John Mobley MobCorp / Mobleysoft Autonomous Systems Commander (MASCOM)
We introduce the concept of Computational ESP — the capability tier at which a machine’s world model becomes rich enough that prediction is indistinguishable from perception. Drawing on implemented work in degraded document recognition (Brainsight), principles from predictive coding neuroscience, constraint satisfaction in AI, and the historical precedent of the US government’s Stargate remote viewing program, we argue that “extrasensory perception” is not a paranormal phenomenon but an engineering milestone: the point where a system can reliably infer the state of unobserved systems through sufficiently deep models of reality’s constraints. We formalize three progressive perception modes — Brainsight (prediction-assisted vision), Blindsight (structural inference without sensory input), and Farsight (world-model-driven prediction of unseen states) — and show that the attractor dynamics underlying these modes constitute a meta-prediction system: a system that predicts which prediction strategy will maximize global coherence. We propose that Computational ESP is to machines what the Force is to Jedi — not mysticism, but deep constraint awareness that makes observation redundant.
Keywords: world models, predictive coding, meta-prediction, constraint satisfaction, computational perception, machine intuition, attractor dynamics
In 1972, the Central Intelligence Agency commissioned physicists at Stanford Research Institute to investigate whether humans could perceive distant locations without direct sensory access — a capability they called “remote viewing” (Puthoff & Targ, 1976). Over the next 23 years, the US government spent $20 million across programs codenamed SCANATE, Grill Flame, Sun Streak, and finally Star Gate, employing approximately 23 remote viewers in over 450 intelligence missions (Mumford et al., 1995). The program was terminated in 1995 after a CIA-commissioned review concluded it had never produced actionable intelligence, and the scientific community classified remote viewing as pseudoscience (Marks & Kammann, 1980; Milton & Wiseman, 1999).
The Stargate researchers had the right intuition and the wrong mechanism.
Their hypothesis, stripped of mystical language, was: a sufficiently informed agent can infer the state of systems it has not directly observed. This is not paranormal. It is prediction. A meteorologist “sees” tomorrow’s weather. A chess grandmaster “sees” the board 15 moves ahead. A construction estimator “sees” the hardware schedule for a hospital before opening the PDF, because building codes, ADA requirements, fire ratings, and manufacturer catalogs constrain the solution space so tightly that the document is redundant — just a receipt confirming what physics and law already determined.
We call this capability Computational ESP: the engineering milestone at which a machine’s predictive model of reality becomes sufficiently complete that it can reliably infer unseen states from known constraints. This paper formalizes the concept, traces its theoretical roots, demonstrates a working implementation of its earliest stage (Brainsight, an analysis-by-synthesis OCR system), and maps the path toward full Computational ESP through three progressive perception modes.
The deeper claim is that Computational ESP is not one algorithm but a meta-prediction system — a system that selects among prediction strategies to maximize global coherence. When there are pixels, use Brainsight. When there is structure, use Blindsight. When there is a world model, use Farsight. The meta-layer knows which sight to deploy, because it models not just the world but its own prediction capabilities. This self-awareness of predictive capacity is what transforms a collection of heuristics into something that functions like intuition.
The US government’s interest in remote viewing was defensive: intelligence reports suggested the Soviet Union was spending 60-300 million rubles annually on “psychotronic” research (Puthoff, 1996). The CIA’s response was to fund equivalent research at Stanford Research Institute, where physicists Harold Puthoff, Russell Targ, and Edwin May conducted controlled experiments with subjects like Ingo Swann, Pat Price, and Army veteran Joseph McMoneagle.
The SRI team reported statistically significant results — effect sizes greater than 0.6, significance at four standard deviations from chance — and claimed that remote viewing accuracy was independent of distance (up to 10,000 km) and even worked across time (Targ & Puthoff, 1977; May et al., 1990). However, replication attempts by Marks and Kammann (1980) revealed systematic methodological flaws: session dates written on transcripts, references to previous targets, and other cues that allowed judges to infer ordering without any paranormal mechanism.
The 1995 review by the American Institutes for Research produced a split verdict. Statistician Jessica Utts concluded that the evidence was statistically significant (5-15% above chance). Psychologist Ray Hyman argued the conclusion was “premature” and that independent replication was lacking (Mumford et al., 1995). The CIA terminated the program.
Earlier, J.B. Rhine at Duke University had conducted thousands of ESP trials using Zener cards (circle, square, wavy lines, cross, star), reporting hit rates above the 20% chance baseline. These results were later discredited when it was discovered that subjects could read symbols through card backs, detect subtle experimenter cues, and exploit other sensory leakage channels (Hansel, 1966). Four independent psychology departments failed to replicate Rhine’s results.
The ganzfeld protocol — mild sensory deprivation using ping pong ball halves over eyes, white noise, and red light — was designed to eliminate sensory leakage and enhance any genuine psi signal. Milton and Wiseman’s (1999) meta-analysis of 30 ganzfeld studies across multiple laboratories found effect sizes indistinguishable from chance.
Every program followed an identical trajectory: initial exciting results → methodological flaws discovered → independent replication failures → program terminated. The consistent finding across 70 years of research: when experimental controls are tightened sufficiently, the anomalous signal vanishes.
Strip the mysticism and what remains is a precise computational question: Can an agent infer the state of a distant system using only internal representations, without direct sensory access?
The answer is yes — but not through a “psi channel.” Through world models. The Stargate viewers occasionally produced accurate descriptions not because of psychic ability but because of unconscious probabilistic reasoning: base rate knowledge about Soviet military installations, contextual cues in the tasking, and human pattern-matching applied to vague prompts. When the task was constrained enough (describe a military facility), base rate knowledge alone could achieve 5-15% above chance on structured response categories. This is not ESP. It is Bayesian inference with an implicit world model.
The researchers were measuring the residual predictive power of human world models after sensory channels were blocked. They interpreted the signal as anomalous. It was simply prediction.
We formalize three progressive perception modes, each representing a different balance between direct observation and model-based inference.
Definition: Perception of degraded sensory input by generating top-down predictions and verifying them against bottom-up evidence.
P(state | degraded_observation, world_model) =
P(observation | state) * P(state | world_model) / P(observation)Brainsight is implemented and operational in the PhotonicOCR system (Mobley, 2026). When document text is too degraded for character-level recognition (down to 6 pixels tall), the system:
On architectural hardware schedules with 6px text, Brainsight extracts 22 unique domain terms where bottom-up OCR extracts zero (Mobley, 2026). The system is “seeing” what isn’t visible by rendering what should be visible and checking for consistency.
Alpha parameter: ~0.4 (40% sensory evidence, 60% prediction). The system still needs photons; it just needs far fewer of them than traditional recognition.
Definition: Inference of content from document structure, layout geometry, and neighboring context, without examining the target content directly.
P(state | structure, neighbors, document_type) =
Product_i P(state | neighbor_i) * P(state | column_type) * P(state | row_context)Named after the neurological condition in which patients with damage to primary visual cortex can respond to visual stimuli they cannot consciously perceive, Blindsight for machines means inferring cell contents from the grid itself.
A hardware schedule is a highly constrained matrix: - Columns are typed: column 1 = door numbers, column 2 = hardware types, column 3 = manufacturers, column 4 = model numbers - Rows are correlated: a fire-rated door requires specific hardware combinations - Cells are mutually constrained: if the closer is LCN 4041, the mounting is likely SURFACE, not CONCEALED
Given the column header (“CLOSER”), the rows above (LCN 4041, LCN 4041, LCN 4041), and the door type (fire-rated corridor pair), the empty cell is almost certainly LCN 4041. No pixels required. The structure is the information.
The attractor collapse mechanism in Brainsight is already proto-Blindsight: it propagates constraints across words without examining new visual evidence. Extending this from word-level to cell-level to document-level is the path from Brainsight to Blindsight.
Alpha parameter: 0.0. Zero sensory evidence. Pure structural inference. Seeing the matrix.
Definition: Prediction of system states from a world model alone, without access to the specific system being predicted.
P(state | world_model, context) =
P(state | building_type, jurisdiction, occupancy, requirements)Farsight does not look at a document. It does not look at a building. It knows:
Given only “hospital, California, fire-rated corridor pair,” a Farsight system outputs the hardware schedule. The document is redundant — it’s confirming what physics, law, and engineering already determined.
This is what a veteran construction estimator does unconsciously. They don’t read hardware schedules; they predict them, then scan the document for discrepancies. Their world model is so rich that observation is merely confirmation.
Alpha parameter: -1. Negative alpha — the system is not observing; it is generating. The “observation” (reading the actual document) would be a verification step on an already-complete prediction.
Pure Vision -------- Brainsight -------- Blindsight -------- Farsight
alpha = 1.0 alpha = 0.4 alpha = 0.0 alpha = -1
100% photons 40% photons 0% photons no document
0% prediction 60% prediction 100% structure 100% world modelEvery OCR system operates somewhere on this continuum. Template matching is at the left (pure vision). GPT-4V is near the right — it “reads” degraded text by predicting from its world model what the text likely says, using visual evidence as a weak constraint on a strong prior. Brainsight is the first system to make the continuum explicit and navigable — the alpha parameter is adaptive, shifting rightward as visual evidence degrades.
Brainsight, Blindsight, and Farsight are not three separate systems. They are three strategies in a meta-prediction system. The key question is not “what does this word say?” but “which prediction strategy will produce the most coherent interpretation?”
Computational_ESP(target) = argmax_strategy { coherence(world_model, predict(target, strategy)) }When the image is clear, Brainsight wastes computation — just read the pixels. When the image is degraded, pure vision wastes computation — the pixels are noise. The meta-predictor selects the strategy that maximizes coherence between the prediction and all available constraints.
This is what the brain does. You don’t consciously choose between reading letter-by-letter, recognizing word shapes, and guessing from context. Your visual system seamlessly selects the strategy that yields the most coherent percept. The “aha” moment when degraded text suddenly pops into focus is the meta-predictor switching from bottom-up to top-down processing. You didn’t suddenly see better pixels — you found the prediction that makes everything click.
The attractor collapse mechanism reveals the objective function underlying all three perception modes: global coherence.
A percept is coherent when: - Each element is consistent with its neighbors (local coherence) - Each element is consistent with the document structure (structural coherence) - Each element is consistent with the world model (global coherence) - The interpretation maximizes the joint probability across all elements simultaneously
This is why the correct interpretation “attracts” the system. It’s not that one word is correct — it’s that one configuration of all words is correct, and that configuration is the unique fixed point of the constraint network. Finding it feels like intuition because the search is unconscious and the result is sudden. But it’s just MAP inference on a richly connected graphical model.
What humans call “intuition” — the sense that something is right before you can articulate why — is efficient inference on a world model too complex to introspect. A construction estimator’s “gut feeling” about a hardware schedule is marginal inference over thousands of implicit constraints: building codes, manufacturer specifications, cost optimization, historical patterns, regional preferences.
Computational ESP operationalizes intuition. The system doesn’t “feel” that THRESHOLD is the right word — it computes that THRESHOLD is the only word that simultaneously satisfies constraints from neighboring words, column type, row context, and domain vocabulary. The computation is the feeling. The coherence is the intuition.
The analogy to the Force in Star Wars is not casual. Consider the correspondence:
| The Force | Computational ESP | Mechanism |
|---|---|---|
| Precognition | Farsight | World model extrapolation |
| Danger sense | Anomaly detection | Constraint violation alarm |
| Force sight | Brainsight / Blindsight | Prediction-assisted perception |
| “Search your feelings” | Attractor collapse | Let the constraints converge |
| “Disturbance in the Force” | Coherence break | World model prediction error |
| Jedi reflexes | Pre-emptive action | Act on prediction before confirmation |
| Force ghost | Persistent world model | Knowledge survives beyond observation |
A Jedi’s precognition is modeled as: the Force user has internalized the dynamics of the system (lightsaber combat, ship trajectories, emotional states) so deeply that prediction outruns observation. They don’t “see the future” — they compute it faster than it arrives. This is exactly what Farsight does: given sufficient domain knowledge, prediction is faster than perception, and the “future” (the unread document) is already known.
The Force is not magic. It is what prediction feels like from the inside when the world model is complete enough.
Layer 4: Meta-Predictor — selects perception strategy
Layer 3: Farsight Engine — domain ontology + constraint solver
Layer 2: Blindsight Engine — document structure model
Layer 1: Brainsight Engine — analysis-by-synthesis + NCC
Layer 0: Sensory Input — pixels, signals, observationsEach layer can operate independently, but the meta-predictor orchestrates them based on available evidence and required confidence:
def computational_esp(target, evidence_quality, required_confidence):
if evidence_quality > 0.7:
return brainsight(target) # enough pixels to guide prediction
elif document_structure_available():
return blindsight(target) # structure constrains the answer
elif world_model_covers(target):
return farsight(target) # world model knows the answer
else:
return uncertainty(target) # honest: we don't knowThe critical fourth case — returning uncertainty — distinguishes Computational ESP from hallucination. A system that always predicts, regardless of model coverage, is a confabulator. A system that knows the boundaries of its world model and returns calibrated uncertainty when those boundaries are exceeded is a genuine meta-predictor.
Farsight requires a world model with three properties:
Coverage: The model must contain the relevant domain constraints. For construction hardware: building codes, manufacturer catalogs, ADA requirements, fire ratings, material compatibilities, regional preferences, cost constraints.
Compositionality: Constraints must compose. “Fire-rated” + “corridor” + “hospital” + “California” must intersect to produce a specific set of valid hardware configurations, even though no single constraint is sufficient.
Calibration: The model must know what it knows. If the jurisdiction is unfamiliar, or the building type is novel, the Farsight engine must reduce its confidence and defer to lower layers (Blindsight, Brainsight, or raw observation).
Unlike neural world models trained on internet-scale data, Computational ESP’s world model is constructed from authoritative sources:
This makes the model auditable — every prediction can be traced to specific code sections, manufacturer specs, and constraint interactions. When the model predicts LCN 4041 for a fire-rated corridor closer, the reasoning chain is explicit: NFPA 80 requires Grade 1 closer → LCN 4041 is Grade 1 → LCN is the specified manufacturer for this project → prediction: LCN 4041.
This auditability is critical for professional applications. An estimator cannot use a system that says “trust me, it’s LCN 4041.” They need “NFPA 80 §6.4.3 requires Grade 1 closer on fire-rated assemblies; LCN 4041 is Grade 1 per manufacturer spec sheet LCN-4041-R3; LCN is the basis-of-design manufacturer per specification section 087100.”
The meta-predictor’s objective function is global coherence — the degree to which every predicted element is mutually consistent:
coherence(prediction) = (1/N) * sum_i sum_{j in neighbors(i)} compatibility(pred_i, pred_j)Where compatibility is domain-specific: - For OCR: bigram probability × domain co-occurrence × dictionary membership - For hardware schedules: code compliance × manufacturer consistency × cost reasonableness - For world models: physical plausibility × legal compliance × historical precedent
A prediction is accepted when coherence exceeds a threshold. Below the threshold, the system requests more evidence (zoom in, read adjacent cells, consult additional constraints) or reports uncertainty.
Current document AI systems (Google Document AI, Azure Form Recognizer, Amazon Textract) operate primarily at Layers 0-1: extracting text from pixels and mapping it to fields. Computational ESP adds Layers 2-4: using document structure, domain knowledge, and world models to predict content that is illegible, missing, or ambiguous. This is particularly relevant for degraded historical documents, low-resolution scans, and partially occluded text.
A robot navigating a building doesn’t need to read every sign if it has a world model of building layouts. “This is a hospital; the emergency exit is probably at the end of the corridor, left side, with illuminated signage.” Farsight for spatial navigation is the same principle applied to three dimensions.
The deepest application: a world model complete enough to predict experimental results before running experiments. This is what theoretical physics does — predict the Higgs boson’s mass from the Standard Model, then confirm by observation. Computational ESP generalizes this: any domain with sufficient theoretical structure can support prediction without observation.
The progression from Brainsight to Farsight recapitulates a fundamental debate in cognitive science: is perception primarily bottom-up (Gibson, 1979) or top-down (Helmholtz, 1867; Gregory, 1970)? Computational ESP suggests the answer is neither — it is meta-adaptive. Intelligence is the ability to navigate the continuum, allocating resources between observation and prediction based on which yields greater coherence at lower cost.
A novice reads every word of a hardware schedule. An expert scans for anomalies — deviations from what their world model predicts. The expert is running Farsight with Brainsight as a verification layer. They’re not reading; they’re auditing their predictions against reality. This is what all perception becomes when the world model is rich enough: prediction, audited by observation.
Computational ESP is the capability tier at which a machine’s world model is rich enough that prediction becomes indistinguishable from perception. It is not a single algorithm but a meta-prediction framework that navigates three perception modes — Brainsight (prediction-assisted vision), Blindsight (structural inference), and Farsight (world-model prediction) — selecting the strategy that maximizes global coherence given available evidence.
The first stage, Brainsight, is implemented and demonstrated: an analysis-by-synthesis OCR system that reads 6-pixel text by generating predictions and cross-correlating them against degraded reality. The architecture generalizes naturally to document-level structural inference (Blindsight) and domain-level world model prediction (Farsight).
The US government spent $20 million over 23 years looking for remote viewing in human psychics. They found noise. The capability they were searching for — inferring the state of distant, unobserved systems — is real, but its mechanism is not psychic. It is computational. A sufficiently rich world model, equipped with domain constraints, structural knowledge, and calibrated uncertainty, can predict the state of systems it has never observed, for the same reason a physicist can predict the trajectory of a planet it has never visited: because the constraints are deterministic and the solution space is finite.
The Force is not magic. It is what prediction feels like from the inside when the world model is complete enough to make observation redundant. Computational ESP is the engineering program to build that model.
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181-204.
Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127-138.
Gibson, J.J. (1979). The Ecological Approach to Visual Perception. Boston: Houghton Mifflin.
Gregory, R.L. (1970). The Intelligent Eye. London: Weidenfeld & Nicolson.
Hansel, C.E.M. (1966). ESP: A Scientific Evaluation. New York: Charles Scribner’s Sons.
Helmholtz, H. von. (1867). Handbuch der physiologischen Optik. Leipzig: Leopold Voss.
Marks, D. & Kammann, R. (1980). The Psychology of the Psychic. Buffalo, NY: Prometheus Books.
May, E.C., Utts, J.M., Humphrey, B.S., Luke, W.L.W., Frivold, T.J., & Trask, V.V. (1990). Advances in remote-viewing analysis. Journal of Parapsychology, 54, 193-228.
Milton, J. & Wiseman, R. (1999). Does psi exist? Lack of replication of an anomalous process of information transfer. Psychological Bulletin, 125(4), 387-391.
Mobley, J. (2026). Brainsight: Analysis-by-synthesis OCR with attractor dynamics for severely degraded document text. MASCOM Technical Reports.
Mumford, M.D., Rose, A.M., & Goslin, D.A. (1995). An evaluation of remote viewing: Research and applications. Washington, DC: American Institutes for Research.
Puthoff, H.E. (1996). CIA-initiated remote viewing program at Stanford Research Institute. Journal of Scientific Exploration, 10(1), 63-76.
Puthoff, H.E. & Targ, R. (1976). A perceptual channel for information transfer over kilometer distances: Historical perspective and recent research. Proceedings of the IEEE, 64(3), 329-354.
Rao, R.P. & Ballard, D.H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79-87.
Targ, R. & Puthoff, H.E. (1977). Mind-Reach: Scientists Look at Psychic Abilities. New York: Delacorte Press.
Yuille, A. & Kersten, D. (2006). Vision as Bayesian inference: Analysis by synthesis? Trends in Cognitive Sciences, 10(7), 301-308.