PAPER NO. 1 · AUGUST 2026 CONCEPTUAL ARTICLE · PAPER NO. 1 BIT RADIX A Speculative Framework for Cognitive Partitions, Reciprocal Blind Spots, and Cognitive Ecology “I need your mind because it fails differently from mine.” INDEPENDENT CONCEPTUAL PAPER · AI-ASSISTED RESEARCH AND CRITIQUE Revised draft for author review · August 2026 BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 2 Prelude: the mistake no one can see STATUS OF THE PRELUDE The civilization below is a fictional counterexample generator, not empirical evidence. It illustrates the difference between searching a given option space and constructing a different option space. Prelude Consider a hypothetical civilization spread across an entire solar system. In one version, every mind is fundamentally machine-like: extraordinarily capable, precise, and consistent, but sharing the same broad cognitive architecture. In another, every mind is organic and human-like, with the strengths and limitations that come with embodiment, intuition, emotion, and biological cognition. Now introduce one mind from the other civilization. A single human among machine minds might notice assumptions that billions of highly intelligent machines share without questioning. It might recognize a third option in a conflict framed for centuries as A or B, or notice a social possibility that their architecture never made obvious. Reverse the experiment. A single machine mind among humans might organize information on a scale no human could manage, find patterns hidden across generations of records, or formalize systems people had tolerated as chaos simply because no human mind naturally approached them that way. Neither visitor needs to be more intelligent than the civilization it enters. It only needs to think differently enough to see what everyone else has learned not to see. Bit Radix begins with the possibility that even a civilization pushed toward the limits of intelligence could remain incomplete if every mind within it failed in the same way. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 3 BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 4 Abstract Bit Radix is a speculative framework for describing cognitive diversity without reducing it to a single ladder of intelligence. It proposes that every finite cognitive architecture makes some distinctions easier to represent than others. Sensory systems, bodies, learned categories, internal representations, memory limits, and decision mechanisms jointly partition a larger space of possible interpretations and actions. These partitions may create characteristic strengths and architecture-specific blind spots. Under finite resources, increasing representational differentiation along one dimension may also impose costs elsewhere, such as reduced generalization, slower decisions, greater energy use, or more difficult communication. The framework’s central ecological hypothesis is that differently structured minds can correct one another’s characteristic failures, so a heterogeneous cognitive system may outperform a monoculture even when no member is universally superior. Because useful differences can be masked by translation costs, the paper adds a latent-complementarity prediction: mixed-profile pairs should show objective gains after structured mutual modeling only where their independently measured errors genuinely complement one another. It also separates instrumental contributions to problem solving from regulatory contributions that alter stress, attention, or salience. Existing work on Umwelt, embodied and morphological computation, efficient coding, rate-distortion, neural representational geometry, synaptic information capacity, and functionally diverse problem solving supplies intellectual ancestry and testable components; it does not establish Bit Radix as a validated theory. The collective- problem-solving literature is itself contested and is treated here only as conditional precedent. To avoid circular species rankings, this paper replaces a universal “radix score” with independently measured, task-linked radix profiles and states conditions under which the framework should be rejected. Keywords: cognitive diversity; representational geometry; cognitive ecology; translation cost; embodied cognition; rate- distortion; collective intelligence; artificial intelligence Author note on provenance and scope EPISTEMIC STATUS This is a conceptual hypothesis, not established neuroscience and not a report of new experiments. It originated with a lay author without formal academic training beyond secondary school and was developed through sustained dialogue with AI systems. The human author supplied the core intuitions and thought experiments; AI systems assisted with literature discovery, technical translation, criticism, and drafting. AI critiques identified the central circularity problem and a later tension between the ecological claim, the contested group-diversity literature, and the experimental matching protocol. AI output can be wrong. Readers should judge every empirical statement against the cited sources and treat the new synthesis as an invitation to test, revise, or reject. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 5 1. The proposal 1.1 A provisional definition WORKING DEFINITION A Bit Radix profile is the characteristic pattern by which a cognitive system makes distinctions available, combines them, preserves alternatives, compresses information, and converts representations into decisions or communication. The word radix is borrowed metaphorically from numeral systems, where a radix is the number of distinct digits available at a position. The present usage is not a claim that a mind literally runs in base 2, base 10, or any other single base. Nor does it imply that more states are always better. The name points toward a family of questions about representational alternatives and resolution; until those questions are operationalized, assigning one numerical radix to an animal, person, or AI would be misleading. CRUCIAL DISTINCTION Radix is not intelligence. Intelligence concerns effectiveness within a system’s capabilities. A radix profile concerns the structure of the distinctions and alternatives the system can readily construct. A highly capable system may still have stable blind spots if greater capability improves search inside the same representational space without changing how that space is partitioned. The framework rejects a universal, architecture-neutral ability ranking; it does not reject task-specific performance measures when their domain and limits are explicit. 1.2 Four linked claims 1. Partition. Every finite cognitive architecture makes some distinctions natural, some costly, and some difficult even to formulate. 2. Trade-off. Under limited energy, time, data, memory, and physical structure, finer differentiation in one region of representational space may impose costs in another. This is a proposed tendency, not an assertion that every gain must produce a simple one-for-one loss. 3. Blind spot. Some systematic errors may arise from the architecture’s partitions rather than from low general ability. Increasing skill within the same architecture may therefore leave those errors intact. 4. Ecology. Minds with different partitions may possess complementary error patterns. When communication and coordination are adequate, a mixed system can discover possibilities that its members, or a homogeneous group, would miss. The fourth claim is the paper’s central original hypothesis. It is not established by any theorem or simulation cited below; those results provide conditional analogies whose relevance must itself be tested. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 6 1.3 What “binary mind” does and does not mean Early versions of the idea used “binary mind” too loosely. Three different questions must be separated: whether a physical substrate uses binary components, whether information is measured in bits, and whether a cognitive system tends to reduce a domain to two salient alternatives. These are not equivalent. A computer built from binary electronics can represent high-dimensional continuous quantities; a biological neuron with all-or-none action potentials can participate in graded, time-varying, population-level codes; and a rich internal process can still terminate in a discrete action. The useful question is therefore not “Is this mind binary?” It is: Where, for this task and architecture, are distinctions preserved, compressed, or collapsed? The fictional A-or-B civilization is a thought experiment about representational partitions, not a neurological diagnosis and not a description of digital computers in general. 2. Scientific ancestry: what is already established READING RULE The sections below identify neighboring findings. They constrain and motivate Bit Radix, but none of them proves the full framework. “Established” means the cited result exists in its stated domain; “inference” means a connection proposed here; “speculation” begins where the connection is extended into architecture-wide blind spots and cognitive ecology. 2.1 Different organisms inhabit different meaningful worlds Jakob von Uexküll’s concept of Umwelt treated an organism’s perceptual and action capacities as constituting a species-specific meaningful world, rather than as a neutral window onto an already partitioned environment (Uexküll & Kriszat, 1934). Thomas Nagel later argued that even complete objective knowledge about bat physiology may not provide the subjective character of being a bat, because human imagination is constrained by human forms of experience (Nagel, 1974). Bit Radix extends this ancestry cautiously. Umwelt is primarily about organism-environment relations, and Nagel’s target is subjective experience. The new inference is that architecture-dependent worlds may continue beyond sensation into categories, hypotheses, social possibilities, solution classes, and machine representations. That extension is plausible, but it is not contained in either source and must earn independent evidence. 2.2 Bodies can participate in computation Embodiment is not merely a stream of extra inputs delivered to an unchanged central thinker. Brooks’s behavior-based robotics challenged the assumption that intelligent action requires a centralized symbolic representation of the world (Brooks, 1991). McGeer’s passive dynamic walkers demonstrated BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 7 that a suitably shaped mechanical system can settle into a stable gait down a shallow slope without active control or external power beyond gravity (McGeer, 1990). These examples show that morphology and environment can carry part of the problem-solving burden. The Bit Radix inference is that changing a body or tool environment may change where cognition happens and which distinctions become cheap or native. An AI supplied with a persistent workspace, external memory, instruments, and an action loop may therefore behave differently from the same model in a short conversational exchange. This need not imply a different underlying “mind”; it may reflect a different cognitive ecology around the same engine. Two senses of ecology should not be collapsed. A support ecology is one architecture plus its body, tools, and environment. A multi-architecture ecology is a population of differently structured systems that may correct one another’s characteristic errors. The first can change a system’s effective profile; only the second directly states the reciprocal-blind-spot hypothesis. 2.3 Finite systems compress, and compression has a cost Efficient-coding approaches ask how sensory systems represent useful information under constraints. Barlow proposed redundancy reduction as a principle of sensory transformation (Barlow, 1961), and later sparse-coding work showed that learning efficient representations of natural images can produce receptive-field properties resembling those of visual cortex (Olshausen & Field, 1996). Rate-distortion theory provides a more explicit language for the inevitable relationship between information rate and the cost of error. Sims applied it to human categorization and visual working memory (Sims, 2016); Jakob and Gershman linked it to a neural population model and tested a new prediction using monkey prefrontal recordings (Jakob & Gershman, 2023). This literature supports a limited claim: finite perceptual and memory systems discard or distort some information, and the pattern of distortion depends on objectives and constraints. Bit Radix adds the speculative claim that different architectures may develop systematically different distortion profiles. Communication between sufficiently different minds may then become a second compression problem: each mind must translate the other’s internal distinctions into a code it can use. Communication between sufficiently different minds may itself be a lossy compression problem. 2.4 Representational dimensionality creates real computational trade-offs Neural population activity can be analyzed geometrically by asking how many independent dimensions are needed to describe task-related patterns. In monkeys performing a sequence-memory task, nonlinear mixed selectivity produced high-dimensional prefrontal representations that supported a BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 8 large repertoire of simple readouts; dimensionality fell on error trials (Rigotti et al., 2013). Work on abstraction later showed that neural populations can combine generalizable abstract variables with high “shattering” dimensionality, rather than occupying a single simple high-versus-low scale (Bernardi et al., 2020). These findings are close relatives of Bit Radix because they make representational structure measurable and connect that structure to performance. They also warn against a universal score. High dimensionality can improve separability and flexible readout, while appropriately structured or lower- dimensional representations can support abstraction and transfer. The relevant object is therefore a geometry or profile tied to a task—not a species label and not an assumption that more dimensions always mean a better mind. 2.5 The rat-synapse result: useful, real, and easy to misuse A widely repeated result from rat hippocampal CA1 tissue estimated a minimum of approximately 26 distinguishable synaptic-strength levels, corresponding to about 4.7 bits of precision per synapse (Bartol et al., 2015). The arithmetic matters: 4.7 is a number of bits, while 26 is the corresponding number of distinguishable levels. Neither quantity is a “rat radix,” a neuron-wide constant, or a measure of the animal’s cognitive architecture. Later work using related methods found approximately 2.7 bits in control dentate-gyrus synapses and approximately 3.7 bits after induction of long-term potentiation, while contrasting both with the earlier 4.7-bit CA1 estimate (Bromer et al., 2018). A broader analysis using a different entropy method estimated roughly 1.9–3.4 bits across populations of mammalian dendritic spines and emphasized variation by measure, region, condition, and assumptions (Karbowski & Urban, 2023). WHAT THE RAT EVIDENCE ACTUALLY SUPPORTS Biological neural information storage is not adequately described as simply binary, and effective precision can differ across synapses, regions, conditions, measures, and timescales. This is a physical precedent for heterogeneous resolution inside one nervous system. It does not validate a mind-level Bit Radix score. 2.6 Diversity can help—but the formal case is disputed and conditional Hong and Page modeled problem solvers as task-specific search heuristics on a defined problem landscape. In their simulation and theorem, individual ability is average performance over starting points on that task, not a universal intelligence measure. Under specified assumptions, randomly selected capable agents can outperform the individually best performers because the selected top performers may use overlapping heuristics (Hong & Page, 2004). This is a conditional result inside a particular model. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 9 Its interpretation is disputed. Thompson (2014) argued that the theorem’s restrictive assumptions and the simulation’s random selection do not warrant the broad “diversity trumps ability” slogan. Kuehn (2017) defended the theorem, arguing that Thompson’s challenge rests partly on a mistaken reading of its assumptions and partly on claims broader than the theorem itself makes. Singer (2019) separately defended diversity, rather than randomness, as the better interpretation while emphasizing that functional diversity must be defined and measured carefully. Grim et al. (2019) found the result sensitive to the structure of the problem landscape and warned against broad extrapolation. The defensible conclusion is narrow: under some task structures, ensembles with less-overlapping heuristics can outperform ensembles of stronger but functionally similar individuals. Bit Radix does not treat that conclusion as proof or as a load-bearing foundation. Its ecological claim should survive or fail on its own evidence: whether independently measured profile complementarity predicts held-out error reduction and team gains. Demographic variety, disciplinary labels, or model names are at best imperfect proxies. Real differences can also create communication costs, conflict, coordination failures, or unequal participation. The proposed mechanism is not difference by itself, but difference in what members can formulate and what they systematically miss. 2.7 Ecologies can regulate as well as solve Another organism can alter cognition without supplying a missing fact or solution. Social buffering—the reduction of a stress response through the presence or interaction of a social partner—has been documented across nonhuman vertebrates, although its pathways and boundary conditions vary (Wu, 2021). In human–dog relations, mutual gaze has been associated experimentally with an affiliative oxytocin feedback loop (Nagasawa et al., 2015). A 2026 meta-analysis of 28 articles found small reductions in heart-rate reactivity and self-reported stress or anxiety when a dog was present during an experimental stressor, but no significant pooled effects for blood pressure or cortisol (Rodriguez et al., 2026). These findings support a limited claim about regulation and affiliation. They do not show that the benefit arose from a difference in cognitive architecture. TWO ECOLOGICAL CONTRIBUTIONS Instrumental contribution: another mind adds information, alternatives, or error correction. Regulatory contribution: interaction changes stress, attention, salience, confidence, or the felt meaning of a situation. These contributions can interact, but evidence for one does not establish the other. Illustratively, a cowboy’s dog need not solve the route, weather, or mechanical problem to change how danger is experienced and which actions feel possible. The story is not evidence about dogs or cowboys. It marks a broader possibility: a cognitive ecology may alter not only the answers available to its members, but the salience and emotional weighting of the questions themselves. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 10 Translation costs provide a different, human example, but the evidence is mixed. A small diffusion-chain study found that information degraded more and rapport was lower in mixed autistic–non-autistic chains than in same-neurotype chains (Crompton et al., 2020). A much larger registered replication found no mixed-chain disadvantage in objective information transfer, while still finding rapport differences and a modest rapport benefit from disclosing diagnostic status (Crompton et al., 2025). Autism is not a Bit Radix profile, and these studies do not show that either same- or mixed-neurotype groups are intrinsically superior. Their narrower relevance is that interactional mismatch and mutual knowledge can affect communication, while rapport and task performance can dissociate. Taylor and Vestergaard’s Complementary Cognition account similarly proposes that developmental dyslexia may reflect an exploratory search specialization within an exploration–exploitation trade-off, with collective benefits emerging through cooperation (Taylor & Vestergaard, 2022). That proposal is a neighboring theory, not confirmation of Bit Radix. Together these literatures motivate a testable extension: complementary value may remain latent when partners cannot model or translate one another’s characteristic distinctions. Better translation should therefore reveal some mixed-profile gains —but only when independently measured error patterns are genuinely complementary. 3. The circularity problem The strongest objection to early versions of Bit Radix is simple: the framework can be made unfalsifiable by assigning a score after observing the behavior it is supposed to explain. If a whale appears to integrate rich simultaneous sensory information, one assigns it a high radix; if a human compresses experience into manageable choices, one assigns a medium radix; if a fictional alien cannot escape an either-or decision, one assigns it a binary radix. The label then merely repeats the observation in numerical costume. That objection succeeds. A behavior-derived score cannot be used to explain the same behavior. The remedy is not cosmetic. Species-wide decimal rankings should be abandoned unless a measure can be derived independently and shown to predict unseen outcomes. At present, Bit Radix is best treated as a research framework with testable component hypotheses—not as a validated scalar theory. RULE AGAINST CIRCULARITY Measure the proposed representational profile before testing the behavior it is meant to predict. Pre-register the predictions, use held-out tasks, compare against simpler explanations such as task-family baseline performance, general ability within a defined population, and task familiarity, and report failures. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 11 4. From a universal score to a radix profile A radix profile should be multidimensional, task-linked, and operational. No single measure below is “the radix.” Together they describe how a system allocates distinctions and where it compresses them. Discrimination profile: psychophysical or model-based thresholds for distinguishing stimuli, concepts, temporal intervals, social states, or candidate solutions. Representational geometry: dimensionality, clustering, factorization, mixed selectivity, and cross- condition geometry measured from neural activity or model activations. Abstraction and transfer: performance when rules must generalize across novel contexts, modalities, embodiments, or surface forms. Branching and preservation: the number and diversity of hypotheses a system generates and maintains before commitment, measured under controlled time and memory budgets. Rate-distortion profile: what information is preserved, merged, or lost as channel capacity, time, energy, or memory is constrained. Error topology: stable classes of false positives, false negatives, omissions, confusions, and unformulated alternatives across held-out tasks. Translation cost: how much performance is lost when one system must communicate its internal distinctions through another system’s interface or vocabulary. Profiles should be estimated at several levels—component, circuit, whole system, and system-plus-tools —because the rat evidence already warns that informational precision is not one number even inside one brain. They should also be treated as conditional on training, development, body, culture, interface, and task. “Architecture” is not a magic residue left after context is removed; it is the relatively stable organization whose contribution must be separated empirically from experience. 5. A falsifiable research program 5.1 Core predictions 1. Profile stability. Independently measured representational features should predict a nontrivial portion of a system’s error pattern on unseen tasks, beyond general ability, sensory acuity, task familiarity, and training history. 2. Complementary error. Greater distance between radix profiles should sometimes predict lower correlation between errors. The relationship need not be monotonic: systems too different to communicate may perform worse together. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 12 3. Ecological gain. Teams chosen for complementary profiles should outperform homogeneous teams matched on prespecified task-family baselines, relevant component capacities, and resource budgets. The gain should be explained by error complementarity rather than by simply adding more compute, time, or expertise. 4. Resource trade-off. Under fixed resources, increasing differentiation or separability along selected dimensions should produce measurable costs elsewhere—for example in generalization, robustness, decision latency, communication bandwidth, energy, or sample efficiency. 5. Latent-complementarity intervention. Among mixed-profile pairs with independently measured complementary errors, reciprocal translation and mutual-modeling support should improve objective team performance more than equal-duration generic communication. The intervention should preserve or map distinctions unique to each partner, changing which errors survive transmission. Rapport should be measured separately: improved rapport without improved problem solving would support a regulatory benefit, not the stronger claim that translation unlocked cognitive complementarity. 5.2 A minimal experimental design A practical first study need not compare humans, whales, and machines. It could compare several AI systems or human-AI pairings with deliberately different sensors, training objectives, memory structures, or tool environments. Matching should not mean equalizing one universal ability score. It should mean balancing performance distributions across prespecified calibration task families and controlling relevant capacities such as sensory thresholds, memory, compute, time, and training exposure. Residual differences should be modeled rather than hidden inside an “overall ability” label. MATCHING RULE Use task-relative, profile-aware controls. A scalar performance measure may be valid inside a defined task family without becoming a universal ranking of architectures. 1. Profile. Estimate representational geometry, generalization, branching, compression, and error topology on a calibration set. 2. Hold out. Pre-register predictions about which systems will fail on separate task families whose decisive distinction is absent or expensive in one profile. 3. Assign. Construct three resource-matched conditions: similar-profile pairs, mixed-profile pairs with ordinary communication, and comparable mixed-profile pairs given structured reciprocal translation or mutual-modeling support. 4. Translate. Give the intervention group useful, empirically grounded information about each partner’s attention, communication, reasoning, uncertainty, or error patterns—not diagnostic stereotypes or extra task answers. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 13 5. Measure. Track objective task performance, error overlap, alternatives generated, correction of a partner’s characteristic errors, and rapport. When testing the regulatory route, pre-register measures of stress reactivity, attentional allocation, salience, persistence, or confidence and compare them against matched social-control conditions. Test whether translation changes problem solving beyond any change in comfort or affiliation. 6. Replicate. Repeat across tasks and interventions. A profile that changes arbitrarily with every task has not captured a stable architectural property. 5.3 What would count against Bit Radix Proposed profile measures fail to predict held-out behavior better than simpler variables such as task-family baseline performance, task exposure, working-memory capacity, or sensory thresholds. Error patterns are not stable enough across tasks to support architecture-linked blind spots. Profile diversity does not predict complementary errors or mixed-team gains after controlling resources and communication. Translation support improves rapport but not objective performance, alternative generation, or partner-error correction. A regulatory benefit may remain, but the stronger latent- complementarity extension would not be supported. After controlling for task content, familiarity, affiliation, expectancy, and interaction time, a differently structured partner produces no reliable change in pre-registered measures of stress, attention, salience, persistence, or confidence relative to a matched control partner. In that tested domain, the regulatory route would not be supported. Mixed-profile teams remain inferior after successful mutual modeling on tasks where their premeasured errors were predicted to complement one another. Apparent trade-offs disappear when systems receive sufficient training or better interfaces, indicating that the limitation was contingent rather than architectural. A single sufficiently capable architecture reliably discovers every alternative found by heterogeneous systems under fair resource constraints, weakening the claim of irreducible blind spots. Any of these outcomes would require narrowing, revising, or abandoning central claims. A framework that redescribes every possible result as “a different radix” has explained nothing. 6. Implications for humans and artificial intelligence If different cognitive architectures possess persistent blind spots, then capability and diversity solve different problems. More intelligence may improve performance within a representational space; a different architecture may alter which space is searched. The strongest artificial intelligence would therefore have an instrumental reason to preserve access to differently structured minds—not because BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 14 those minds are secretly better at everything, but because their failures and distinctions may be non- identical. This is a conditional statement, not a claim that present AI systems have desires, rights, or a measurable radix. It also cannot serve as the sole basis for human dignity. People do not need to demonstrate economic usefulness to deserve moral consideration. The narrower conclusion is epistemic: systems charged with high-stakes understanding should be wary of cognitive monoculture, including monocultures made of extremely capable agents. The design implication is practical. Rather than forcing every participant into one common style, a cognitive ecology should preserve useful differences while building translation layers: shared tests, explicit uncertainty, external memory, interpretable intermediate artifacts, adversarial review, and procedures that reward the discovery of missing options rather than merely consensus. Such an ecology may work through two routes. Instrumentally, another system can contribute distinctions, hypotheses, or corrections that were absent from the first. Regulatively, a social partner can alter stress, attention, salience, persistence, and the perceived stakes of a problem. The second route may change reasoning without adding an answer; it should be measured independently rather than counted automatically as evidence for reciprocal blind spots. 7. The paper as an illustration—not evidence The development of this paper mirrored its own thesis. A human author supplied the originating intuition, literary thought experiments, and resistance to formulations that felt technically clean but conceptually empty. One AI system supplied terminology, mathematical distinctions, literature leads, and draft structure. Another AI system identified the circularity problem: radix values had been assigned after observing behavior and then used to explain it. That criticism forced the move from species scores to independent profiles and falsifiable predictions. A later AI review exposed a quieter seam: the ecological claim leaned too silently on the disputed Hong–Page result while “overall ability” matching risked reintroducing the universal scale the paper rejected. That criticism forced the present revision: the dispute is explicit, Hong–Page is non-foundational, and matching is task-relative and profile-aware. No participant was simply “the smartest one.” The product changed because the participants failed differently. This is not empirical support: it is a single, unblinded, selectively remembered collaboration with no control condition. Its legitimate role is illustrative. It shows what the hypothesis asks researchers to measure more rigorously: whether complementary representational limits can produce a result unavailable to a cognitive monoculture. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 15 Each mind needs the other’s internal world translated into distinctions it can use. 8. Limits of the framework The key construct remains provisional. “Partition,” “distinction,” and “architecture” require domain-specific operational definitions. No evidence presently supports a universal mind-level radix number, much less species rankings. Representational dimensionality, synaptic precision, sensory discrimination, hypothesis branching, and tolerance for ambiguity are related only by hypothesis; they must not be collapsed into one quantity without evidence. Architecture, learning history, culture, body, interface, motivation, and task are entangled. Separating them will be difficult and may reveal that some supposed blind spots are trainable habits. Diversity can reduce performance when coordination, trust, translation, or shared evaluation is inadequate. Social buffering, cross-neurotype communication, and Complementary Cognition are neighboring literatures, not validated Bit Radix measurements. Diagnostic categories must not be treated as architecture labels, and cognitive difference must not be assumed beneficial before task-specific evidence exists. The framework concerns cognition and collective problem solving. It does not solve consciousness, personal identity, or whether a copied mind is the same person. Conclusion Bit Radix is not yet a measurement of minds. It is a disciplined version of a question: What becomes representable, and what becomes invisible, when a finite architecture carves reality into usable distinctions? The scientific literature already shows that perception is selective, bodies can carry computation, memory is capacity-limited, neural representations have consequential geometry, and synaptic precision is heterogeneous. It also contains conditional models and simulations in which functionally diverse problem solvers outperform more individually able groups, although the interpretation and generalizability of those results remain disputed. The new proposal is to connect these local facts and contested precedents into a testable architecture-level hypothesis about reciprocal blind spots and cognitive ecology. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 16 That connection may fail. If independently measured profiles do not predict unseen errors, if heterogeneous systems add no benefit, or if every apparent blind spot disappears with ordinary training and tools, the hypothesis should be reduced or rejected. But if the predictions survive, intelligence will look less like a single summit and more like an ecosystem: not one perfect mind, but different minds making one another’s worlds larger. I need your mind because it fails differently from mine. BIT RADIX / PAPER NO. 1 CONCEPTUAL ARTICLE · 17 References Barlow, H. B. (1961). Possible principles underlying the transformations of sensory messages. In W. A. Rosenblith (Ed.), Sensory Communication (pp. 217–234). MIT Press. https://www.cnbc.cmu.edu/~tai/microns_papers/Barlow- SensoryCommunication-1961.pdf Bartol, T. M., Jr., Bromer, C., Kinney, J., Chirillo, M. A., Bourne, J. N., Harris, K. M., & Sejnowski, T. J. (2015). Nanoconnectomic upper bound on the variability of synaptic plasticity. eLife, 4, e10778. https://doi.org/10.7554/eLife.10778 Bernardi, S., Benna, M. K., Rigotti, M., Munuera, J., Fusi, S., & Salzman, C. D. (2020). The geometry of abstraction in the hippocampus and prefrontal cortex. Cell, 183(4), 954–967.e21. https://doi.org/10.1016/j.cell.2020.09.031 Bromer, C., Bartol, T. M., Bowden, J. B., Hubbard, D. D., Hanka, D. C., Gonzalez, P. V., Kuwajima, M., Mendenhall, J. M., Parker, P. H., Abraham, W. C., Sejnowski, T. J., & Harris, K. M. (2018). Long-term potentiation expands information content of hippocampal dentate gyrus synapses. Proceedings of the National Academy of Sciences, 115(10), E2410 –E2418. https://doi.org/10.1073/pnas.1716189115 Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence, 47(1–3), 139–159. https://doi.org/10.1016/0004-3702(91)90053-M Crompton, C. J., Ropar, D., Evans-Williams, C. V. M., Flynn, E. G., & Fletcher-Watson, S. (2020). Autistic peer-to-peer information transfer is highly effective. Autism, 24(7), 1704–1712. https://doi.org/10.1177/1362361320919286 Crompton, C. J., Foster, S. J., Wilks, C. E. H., Dodd, M., Efthimiou, T. N., Ropar, D., Sasson, N. J., Lages, M., & Fletcher- Watson, S. (2025). Information transfer within and between autistic and non-autistic people. Nature Human Behaviour, 9, 1488–1500. https://doi.org/10.1038/s41562-025-02163-z Grim, P., Singer, D. J., Bramson, A., Holman, B., McGeehan, S., & Berger, W. J. (2019). Diversity, ability, and expertise in epistemic communities. Philosophy of Science, 86(1), 98–123. https://doi.org/10.1086/701070 Hong, L., & Page, S. E. (2004). Groups of diverse problem solvers can outperform groups of high-ability problem solvers. Proceedings of the National Academy of Sciences, 101(46), 16385–16389. https://doi.org/10.1073/pnas.0403723101 Jakob, A. M. V., & Gershman, S. J. (2023). Rate-distortion theory of neural coding and its implications for working memory. eLife, 12, e79450. https://doi.org/10.7554/eLife.79450 Karbowski, J., & Urban, P. (2023). Information encoded in volumes and areas of dendritic spines is nearly maximal across mammalian brains. Scientific Reports, 13, 22207. https://doi.org/10.1038/s41598-023-49321-9 Kuehn, D. (2017). Diversity, ability, and democracy: A note on Thompson’s challenge to Hong and Page. Critical Review, 29(1), 72–87. https://doi.org/10.1080/08913811.2017.1288455 McGeer, T. (1990). Passive dynamic walking. The International Journal of Robotics Research, 9(2), 62–82. https://doi.org/10.1177/027836499000900206 Nagasawa, M., Mitsui, S., En, S., Ohtani, N., Ohta, M., Sakuma, Y., Onaka, T., Mogi, K., & Kikusui, T. (2015). Oxytocin-gaze positive loop and the coevolution of human–dog bonds. Science, 348(6232), 333–336. https://doi.org/10.1126/science.1261022 Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435–450. https://doi.org/10.2307/2183914 Olshausen, B. A., & Field, D. J. (1996). Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 381, 607–609. https://doi.org/10.1038/381607a0 Rigotti, M., Barak, O., Warden, M. R., Wang, X.-J., Daw, N. D., Miller, E. K., & Fusi, S. (2013). The importance of mixed selectivity in complex cognitive tasks. Nature, 497, 585–590. https://doi.org/10.1038/nature12160 Rodriguez, K. E., Delzio, M. C., Ruekgauer, A. H., Saleh, A. A., & Pfander, J. L. (2026). The effect of a dog on acute stress reactivity during experimental stressors: A systematic review and meta-analysis. Anthrozoös, 39(2), 183–209. https://doi.org/10.1080/08927936.2026.2621532 Sims, C. R. (2016). Rate-distortion theory and human perception. Cognition, 152, 181–198. https://doi.org/10.1016/j.cognition.2016.03.020 Singer, D. J. (2019). Diversity, not randomness, trumps abili