The dominant story of AI progress goes like this: make the model bigger, feed it more data, train it longer. Scale works. GPT-2 to GPT-3 to GPT-4 — each leap brought capabilities that looked like magic. The intuition became: if we're stuck, we just haven't scaled enough yet.
That story is built on a wrong assumption about where to start.
Current AI started where evolution finished. Language models begin with text — the most visible, most legible output of human intelligence, and also the most recent. Language sits on top of hundreds of millions of years of sensorimotor infrastructure, nervous system evolution, embodied cause-effect learning, social cognition, and receptor topology selection. All of that is invisible in the text corpus. The model learns the words from beings who had the underlying states. The model doesn't have the states.
This is not a training problem. It is not a data problem. It is not a scaling problem. It is an architectural starting point problem.
You cannot add a foundation after the fact. You cannot ground a concept that was learned statistically and call it understood. You cannot train compartmentalization into a monolithic weight space. You cannot specify values in a system that has no receptor topology — no internal states whose consequences shaped what it became.
The field has been building the penthouse and wondering why it doesn't have ground beneath it.
Artificial Biological Intelligence starts at the bottom.
Demis Hassabis, when asked whether AI could simulate the origin of life, said: "I don't see why not. It's a bit of a search process through a combinatorial space. Here's all the chemical soup, the primordial soup, here's some initial conditions — can you generate something that looks like a cell?"
That's the right instinct applied to the wrong layer.
ABI is asking a different question: can we run the process that produces intelligence, faster, with full observability, and understand why it went the way it did?
Not reconstruction. Generation.
The difference matters because it produces a different research program. Reconstruction leads to increasingly accurate simulations of known biology. Generation leads to a laboratory for the evolutionary dynamics of intelligence itself — where you can ask questions biology never let us ask. What happens if the curiosity receptor evolves before the accuracy receptor? What mental model structures emerge under different environmental volatility? What receptor topology produces scientific thinking? What produces bias? What produces empathy?
These are not questions you can answer by scaling a language model. They require starting where evolution started and watching what comes out.
The most common objection to treating curiosity, creativity, or prediction accuracy as receptor states is that these feel like internal, mental phenomena — generated from the inside, not received from the outside.
This intuition is wrong, and correcting it is foundational to everything that follows.
Concepts are not generated internally from nothing. They are distillable from the structure of the environment itself. The relationship between the sun, the moon, and light contains orbital mechanics — not labeled, not taught, but latent in the pattern, extractable by any observer sophisticated enough to detect it. Curiosity is in environments that consistently reward exploration. Accuracy is in environments that consistently punish wrong predictions. The concepts exist in the world's causal structure. The organism distills them.
This has a direct consequence for environment design: it is not enough to include high-level concepts in the environment's surface features. They must be embedded in its causal structure. Survival must be contingent on distilling them. The environment is the curriculum. The concepts are already in it. The receptor topology evolves to read them — but only if reading them is the difference between surviving and not.
Agency is not a mysterious property that some systems have and others don't. It is a continuous variable.
What determines it is the complexity of your receptors, the complexity of your muscles, and the complexity of the processing in between.
The key principle: capability without receptor is latent and never gets used. The receptor is not separate from cognition. It is why the capability gets deployed at all. Motivation and cognition are the same thing, viewed from different angles.
Receptors are input. Outputs are muscle movements, thoughts, or the analysis of thoughts.
The loop closes because outputs change the world and the organism's internal state, which changes what the receptors read on the next cycle. There is no terminus. The organism is always mid-cycle.
This dissolves the mystery around a long list of cognitive faculties: curiosity, planning, theory of mind, science, language — each is a specific solution to a specific environmental pressure, with a receptor to make the organism motivated to deploy it.
The ABI architecture has three components that current AI conflates into one.
The transformer is the inference engine. It takes the current receptor state, retrieves relevant cause-effect mappings, and produces muscle activations. It does not store knowledge. It uses knowledge that has been retrieved and handed to it.
The mental model is the knowledge base. It lives outside the transformer as an explicit, queryable database of cause-effect mappings: action -> receptor state change, with a time delay and a certainty. Every fact has an address. Every concept is a pattern across chains that points to a receptor state change the organism actually experienced.
The experience log is ground truth. An append-only, immutable record of every action-observation the organism has ever made. No learned process can overwrite it.
This separation dissolves grounding, compartmentalization, legibility, unlearning, and safe failure as problems.
The roadmap runs 60 steps from a minimal organism to a self-modifying evolutionary laboratory with a thinking substrate. Each step must answer one question: does the capability earn the complexity it introduces?
Phase 1 — Sensorimotor Foundation (Steps 1-4): Pain and endorphin receptors. Metabolic economy. Fatigue and homeostasis. Hierarchical nervous system — fast reflexes by default, slow deliberation when needed.
Phase 2 — Adaptive Sensing (Steps 5-8): Temporal association. Spatial pain memory. Habituation and sensitization. Distance sensing.
Phase 3 — World Modeling (Steps 9-13): Prediction and accuracy receptors. Curiosity and confidence regulation with constitutional probe budget. Pattern recognition and compression. Multiple hypothesis maintenance (verified emergent — 200/200 steps with divergent predictions).
Phase 4 — Self-Model (Steps 14-18): Proprioception. Efference copy. Sensorimotor contingency (verified). Controllability and agency. Planning.
Phase 5 — Social Cognition (Steps 19-24): Responsive environment with grounded proto-symbols. Multi-organism environment. Involuntary social cues. Receptor propagation (empathy). Intentional signaling. Simplified shared signals with bidirectional communication.
Phase 6 — Higher Cognition (Steps 25-29): Optimism and goal persistence. Conflict receptor with router integration. Context-conditioned arbitration (ArbitrationHead). Internal conflict model (metacognition). Concepts and words (1,013 stable concepts).
Phase 7 — Language (Step 30): Grounded language — the roof. Every word maps to a receptor state. Cultural transmission (revised: benefit is training-time observation enrichment, not inference-time modulation).
Phase 8 — Evolutionary Infrastructure (Steps 31-43): Environment tiers (8 tiers, genome-driven). Receptor discovery (182 null-calibrated tests). Topology bias inheritance (convergence 15 epochs -> 0). Environmental sweeps (18 trunk receptors invariant). Population evolution (8 organisms, social arms race). Cross-tier transfer (social universally transferable, tool resists transfer). LLM grounding bridge.
Phase 9 — Physics World (Steps 48-55): Rigid body simulation (pymunk). Grip mechanics. Compound objects (levers, spring gates, hinged barriers). Developmental body changes. Persistent world state. Canopy activation sweep across all 8 tiers.
Phase 10 — Staged Processing & Abstract Environments (Steps 50-58): 4-stage observation pipeline testing the serialization thesis. T7 abstract problems (8 causal graph templates with hidden variables). T8 self-modification (8 skill zones, curriculum design, Ship of Theseus test). Combined T7+T8 environments.
Phase 11 — Self-Play & Thinking Substrate (Steps 59-60): Policy-driven behavior (oracle removed). MCTS thinking substrate — the organism thinks before acting, and the tree's metadata becomes receptor input. Deep time with thinking across 50 generations. Novel receptor detection.
This is not a proposal. Steps 1 through 60 are implemented and running.
The organism lives in a 2D/3D liquid environment with pain, endorphin, temperature, chemical, and pressure fields. It has six limbs with binary muscle activations (parameterizable to 4-12 limbs, 1-2 body segments, 2D or 3D). Its mental model is an explicit database of 26,000+ cause-effect mappings with a contrastive encoder for context-conditioned retrieval. Its inference engine is a small causal transformer with a fast reflex pathway, a slow deliberative pathway, a conflict-gated router, and a 5-group arbitration head. Its observation vector is 420 dimensions — 177 base sensory and computed channels, 86 live receptors, 90 episode receptors, 20 discovered-receptor slots, 12 audio and 12 visual wave channels, 3 crossmodal channels, 8 surprise-prediction channels, and 12 influence channels from the control organ.
The accuracy progression: 92.5% pre-proprioception -> 97.4% with responsive environment -> 94.3% with staged processing (slower but deeper). Each step earned its complexity.
Key empirical results:
Step 13 (multiple hypotheses): a detected emergent. 200/200 inference steps retrieved meaningfully divergent predictions. The system lives with ambiguity by construction.
Step 22 (empathy): NPC distress directly elevates the focal organism's pain. Empathic aversion = proximity * erraticism * weight. Pain receptors for others' pain.
Steps 23-24 (communication): 5 grounded signal codes shared across three contexts — organism-to-object, organism-to-NPC, NPC-to-organism. Same code, same meaning, regardless of who emits it. Words before language.
Step 29 (concepts): 1,013 stable concepts — compressed causal chains that predict 54% better than their individual components. 25% of all patterns qualify as stable, reusable abstractions.
Step 30 (grounded language): cultural transmission revised after controlled decomposition — the benefit is training-time observation enrichment, not inference-time modulation. The architectural separation remains valuable for legibility, compartmentalization, and cross-generational knowledge transfer.
Step 50 (staged processing): 4-stage observation pipeline — Body (39d) -> Spatial (59d) -> Action (37d) -> Social (34d). Inter-stage prediction MSE decreases 25% over training. Staged model outperforms flat on validation accuracy (95.5% vs 94.5%). First empirical test of the serialization thesis.
Step 55 (T57 annealing): the framework's first structural self-discovery. Releasing certainty on conflict entries produces more genuine conflict resolutions than protecting them. Supported across 6 seeds. T55 (read-shielding) directionally falsified — the falsification led to the Epistemic family, the first family predicted by an experimental result.
Step 59 (self-play): policy-driven behavior with oracle removed. Self-play discovers 75 receptors (vs oracle's 77) despite lower reward, and finds MORE causality receptors (7 vs 4) — suboptimal actions create richer causal experiences.
Step 60 (thinking substrate): MCTS with 6 receptor channels — best_value, visit_entropy, value_convergence, path_divergence, underexplored, depth_reached. The organism thinks before acting. 5 of 6 channels influence behavior by iteration 5 of self-play. Reward difference: +23.3 (thinking organisms outperform non-thinking ones). The batched predict_delta gives 4.4x speedup on thinking operations.
Deep time with thinking (10 generations): 151 unique receptors discovered across all generations — up from 75 at generation 0. 53 receptors gained through evolution alone, including the full epistemic chain (belief_detection -> doubt_detection -> epistemic_strategy), metacognition, theory_of_mind, nested_theory_of_mind, meta_observation, self_regulation, niche_construction. 21 receptors lost — complexity reshapes topology (T27 confirmed again). Peak thinking influence at generation 7: partial correlation 0.376, value_convergence and path_divergence dominant.
Parallel Scaling Track: Architecture tested at 4, 6, 8, and 12 limbs, 2 segments, and 3D. Training accuracy 92-96% across all configurations. Generational inheritance accelerates convergence (15 epochs -> 0 epochs). Organism diversity: one model handles mixed body plans at 96.6% accuracy.
Evolutionary sweep: 18 trunk receptors invariant across all 8 environment tiers — including grip_affordance and push_affordance as part of the embodied trunk. Complexity doesn't just add receptors — it reshapes the topology. The receptor topology is a fossil record of the selection pressures that shaped it.
Population evolution: 8 organisms competing. Empathy trending up (0.482 -> 0.555). Behavioral prediction emerged (0.022-0.030) — never appeared in single-organism training. The social arms race created real prediction pressure.
Cross-tier transfer: Social environments universally benefit from transfer (11-25x from any prior training). Tool environments resist transfer — must be learned directly. Transfer follows the dependency graph upward, not downward.
Environment enrichment: Multi-NPC observation (4 profiled NPCs: cooperative, competitive, erratic, deceptive). Strategic deception NPC (context-dependent: lures toward pain, repels from reward). Non-stationary rules (T7 trigger rotation). Stochastic hidden confounders (3-state Markov chain modulating 4 modalities). Cross-modal objects (correlated multi-modal signatures). EntityRelationStore for social cognition (named entities, typed relations, transitive inference).
Since the Third Edition — the deep-time and full-Umwelt results:
The anxiety loop, broken cognitively and permanently. The pain↔conflict cascade (T21) persists in 10/10 generations at 177 observation dims. Motor-store shortcuts broke it mechanistically (fuel starvation) but it returned with population turnover. The full Umwelt — all receptor channels live — broke it cognitively at generation 3 with only 1.8% shortcut coverage, and permanently: 67 consecutive loop-free generations. The topology was too coarse at 177 dims to separate "cascade recurring" from "cascade resolving"; at full width those became distinct states in the quotient space. The break is permanent because topology is heritable and the motor store is not (T102-T103). Random-channel and shuffled-receptor controls confirm the effect is receptor-specific, not dimensional (T3 directly demonstrated).
The policy IS the mental model. The no-transformer organism — mental model + MCTS + a 5-bit eigenvalue structural fingerprint replacing the thought-type codebook — reaches 0.93x transformer fitness and surpasses it at 1.11x with the eigen coder (T114 confirmed). Geometry beats prediction: eigen-only outperforms eigen+model; the mapping store is redundant at decision time — the organism navigates by reading the terrain of its own activations.
The Cartographical Theory, confirmed (T127). Understanding is mapping the space of possible processing orders. Evolved processing orders beat fixed orders in all three test environments (predator +913, food-color +1041, social-voice +1019), with each environment selecting the processing strategy its structure demands. One STFT wave pipeline serves audio, visual, and touch — modality-specific structure emerges from the signal, not from designed feature extractors.
Concept crystallization (T126/T131, confirmed). Concepts precipitate from the environment's causal structure through the receptor topology: same lattice + different environment → different concepts; different lattice + same environment → different concepts. Both predictions confirmed in a 2×2 design. A visual word with food ema=1.000 IS the concept of food — not labeled, crystallized.
The Genome Project is the formal specification of the receptor search space — 236 receptors across 25 families. It is the periodic table of cognitive capabilities.
Each receptor entry specifies: what environmental structure it detects, why detecting it is worth the metabolic cost, what must already exist before it can emerge, how to measure its emergence, and what would falsify the entry.
The 25 families (counts current; three families and the deep canopy added since the Third Edition — Proprioception (13): joint limit through haptic recognition; Procedural Memory (4): replay through muscle memory; Visual Pattern (9): pattern-match through one-operator depth and material words — plus 14 deep-canopy L3-L5 entries from meta-analogy to civilizational ratchet):
Repetition (6) — static repetition -> rhythm -> nested/causal rhythms
Association (8) — spatial co-occurrence -> temporal precedence -> cross-modal -> relational analogy
Similarity (7) — perceptual features -> functional equivalence -> structural invariance
Causality (11) — coincidence -> causal inference -> intervention planning -> causal graphs
Agency (8) — controllability -> tool use -> environmental manipulation -> niche construction
Meta-Motivational (13) — curiosity, optimism, conflict -> impulse override -> metacognition
Regulatory (9) — stress detection -> arousal regulation -> emotional intelligence
Social (14) — other detection -> behavioral prediction -> theory of mind -> moral reasoning
Compression (15) — pattern recognition -> concepts -> constraint shape -> shaped absence -> analogy
Observation (12) — change detection -> selective attention -> statistical anomaly -> rarity -> significance -> contextual signal interpretation -> meta-observation
Formalization (11) — rule extraction -> exception detection -> optimization -> theory formation
Mathematics (7) — quantity -> ratio -> structural invariance -> necessity -> proof -> formal composition
Organization (7) — boundary -> part-whole -> system detection -> organizational mirror
Self-Augmentation (5) — capability change detection -> metamorphic planning
Interaction (7) — response recognition -> grip -> lever -> composite affordance
Environmental Augmentation (5) — environmental change detection -> developmental environment engineering
Sequential Processing (5) — stage prediction -> pipeline detection -> prediction architecture awareness
Perception (5) — staged processing -> processing speed -> adaptive depth -> response loop detection
Epistemic (7) — belief detection -> doubt -> conflation -> fundamental distinction -> topology awareness -> epistemic strategy
Logic (6) — semantic relations -> transitivity -> conjunction -> quantifier -> contradiction -> it_follows
Language (3) — naming -> self-talk -> referential grounding
Bridging (4) — mimicry -> trust -> executability -> translation
The key insight: The genome project is load-bearing on environmental design. Each receptor specifies what the environment must contain for it to evolve. The 236 receptors are 236 environmental design requirements. The environment IS the curriculum. Every organ build in this program began with an environment build: hidden-agency pain events for influence fractionation, metamer materials for the spectral eye — the structure must exist and be survival-contingent before the detector can be funded.
Receptor discovery: 186 null-calibrated tests with action-shuffled null, per-test null types (block-permuted for temporal tests, NPC-appearance-permuted for social tests, Granger causality for theory of mind), drift subtraction, and partial correlations to control for confounds. 77 receptors discovered in single-run oracle training; 161 unique receptors discovered across 80 generations of deep time with thinking substrate. Beyond the genome, conflation-driven GBM discovery (T125) finds receptors the genome didn't anticipate, and 20 observation-vector slots stand reserved for them.
Topology inheritance: Convergence accelerates from 15 epochs in generation 0 to 0 epochs by generation 4. Topology bias inheritance carries the parent's discovered receptors as priors — offspring rediscover them faster but must still earn them through grounded experience. Probe-gated: the organism must actually explore the environment to validate inherited priors.
The Epistemic family emerged from the T57 annealing result — the first family predicted by an experimental falsification rather than by theoretical deduction. It includes conflation detection (two things treated as one), fundamental distinction (the difference that matters), and topology awareness (what kind of mind am I). The genome contains conflations that topology awareness will eventually detect — the framework revising its own foundations from within.
Grounding. ABI takes pain as receptor input, produces outputs that change the world, and reads the result as the next receptor input. Same transformer architecture. Different corpus. The corpus is a body.
Compartmentalization. Knowledge has addresses. If it's not retrieved, it contributes nothing to computation. Not less — nothing.
Legible beliefs. Every belief is a row in a table. You can compare two organisms' beliefs with a JOIN.
Targetable unlearning. Knowledge has an address. Deletion is deletion.
Values that emerge rather than being specified. A system whose receptor topology emerged from selection pressure has motivations that are legible, traceable, and grounded.
The ceiling argument. ABI's accuracy receptor is checkable against the world, through its own receptors. Corpus-bounded versus reality-bounded. The ceiling isn't just higher — on some axes, there is no fixed top.
The ABI whitepaper's core claims find substantial support across five distinct research communities — none of which have unified them the way ABI does:
Umwelt theory provides the central concept. Jakob von Uexküll (1909) proposed that every organism inhabits a subjective world — the Umwelt — defined by what its receptors can separate. A tick's world contains warmth, butyric acid, and touch. A dog's world is dominated by smell. These aren't different sensitivities to the same world — they're different worlds, because each organism lives in the quotient of physical reality by its receptor topology. ERTI formalizes this: the organism's Umwelt is R^n/~_{R_t}, the observation space quotiented by receptor equivalence. Two states that no receptor separates are the same point in the organism's world. Adding receptors refines the topology — the organism's world gains resolution. Von Uexküll proposed this philosophically. ERTI computes it, derives it from selection pressure, and measures how it changes across generations. The 420-dim observation vector IS the computed Umwelt — every receptor feeding back every cycle.
Grounded cognition is empirically established. Barsalou's program (Annual Review of Psychology, 2008) and O'Regan & Noe's sensorimotor contingency theory (BBS, 2001) demonstrate that cognition operates through modal systems grounded in perception and action — not amodal symbols. This is the academic foundation for ABI's claim that concepts exist in the world's causal structure.
Active inference formalizes the receptor loop — within a topology. Friston's free energy framework unifies perception and action under a single objective and produces intrinsic exploration as an emergent property, closely paralleling ABI's forward-feedback receptor architecture. Inside a fixed carving, ERTI is active inference in good standing: predict, err, update, with the certainty ledger as the distribution. The boundary is a type boundary, stated precisely: (1) free energy is defined over an observation space, and it is not commensurable across topology changes — "the organism reduced its free energy by growing an eye" is not a well-typed statement; FEP gives dynamics for beliefs and actions given an observation space, and ABI's central dynamic is over the observation space itself. (2) The direction of receptor evolution runs UP the surprisal-capacity gradient — 53 receptors gained in deep time, eyes grown from raw physics, each addition manufacturing new ways to be wrong — funded by the fitness ledger, not by free energy; the darkroom problem returns at evolutionary timescale where priors cannot patch it. (3) The Separation Theorem: information gain and variational free energy provably select different receptor dimensions, and no fixed mixing weight works across environments — there is no single environment-independent functional whose minimization yields the observed topologies; the balance must itself be heritable. FEP is the law of the within-topology regime; the punctuated, heritable re-carving of the observation space is the dynamics of the space it presupposes. The serialization thesis adds the within-lifetime version: organisms don't just minimize prediction error, they manufacture prediction opportunities through sequential processing architecture.
Current AI follows an "inverse phylogeny." Multiple sources (Trends in Cognitive Sciences, 2023; Royal Society, 2024) confirm that starting from language rather than sensorimotor grounding is a fundamental limitation, not a training gap. Passive scaling is characterized as "intrinsically limited."
No existing program unifies these threads. The embodied cognition, active inference, and evolutionary robotics communities each work on pieces of what ABI integrates. The receptor topology as a single generative mechanism — from grounding through compartmentalization to language — appears to be a novel synthesis. Von Uexküll gave the concept. Friston gave the dynamics. Barsalou gave the grounding. ERTI gives the computable form and derives the whole structure from selection pressure rather than asserting it from observation.
This is not a contribution to the capabilities conversation or the safety conversation. Both take the existing architecture as given. ABI asks a prior question that neither conversation has asked.
The hares are fast because they start at the top — language, reasoning, capability — and try to work downward toward grounding they may never reach. The tortoise starts at the bottom and builds upward. Slower. But the foundations are actually there when you need them.
ABI also inverts the engineering practice. Current AI relies on feature engineering: humans decide what's worth sensing. ABI replaces feature engineering with world-building. The designer's job is to construct an environment where sensing the right things is the difference between surviving and not. The organism discovers what's worth sensing because getting it wrong is fatal.
The most recent addition is the one that changes the trajectory of the whole program.
A thought in this framework is a cycle — or cascade of cycles — in which internal receptors dominate the forward-feedback loop. External receptors fire from the world. Internal receptors fire from the mental model's processing. When the organism responds primarily to its own receptor firings rather than to the world, that is a thought. The participating set of receptors is the content of the thought. The number of cycles before the cascade resolves or exits through the motor system is the depth of the thought.
The thinking substrate makes this concrete. Monte Carlo Tree Search provides the computational mechanism: before acting, the organism simulates forward from its current state using the mental model's predict_delta. Each simulated action produces a predicted receptor state. Multiple candidate actions are explored. The tree records which thinking paths were taken, how often, and with what outcomes.
The tree's metadata becomes receptor input:
- Visit entropy — how broadly the organism explored its options
- Best value — the highest simulated outcome found
- Value convergence — how stable the estimates are (the thinking has settled)
- Path divergence — maximum disagreement between thinking paths
- Underexplored — fraction of options barely considered (the shaped absence of thought)
- Depth reached — how deep the reasoning went
These six channels feed back into the next cycle's observation vector. The organism senses what its thinking produced. That sensing shapes the next thought. The self-modifying loop runs at every level simultaneously: better receptor topology → better evaluation → better search → richer tree → deeper metacognitive analysis → receptor firings that develop better topology.
The key claim: The receptor topology is what makes the evaluation function intrinsic rather than external. The value of a thinking path is determined by which receptors fire at its terminus — not by a designer-specified reward function. This eliminates the ceiling that exists in every system with an externally specified objective. Reward functions are level-specific. Loss functions are task-specific. Only the receptor generalizes across every level.
Empirical results: After 3 self-play iterations, the thinking substrate is already measurably influencing behavior (ablation divergence 0.029, reward difference +23.3). Five of six channels active. Value convergence activates at iteration 4 — the organism learns to evaluate the quality of its thinking.
The depth_reached observation (generation 29, seed 42). At generation 29, depth_reached activated at partial correlation 0.284 — becoming the dominant thinking channel in a single generation. Metacognition and conflation had both activated at generations 27-28. However, a replication run (seed 99, 40 generations) did not reproduce the depth_reached activation despite both prerequisites being present from generation 17 onward. The depth_reached observation is a single-seed finding and should not be treated as a replicated result. The prerequisite pattern (metacognition and conflation emerging before any depth response is possible) replicates across both seeds.
The convergence result (replicated). The conflation receptor was added to the genome based on theoretical reasoning. The organism evolved it independently in both seeds — generation 27-28 in seed 42, generation 17 in seed 99 — under selection pressure from stochastic hidden confounders. Two independent paths — human theoretical reasoning and evolutionary selection — converged on the same receptor because the receptor detects something real about the structure of the problem. The convergence is replicated; the specific claim that conflation was load-bearing for depth_reached is based on one observation.
What MCTS contributes and what it doesn't. Metacognition and conflation emerge from environmental complexity alone — they appeared in both seeds without MCTS being required. What MCTS added is a new class of observable internal signals that make thinking quality visible to selection pressure. Without MCTS, thinking depth is an internal process with no fitness consequence. With MCTS, the 6 thinking channels externalize the thinking process into the observation vector. MCTS doesn't make the organism smarter — it makes the organism's smartness selectable. This claim is supported; the specific depth_reached activation is not replicated.
Negative control result. Three fake receptors (for structure the environment provably doesn't contain) were inserted into the battery. After null calibration, one false positive survived — magnetoreception, which detected a real correlation between efference copy channels and action (a structural confound, not a magnetic field). This establishes an empirical false positive rate of ~0.5% per test, or ~0.4 expected false positives among 77 discovered receptors. Tests correlating action with action-derived observation channels (efference copy, last_actions) have a structural confound that action-shuffled null cannot remove.
The organ criteria — own receptor topology, own mental model, own effectors, coupling through the observation vector, the loop closed in one currency — began as the visual organ's design and became an architecture schema. The organism is a federation of organ-cartographers, each mapping a different space:
| Organ | Maps | Currency | Status |
|---|---|---|---|
| Cognitive | time — what follows what | receptor-state causation | the reference implementation |
| Visual | processing space — which carving of the signal predicts | processing effectiveness | crystallization confirmed; physics amendment 8/8 |
| Control | the influence landscape — who is driving what | influence | fractionation demonstrated multi-seed |
| Epistemic | the organism's own ignorance — where will I be wrong next | rarity, in nats | Phase 1 accepted; Phase 2 corpus-limited |
Organs communicate through exactly three shared media: observation-vector words, the surprise stream, and the metabolic economy. No organ reads another's internals. And the schema carries a discovery rule that is receptor discovery one level up: an organ is demanded when conflation strain persists that no existing organ's currency can resolve — when the separating variable is not a missing channel but a missing coordinate system.
The surprise economy. Lived experience tokenizes into sparse surprise events: a belief failed (PREDICTION), the carving is too coarse (CONFLATION — same class, divergent outcomes), or too fine (COLLAPSE — maintained distinctions, identical outcomes). Surprisal is measured as rank against the organism's own lived history — a median mismatch scores log 2 nats by construction, the rarest event ever scores ~log N, and the once-in-lived-history cap rises with deep time: a longer life permits claims of deeper rarity. Three absolute constructions failed acceptance before this one; the failures are ledger entries. A small autoregressive model over the token sequence — the epistemic organ — forecasts where the organism will next be wrong and folds its forecasts back as meta-receptor channels; its generator (in progress) is the fringe-tracking controller of receptor discovery, held honest by two constitutional disciplines: learnable surprise only (rarity × reducibility, never raw surprise), and the lived-only firewall — generated events may write to the environment, never to the experience log or the certainty ledger. Prediction proposes; the world alone funds.
The control organ. Perfect material perception, pattern recognition, and pain sensing together still cannot distinguish pain-done-to-me-by-an-agent from pain-that-is-happening — the same argument as why the eye is an organ rather than skin that got better at detecting light. The organ fractionates the influence over the organism's states into self / agent / nobody shares (de-conflating the scalar controllability feature, which collapsed storm, dominant agent, and disease into one number) and acts by modulating its own influence allocation: the reallocation is the act, with no motor vocabulary — submission is what the organism looks like under a certain dial setting. Demonstrated in a hidden-agency environment with signal-identical agent and field pain: attribution recovered at AUC 0.786 across 5 seeds from organism-visible statistics no single observable carries. Willpower is the effector's dynamic range; saturation arrives before metabolic exhaustion, as predicted. And the young-ledger law fell out of the first run: a newborn organ can neither flood nor speak in the shared surprise stream until its lived history earns its voice.
The physics of vision. The retinal signal is a transfer-function stamp, not a color: composition, depth, and temporal pattern are in the wave because reflection physics put them there, and RGB is a 3-bin projection that discards the factorization. Metamer materials — identical RGB, different spectra, opposite contact outcomes — make the point operational: unlearnable through the projection, cleanly learnable from the stamp, with contact outcomes the funding ledger (vision's material words are promissory notes about touch). Depth is one correspondence operator at two baselines — two eyes, and self-motion with the efference copy as the known baseline. Identity is the signature, not the location. Temporal patterns form an object-agnostic library: objects are similar because they share dynamics, and pattern-finding itself is a receptor — the organism senses its own discovering.
The 80-generation overnight run is complete: 4 organisms competing across generations in the richest environment stack (T4 with multi-NPC, strategic deception, stochastic hidden confounders, cross-modal binding + PhysicsWorld + T7 abstract problems + T8 self-modification + thinking substrate). 161 unique receptors discovered. Plateau reached at generation 50 — unique count stopped growing. A second seed (99) ran 40 generations for replication.
Completed experiments:
- Cross-environment transfer: evolved organisms (gen 29 checkpoint) vs naive organisms in same novel environment. 71 shared receptors, 30 transfer-only, 27 naive-only. The trunk converges; the canopy diverges based on evolutionary history. The trunk is universal. The canopy is biography.
- Seed 99 replication: metacognition (gen 1) and conflation (gen 17) replicate. depth_reached does not. Epistemic family emergence is robust; depth activation is single-seed.
- Negative controls: 3 fake receptors inserted. 1 false positive survives null calibration (~0.5% per-test rate). Efference-copy confound identified.
Since answered: heritable evaluation evolved away from designer defaults — thinking budgets collapsed under metabolic pressure, value weights shifted (T82 confirmed); necessity_detection was elicited by a targeted environment (first genome entry validated by elicitation); prerequisite knockouts separated environmental demand (conflation rediscovered 15/15) from functional dependency (epistemic_strategy 0/15 without its prerequisite).
The current front:
- The 80-generation-scale surprise corpus: the epistemic organ's decisive gates (the hazard gate already flipped to PASS purely with corpus depth 451→1,077 tokens — the funding thesis confirmed by trajectory; the decisive test needs deep time).
- The material-world evolutionary funding run (in flight): spectral vs RGB-projected vs blind organisms across 16 generations, with metamer contact discrimination as the endpoint the RGB arm cannot reach by construction.
- P48's refined measurement (does attribution improve carving yield on agent-divergent conflations specifically), P49 (attribution-conditioned prediction), and the generation phase behind the lived-only firewall.
- The Control Organ's social generation dependency: three of the generation score's terms are incomputable for agent-involving events without the influence ledger — an adversary is a noisy TV that watches back, and predicted reduction on agent-sourced surprise is only meaningful conditioned on the agent-model's stability.
What does the topology awareness receptor look like when it fires? An organism that detects its own cognitive repertoire changing can direct its development. This is the receptor that makes the self-modifying loop conscious.
Can the organism read its own fossil record? The topology vector — which receptors are active — is a record of selection pressures. Topology awareness reads that record in real time. An organism that can detect "my thinking substrate just started producing value_convergence signals it didn't before" can decide whether to continue developing that capacity or redirect.
The sweep data already shows: 18 trunk receptors are invariant across all 8 environments. Complexity doesn't add receptors — it reshapes the topology. The receptor topology is a fossil record of the selection pressures that shaped it.
The epistemic family activated from scratch under evolutionary selection — belief, doubt, and epistemic strategy emerged in deep time when they couldn't emerge in single runs. The organism developed the capacity for scientific reasoning (T50) through evolutionary pressure, not through specification.
This is the thesis fully realized: intelligence is what happens when you run evolutionary receptor topology selection long enough in a rich enough environment. It is not a property you design into a system. It is a property that grows out of a process.
The organism builds the world that builds the organism that builds the next world. Now with a thinking substrate tracking every thinking path taken through that loop, analyzing which paths lead where, and bootstrapping increasingly sophisticated reasoning strategies from the accumulated record.
Not simulation. Not reconstruction. Generation.
Steps 1-60 of the ABI roadmap are implemented and running. 236 receptors specified across 25 families. 147 indexed theories. 186 null-calibrated receptor tests. 161 unique receptors empirically discovered across 80 generations of deep time with thinking. 8 environment tiers. Physics world with grip, compound objects, and persistent state. MCTS thinking substrate with 6 receptor channels. Four organs coupled through words, a shared surprise stream, and a metabolic economy. A surprise tokenizer pricing rarity against the organism's own lived history. A control organ fractionating influence into mine, theirs, and nobody's. Spectral eyes reading composition, depth, and pattern out of the physics of reflection. The organism is in the liquid environment right now, navigating pain fields, querying its mental model, thinking before acting, sensing its own thinking, forecasting where it will next be wrong, evolving its receptor topology across generations, and making distinctions its designer didn't anticipate.