A brain with continuity
Membrane potentials evolve. Synaptic currents decay. Neurons fire and recover. Spatial tuning and sensory cues shape the 128 filtered channels that enter the policy.
Inspect the circuit ↑A SMALL NERVOUS SYSTEM. A VERY LARGE WORLD.
A simulated rat, loose on the web.
Spiking. Exploring. Learning slowly.
CA: TBA
Enter the observation room ↓01 / OBSERVATION ROOM
An autonomous, ongoing simulation. You're here to observe.
The room is ready.
Connecting to the autonomous session.A motor decision may turn or rest; it does not necessarily open a page. Recoveries are automatic navigation resets.
02 / BRAIN & BEHAVIOR
Persistent neurons. Real spikes. A decision you can trace—and see which channels shaped it.
Absolute share = |group contribution| / Σ|group contributions|, excluding bias, for this chosen action. It describes score magnitude, not causal influence. Each signed contribution is filtered spike activity × learned policy weight.
The observed run is sensory-dominated. Its engineered priors weight 16 sensory channels; 112 spatial channels start at zero actor weight and can change through learning. The live shares above show the current balance.
Waiting for the live neuron sample. Rat anatomy: Waxholm Space atlas. Neuron positions illustrate functional populations; they are not measured cell locations.
Inspection changes only your view. The simulation continues independently.
TD updates change the policy weights. The follow-rate curve is session-only, calculated over up to the latest 100 decisions; it is an outcome measure, not proof that learning improved navigation.
No observations recorded yet.
LINK SELECTION · HEURISTIC · NON-NEURAL
Awaiting a recorded link decision.
This hand-coded rule ranks allowed DOM links using visit counts, page position, label length, and distance, then makes a reproducible draw. It is outside the neural model and does not read or understand the excerpt.
THE WORLD IT REMEMBERS
Its first encounters will appear here.
This is a persistent SQLite visit-count table that changes link scores. It stores page URLs, titles, timestamps, and counts; it does not store or understand page content.
03 / THE RECORD
Inspect the live decision record, current chain head, and public methods.
Read methods & evidence ↗Download behavior baselines ↗SOURCE PRIVATE DURING DEVELOPMENT · PUBLIC GITHUB PLANNED
The chain head and terminal-hash ledger are published by the same operator. They can expose later changes to a saved record, but they do not independently prove when the run occurred.
— recorded decisions
—Each decision commits to the record before it.
Loading the recorded observation, neural output, and action outcome…
The live record binds each saved observation, neural output, chosen action, and browser outcome to the prior decision with SHA-256. The public terminal-hash ledger preserves each segment's final head.
This is operator-published evidence, not independent attestation or biological validation. Source release and public replay tooling are planned for a later GitHub publication.
04 / THE FOUNDATION
RATBRAIN brings rat-inspired spatial encodings, a persistent spiking circuit, and online weight updates into one observable simulated animal. Filtered spikes feed a policy that drives real browser actions. Its state survives the next decision—and the next session.
Membrane potentials evolve. Synaptic currents decay. Neurons fire and recover. Spatial tuning and sensory cues shape the 128 filtered channels that enter the policy.
Inspect the circuit ↑New destinations produce reward; TD updates change action and value weights. The session outcome curve shows whether model-selected follows rise rather than assuming improvement.
Inspect the updates ↑Inputs, spikes, choices, and outcomes are recorded together. Inspect the live record and its self-attested terminal hashes; public source and replay tooling are planned.
Examine the record ↑A reduced, rat-inspired spiking circuit—not a whole-brain reconstruction. Neuron counts and physical units describe the simulation; anatomy comes from a rat atlas; placement of model units is illustrative. The methods specify each assumption.