechanism Factor high-level concept signals and low-level image guidance, then examine their conditional contribution. A mixture can improve appearance while making feature attribution harder unless each component is ablated. Exact engineering operation Implement the static baseline as a separate adapter and log modality-specific decoder outputs before generative fusion. Freeze the pretrained image generator during held-out neural-dependence tests. Experiment that could reject the claim Remove text guidance, high-level visual guidance and low-level guidance one at a time; score which feature types remain supported by neural evidence. FORMAL RELATION / Source dependence Ablation gain = correct-signal score minus prior-only score viewed images > shared features > imagined test Measurement > inference > independently tested result SOURCE LINKS C04, C05, C22