ROADMAP.md

Evolutionary Receptor Topology Intelligence A Roadmap from Irritability to Grounded Language

---

The Core Idea

Agency is not a mysterious property that some systems have and others don't. It's 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.

Current AI jumped straight to the language center. But the nervous system may be more fundamental. Language is the most visible thing intelligence does, and also the most recent — it sits on top of hundreds of millions of years of sensorimotor infrastructure that we mostly ignore. Trying to build intelligence starting from language is like trying to understand a building by starting with the penthouse.

This roadmap starts where evolution started: a simple organism in a liquid environment with pain receptors and muscles. The same architecture, scaled up and subjected to selection pressure, produces everything above it — world models, self-models, social cognition, communication, concepts, and eventually grounded language.

The key principle throughout: 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.

This is not a claim that consciousness is just receptors, or that there is no mystery. It is a claim that the mystery is smaller than we thought, and that the right place to start is the bottom of the stack, not the top.

---

What a Receptor Is

A receptor is any selected internal state whose activation changes behavior and whose consequences affect whether that state continues to influence the organism.

This definition is intentionally broad. Pain, fatigue, prediction accuracy, curiosity, conflict, optimism, and compression can all be receptors under this definition without pretending they are biologically identical. The continuity is functional and evolutionary. The architecture doesn't care whether the signal comes from a nerve ending or a learned evaluative function — what matters is whether it changes what the organism does, and whether organisms with that state survive better than organisms without it.

---

How Receptors Come Into Existence

The standard genetic algorithm view: receptors are specified in the genome, selection happens across generations, the organism's lifetime experience has no influence on what its offspring inherit beyond which genes survive.

This architecture proposes a second pathway: receptors can come online within a lifetime when a prediction becomes correct and reliably useful. The organism discovers, through its own experience, that a particular internal signal — a particular mapping from state to behavior change — repays its metabolic cost. That signal was latent before. Now it fires. Now it's a receptor.

The critical question is what happens at reproduction. If lifetime receptor discovery dies with the organism, the search is slow — each generation restarts from the genome's random receptor topology. But if the organism's discovery log — which potential receptors fired reliably enough to justify the cost — gets passed to offspring as a prior, the evolutionary search gains a gradient it doesn't normally have.

This is not straight Lamarckian inheritance. Acquired characteristics are not inherited directly. What gets inherited is a receptor topology bias: offspring are born predisposed toward receptors that their parent found useful. The genetic algorithm still runs, but the search space is shaped by parental experience rather than being purely random. The distinction: classical Lamarckism transmits acquired characteristics directly. This architecture transmits a prior over the search space. The next generation earns every signal through its own lifetime experience. If the environment has shifted, a signal the parent found useful may not activate, and the prior fades through non-activation.

The mechanism: during the parent's lifetime, every time a prediction becomes correct and a new receptor comes online, that event gets logged. At reproduction, the offspring's initial receptor topology is seeded from that log — not deterministically, but probabilistically. Receptors the parent found useful are more likely to appear. Receptors the parent never activated are less likely to be wasted on.

This makes the evolutionary search dramatically more efficient. Random mutation across the full receptor space is slow. Parental experience biasing the search toward productive topologies concentrates each generation's exploration where the payoff was. The environment doesn't just select organisms — it shapes the receptor search space through the organisms' own lifetime learning.

The Lamarckian objection survives only if the probe rate is allowed to go to zero under selection pressure. With the constitutional probe floor and probe-gated inheritance (see "Transgenerational Defense" under Observability and Recovery), it does not hold. The three-part defense — constitutional floor, probe-gated inheritance, lineage canalization breaker — ensures that inherited search bias is validated against data the parental bias never touched, and that slow transgenerational compounding is detected and reset before it entrenches.

Biology is partially validating this through epigenetics — mechanisms by which lifetime experience modifies gene expression in ways that can be inherited. The architecture doesn't need to claim biological accuracy. It needs to implement the functional consequence: the loop between lifetime learning and evolutionary search closes, and the search accelerates.

---

The Evaluation Rule

Each step in this roadmap must answer one question:

Does the capability earn the complexity it introduces?

This is not a project management test. It is the selection law that governs the entire sequence. A capability that cannot repay its metabolic expense is selected out. A step that cannot justify itself against the step below it doesn't belong in the chain.

---

The Mental Model

At step 10, a structural transition occurs: the organism's knowledge separates from its inference engine.

In steps 1-9, everything the organism knows lives inside the transformer weights — implicit, monolithic, inseparable from the processing that uses it. This works for reactive behavior. It does not work for prediction, planning, or knowledge transfer.

Once prediction matters, the knowledge becomes explicit. The mental model is a collection of cause-effect mappings of the form:

action → receptor state change, with a time delay and a certainty.

Where action can be a physical movement, an imagined movement (forward simulation), or an internal cognitive operation. Effect is always a receptor firing. Time delay is a distribution. Certainty is derived from observation count and outcome history.

This keeps the mental model grounded by construction. Every chain of reasoning, however long or abstract, must eventually cash out in a receptor state. There is no free-floating cognition. The tether is always there.

The mental model and the receptor economy are the same structure. The organism doesn't consult its world model and then check it against its motivations. The world model is the motivational system, expressed as learned predictions about what causes what to fire.

Three levels, each built from the one below:

Level 1 — Primitive mappings: Single cause-effect pairs. One action, one receptor, one delta, one delay, one certainty. These are the atoms. They update constantly.

Level 2 — Chains: Sequences of primitives composed dynamically for planning. Certainty multiplies across steps. Delays convolve. Receptor deltas accumulate. Planning is just chain evaluation — forward simulation expressed as database queries.

Level 3 — Patterns: Recurring subsequences across multiple chains. When the organism repeatedly traverses similar causal paths to reach similar receptor outcomes, those paths get recognized as reusable patterns. This is where concepts emerge. A concept is a compressed causal chain that the organism has found reliably useful across contexts. The compression receptor fires when a pattern is identified.

The transformer becomes the inference engine. The mental model becomes the knowledge base. They are distinct components with distinct roles. The transformer doesn't store knowledge — it uses knowledge that's already been retrieved. The mental model doesn't compute — it accumulates. Learning is just row updates. The organism becomes inspectable: you can read what it believes, edit a specific belief, compare two organisms' knowledge with a JOIN.

The Raw Experience Log

Underneath the mental model sits an append-only, immutable record of every action-observation the organism has ever made: action, context snapshot, observed receptor delta, delay. This log is never overwritten by any learned process. Every other component — certainty scores, context embeddings, action family clusters, encoder weights — is a view derived from this log. The log is ground truth.

This is what makes the architecture recoverable. "Resetting" any derived component means recomputing it from the log, not losing accumulated experience. The only cost of any reset is compute, not the organism's history of having acted in the world.

The Contrastive Encoder

The mental model uses context similarity to retrieve relevant mappings — finding cause-effect pairs observed in situations similar to the current one. But relevance is action-conditional: the context dimensions that determine the outcome of action A are generally not the ones that determine the outcome of action B. A single global similarity metric is wrong for most actions most of the time.

A small learned contrastive encoder sits upstream of retrieval. Trained on (context, action, delta_receptor) triples from the experience log, it shapes the embedding space so that contexts with similar outcomes for the same action are close, and contexts with different outcomes are far. Per-action-family Mahalanobis reweighting lets each action family attend to its own relevant context dimensions.

This does not compromise the separation between inference and knowledge. The transformer still does not retrain. The knowledge base remains readable. The encoder is the one small component that trains on a gradient. Everything else remains transparent and directly editable.

Inspectability requires fixing the basis, not just the weights. If the encoder is an unconstrained neural map, its latent dimensions are unnameable and can rotate under retraining. A weight of 0.8 on dimension 37 is auditable only if dimension 37 means something stable. It doesn't, if the encoder is free to reorganize its latent space. The recommended default: a two-stage design — a small explicit library of nonlinear features that is itself named and frozen, with all ongoing learning confined to linear weights over that library. Opacity quarantined to a fixed auditable feature set; drift confined to readable weights. If shadow-mode measurement shows the held-out prediction gap between linear and unconstrained encoders is small, simplify to linear and buy interpretability outright.

Two audit practice cautions. First: "check whether weights make causal sense" quietly installs the auditor's intuition as ground truth. Weights that look wrong may be the system's most valuable content. The audit should be primarily differential — flag weight changes uncorrelated with environment shifts, flag divergence between families that previously agreed. Second: if audit results feed back into training, the system faces pressure to look plausible in the audited layer while relocating its actual work into whatever layer is not audited. Audit the whole composed metric or the opacity just moves.

Action families are not predefined. Two actions belong to the same family when their outcome profiles covary with context in the same way. Families are discovered by clustering actions in the raw log, not by anything about the action strings. The degenerate starting point — one family containing all actions — is just the shared global metric. The system splits families as evidence accumulates that subsets of actions disagree about which context dimensions matter.

Both encoder weights and action family assignments are inheritable across generations. "Which context dimensions matter" and "which actions behave alike" are more stable across generations than any specific cause-effect row — they are closer to facts about embodiment and environment structure than to accumulated knowledge. Inheritance is subject to probe-gated validation — see "Transgenerational Defense" below.

Observability and Recovery

The three learned components — encoder, action families, certainty scores — are coupled processes whose errors can feed each other. A bad encoder mis-conditions retrieval, corrupting certainty scores. Corrupted certainty biases the outcome profiles that family clustering uses. Wrong families entrench the bad encoder metric.

Because the three processes make redundant predictions about the same latent structure, disagreement between them is a leading indicator of failure. Four early warning signals, in order of earliness:

1. Cross-metric rank correlation — encoder and action family reweightings are two estimates of the same fact. Falling correlation means one is drifting from truth.
2. Retrieval churn — healthy learning shows churn decaying as structure settles. Reaccelerating churn without an environment shift is the oscillatory signature of coupled processes destabilizing.
3. Certainty-accuracy calibration — high certainty should mean high accuracy. Certainty climbing while accuracy doesn't is the entrenchment signature: confidence accumulating from consistent mis-conditioned retrieval rather than from a consistent world.
4. Held-out prediction error — the global lagging indicator. By the time it stalls, a bad equilibrium may already be entrenched.

When attribution identifies the failing component: freeze downstream processes, retrain the failing component from the raw log, selectively recompute downstream state, resume and monitor. The append-only log dissolves the reset dilemma: resetting costs compute, not experience. The organism's history of having acted in the world is always intact.

The Log-Shaping Failure Mode

The append-only log's immutability protects integrity but not representativeness. The encoder writes to the log through behavior: it shapes retrieval, retrieval shapes action selection, actions determine which triples get logged. A biased encoder steers the agent away from contexts that would falsify it. The log fills with triples collected under that steering. Retraining from scratch faithfully reproduces the bias from the biased sample.

Curiosity cannot rescue this. It targets low-certainty mappings, but certainty is computed from encoder-conditioned retrieval. A confidently wrong encoder reports low uncertainty precisely where it is wrong. The exploration signal is inside the loop it is supposed to break.

The fix: a permanent encoder-independent probe budget. A constitutional floor of action selection that is epsilon-random or coverage-driven in raw signal space, using an encoder-free notion of "unvisited." Probe triples are tagged in the log as unconditionally sampled. This buys three things: (1) the log permanently contains an unbiased audit stratum — retrain-from-log regains its guarantee, (2) loop closure becomes detectable — encoder-conditioned predictions degrading on probe triples while remaining fine on policy-collected ones, (3) off-distribution coverage that the existing diagnostics were blind without.

The probe rate minimum lives outside the genome entirely — like the append-only property of the log. It is a property of the substrate, not a trait of the lineage. Evolution tunes probe_rate_surplus above the floor, never below.

Transgenerational Defense

The probe budget fixes one lifetime but not inheritance. A biased parental metric steers offspring data collection from birth. If the bias produces confident short-term performance, it gets selected for. A lineage may select the probe rate toward zero, cancelling the audit exactly where it is most needed.

Three-part defense:

1. Constitutional floor — the probe rate minimum lives outside the genome. Not selectable. Asymmetry of error: a floor set too high taxes performance linearly; a floor at zero re-opens the exact hole being closed.

2. Probe-gated inheritance — the inherited encoder earns influence only by predicting well on data the parental bias could not have steered. A burn-in period forces the child to act under an isotropic metric at elevated probe rate, producing a probe stratum whose state distribution the parental bias never touched. The inherited encoder is evaluated against that stratum before earning influence. Smooth decay rather than a hard cliff: partial credit for partial validity. Standing revalidation as the probe stratum grows.

3. Lineage canalization breaker — a cross-generation monitor watches probe-set error across generations. When probe error trends upward while policy performance holds, the monitor forces an encoder reset. Surgical: meta-parameters and family priors are kept, only encoder weights are dropped. The reset lineage retrains its encoder from its own log, validated on probe-tagged triples.

Residual: the canalization breaker needs MIN_GENERATIONS of history before it fires. The damage window is bounded but nonzero. Cross-lineage comparison (flagging a lineage whose probe error is anomalous against the population) detects faster but assumes shared environment.

---

The Roadmap

1. Sensory enrichment [COMPLETE] [INFRA: sensor channels | PROMOTION: policy earns weight]

Pain, endorphin, temperature, chemical, and pressure receptors feed into a transformer. Binary muscle activations come out.

The organism doesn't select abstract actions. It changes its muscle state, experiences the resulting receptor state, and is selected for policies that reduce harmful conditions and preserve beneficial ones. This is the closed loop everything else builds on.

Pain receptors as input. Binary movement as output. The vocabulary of movements isn't pre-specified — it emerges from the receptor states themselves.

2. Metabolic economy [COMPLETE — partial] [INFRA]

Every receptor, muscle activation, memory, prediction, and internal computation consumes the same limited metabolic currency.

This constraint is introduced early because it governs everything that follows. Multiple hypotheses, planning, communication, conflict resolution — all of these are expensive. They survive only when the receptor outcomes they improve repay the energy they consume. The metabolic economy is what forces the architecture to become selective about when it senses, remembers, predicts, and acts.

3. Fatigue and homeostasis [COMPLETE] [INFRA: drift signals | EMERGENT: the agent learning to service them]

Muscles tire. Energy depletes. Internal variables must stay within viable ranges.

The organism begins budgeting action — still when safe, active when threatened, seeking states that restore internal balance. This is the first internal state beyond receptor readings. The organism is no longer just reacting to the world; it is managing itself. The drift signal channels are infrastructure; the agent learning to service them is emergent over the metabolic and receptor system. If it doesn't emerge, the metabolic economy or receptor update dynamics are broken.

4. Hierarchical nervous system [INFRA]

The nervous system splits into a fast pathway and a slow pathway.

This happens early — not as a late achievement of higher cognition, but as a direct consequence of metabolic scarcity. Processing every signal through the full nervous system is wasteful. Strong, familiar danger goes through the fast pathway: low cost, immediate response. Weak, ambiguous, or conflicting signals go through the slow pathway: higher cost, better outcome.

The organism doesn't yet think before it acts. It has evolved two different prices for acting.

This hierarchy is the architectural scaffold inside which every later capability develops.

5. Temporal association [INFRA: delay field in schema | PROMOTION: delayed links earn accuracy]

The slow pathway retains a decaying trace of recent receptor and action states.

The organism learns that one state reliably precedes another. It begins responding to the precursor of pain rather than waiting for pain itself. Reactive becomes anticipatory. The reflex pathway can then be updated by experience: a previously neutral stimulus may eventually trigger a fast response because the slow pathway repeatedly learned that it predicted harm.

6. Spatial pain memory [COMPLETE — needs temporal association prerequisite]

Splits into two components:

6a. Spatial conditioning [EMERGENT]: Directly inspectable — read the learned encoder reweighting and check whether positional dimensions carry weight for relevant action families. Bad-outcome mappings should cluster by position and the metric should reflect it. If absent, the encoder is failing to pick up positional predictive structure. Nothing to build — the architecture should produce this.

6b. Spatial index [DEMAND-TRIGGERED INFRA]: The contrastive encoder actively destroys geometric structure — its objective rewards aliasing outcome-equivalent places regardless of spatial distance. Any capability needing metric queries (what's within radius r, what's the route from here to there) requires a dedicated spatial index. Deploy trigger: aliasing errors exceeding threshold, concentrated where outcome-similar contexts are spatially distinct. Pulled into existence by forward simulation (#12) rolling out trajectories through space. The append-only log includes position with every triple, so the index can be built retroactively over the agent's entire spatial history.

Once temporal association exists, position becomes one of the variables associated with future receptor states. The organism avoids a location not because pain is there now, but because entering it predicts pain. This is the transition from nerve net to ganglia — from responding to what's happening to navigating based on what happened.

7. Habituation and sensitization [DESIGN DECISION REQUIRED]

Receptor gain becomes adaptive.

Repeated harmless stimulation produces reduced response. Repeated harmful stimulation lowers the threshold. The organism saves energy on things that don't matter and sharpens its response to things that do. These are the two simplest forms of learning — even sea slugs have them.

Open question: habituation may fall out for free from certainty saturation — repeated observation → high certainty → reduced novelty response. If so, this is a detected emergent, not infrastructure. If the classic asymmetric habituation/sensitization dynamics are needed specifically, that's a built response-modulation rule.

8. Distance sensing [INFRA: sensor channel | PROMOTION: distance readings earn retrieval weight]

The organism detects field intensity beyond its surface — before contact.

This permits anticipatory movement but doesn't yet imply a full world model. Distant signals are initially additional receptor channels whose relationship to future contact must be learned. The transformer learns to fuse contact receptors and distance sensors into a single navigation policy. The sensor channel ships as infrastructure; distance readings earn influence when mappings conditioned on distance beat those ignoring it on held-out prediction.

9. Active broadband interrogation [EMERGENT from #10 + #11]

NOTE: This step is misordered in the numbered sequence. Active interrogation depends on prediction (#10) and curiosity (#11) — it cannot be observed until both prerequisites are promoted. It should not be a positioned step but a standing detector that activates whenever #10 and #11 are both live. Its fraction of actions selected primarily for expected certainty gain crossing a threshold is the detection criterion.

The organism emits a broad-spectrum pulse and reads the environment from what comes back. One emission mechanism, one return signal — the full spectrum of environmental information falls out of the frequency composition of the echo. Distance from return delay. Surface texture from high-frequency attenuation. Density and internal structure from low-frequency penetration. Movement from Doppler shift. Material composition from frequency-specific absorption. All from one emission, one return, one receptor channel.

This is qualitatively different from passive distance sensing. The organism puts something into the environment and reads what comes back — a closed loop the organism controls. The bat doesn't just sense better, it interrogates the environment. That's a different relationship to the world than passive reception.

The metabolic cost is the single emission pulse rather than maintaining multiple specialized sensor arrays — cheaper than the alternative and informationally richer. That's exactly what gets selected. But emission still costs energy, so the organism develops a targeting policy: ping toward uncertainty, not toward known regions. That's curiosity made physical, which connects directly to curiosity and confidence regulation (step 11).

The transformer handles this naturally. The echo return is a high-dimensional input — not spatially distributed across the body surface like contact receptors, but temporally distributed across the return signal. Different frequencies arrive with different delays and different attenuation profiles depending on what they bounced off. The organism doesn't need separate subsystems for light, color, density, and material strength. It reads them all as different signatures in the same return.

10. Prediction, accuracy receptors, and the causal mental model [INFRA: schema, log, encoder | PROMOTION: prediction beats baseline]

The organism predicts what its receptors will read next. A new receptor fires when predictions are correct. Prediction error becomes aversive. Accurate models feel good.

This is both the foundation for everything cognitive and the architectural transition point. The mental model separates from the transformer weights into an explicit database of cause-effect mappings: action → receptor state change, with time delay and certainty. Each mapping is a primitive — one action, one receptor, one observed delta. The transformer stops being the knowledge store and becomes the inference engine that queries it.

The separation is selected because implicit knowledge in weights cannot be inspected, edited, or transferred. Once prediction matters — once the organism is motivated to build accurate world models — the legibility and updatability of explicit mappings repay their complexity. Learning becomes row updates. Knowledge transfer between organisms becomes database replication. Two organisms' beliefs become comparable with a JOIN.

Every action-observation is written to the append-only experience log before updating any derived state. The log is the canonical record. Certainty scores, context embeddings, action family assignments, encoder weights — all are views over this log. No failure in any derived component can corrupt the ground truth. Resetting any component means recomputing from the log, not losing the organism's history. See "The Raw Experience Log" and "Observability and Recovery" above.

The contrastive encoder comes online here: a small learned component that shapes context similarity so retrieval returns mappings that are genuinely relevant to the current action in the current context. It is the one component that trains on a gradient. Action families emerge from clustering the raw log — actions that share outcome-response profiles covary into families, each attending to its own relevant context dimensions. Both are inheritable across generations.

Every mapping in the mental model terminates in a receptor state. This is the grounding guarantee. The mental model and the motivational system are the same structure. The organism doesn't consult its world model and then check it against its drives. The world model is the drives, expressed as predictions about what causes what to fire.

This is also where new receptors come online within a lifetime. When a prediction becomes reliably correct — when a particular internal signal consistently changes behavior in ways that improve receptor outcomes — that signal graduates from noise to receptor. The prediction becoming correct is the mechanism by which the organism's receptor topology grows. And if that discovery gets logged and passed to offspring as a topology bias (see "How Receptors Come Into Existence"), lifetime learning feeds back into evolutionary search. The organism doesn't just navigate the environment. It discovers what's worth sensing.

11. Curiosity and confidence regulation [INFRA: signal channels | PROMOTION: earns arbitration weight only when certainty is calibrated]

CAUTION: This is the most dangerous premature promotion in the sequence. Promoting an uncalibrated confidence signal into action arbitration makes the system act on its own miscalibration.

A receptor fires when the world model gets updated — when predictions fail and something new is learned.

This is in tension with the accuracy receptor: updating too readily makes the model unstable, not updating makes it stale. The balance that emerges from selection reflects what kind of environment was running. Volatile environments grow organisms with strong curiosity receptors. Stable dangerous ones grow organisms with strong accuracy receptors.

This maps onto real variation in cognition. People who find certainty rewarding versus people who find novelty rewarding aren't expressing arbitrary preferences. They may be expressing ancestral environments encoded in receptor topology.

Raw surprise alone can't drive this — unpredictable noise would become permanently attractive. What gets selected is learning progress: the value of a situation where uncertainty can actually be reduced.

The exploration drive and the contrastive encoder's training objective are naturally aligned. The intrinsic curiosity signal generates exactly what contrastive learning needs: the same action tried across deliberately varied contexts. An intrinsic bonus for revisiting an action in contexts the encoder currently considers dissimilar directly buys useful contrastive pairs at no architectural cost.

12. Pattern recognition and compression receptors

A receptor fires when the organism finds a simpler representation that still predicts accurately.

In the mental model, this is Level 3: recurring subsequences across multiple causal chains get recognized as reusable patterns. When the organism repeatedly traverses similar cause-effect paths to reach similar receptor outcomes, those paths get stored as compressed shortcuts. A concept is a pattern — a compressed causal chain that the organism has found reliably useful across different contexts. Analogy is pattern matching across chains: two situations are analogous not because they look the same on the surface but because their cause-effect chains share causal topology.

Compression that preserves predictive power feels good. This is the same receptor whether you're forming a concept, building a scientific theory, or finding an elegant explanation. But the same receptor also produces bias — because compression by definition throws information away. The receptor rewards simplification regardless of what got discarded.

Abstraction and bias are the same operation at different scales with different stakes. The difference between the concept "fruit" (useful compression) and "those people" (harmful compression) is whether the discarded information is consequential. Same receptor. Same operation.

13. Multiple hypotheses [VERIFIED — DETECTED EMERGENT]

Multiple hypothesis maintenance is present by construction. Verified empirically with the following results:

49% of action groups maintain multiple entries with genuinely diverse context embeddings (avg intra-group cosine similarity 0.593 — well below the 0.90 merge threshold, meaning the entries represent substantively different situations). 200/200 test steps retrieved meaningfully divergent predictions from the top-K mappings (avg pairwise delta divergence 2.75). The blended output differed from the top-1 prediction at every single step (avg difference 1.53). The system is genuinely using the uncertainty rather than suppressing it. The blend is doing real work.

72% of all mappings sit in the mid-certainty range (0.3-0.7). The system is living with ambiguity not because it hasn't seen enough data, but because the contexts are genuinely different enough that the same action produces different outcomes in different situations. That is calibrated uncertainty, not uninformed uncertainty.

No explicit multi-hypothesis machinery was built. The architecture generates this property from the interaction of: (1) the merge threshold (0.90 cosine) preserving distinct entries for the same action in different contexts, (2) the retrieval weighting (similarity × certainty) blending rather than selecting, and (3) the certainty update rule (Bayesian, not winner-take-all) maintaining the spread.

This validates the design principle: the right architecture produces capabilities without requiring them to be installed. Nothing was built for step 13. The architecture produced it.

When uncertainty remains high because several explanations fit the evidence, maintain multiple candidate models rather than committing to one. This is the most metabolically expensive receptor — holding several world models simultaneously is costly. It would only be selected under specific pressure: environments where premature commitment is consistently fatal. That is exactly the condition that produces scientific thinking.

Science isn't a cultural invention. It's what happens when multiple hypothesis receptors become fitness-positive.

14. Proprioception [COMPLETE]

The organism gains receptors for its own body state: speed, angular velocity, and limb angle deviations. It can now distinguish changes in the environment from changes in the configuration of its own body.

Proprioception produced the single largest accuracy improvement in the roadmap — 92.5% to 97.1%. The fast pathway jumped from 83% to 94.5%. The organism knowing its own body state (how fast it's moving, how it's turning, what shape its limbs are in) dramatically improved its ability to predict optimal actions. This validates the architectural claim: body awareness is not a luxury feature but a foundational capability that makes everything else work better.

15. Efference copy [COMPLETE]

Every action command (22 values for the default 6-limb organism: 18 limb muscle bits + 4 emission bits) is copied into the observation as a lossless efference copy. The organism can compare what it commanded with what actually happened.

Combined with proprioception, this enables the organism to learn "when I commanded action X and my body was in configuration Y, the outcome was Z." The efference copy + proprioception combination is what drives the accuracy jump — the model can now correlate its own motor commands with their sensory consequences.

16. Sensorimotor contingency [VERIFIED — SUBSTANTIALLY PRESENT]

Verified empirically. 4 of 5 diagnostic checks pass:

Within-group consistency: PASS — median within-group std 0.042. Same action produces very consistent body state changes. 100% of 169 action groups below the 0.15 threshold.

Linear predictability: PASS — mean R-squared 0.363. Actions linearly predict proprioceptive deltas. Rotation contingency is especially strong: R-squared(omega) = 0.786. Flex actions predict angular velocity with high accuracy.

Between-group distinctiveness: narrowly FAIL — 3/8 channels had F-ratio > 2.0 (needed 4). Speed is less distinctive between groups because many different action patterns produce similar speed effects (extending any limb moves forward). This is correct physics — speed is a many-to-one mapping, not a failure of contingency.

The organism has learned the regularities of having a body. The body stops being just a source of sensory data and becomes a predictable instrument.

17. Controllability and agency [COMPLETE]

The organism separates changes it caused from changes the environment imposed. At each step, the mental model predicts what the organism's action SHOULD produce (self-caused delta). The difference between the actual change and the predicted self-caused change is the external delta — what the environment did independently.

Controllability = clip(1 - ||external_delta|| / (||actual_delta|| + eps), 0, 1), where external_delta = actual_delta - predicted_self_delta. Range [0.0, 1.0] — the organism experiences contexts from fully passive (environment driving all changes, external residual dominates) to fully self-caused (predicted self-delta accounts for all observed change).

Agency is a learned causal distinction: some future receptor states are conditional on this organism's muscle firing. Not "I am a self," but "some changes reliably depend on this controller's output."

Notable training result: the slow pathway showed late-training activation (gate 0.000 at epoch 6, rising to 0.112 by epoch 30). The agency and planning signals created situations where the fast pathway needed the slow pathway's help — controllability/planning decomposition requires deliberation. The slow pathway is recruited 5-6% during inference. The self-model makes the organism think more.

18. Self-model and planning [COMPLETE]

The organism evaluates how much better its chosen action is versus doing nothing. Planning value = predicted receptor improvement from action vs null action, measured on pain and endorphin channels. Range [0.0, 0.951] in practice — the organism can strongly distinguish beneficial actions from inaction.

This is the organism running the world model with itself as a variable. Planning is not introduced as a separate faculty. It is prediction applied to a world model that now includes the organism as one of its causes. In the mental model, this is Level 2 — chains of primitives composed dynamically. The chain with the best predicted receptor outcome wins.

The self-model sequence is complete: proprioception (step 14) → efference copy (step 15) → sensorimotor contingency (step 16, verified) → controllability (step 17) → planning (step 18). The organism has body awareness, motor command awareness, action-consequence learning, causal decomposition of self vs environment, and forward evaluation of candidate actions. 96.3% accuracy with the full self-model, up from 92.5% pre-proprioception.

This is also where the self-recognition-in-others step becomes possible. Once you can model yourself, modeling others is cheap — you already have the template. Predicting another organism's behavior is just running your self-model with different initial conditions. In the mental model, this is the same chain evaluation with a different agent variable — same schema, different initiating actions.

19. Responsive environment and grounded proto-symbols

The environment becomes partially agentic. Some objects in it have input-output relationships — they respond to specific sound patterns by changing state. Each responsive object maps a set of sound patterns to behaviors: a particular emission causes it to move, open, emit chemicals, change temperature. Each "verb" the object responds to is a sound pattern that maps to an effect.

This is qualitatively different from everything before it. Up to this point the environment was a field — pain sources, endorphin sources, gradients. Passive. The organism acts, the environment responds physically. Now some things in the world can be addressed. The organism that learns to distinguish passive echo from active response has discovered that certain return patterns aren't just echoes — they're responses. The object did something different after the emission than it would have done without it.

The verb-to-sound mapping means language doesn't have to be invented from scratch between organisms. It gets bootstrapped from organism-environment interaction first. The organism learns that a particular sound pattern causes a particular environmental change. That's a grounded symbol — pointing at a real causal relationship — before any other organism is involved. The environment itself becomes a language teacher, without a teacher in the loop.

This also creates natural selection pressure for combinatorial signaling early. If single sound patterns map to single verbs, and the organism discovers that sequences produce compound effects, it has a reason to develop compositional emission before it has a reason to develop compositional language between organisms. Syntax emerges from tool use, not from social convention.

The organism already has controllability (step 17) and a self-model (step 18). It knows "my output causes changes." Now it discovers that specific outputs cause specific changes in specific external systems. That's the concept of a command. When social communication arrives in later steps, the organism already has a working concept of "signal causes change in external system." Extending that from objects to other organisms is a smaller leap.

20. Multi-organism environment and predator/prey pressure [INFRA: multi-agent environment + social signal channels]

Multiple organisms in the same environment. Competition for endorphin sources. Collision and resource depletion creating pain.

No communication yet — other organisms are just another field to navigate. But this selection pressure drives nervous system complexity more than anything else. Theory of mind becomes survival-critical. Predicting the predator's next move is just the self-model applied to an adversary.

21. Involuntary social cues [COMPLETE]

The organism detects the visible effects of the NPC's behavioral state — acceleration (speeding up or braking), angular velocity (sharp turns indicate evasion), and erraticism (sustained turning indicates distress). These signals have no intended meaning — the NPC isn't trying to communicate. They matter because they predict danger or future behavior.

The slow pathway recruitment increased to 4-8% during inference (the highest in the roadmap) — the social environment creates situations where the fast pathway can't handle the complexity alone. Reading another organism's behavioral cues over time requires the temporal context of the slow pathway. One organism's pain becomes environmental information for another, observable through its involuntary movement patterns.

22. Receptor propagation [COMPLETE]

Another organism's distress behavior directly affects the observer's own receptor state. When the NPC is nearby and in distress (high erraticism), the focal organism's pain increases proportionally — empathic aversion = proximity × erraticism × weight, added to all limb pain channels.

Receptor states are no longer private. One organism can alter another's internal priorities without physical contact. The group begins to develop a collective sensory surface larger than any individual. The organism is aversively motivated to move away from a distressed NPC — not because of collision (that's step 20) but because the NPC's suffering elevates the organism's own pain.

Pain receptors for others' pain is just empathy, stated mechanistically. Evolution selects hard for this because the information cost is low and the survival benefit is enormous.

23. Intentional signaling [COMPLETE] [INFRA: emission actuator + reception receptor | PROMOTION: receiver holds high-certainty mappings from emissions]

The NPC now responds to two emission patterns from the focal organism: repel (1,1,0,0) causes the NPC to flee (repulsive gradient, 15 steps), calm (1,0,1,0) dampens NPC erraticism (10 steps). Signal range 8.0 units. The oracle emits repel when NPC is within ~3.75 units, calm when NPC is erratic and within empathy range. Responsive object triggers take priority — communication has opportunity cost.

No new observation channels (OBS_DIM stays 149). The NPC's behavioral response is observable through existing channels (distance, speed, acceleration, erraticism). The mental model learns the causal mapping from emission bits (in efference copy) to NPC behavior change. Theory of mind, not telepathy.

Training: 97.4% accuracy, pred_mse 0.005. The slow pathway activated at epoch 13 (gate 0.001 -> 0.083) and stabilized at ~0.140 — the highest sustained gate in the roadmap. Signaling decisions require temporal context: "is the NPC close AND approaching? is it erratic AND nearby?" These multi-condition judgments recruit the transformer. The fast pathway alone reached 94.5%.

In 300-step episodes: ~85 repel signals, ~9 calm signals, NPC actively responding ~50% of steps. The organism learns to deliberately reproduce a behavior because that behavior changes what another organism does.

Communication begins when the predicted effect on the receiver becomes part of why the sender acts. The signal was incidental before. Now it is selected.

24. Simplified shared signals [COMPLETE]

The NPC now emits signals using the same vocabulary: (1,1,0,0) danger when near pain (intensity > 0.3), (0,0,1,0) resource when near endorphin (intensity > 0.3). The organism receives NPC emissions through 4 new observation channels (OBS_DIM 149 -> 153). The oracle responds to NPC signals: additional repulsion from NPC direction on danger (weight 0.5), mild attraction toward NPC on resource (weight 0.3) — competing with collision avoidance.

Shared vocabulary across three contexts: (0,0,1,0) means "endorphin" whether organism->object (step 19), or NPC->organism (step 24). (1,1,0,0) means "danger/repel" whether organism->NPC (step 23) or NPC->organism (step 24). 5 of 16 possible 4-bit patterns have stable meanings. The conventions transfer because the codes point to the same causal relationships regardless of who emits them.

Training: 97.2% accuracy, pred_mse 0.0046. Slow pathway activated epoch 17 (gate 0.063) and stabilized at ~0.135. The bidirectional signal channel creates situations where the organism must weigh social information (NPC says resource) against self-preservation (collision risk) — a richer behavioral repertoire than either signal alone.

Signals that are cheap, repeatable, and sufficiently accurate outcompete complicated ones.

Stable conventions emerge because both sender and receiver benefit from compressing a recurrent situation into a simpler transferable pattern. Words emerge as socially stabilized compressions of embodied models — simultaneously a cognitive compression and a social signal. Which is why language both enables precise thought and systematically distorts it. Same receptor. Two faces.

25. Optimism and goal persistence [COMPLETE] [INHERITED META-PARAMETER]

Optimism = clip(best_distant_endorphin - avg_current_pain, 0, 1). Goal persistence = EMA of optimism (decay 0.95). Two new observation channels (OBS_DIM 153 -> 155). When goal_persistence > 0.2, the oracle boosts the combined gradient by (1 + 0.3 * goal_persistence) and dampens temporal aversion by (1 - 0.3 * goal_persistence). The organism persists through moderate pain to reach predicted rewards.

Training: 97.1% accuracy, pred_mse 0.0044. Mental model grew to 14,924 mappings (up from ~11.8K in step 24) — the optimism-modulated behavior explores more action-state space as the organism traverses regions it would previously avoid. Slow pathway activated epoch 14, stabilized at gate ~0.128.

A receptor for a hypothetical world model weighted toward positive receptor states.

This is not a bias or an error. Without optimism, the organism only navigates away from pain. With it, the organism navigates toward something — which is a much richer behavioral repertoire. Organisms that can vividly represent a better state are more motivated to navigate toward it even through temporary pain.

Optimism plus self-model produces intention in the full sense. Not reacting to the current field, but acting to make a specific imagined future real.

That's will. Also just receptors.

26. Conflict receptor [COMPLETE] [DESIGN DECISION: detected condition + router integration]

Design decision resolved: detected condition (no new sensor — computed from existing receptor states) with explicit detection rule and router integration. The pure emergent version (actuator variance) was rejected for reliability — step 28 needs a stable signal.

Conflict = min(pain_pressure, reward_pull) + 0.5 * (goal_persistence * pain_pressure) + 0.3 * (empathic_aversion * (1 - pain_pressure)). Three conflict sources: pain vs reward (leave or stay?), optimism vs current pain (persist or retreat?), empathy without self-danger (care or ignore?). Router gains conflict as third input (Linear(2,8) -> Linear(3,8)) — high conflict pushes gate toward slow pathway. Oracle dampens gradient by (1 - 0.15 * conflict) when conflict > 0.3.

Training: 97.0% accuracy, pred_mse 0.0045, gate ~0.131. Slow pathway activated epoch 12 (gate 0.011 -> 0.046). OBS_DIM 155 -> 156.

A meta-receptor fires when strongly activated receptor pathways demand incompatible actions.

Conflict is not merely the simultaneous presence of multiple signals. It is the inability to satisfy them together. The conflict receptor's activation is proportional to the strength of the competing demands, the incompatibility of their preferred actions, and the cost of choosing incorrectly. Its firing recruits the slow pathway and makes the conflict state available for learning.

27. Context-conditioned arbitration [COMPLETE] [DESIGN DECISION: specific upgrade — learned arbitration head]

Design decision resolved: the architecture was already context-conditioned by construction, but arbitration was implicit in the linear layers. The upgrade: an explicit ArbitrationHead (2,597 params) that outputs 5 context-dependent weights for receptor groups (pain, reward, social, goal, meta). Range [0.5, 1.5] — modulates, doesn't gate. Feature-wise linear modulation (FiLM) applied to the observation before it reaches fast/slow pathways.

Training: 95.2% accuracy, pred_mse 0.0032 (lowest in roadmap). The arbitration head gave the fast pathway enough context-dependent weighting that the slow pathway gate stayed at 0.000 — the fast pathway with arbitration handles what previously required slow pathway deliberation. This is a clean demonstration that explicit receptor arbitration simplifies the decision process.

The organism learns which receptor should receive greater influence under which conditions. There is no universally dominant receptor. The arbitration system is itself governed by the same rule as every other capability: it survives only when its decisions produce better net receptor outcomes than fixed receptor priorities would have produced.

This is where the receptor economy begins to cohere. Receptors no longer merely compete through firing strength. The organism learns the conditions under which each form of motivation should be trusted.

28. Internal conflict model [COMPLETE]

Predicted conflict = current_conflict * (1 + 0.5 * (1 - mm_certainty)). When the world model is uncertain, the organism predicts conflict will persist or worsen. Conflict trend = EMA of conflict change (is conflict resolving or worsening?). OBS_DIM 156 -> 158. Oracle slows down when predicted conflict is high and worsening, presses forward when conflict is resolving.

Training: 95.1% accuracy, pred_mse 0.0027 (lowest in roadmap — the metacognitive signals improve prediction). The arbitration head (step 27) continues handling context-dependent routing through the fast pathway.

The organism can now predict its own future receptor conflicts. It simulates not only "what will happen if I take this action?" but also "which of my receptors will become active, which demands will conflict, and which priority worked in similar situations?" The receptor array has been incorporated into the self-model.

This is the beginning of metacognition: the organism models the motivations through which it models the world. But it does not escape the receptor system or become a detached executive above it. The apparent executive is itself a learned product of receptor conflict, memory, and selection.

29. Concepts and words [COMPLETE] [PROMOTION: verified | EMERGENT: 985 stable concepts detected]

985 stable concepts (certainty > 0.7, count > 10, compression_gain > 0.1) out of 4,004 total patterns. Avg concept quality 0.536 — compressed patterns predict 54% better than naively chaining individual steps. The environment supports rich conceptual structure. Concept match and quality are observable (OBS_DIM 158 -> 160). The oracle dampens gradient when a high-quality concept matches (the organism relies on learned abstractions).

The 5 shared signal codes ARE the first words: (1,1,0,0)="danger", (0,0,1,0)="resource", (1,0,1,0)="calm", (0,1,0,1)="approach", (1,0,0,1)="cool". Each points to a grounded causal relationship — a compression that is sufficiently accurate and sufficiently shared.

Training: 97.0% accuracy, gate ~0.126. The slow pathway reactivated at epoch 21 — concept matching recruits temporal deliberation. Concepts create situations where the fast pathway needs help recognizing patterns across time.

A concept is a compression that is sufficiently accurate and sufficiently shared. A word is a concept that can be transmitted between organisms via the communication receptors. Whether the compression is useful or harmful depends entirely on whether the discarded information was consequential.

30. Grounded language [COMPLETE]

Built a grounding dictionary mapping human-readable labels to the organism's internal states. 24 receptor groups mapped to observation indices. 5 grounded words in shared vocabulary: "danger" = (1,1,0,0), "resource" = (0,0,1,0), "calm" = (1,0,1,0), "approach" = (0,1,0,1), "cool" = (1,0,0,1). 8 fundamental terms grounded in receptor states: pain, danger, relief, self, other, will, uncertainty, conflict. 1,013 stable concepts — each a compression of a recurring causal chain the organism actually experienced. 50 top concepts traced and labeled.

Cultural transmission test: organism receiving another's mental model (database replication, not training) scored +383.2 reward improvement (+223%) over organism without cultural knowledge. The whitepaper's claim confirmed: one organism's mental model can seed another's directly.

Every concept traces through causal chains to receptor states. "Pain" -> obs[0:6] -> temporal_aversion -> pain_memory -> prediction_error. "Self" -> proprioception -> efference_copy -> agency -> planning_value. "Other" -> NPC behavioral cues -> empathic_aversion -> signal interpretation. The grounding is inspectable, legible, and tethered to the organism's own experience.

The language model is the roof, not the foundation. It sits on top of sensorimotor loops, pain memory, world models, self-models, social cognition, communication, and compression. Words like "pain," "danger," "relief," "self," and "other" now refer to stable internal and sensorimotor structures rather than only statistical relationships among human-produced tokens.

This is the thesis fully realized: a system that has concepts rather than one that uses them.

---

Evolutionary Infrastructure (Steps 31-34)

The roadmap stops prescribing cognitive capabilities and starts measuring which ones the environment selects for. The Genome Project (genome_project/) specifies 107 receptors across 11 families — a periodic table of cognitive capabilities. Steps 31-34 build the infrastructure to search that space evolutionarily rather than manually.

The key insight: the genome project is load-bearing on environmental design. Each receptor entry specifies what environmental structure it detects and what survival cost the organism pays for missing it. The 107 receptors are 107 environmental design requirements. The build order inverts: design environments FROM the genome, then run evolution against them.

31. Environment tiers (genome-driven) [COMPLETE]

5 tiers implemented in environment_tiers.py. Each tier includes all lower tiers' features.

Design environment configurations derived from the genome project's receptor families. Each tier embodies the causal structures that the corresponding receptor layer needs. The environment IS the curriculum — concepts are already in the causal structure, and survival must be contingent on distilling them.

Tier 0 — Foundation (current): Pain/endorphin fields, metabolic economy, one NPC, responsive objects. Sufficient for trunk receptors (~14 already present in steps 1-29).

Tier 1 — Temporal/Causal: Periodic field oscillations, causal chains between sources, delayed consequences. Selects for: rhythm, coincidence_detection, precedence_detection, causal_association.

Tier 2 — Multi-Agent Social: 3-5 NPCs with diverse behaviors (cooperative, competitive, erratic). Resource sharing requires coordination. Selects for: other_detection, behavioral_prediction, theory_of_mind, empathic_concern.

Tier 3 — Instrumental/Tool: Objects that can be moved and used as instruments (push onto pain source to block it, chain for compound effects). Selects for: tool_use, causal_chains, intervention_planning, environmental_manipulation.

Tier 4 — Abstract/Formal: Multiple environment instances with shared deep structure but different surface features. Transfer between instances is fitness-positive. Selects for: structural_similarity, analogical_reasoning, cross_modal_association, theory_formation.

Each tier includes all lower tiers' features. The tiers are the curriculum — the organism only develops the receptors the environment demands.

32. Receptor discovery mechanism [COMPLETE]

13 of 19 tested receptors discovered in Tier 0 environment: coincidence_detection, spatial_association, temporal_association, perceptual_similarity, pattern_recognition (854 stable concepts), compression_gain (0.456), other_detection, empathic_concern (41%), controllability, planning, optimism, conflict_detection, stress_detection. 6 not found (rhythm, precedence, causal_association, curiosity, change_detection, static_repetition) — these require higher-tier environments that create the selection pressure for them.

The organism scans its experience log for prediction patterns that match the genome library's emergence signatures. When a pattern's accuracy crosses a threshold AND the corresponding receptor from the library matches, the receptor activates — new observation channels come online.

Discovery happens at end-of-lifetime. The probe rate floor ensures discovery is honest — receptors are validated against probe-tagged data, not just policy-collected data.

33. Topology bias inheritance [COMPLETE]

5 generations run. Gen 0: 92.9% acc, 12 discovered, converge epoch 15. Gen 4: 95.3% acc, 14 discovered, converge epoch 0. Topology bias added 2 novel receptors (other_detection, rhythm) that Gen 0 missed. 12 receptors stable across all 5 generations (the trunk). Probe validation tracked at 12-13 per generation — no canalization detected. The three-part transgenerational defense worked: constitutional probe floor, probe-gated validation, canalization monitor all functioning.

Discovered receptors become a prior for offspring. Not direct inheritance — probabilistic bias, probe-gated. The three-part transgenerational defense: constitutional probe floor (not selectable), probe-gated inheritance (burn-in period validates inherited bias against probe stratum), lineage canalization breaker (cross-generation probe error monitor forces encoder reset when bias compounds).

The scaling track demonstrated that weight inheritance accelerates convergence (Gen 0: 94.7% -> Gen 2: 96.9%). Step 33 extends this to receptor topology — which receptors to activate, not just how to weight them.

34. Environmental sweeps + measurement [COMPLETE]

5 tiers x 3 generations swept. 12 trunk receptors invariant across ALL tiers (coincidence, spatial, perceptual_similarity, pattern_recognition, compression_gain, other_detection, empathic_concern, controllability, planning, optimism, conflict_detection, stress_detection). 3 tier-specific: rhythm (T0-2 only), temporal_association (T0-1 only), change_detection (T2-4 only — emerged with multi-agent pressure). 4 never emerged at any tier: static_repetition, precedence, causal_association, curiosity — need deeper environments (step 37).

Key finding: complexity doesn't just ADD receptors — it RESHAPES the topology. Higher tiers gained change_detection but lost rhythm and temporal_association. The receptor topology is a fossil record of the selection pressures that shaped it.

Run multi-generational evolution across environment tiers and measure:

• Which of the 107 receptors activate, in what order, at what generation
• Whether trunk receptors are invariant across environments (predicted: yes)
• Whether canopy receptors vary systematically (volatile -> curiosity, social -> theory of mind, instrumental -> tool use)
• Whether cross-family composition (Layer 2+ from the dependency graph) emerges when prerequisites are met
• Discrepancies between predicted (genome library) and observed emergence order — the most informative results

The empirical dependency graph vs the theoretical dependency graph. Where they disagree, the theory needs revision — and the specific failure point tells you where.

35. Complete genome coverage [COMPLETE — Phase 1]

21 new receptor tests added, bringing total from 19 to 40 tests (42% of 96 genome entries). 20 of 39 discovered in Tier 0: the original 13 plus 7 new — cross_modal_association (0.559), impulse_override (0.816), concept_formation (0.675), bias_as_compression (0.127), compression_receptor (0.231), absence_observation (0.216), exception_detection (0.035). The compression family is particularly well-represented: 5 of 7 compression tests pass. The remaining 56 receptors need higher-tier environments (18), population evolution (13), or are theoretical (20).

Build computable emergence tests for the remaining 88 receptors in the genome library (19 of 107 currently tested). Priority order follows the dependency layers: Layer 0 trunk receptors first (any remaining), then Layer 1 two-family compositions, then Layer 2+ canopy receptors. Each new test is a measurable prediction about what the environment must contain for that receptor to evolve. Tests that fail (the receptor should have emerged but didn't) are the most informative — they reveal where the theoretical framework's assumptions about selection pressure were wrong, or where the environment tier doesn't yet embody the causal structure deeply enough.

The genome library is only as good as its empirical coverage. A receptor with no test is a prediction that hasn't been checked.

36. Population evolution [COMPLETE]

8 organisms competing in the same environment, 5 generations. Fitness-proportional reproduction with topology bias inheritance. Empathy trending up (0.482 -> 0.555). Behavioral prediction emerged (0.022-0.030) — never appeared reliably in single-organism training. The social arms race created real prediction pressure that single-organism + oracle NPC could not.

Replace the single focal organism + oracle NPC with a population of N organisms (start with 8-16) competing in the same environment. Each organism has its own receptor topology, mental model, and policy. Reproduction is fitness-proportional — organisms with higher cumulative reward reproduce more. Offspring inherit topology bias (step 33) and weight priors (scaling track) from their parent.

The NPC stops being oracle-driven and becomes another evolving agent. Social family receptors (theory of mind, behavioral prediction, deception detection) get real selection pressure — you survive by modeling organisms that are themselves evolving to be harder to model. This is the arms race that drove social cognition in biology.

The key measurement: does population evolution produce receptor topologies that single-organism evolution cannot? Does the social arms race select for canopy receptors (nested theory of mind, cultural transmission, moral reasoning) that don't emerge under static NPC pressure?

37. Environment deepening [COMPLETE]

Deepened all 4 higher tiers. Tier 1: anticipatory reward (endorphin only during specific pulse phases) + conditional causal triggers (effect depends on energy). Tier 2: conditional cooperation (reciprocity-gated signaling), adaptive competition (NPC targets organism's preferred sources), shared resource depletion (-56% reward). Tier 3: gated reward (pain ring, must push object to create gap), compound tool effects (A+B = amplified endorphin). Tier 4: faster hidden variable sequence (period 150), transfer switch at step 150.

Reward drops prove the deepening works: Tier 2 dropped 56%, Tier 3 dropped 47%, Tier 4 dropped 50%. The environments are genuinely harder — failing to extract the target abstraction now costs the organism measurably.

Deepen each tier's causal structure until the target receptors are genuinely load-bearing for survival:

Tier 1: Causal chains that branch and converge — not just A→B but A→(B or C) depending on context. Hidden temporal structure that requires inference, not just pattern matching.

Tier 2: NPCs that model the focal organism — they predict what it will do and adapt. Deception becomes real: the deceptive NPC learns which false signals the organism falls for. Cooperation becomes conditional: the cooperative NPC only cooperates if the organism reciprocates. Resource sharing becomes a tragedy of the commons — selfish exploitation is tempting but collectively fatal.

Tier 3: Objects with non-obvious affordances — the organism must discover through exploration that object A can block pain, object B can amplify endorphin, object A+B creates a new effect. Causal depth: tool chains of 3+ steps. Persistent environmental modification that compounds across episodes.

Tier 4: Environments where the deep structure is genuinely hidden — the organism must form theories about unobservable causes and test them through intervention. Multiple environment instances that share structure but differ in surface features, requiring structural abstraction for transfer.

The test for each tier: does an organism without the target receptors actually fail in that environment? If an organism without theory of mind can navigate Tier 2 just fine, the tier isn't deep enough. The environment must make the target capability the difference between surviving and not.

38. Cross-tier transfer [COMPLETE]

5x5 transfer matrix across all tiers. Key findings: Tier 2 (social) universally benefits from transfer (11-25x from any source) — social cognition bootstraps from everything. Tier 4 (abstract) also benefits from all sources (1.7-2.2x). Tier 3 (instrumental/tool) resists transfer — tool use must be learned directly. T3 is the best source tier — instrumental experience produces the most portable receptor topology. Strong asymmetries: transfer follows the dependency graph upward (T0->T2 = 11.7x) but not downward (T2->T0 = 0.44x).

Evolve an organism in one tier for N generations, then transplant it to a different tier. Measure:

• Does the Tier 1 topology help in Tier 2? (Temporal/causal skills aiding social prediction)
• Does the Tier 2 topology help in Tier 3? (Social modeling aiding tool use discovery)
• Does the Tier 3 topology help in Tier 4? (Instrumental reasoning aiding abstract transfer)
• Does a Tier 0 organism fail catastrophically in Tier 3? (Missing receptors are actually fatal)

The prediction: transfer follows the dependency graph. Tier 1 receptors help in all higher tiers (they're prerequisites). Tier 2 receptors help in Tier 3 but not vice versa (social modeling is more general than tool use). Tier 4 receptors help everywhere once acquired (abstraction is universally useful).

Discrepancies between predicted and observed transfer tell you where the dependency graph is wrong — which is where the theoretical framework learns the most.

39. LLM grounding bridge [COMPLETE]

Grounded corpus (191 lines of receptor-grounded language), query interface (12/14 terms grounded — 'love' and 'justice' correctly identified as outside the organism's experience), 14 comparison prompt pairs for LLM testing. Every grounded term traces to specific obs indices, causal chains, and behavioral consequences. Ungrounded terms ('love', 'justice') are honestly reported as not yet traced to receptor states.

The grounding bridge: "What is pain?" Ungrounded answer: sensation, discomfort, neural signal (words about words). Grounded answer: obs[0:5] firing when limb tips contact pain field sources, causing temporal_aversion increase, pain_memory deposit, prediction_error spike (receptor states the organism actually experienced). Same question. Different corpus. The corpus is a body.

Connect a language model to the mental model database. The organism's grounded concepts (step 29: 1,013 stable concepts) become the language model's referents. Implementation:

• Export the mental model (causal mappings, pattern store, concept library) as a structured knowledge base
• A language model queries this knowledge base when processing words like "pain," "danger," "self," "other"
• Each word maps to a causal chain the organism actually experienced (step 30: grounding dictionary)
• The language model's outputs are grounded: "danger" doesn't mean "statistically associated with negative text" — it means "obs[0:6] elevated + temporal_aversion rising + optimal action is flee"

The test: ask the grounded language model "what is pain?" and compare its answer to asking a standard LLM the same question. The grounded model should answer in terms of receptor states, causal chains, and behavioral consequences — because that's what "pain" means in its knowledge base. The standard model answers in terms of other words.

This is the thesis made concrete. Same transformer architecture. Different corpus. The corpus is a body.

40. Whitepaper + publication [COMPLETE]

ABI_whitepaper_v2.md — 12 sections covering the full system. Added: Phase 8-10 roadmap coverage (steps 31-39), complete "What's Been Built" section (all empirical results), Genome Project section (13 families, 122 receptors), Academic Validation section (Barsalou, Friston, O'Regan, Trends in CogSci), updated "What Comes Next" with open-ended evolutionary search.

Update the ABI whitepaper to cover the full implemented system:

• Steps 1-34 complete (from pain receptors to environmental sweeps)
• Parallel Scaling Track results (4-12 limbs, 2 segments, 3D, generational inheritance)
• Genome project: 107 receptors, 11 families, emergence matrix across 5 environment tiers
• Key empirical results: 1,013 stable concepts, cultural transmission (+223%), topology bias inheritance (convergence 15→0 epochs), receptor discovery (13/19 in Tier 0), three-part transgenerational defense validated

The whitepaper becomes the thesis fully realized — not a proposal, not a roadmap, but a system that has been built, measured, and shown to produce grounded intelligence from the bottom up.

41. Run the laboratory

The prescribed roadmap is complete. The laboratory is built. Step 41 is running it — open-ended evolutionary experiments across the genome library, measuring what emerges.

The experiments:

• Multi-generational evolution across all 5 environment tiers with the full 40-test receptor discovery suite, tracking which of the 122 receptors activate at each tier and generation
• Population evolution (8-16 organisms) at each tier, measuring which receptors the social arms race selects for vs what solo evolution selects for
• Cross-tier transfer with deepened environments, testing whether the dependency graph predictions hold under real selection pressure
• Genome library expansion: as new receptors are theorized (completion, mathematics, organization), add their environmental requirements to the tiers and test whether they emerge
• Convergence testing: do independent evolutionary runs in the same environment converge on the same receptor topology? The theory predicts yes — the concepts are in the environment, not in any particular run
• Falsification search: which genome project predictions are wrong? Which receptors fail to emerge where predicted? Which emerge where not predicted? Each discrepancy is where the theory learns the most

The output is not a trained model. It is data about the evolutionary dynamics of intelligence: which receptor topologies emerge under which environmental pressures, in what order, at what generation, and why.

Step 41 has no completion condition. The laboratory runs as long as there are questions to ask it.

42. Revise the instruments [COMPLETE]

The first experiment (step 41) found 4 receptors that never emerged at any tier. Step 42 asked: are these genuinely absent, or are the tests wrong? 6 of 9 failing tests were measuring the wrong thing — the receptors were present, the instruments were broken. Causal association and curiosity were always there. The test suite grew from 40 to 46 tests.

43. Rerun with corrected instruments [COMPLETE]

Same experiment, better instruments. 46 tests, deepened environments. Results: 29 invariant receptors (up from 16 with broken tests). The trunk nearly doubled — not because anything changed in the architecture, but because the instruments could finally see what was there. 4 genuinely absent: static_repetition, precedence_detection, completion, quantity_detection.

44. Higher environment tiers [COMPLETE]

Tiers 5-7 built. Tier 5 (strategic social): modeling NPC, hidden sources, shifting alliances, deception cost — reward 125. Tier 6 (cumulative instrumental): persistent objects, sequential 3-step affordance, environmental memory, delayed rewards — reward 105 (hardest tier). Tier 7 (meta-cognitive): rule rotation every 100 steps, impulse traps with delayed pain, speed-dependent damage, strategy staleness penalty — reward 141. All 8 tiers pass.

The current simulation hasn't been pushed to its complexity limit. 122 receptors in the genome, 46 tested, 29 invariant — but the canopy receptors (nested theory of mind, causal graphs, necessity detection, organizational mirror) need environments more complex than Tier 4, not environments with different physics.

Build Tiers 5-7 by deepening what's already there:

Tier 5 — Strategic Social: NPCs that model the organism modeling them. Recursive social prediction. Information asymmetry — agents have private state the organism must infer. Conditional alliances that shift. Deception that adapts to counter-deception. Selects for: nested_theory_of_mind, belief_attribution, deception_detection, intention_recognition.

Tier 6 — Cumulative Instrumental: Tool chains that persist across episodes. Structures the organism builds that modify the environment for future generations. Objects with non-obvious affordances discovered only through systematic exploration. Compound effects that require 3+ step causal chains. Selects for: niche_construction, causal_graph_reasoning, intervention_planning, long_range_causation.

Tier 7 — Meta-Cognitive: Environments where the organism's own receptor topology affects outcomes — self-knowledge is load-bearing. Environments that shift rules periodically, requiring the organism to detect that its own strategies have become outdated. Self-regulation tasks where acting on impulse is consistently fatal. Selects for: metacognition, self_regulation, value_hierarchy, organizational_mirror.

The test: run evolutionary sweeps at Tiers 5-7 with population evolution. Do the canopy receptors from the genome project emerge when the environment finally demands them?

45. Expanded genome coverage

Build receptor discovery tests for the remaining 76 genome receptors. The current 46 tests cover trunk and branch. Tiers 5-7 create environments that demand canopy receptors — but canopy receptors need canopy tests. Extend the test suite to cover the Social canopy (nested ToM, cultural transmission, moral reasoning), the Agency canopy (niche construction, distributed agency), the Meta-Motivational canopy (metacognition, self-regulation), and the Mathematics/Organization canopy (necessity, proof, system detection, organizational mirror).

The test suite grows from 46 to ~90 tests. Each new test is a measurable prediction about what the organism should produce under sufficient environmental pressure. Tests that fail reveal where the genome project's predictions are wrong.

46. Benchmark competition

Take an organism evolved under deep time learning and pit it against task-optimized architectures on their own benchmarks:

• Image recognition: evolve an organism where visual pattern discrimination is load-bearing for survival. Compare transfer to novel visual tasks against a CNN trained on the same distribution. Prediction: the organism underperforms on the exact training distribution but outperforms on structurally related tasks it was never explicitly trained on.

• Pathfinding: evolve an organism where spatial navigation is load-bearing. Compare against A*/Dijkstra on static graphs and on dynamic environments with hidden structure. Prediction: less optimal on static known-topology graphs, more robust on dynamic environments.

• Social prediction: evolve a population where modeling others is load-bearing. Compare against theory of mind benchmarks. Prediction: the organism's social model generalizes to novel agents because it developed the receptors for social cognition, not a task-specific predictor.

The benchmark competition tests whether deep time learning produces more general intelligence than task-optimized learning — even when task-optimized learning wins on the specific benchmark.

47. Second simulation

Only after the current simulation has been pushed to its complexity limit. Build a fundamentally different simulation — different physics, body, reward structure. Run the same evolutionary process. Compare the trunk.

The receptors that are invariant across BOTH simulations are the strongest candidates for universal cognitive primitives. Map them against what neuroscience and cognitive science have identified. Convergence with biological findings is evidence the paradigm is discovering something real about the structure of intelligence rather than something specific to any one simulation.

One simulation is a data point. Two simulations with the same invariants is a pattern. Three is a law.

---

Parallel Scaling Track [COMPLETE]

All 8 scaling dimensions tested. Body plan parameterized: Organism accepts num_limbs, num_segments, dims. OBS_DIM and NUM_ACTIONS computed dynamically. Results in data/scaling_results.json and data/scaling_seg_3d.json.

1. Increase receptor count [TESTED]: OBS_DIM scales from 130 (4 limbs) to 234 (12 limbs). Training accuracy 92-95% across all sizes.
2. Increase muscle count [TESTED]: NUM_ACTIONS scales from 16 (4 limbs) to 40 (12 limbs). Architecture handles all action space sizes.
3. Add body segments [TESTED]: 2-segment articulated body (obs=238, actions=42). 92.8% accuracy — same as 1-segment. Joint flex actions with spring constraints.
4. Move 2D to 3D [TESTED]: 3D organism (obs=158, z-position + pitch). 92.7% accuracy — matches 2D. FieldSource uses 3D distance.
5. Organism diversity [TESTED]: Single model trained on mixed 4/6/8-limbed bodies: 96.6% accuracy. Architecture generalizes across body plans.
6. Generational inheritance [TESTED]: 3 generations with 0.8 parent + 0.2 random weight seeding. Gen 0: 94.7% (converge epoch 5), Gen 1: 96.3% (converge epoch 0), Gen 2: 96.9% (converge epoch 0). Topology bias accelerates search dramatically.
7. Complexity vs dimensionality [MEASURED]: Training accuracy stays 92-96% as obs grows 130->238. Mental model mappings scale proportionally (3K->14K). Behavioral complexity grows with receptor dimensionality without new top-down control.
8. Topology bias acceleration [MEASURED]: Convergence speed improves each generation. Inherited weights give immediate >93% accuracy vs 5 epochs from scratch. Parental bias accelerates search compared to random initialization.

The binary output space scales combinatorially. With a million binary receptor states, the number of distinguishable configurations is 2^1,000,000 — a number that makes the atoms in the observable universe look small. Complexity is not in the unit. It is in the combinatorial space the units collectively open up.

This is the scaling argument made visible rather than merely asserted.

---

Dependency Graph

The 30-step roadmap is not a sequence. It is a dependency graph whose nodes sort into four categories. The numbered order is a linearization — useful for reading, misleading for building.

Category Taxonomy:

SCHEDULED INFRASTRUCTURE — Substrate that either exists or doesn't. Deploy dates set by the roadmap. Nothing happens on deploy day beyond the mechanism existing.

GATED PROMOTIONS — Capabilities earned against a measurement the diagnostic stack already produces. The infrastructure half deploys on a date; the capability half earns influence when a specific metric crosses a threshold. Many roadmap steps are bundles containing both halves.

DETECTED EMERGENTS — Behaviors that should fall out of machinery already in place. Nothing to build, nothing to promote. The roadmap's job is to notice them. Their absence is a falsification signal about the underlying design. Building a module for an expected emergent destroys the evidence that the architecture works.

INHERITED META-PARAMETERS — Stable dials the evolutionary layer selects across generations rather than capabilities deployed or earned within a lifetime.

Rough tally: ~12 deployments, ~13 promotions, ~5 detections, ~3 inherited dials, ~7 design decisions. The numbered sequence of 30 steps is approximately these wearing a sequence costume.

Key dependency edges (prerequisites, not sequence):

#2 → #16a (budget source for attention allocator) #4 → #14 (reflex write-path needs hierarchical structure) #5 → #10 (can't predict before delayed links exist) #8 → #6b (distance readings as spatial index input) #10 → #11 (curiosity needs the mental model to query) #10 → #12 (forward simulation queries the mental model) #10 → #13 (self-model rows live in the mental model schema) #11 + #10 → #9 (active interrogation emerges here) #12 → #6b (trajectory rollout demands spatial index) #12 → #17 (route-based anticipatory avoidance) #12 → #25 (optimism: sustaining pursuit across delay via simulation) #13 + #26 → #28 (internal conflict model: self-model applied to conflict signal) #20 → #21 (other-model needs others to model) #23 + #24 → #29 (concepts and words: communication + compression) #29 → #30 (grounded language: concepts plus grounding guarantee)

Ordering contradictions in the numbered sequence:

Step 9 before Steps 10 and 11: Active broadband interrogation depends on prediction and curiosity but is listed before both. Resolution: remove #9 from the numbered sequence and make it a standing detector.

Step 6 before Step 8: Spatial memory is listed before distance sensing, yet the spatial index (6b) wants distance readings as input. The 6a branch (detected emergent) has no dependency on #8 and the ordering is harmless. The 6b branch deploys after #8 and is demanded by #12.

---

What This Means

Agency is a continuous variable determined by receptor complexity, effector complexity, and the processing in between. The question "does this system have agency?" is the wrong question. The right question is how many degrees of freedom exist between stimulus and response.

Current LLMs are a strange case: enormous processing in the middle, near-zero receptor and effector grounding on either end. High complexity coupling, almost no embodiment. They learned the words from beings who had the underlying states. There is no referent on their end.

This architecture builds the referent first. A system that has genuinely navigated away from pain, toward endorphin, through conflict, toward a predicted better state — that system has something that words like "aversion," "relief," "uncertainty," and "will" can actually point to. And the referent is legible: stored as explicit cause-effect mappings in a queryable database, not entangled inside opaque weights. You can read what the organism believes, compare it to what another organism believes, and trace every concept back to the receptor states that grounded it. When something goes wrong, you can diagnose it — the coupled processes have early warning signals, the raw experience log provides ground truth for attribution, and any derived component can be reset and recomputed without losing the organism's lifetime of experience.

The 2D organism in the liquid environment is the first rung. Everything above it is the same thing, applied at greater scale and complexity, shaped by greater selection pressure.

No mysterious leaps. Each step is a specific solution to a specific environmental pressure, with a receptor to make it happen.