Research mode.
Biology, models
and explicit assumptions.
The colony runs a JavaScript adaptation of selected published neural equations, not a complete bee brain. Its 50 individuals are separate numerical models, not demonstrated conscious minds. We distinguish biological evidence, computation and our software policy.
Recorded activity / a separate experimental layer
The Neural View includes the original FU Berlin ALF-1 reconstruction (170,494 source triangles) alongside the ACT atlas. Coordinates and topology are preserved; duplicated source geometry is not counted as separate recorded cells.
Strube-Bloss & Rössler (2018), CC0 recordings supply 92 recorded units across 23 animal filename IDs and six stimulus conditions. The viewer preserves 449,208 original spike timestamps within windows from −1 to +4 native source units around all supplied stimulus markers. Select an animal, stimulus and repeat, then play or scrub the raster.
The source description does not explicitly establish the timestamp unit. Playback uses one source unit per wall-clock second as a display rate, not verified biological real time. These recordings are not registered to ALF-1: channel positions are a raster layout, not anatomical coordinates. No experimental impulses are projected onto unrelated reconstructed neurons. This is recorded biology, not live BNB activity or a complete bee connectome.
00 / Apis mellifera brain anatomy
The familiar “roughly one million neurons” description follows Menzel (2012). It is an approximate contextual figure, not the size of this computational core. Counts vary by preparation and method; newer whole-brain cell-count estimates differ and should not be treated as identical neuron-count measurements.
Habenstein, Grübel, Pfeiffer & Rössler (2023) demarcated 35 neuropils and 25 fiber tracts in the honeybee cerebrum. Here “cerebrum” excludes the optic lobes. This anatomical atlas does not specify all neuron-to-neuron connections or their dynamics.
The Neural View now loads original Honeybee Standard Brain ACT geometry from FU Berlin (Brandt et al., 2005; DOI 10.1002/cne.20644). Original coordinates and topology are preserved: 74,364 vertices and 148,120 triangles. Mesh object IDs are retained without guessed neuropil assignments. No calculated neural activity or generated wiring is overlaid. The flying x-ray bee artwork elsewhere remains artistic, not microscopy.
| Biological system | Observed role | Digital analogue / current status |
|---|---|---|
| Optic lobes | Visual processing; lamina, medulla, lobula | Visual perception adapter planned; inactive |
| Antennal lobes | Primary olfactory processing | Numeric transaction encoder active; not an AL neural model |
| Mushroom bodies | Learning, memory, integration | Sparse familiarity model active; economic outcome learning not connected |
| Central complex | Orientation and navigation | Heading subcircuit active; environment-graph navigation incomplete |
| Waggle dance | Resource-location communication | Software recruitment protocol active; not a biological dance simulation |
01 / Executable foundation
Sun, Yue & Mangan (2020), A decentralised neural model explaining optimal integration of navigational strategies in insects. The authors provide open model code under the MIT license. Our adapter is pinned to commit 470c7d783dc5678e22170f582b09a72b265071cb.
| Component | Implemented in the colony | Boundary |
|---|---|---|
| Mushroom Body | 81 projection inputs, 4,000 thresholded KC units, one plastic novelty output; rectified PN and EN. | Insect visual-familiarity model adapted to numeric chain features. Not a measured honeybee wiring matrix. |
| Learning | KC→EN weights start at 1. Active connections depress by 0.2, clipped at 0, when an observation is stored. | Familiarization, not verified profitable behavior. No outcome-reward oracle. |
| Central Complex | 16 TL2, 16 CL1, 8 recurrent TB1 rate units. Sourced sigmoid slopes, biases and sinusoidal inhibitory connectivity. | Heading circuit only. No CPU4 path integrator or complete CPU1 motor pathway implemented. |
| Individuality | Distinct seeded projections, threshold, state and memory for each Bee. | Engineering parameter variation, not 50 biological specimens. |
02 / Declared modifications
We use the author's alternate fixed-10-unique-PN-per-KC generator, rather than their constructor's default fan-out generator. A seeded JavaScript PRNG replaces Python's RNG. Thresholds vary from 6.2 to 6.6; learning rate is 0.2. These settings are ours, not an experimentally fitted bee parameter set. Novelty is the EN output divided by active KC count (zero when no KC is active).
The compass preserves TL2 = sigmoid(6.8 cos(θ−preference)−3), CL1 = sigmoid(−3 TL2+0.5), and TB1 = sigmoid(5(0.667 CL1 input−0.333 recurrent inhibition)). Ten update steps per second are interface time, not biological time. Flight positions and target selection are an engineered spatial metaphor, not physical navigation through a blockchain.
03 / The BNB adapter
A read-only server endpoint verifies chain ID 56 and samples up to the first 16 transactions from a block three blocks behind the latest head. This is neither a complete chain index nor a finality guarantee. Hexadecimal address/calldata features plus gas, value, nonce and input-size features become 81 numeric inputs. Input features are not an understanding of a protocol, audit result or financial-risk rating.
Each Bee processes these features separately. Shared RPC observations are not independent evidence sources. Novel objects recruit peers; a software rule records a local watch decision after at least eight distinct votes and at least 75% support. The rule uses model novelty, not biological consensus equations. No transaction is submitted.
This deployment runs one shared, always-on read-only colony. Learned weights are atomically checkpointed on the server and runtime events are appended to an operator-managed log. Closing or backgrounding an observer does not stop the worker. Export retrieves the full current checkpoint; the browser receives a reduced presentation view during normal observation. A restored checkpoint still requires fresh chain observations before processing resumes.
04 / Actual bee connectivity: the next research track
Sayre et al. (2026), Deep conservation of head direction circuits in bees, ants and flies is a preprint with a public repository of synapse tables, extrapolation code and rate models. Its bee model uses the sweat bee Megalopta: regional synaptic data are extrapolated to a full head-direction circuit. That is not a complete Apis mellifera connectome. This dataset is identified for further work; it is not loaded into the current colony.
Stone et al. (2017) supplies the anatomically constrained navigation foundation used by the Sun model. Open honeybee MBON recordings provide experimental responses, not a full synaptic network. Anatomical atlases likewise do not by themselves specify executable dynamics.
05 / Still required for the full product
Actual ecosystem graph exploration; protocol/DEX/liquidity and website adapters; reproducible ingestion and independent evidence verification; a calibrated feedback objective; permissioned execution contracts, spending budgets and transaction receipts. These are not represented as complete. The current votes authorize only local observation, never spending.
The current Genesis colony has 50 models. Each model keeps its own projections, learned weights and experience. The contribution score is 100 × (3 × discoveries + recruitments + retained observations) / max(1, 5 × observations), capped at 100. Each 600-step active-runtime epoch allocates 10 + floor(score/10) inspection credits. Credits are software work quotas, not money or a biological fitness measure. Parameter mutation and resource-funded population growth are not active.
The flying specimens are artistic x-ray-style imagery. Brain positions and filaments in the artwork are illustrative and should not be used as scientific anatomy.
06 / Deployment and reproducibility
The VPS deployment connects a shared worker with atomic checkpoints and an append-only event log. Browser-local mode remains a fallback for deployments without that runtime configuration. No private key is accepted by the website, and every financial preflight remains blocked until a separately reviewed executor exists.
FENGZHI uses a versioned 50-Bee checkpoint. Export the current colony from the Hive to inspect its learned state and evidence references.
References and research material
- Habenstein et al., 2023 — cerebral neuropil atlas
- University of Konstanz — bee brain morphology resources
- Freie Universität Berlin — honeybee brain registration
- Peng & Chittka, 2017 — associative-learning model (not implemented here)
- Sun, Yue & Mangan, 2020 — implemented equation source
- Stone et al., 2017 — navigation foundation
- Open MBON electrophysiology — not a connectivity matrix
07 / Execution-status feedback & economic identity
The read-only receipt adapter checks the canonical block hash and requires at least 12 confirmations. Each Bee learns only receipts for transactions it inspected. A separate engineering memory groups outcomes by destination and method selector; after at least three observations, more failures than successes suppress support for another local watch of that pattern. This is not a published bee-learning equation, financial safety score or economic reward. Deduplication and checkpointing prevent replayed evidence from adding experience twice within retained history.
Two public configuration values identify the token and dev-wallet. A chain-56 and bytecode check gates the market adapter. Indexed token-as-base pools come from DEX Screener's documented token-pairs API. Primary-pool data are provider aggregates, not a complete ecosystem census. Dev-wallet balance is not creator revenue. Token trades and financial actions are not enabled.
For real anatomy, consult the publisher's Honeybee Standard Brain download collection and the Habenstein atlas viewer. The FU Berlin standard-brain geometry is included here; the Habenstein atlas is a distinct reference and is not the displayed geometry. Commercial reuse permissions for FU Berlin meshes and semantic mesh identification remain unresolved. The CC0 license applies to the electrophysiology dataset, not these meshes.