RATBRAIN v0.4.2: Methods, Scope, and Evidence

RATBRAIN project maintainers · Independent experimental prototype · 11 September 2026 · Public contact not yet provided

RATBRAIN is a persistent spiking-agent prototype in a real browser. It integrates 3,145,729 individually stateful simulated units and records how their 128 aggregate channels feed a seven-action controller.

Abstract

RATBRAIN v0.4.2 is a closed-loop software instrument built to expose the passage from engineered observation to action. A constrained browser supplies DOM link geometry, position, heading, two screenshot statistics, boundary state, and prior experience. Those values drive 3,145,729 heterogeneous leaky integrate-and-fire (LIF) units connected by 12,582,916 deterministic directed edges. Spikes are aggregated into 128 functional channels and filtered before entering a seven-action actor–critic.

The neural kernel carries every membrane voltage, refractory counter, and synaptic current across decisions. It also persists 128 population filters, the learned policy and value parameters, eligibility traces, familiarity, fatigue, simulated time, and random-generator state. The implementation uses a compact C++ numerical kernel loaded by Python through ctypes; the behavioral controller, browser, evidence recorder, and public server remain small, readable Python modules.

The public observatory is read-only. It shows 8,192 sampled model units and recorded event trains from 256 raster units, while population statistics use spikes counted across all 3,145,729 units. A Waxholm Space rat-brain atlas supplies geometry for the visualization; simulated-unit placement and functional mapping are illustrative. Audited snapshots were dominated by the 16 sensory channels. The 112 spatial-channel weights began at zero and were near zero in those snapshots; later live attribution shows small nonzero spatial contributions. Online parameter updates run, but improved navigation has not yet been established against controlled baselines.

Scope and non-claims

RATBRAIN contains simulated states, not biological cells. Its atlas-derived shape uses real anatomy, while unit placement, channel assignments, recurrent wiring, and the count 3 × 2²⁰ + 1 are engineering choices. The non-round count exercises uneven channel boundaries and display mapping at scale; it is not an empirical estimate of a rat brain.

The system does not establish a whole-brain reconstruction, biological validation, rat-like behavior, language understanding, sentience, vision-based link discovery, or improved navigation from its spiking scale. Link selection is a separate recorded heuristic. Page excerpts are shown to the reader and are not model inputs. Current hashes and replay reports are published by the project itself and have no independent provenance anchor.

What would falsify the stronger mechanism claims

• If a 128-unit rate model with identical observations and priors matches behavior within a prespecified interval, the large spiking layer has no demonstrated behavioral advantage in that test.
• If disabling the 112 spatial channels leaves behavior unchanged, those channels should be described as simulated telemetry rather than a demonstrated navigation mechanism.
• If recurrence-disabled runs match intact runs, the 12,582,916 fixed edges add no measured behavioral value under the tested conditions.
• If held-out follow rate or reward does not improve over frozen priors, parameter updates are active but a learning benefit is not demonstrated.

1. The instrument we built

RATBRAIN is one continuous implemented chain. A real Chromium page supplies DOM geometry plus JPEG luminance and contrast. The browser adapter turns that state into a bounded observation. Spatial and sensory encoders provide 128 tuning values, each broadcast across more than twenty-four thousand LIF units. Population spike counts update 128 temporal filters, and those filters enter the actor and critic. A sampled motor action changes the browser or pose; its explicit result supplies the next parameter-update transition.

The observatory reads this same computation. It does not generate a second set of decorative spike counts. The population rates are calculated from every emitted event. The 8,192 displayed voltages, thresholds, and per-window spike counts are copied from evenly spaced simulated-unit indices. The raster is the event history of 256 model units. A digest of the full neural state travels with every model output, so even unrendered state contributes to reproducibility.

2. The closed loop and its two clocks

v0.4.2 cloud observation, 11 September 2026: ten distinct live states (decisions 84–100) reported 443–659 ms to compute each 20 ms neural window, with a 566 ms median: about 28 times slower than model time. This includes the controller, telemetry, and state hashing; browser work is additional. It is a short operational sample, not a controlled throughput benchmark. The observatory shows the current ratio beside the neural clock.

One normal decision integrates 20 ms of simulated model time as twenty exact 1 ms steps. At every millisecond the kernel visits the full population, applies membrane leak and current, resolves refractory state and threshold crossings, routes recurrent events, and updates the population filters. Browser capture, navigation, persistence, and rendering occur in wall time around that computation.

These clocks are reported separately. In an earlier v0.3 local measurement, five production-path samples had a median of 162 ms for a 20 ms window, about 8.1 times slower than model time before browser work. On the resized v0.3 cloud worker, 16 sampled states measured 335–452 ms per 20 ms window (median 366 ms), about 18 times slower. These are operational measurements, not controlled v0.4.2 throughput results, and the system makes no real-time claim.

1. The browser finds visible, policy-eligible links and refreshes the tracked target’s screen coordinates. New persistent destinations receive a selection bonus; known destinations remain eligible.
2. The adapter combines target direction and distance with page-relative position, heading, scroll state, visible-link density, luminance, contrast, blocked state, prior outcome, and model time.
3. Analytic spatial and sensory functions produce 128 bounded tuning values. The values are broadcast to the corresponding neural populations with per-unit gain variation.
4. The native kernel simulates all 3,145,729 units and 12,582,916 fixed connections for 20 steps. Full-population spikes update the 128 filtered rates.
5. The actor–critic samples forward, turn left, turn right, sniff, scroll, retreat, or rest from the displayed probability distribution.
6. The browser adapter attempts the selected action. The actual outcome, resulting observation, model output, and state hash enter the live view and evidence record; the next step updates parameters from that transition.

Supervisory recovery stays outside model credit. After a navigation failure or dwell ceiling, a supervisor may move to an approved seed. The following observation clears the pending learning transition and traces and suppresses intrinsic novelty for that step. The event and outcome remain visible as supervision.

3. Encoding motifs borrowed from rodent spatial neuroscience

Spatial and directional encoding

The controller begins with three well-established motifs in rodent spatial neuroscience. O’Keefe and Dostrovsky reported hippocampal units related to a freely moving rat’s position.¹ Hafting and colleagues described medial entorhinal cells whose firing fields form a triangular grid.² Taube, Muller, and Ranck characterized postsubicular cells tuned to head direction.³ RATBRAIN uses localized position, multiscale periodic position, and circular orientation as functional encoding motifs.

The analytic forms follow compact spatial-neuron models documented by RatInABox.⁴ The web page becomes a normalized two-dimensional environment. Sixty-four place channels use an 8×8 center lattice with deterministic page-specific remapping. Thirty-six grid channels span three spatial scales and twelve phases. Twelve head-direction channels tile the circle. Sixteen sensory and internal channels represent target geometry, visible-link density, screenshot luminance and contrast, scroll position, boundary and blocked state, novelty, familiarity, fatigue, and bias.

These functions are externally evaluated tuning inputs. They do not establish path integration: the adapter supplies normalized pose and heading. Link locations are DOM-derived engineered affordances, while luminance and contrast come from the latest real JPEG. The model neither discovers links from pixels nor interprets their words. This explicit observation boundary makes the circuit’s contribution testable.

4. The 3,145,729-unit circuit

The v0.4 neural substrate contains exactly 3,145,729 individually updated LIF units. The intentional engineering count is 3 × 2²⁰ + 1; it tests non-round channel boundaries and display mapping at scale and is not an anatomical estimate. Contiguous index ranges divide the units across 128 tuning channels, so neighboring channel sizes differ by one. Each unit receives its channel’s current tuning with deterministic gain and threshold variation derived from its index and seed.

Every source has four fixed outgoing connections within its own tuning channel. Postsynaptic indices are computed from ring offsets 1, 19, 127, and 1,021, so the 12,582,916-edge graph is exact but requires no adjacency list. Source IDs divisible by five deliver −7 mV; the remaining sources deliver +2.5 mV. Across IDs 0 through 3,145,728, this gives ceil(3,145,729 / 5) = 629,146 inhibitory and 2,516,583 excitatory sources. The 80/20 partition and ring wiring are transparent engineering choices, not a reconstructed rat connectome.

The update follows the standard LIF structure: voltage leaks toward rest, drive and synaptic current move it toward threshold, a crossing emits a spike, and voltage resets for a refractory interval.⁵ The Brian 2 neuron and synapse tutorials provide independent executable references for threshold/reset/refractory dynamics and event-driven synaptic changes.⁶⁷ RATBRAIN implements the release kernel directly in standard-library C++ and exposes a narrow C interface to Python.

The readout boundary is z. All spikes in each functional channel contribute to its filtered rate, and the resulting 128 values enter the actor and critic. Analytic tuning has no direct path around this filter. Audited action-attribution frames were sensory dominated. The spatial weights began at zero, were near zero in those frames, and later produced small nonzero group contributions; whether they materially affect behavior remains an ablation question. The display sample reduces transport and drawing cost; it does not change the simulated population.

The raw neural-state payload is 28,312,105 bytes: a validated header, 3,145,729 voltages, 3,145,729 synaptic currents, 3,145,729 refractory counters, and 128 filtered rates. Export appends a 32-byte SHA-256 digest, producing a 28,312,137-byte neural checkpoint. The seed reconstructs static thresholds, gains, population membership, connectivity, and sample indices. Restore validates the digest, format, size, model count, degree, seed, finite ranges, and refractory bounds before accepting state.

V3 migration validates the old checkpoint, including its native binary, before replacing that population with a newly seeded v0.4 population. Controller weights and biases, RNG, memory, model time, and learning-update count are preserved. Native neural state and its clock, the pending transition, and eligibility traces are cleared; metadata discloses neural_state_initialization: reinitialized_for_v4.

5. Learned policy and persistent familiarity

The earlier 350 ms input filter kept an old target direction active across several turns, and after about 20,000 updates the learned policy had drifted into a turning loop. v0.4.1 shortened the filter to 20 ms and reset the actor and critic to engineered priors once under learning revision 1. Neural state, page memory, random-generator state, model time, and lifetime update count were retained; the reset boundary is recorded. Subsequent restarts retain later parameter updates.

The controller above the kernel remains a seven-action softmax actor and linear critic. Sparse initial weights express useful geometric priors: target bearing supports turns, alignment supports forward motion, proximity supports sniffing, a page boundary supports scroll, and blockage supports retreat. The seeded random generator then samples the complete displayed distribution; no supervisor replaces an inconvenient model action.

This is online temporal-difference updating with eligibility traces.⁸⁹ The model reports action logits and probabilities, sampled action, value, external and intrinsic reward, TD error, update count, and update magnitude. “TD error” identifies an algorithmic quantity; it is not a measurement of dopamine. Passing update checks establishes that parameters change, not that navigation improves.

For the chosen action, action_drive exposes 128 signed channel contributions, each equal to that channel’s learned weight times its filtered spike feature. It also reports four group totals, bias, chosen logit, and six dominant channels with their feature and weight. Summing the 128 contributions and bias reproduces the chosen logit.

Two memories have distinct roles. The controller’s bounded 256-context store keys experience by canonical page, stable target identity, and quantized page position. Familiarity grows on encounter and decays in model time; novelty contributes a small intrinsic reward. A memory ablation removes immediate familiarity and its residual filtered features. Separately, SQLite records canonical pages, titles, timestamps, and visit counts across processes. Durable page familiarity changes target priority and first-discovery reward while session coverage remains a separate statistic.

These mechanisms update action-value parameters, repeated-context familiarity, and route preferences. They do not extract propositions or store English facts. Whether the updates improve navigation must be measured against frozen priors on held-out tasks.

6. An anatomical observatory

The three-dimensional brain view derives its exterior and ten gross regional meshes from the Waxholm Space atlas of the Sprague Dawley rat brain v4.01 (RRID:SCR_017124). The underlying reference is a 39 μm isotropic ex vivo MRI atlas of an 80-day-old male Sprague Dawley rat; the version 4 publication describes full-brain coverage and 222 anatomical structure annotations.¹⁰ The included generator reads the v4.01 NIfTI labels and affine, extracts atlas-derived surfaces, preserves point label IDs and coordinate transforms, and exports a compact GLB.

For live activity, the generator uniformly samples 8,192 locations inside labelled atlas tissue with a fixed seed. Those locations provide an anatomical canvas for the equally sized simulated-unit sample. Each point is tinted by whether its signed channel contribution raises or lowers the chosen action’s logit. The 256 raster units use recorded events from the latest 20 ms model window; the viewer replays that window once over 900 ms and does not synthesize firing or loop stale activity.

The display also draws 256 representative recurrent connections using the exact offset-1,021 target formula. Their endpoints use nearest sampled illustrative atlas positions, so the lines are circuit readouts rather than biological projection claims. Atlas sample coordinates are ordered by coarse atlas group and lateral position to make model channels visually coherent. Placement of the 128 functional channels remains schematic. The atlas constrains surface and regional volume; the points are not measured cell coordinates, and the functional colors and connections are not atlas claims.

The moving animal is an original organic, translucent procedural digital rat, replacing the earlier armored robot presentation. It has a short curved muzzle, smooth rounded ears, and separate small eye and nose materials. Its editable Blender scene, generator, and 267,920-byte GLB contain seven mesh nodes, eight primitives, 6,542 exported vertices, and named body, head, leg, and tail pivots for runtime movement. The asset is released under CC0-1.0. Brain-signal color and recorded motor outcomes illuminate its body illustratively; the body points are not peripheral or body neurons.

7. A real browser under a narrow motor contract

RATBRAIN launches Chrome locally and bundled Chromium in the cloud, always in a fresh nonpersistent Playwright context. Browser contexts isolate cookies, local storage, and session storage.¹¹ The agent receives no user profile, credentials, or existing session. Service workers are blocked, downloads cancelled, additional pages closed, and dialogs dismissed. Request routing rejects methods other than GET and HEAD and checks every main-frame destination, redirect, and final URL.¹²

The model never clicks arbitrary scriptable elements. It may follow only the canonical URL of an already validated visible target, using page.goto. Live navigation is limited to narrow reading paths on exact approved hosts, without credentials or alternate ports; subresources have a separate policy. Explicit calibration mode serves only local fixture pages and never silently replaces a failed live page.

Turn changes heading by 0.3 radians. Forward advances pose and pointer. Scroll moves the real viewport. Sniff follows the tracked target only inside the reach threshold. Retreat uses validated history. Rest leaves the environment unchanged. Every outcome records whether execution occurred, action kind, pose and URL before and after, movement distance, and scroll displacement where relevant.

Non-neural link-selection heuristic

The link-selection heuristic operates above the spiking motor controller. It scores allowed visible links using recorded novelty, recent encounters, content placement, label length, and travel distance, then samples a distribution reproducibly. This rule is hand-written and outside the neural model. The selected target remains committed while the motor policy approaches it. Recovery is explicitly recorded supervision and never increments model-selected follows.

Each decision publishes candidate inputs, scores, probabilities, deterministic draw, selected destination, and a computed scoring explanation. A short page excerpt is shown to the viewer but is not a model input. The private verifier recomputes the choice from captured inputs. Nothing in this layer demonstrates semantic comprehension.

Session counters distinguish motor decisions, visited pages, successful follows and scrolls, and automatic recoveries. Learning-update totals and durable page encounters continue across sessions. A turn or rest consumes a motor decision without opening another page.

Viewers observe; operators choose the process

The public site has no steering controls. Every HTTP mutation method returns 405, including loopback requests. Mode and process lifetime are terminal or deployment choices. Read-only endpoints expose current state, JPEG frame, server-sent events, model metadata, proof summaries, and the code-free terminal-head ledger. Source and replay downloads remain private. A long worker token protects worker data while the public health response reveals only service status.

8. Evidence and replay

Source snapshots and replay artifacts are retained privately until the planned GitHub release. The public /api/proof/heads ledger exposes code-free terminal hashes, counts, and segment continuity. It is useful for detecting later changes relative to a recorded head, but it is published by the same project and is neither independently signed nor independently timestamped. Independent replay is not currently possible without access to the private source and retained segment.

Decision evidence stores large neural state efficiently. A segment manifest contains compact controller state and commits to a raw neural.bin snapshot of the 3,145,729-neuron kernel. Source entries include the exact Python model, C++ kernel, native wrapper, recorder, and browser adapter with their SHA-256 hashes. Each decision payload contains its complete observation, external reward, ablation, brain output, neural state digest, explicit motor outcome, frame digest, and metadata.

Segments rotate after a bounded record count without resetting the brain. A continuation manifest commits to the preceding segment’s run ID and head hash. Canonical JSON uses UTF-8, sorted keys, compact separators, and rejects non-finite numbers. SHA-256 follows the standardized digest specification.¹³

The standalone verifier checks source and neural-state hashes, the immutable anchor, sequence and run continuity, every chain link, record count, and final head. It builds or loads the snapshotted native kernel, reconstructs the starting controller and raw neural state, reruns each recorded step, and requires the complete canonical output to match. It remains able to inspect older v0.1/v0.2 evidence that embeds its smaller state directly. The preserved v0.3 production evidence targets native Linux x86-64: an Apple ARM Docker-emulation check reproduced the full neural-state hash but differed by 2 × 10⁻¹⁸ in one controller probability. The strict verifier rejected that difference; matching container versions alone does not establish cross-CPU bit identity.

Private replay checks deterministic model computation and chain integrity relative to an included source snapshot, starting state, compatible runtime, and chosen head. It does not independently attest earlier page pixels, browser effects, run time, or biological accuracy. Without an external signature or timestamp, an operator could replace an entire manifest and chain. The present public record is therefore self-attested.

9. Software checks and measured evidence

A private native Linux x86-64 replay of the first three deployed v0.4.2 decisions reproduced all three neural outputs and four heuristic-selection records exactly, including the new current telemetry. Captured source hashes matched this release. This was an operator-run software check; source and state remain private until the planned GitHub release.

Software checks cover release-scale metadata, a 3,145,729-unit smoke step, the one-millisecond clock, population aggregation, recorded raster bounds, export/restore, tamper rejection, policy arithmetic, parameter updates, browser motor consequences, and deterministic continuation. Passing them establishes implementation behavior. Scientific claims require controlled interventions and held-out behavioral comparisons.

A privately retained v0.4 segment contains two decisions and occupies 11,038,337 bytes. Self-attested head: 387fc2af020aa99ccca9c58a1546c9c4b2caa0affc95070eeccf89b7f4e90432. Archive SHA-256: ac640ddf6b89c2752624f8d3925c44449d1d2081ffdb889b461e9f6f4c0dcae8.

10. Research and validation roadmap

1. Extend dynamics measurements. Measure rate distributions, refractory occupancy, E/I sensitivity, and filter stability across more observations and seeds.
2. Extend paired baselines. Add spatial-channel and recurrence-disabled interventions, larger seed counts, Chromium fixtures, and held-out page graphs under equal decision budgets.
3. Quantify scaling. Report initialization, step time, memory, checkpoint, and replay measurements across unit count, degree, CPU architecture, and compiler, with model and wall time separate.
4. Test held-out environments. Freeze parameters, introduce unseen page graphs and layout changes, and measure follows, unique destinations, revisits, action entropy, and supervisor recovery.
5. Connect anatomy cautiously. Replace illustrative ensemble placement only when cell-type, projection, or region-level data support an explicit mapping and controlled experiment.
6. Extend provenance. Publish independently signed or timestamped heads and retain selected referenced frames.

11. Primary sources and standards

1. O’Keefe, J.; Dostrovsky, J. The hippocampus as a spatial map: preliminary evidence from unit activity in the freely-moving rat. Brain Research, 1971, 34(1):171–175. DOI 10.1016/0006-8993(71)90358-1.
2. Hafting, T.; Fyhn, M.; Molden, S.; Moser, M.-B.; Moser, E. I. Microstructure of a spatial map in the entorhinal cortex. Nature, 2005, 436:801–806.
3. Taube, J. S.; Muller, R. U.; Ranck, J. B. Head-direction cells recorded from the postsubiculum in freely moving rats. I. Description and quantitative analysis. Journal of Neuroscience, 1990, 10(2):420–435. DOI 10.1523/JNEUROSCI.10-02-00420.1990.
4. George, T. M. et al. RatInABox, a toolkit for modelling locomotion and neuronal activity in continuous environments. eLife, 2024, 13:e85274.
5. Gerstner, W.; Kistler, W. M.; Naud, R.; Paninski, L. Neuronal Dynamics, §1.3: Integrate-and-Fire Models. Cambridge University Press online edition.
6. Brian 2 project. Introduction to Brian, part 1: Neurons. Official threshold, reset, and refractory tutorial.
7. Brian 2 project. Introduction to Brian, part 2: Synapses. Official event, weight, delay, and synaptic-state tutorial.
8. Sutton, R. S. Learning to Predict by the Methods of Temporal Differences. Machine Learning, 1988, 3:9–44.
9. Sutton, R. S.; Barto, A. G. Reinforcement Learning: An Introduction, second edition. MIT Press, 2018.
10. Kleven, H. et al. Waxholm Space atlas of the rat brain: a 3D atlas supporting data analysis and integration. Nature Methods, 2023, 20:1822–1829. Atlas files: official NITRC project.
11. Microsoft. Playwright Python: browser context isolation. Official documentation.
12. Microsoft. Playwright Python: network interception. Official documentation.
13. National Institute of Standards and Technology. FIPS PUB 180-4: Secure Hash Standard. SHA-256 specification.