CAKE: Connectome-Anchored Kinetic Emulator

Status: Draft spec v0.11 (2026-10-07). Version history and design rationale: spec/decisions.md.

Goal: Build a physically grounded, closed-loop simulation of a whole nervous system, constrained by the connectome. It should predict activity and behavior under conditions it was never fitted on, including perturbations, mutants and new environments, with error no larger than the variability between animals.

Summary

The connectome gives the wiring of a nervous system but almost none of the cellular parameters (channel densities, receptor kinetics, gap junctions, neuromodulation) that determine its dynamics. Fitting those parameters per neuron is hopeless at scale: ~10⁹ unknowns in the fly, with degenerate and sloppy solutions.

CAKE's central hypothesis (Regulatory Closure, H*) is that these parameters are not free. They are the outcome of activity-dependent regulation that each cell type runs, inside its circuit, inside its body: an equilibrium of that regulation, selected by each neuron's initial state and developmental history. CAKE therefore infers per-cell-type regulatory programs (~10⁴ types × tens of parameters) and lets the parameters of every neuron emerge from simulated development. Two companion hypotheses compute synaptic efficacy from EM morphology (H2) and use structured model errors to map invisible connections such as gap junctions and neuropeptides (H3). All three have pre-registered falsification tests.

CAKE is planned as a large open-source scientific code, organized like an Earth system model: physical components (neural tissue, chemistry, body, environment, observation) joined by a coupler, organisms as configurations, and a community benchmark (the Emulation Ladder, NeuroMIP and blind prediction rounds) as the measure of progress. It starts with C. elegans, where every approximation can be checked against an unapproximated reference, and moves to the fly CNS once H* is decided.

Documents

Document	Contents
1. Problem	The nervous system as a physical system; why the connectome underdetermines dynamics; state of the art
2. Hypotheses	H*, H2, H3; why they are plausible and new; the nested rival family; falsification tests F1–F7 with three-way outcomes, power requirements and the decision rule
3. Evaluation	The Emulation Ladder (E1–E6), normalized scores and task specifications, generalization axes, the Necessity Map, ablation vs. reduction
4. Physics	Physical layers L1–L8 at reference fidelity: cable, channels, synapses, gap junctions, volume transmission, ions, regulation (dynamics, equilibrium selection, identifiability), body and environment
5. Inference	The hierarchical generative model, fitting methods, identifiability, experiment design
6. Compute	Data model, fidelity lattice, execution model, regulation solver and its validity conditions, gradients, caching, error control, budget
7. Platform	Projects we learn from, architecture, CKL mechanism language, repository layout, standards, verification, organism releases, benchmarks, workflow, deployment
8. Community	Governance, working groups, contribution model, scale and timeline
9. Roadmap	Organism phases P0–P4; milestones M0–M7 for C. elegans
10. Risks	Risks and mitigations
11. Prediction and history	Scientific parallels and their limits; history-dependent latent state; A0 evidence; forecast horizons, assimilation cutoffs and calibration
12. Information budget	One observation-space tolerance (nats a dataset can detect) for reductions, solves, sampling and messages; Fisher-weighted compute allocation; sensor-targeted Markov state models; data-rate gate for trajectory tasks
13. Proposals	Proposed theories for the main bottlenecks (not current design); two issues found in the current spec
14. Enabling programs	Twelve standalone computational programs (data products, simulation infrastructure, inference tools) shared with the molecular compiler, WormSim and fly-brain; what those projects have already measured; literature novelty check, MVP and test with kill condition for each (proposals, not current design)
References	Prior work and bibliography
Decision log	Spec versions; every major design decision with rationale, rejected alternatives and revisit triggers

Reading paths

Scientists: Summary → 1 → 2 → 3 → 11 → 9.
Simulation and HPC engineers: Summary → 4 → 6 → 12 → 7.
Contributors and partners: Summary → 7 → 8 → 9.