CIONA

The anatomy is measured.
The dynamics are a model.

CIONA is an independent, reproducible experiment using a published Ciona larva connectivity matrix. It does not reproduce a living brain or demonstrate language, cognition, learning or trading ability.

Original data

Kerrianne Ryan, Zhiyuan Lu and Ian A. Meinertzhagen (2016). The CNS connectome of a tadpole larva of Ciona intestinalis (L.) highlights sidedness in the brain of a chordate sibling. eLife 5:e16962. Original publication, Figure 16, source data 1. Article/data attribution: CC BY 4.0.

The paper describes 177 CNS neurons. The imported chemical-contact matrix contains 222 unique identifiers and 3,105 directed nonzero edges, including peripheral cells, muscle pools, ambiguous cells and basal-lamina targets. We preserve the union of source and target identifiers, rather than constructing an artificial 177-node matrix. Rows are presynaptic sources, columns are postsynaptic targets; their ordering differs.

Weights represent cumulative synaptic contact depth in micrometres, not counts of individual synapses or measured conductances. Gap junctions are not included. Coordinates in the graph are schematic, grouped positions, not measured anatomical coordinates.

Source workbook SHA-256:
ba0903d44fc14e960654d6f43025b89d222b0dd5461f3988b486824e29a2b7cb

What runs

A continuous rate model uses the directed anatomical graph. The server version runs in Python; the light chamber runs the same recurrence and update equations in JavaScript. Activity persists between integration steps. Resetting restarts a reproducible run. The two runtimes use different seeded random initializers, so cross-runtime comparisons must begin from the same initial vector.

drive[j] = 0.008 + 0.9 × sum(W[i,j] × activity[i]) + sensory[j]
target[j] = tanh(max(0, drive[j] − 0.15 × adaptation[j]))
activity[j] += (1 − exp(−dt / 0.18)) × (target[j] − activity[j])
adaptation[j] += (1 − exp(−dt / 1.5)) × (activity[j] − adaptation[j])

Weights are normalized by each target’s total incoming contact depth. Every contact is treated as positive coupling because this matrix alone does not specify all neurotransmitter effects or connection signs. Gains and time constants are engineering choices. This is an illustrative anatomical-network model, not a validated simulation of Ciona behaviour. Basal-lamina targets remain inactive; muscle pools are output units, not neurons.

Sensation and movement

Ciona has an asymmetric ocellus, not two eyes viewing the halves of a webpage. Our artificial adapter distributes left and right darkness signals across alternating PR-I labels; PR-II receives mean darkness. This is a documented screen interface, not reconstructed biological optics. The light chamber uses the same adapter, applied to a simulated light field.

Mean activity of MN1L–MN5L and MN1R–MN5R drives turning and forward movement. The browser software maps these outputs to cursor displacement, periodic scrolling and safe link selection. The chamber adds boundary reflection and a schematic body. Neither the body outline nor the cursor is a microscopy reconstruction.

The browser has no wallet, accounts or text-entry ability. Public pages are restricted to an allowlist. Link validation, timeouts and navigation recovery are software outside the neural model. No language model writes the displayed events; they are factual messages from the program.

Reading the dashboard

Activity is normalized model activity, not recorded biological firing. “Active units” counts units above 0.05, including non-neuronal outputs. Model time is simulated integration time. Browser counters refer to the shared server run; light-chamber counters refer to your local run. A live server response is telemetry from the operator, not cryptographic proof of a remote process.

Nothing is labelled as learned behaviour. Slow adaptation is a numerical state variable, not a claim of biological memory. Token fields remain empty until explicitly configured. The journey performs simulated financial operations only. No funds, wallet or real exchange order is involved; returns are computed from the paper ledger.

The $100 paper-trading journey

The shared journey starts with $100 in virtual USD. A separate instance of the same connectome rate model receives numeric market observations. It does not read a website, interpret a chart, use a language model or understand a token. The market panel displays the actual quote and derived motor signals used by the trading adapter; selecting another market only changes your inspection view.

Market data comes from the DEX Screener public API. Contract addresses identify assets, not ticker symbols. Curated examples and bounded discovery are filtered by available price and liquidity. Discovered profiles are unverified and are not necessarily memecoins. Displayed timestamps record when the server received a provider observation, not a guaranteed exchange trade time. Charts contain observations collected by this run, not invented historical candles.

Recent price changes are transformed into artificial left/right light inputs. The recurrent network produces motor outputs; their relative difference, compared with a neutral calibration, is mapped to buy, sell or hold. The policy, thresholds, limits and software risk exits are exposed in the public journey state. These are engineered rules, not innate animal behaviour or demonstrated financial intelligence.

Fills are simulated at the observed price with adverse slippage and a fee. Quotes that are unavailable or too old cannot generate fills. Liquidity filtering does not prove that an actual order would execute, or model transfer taxes, rug pulls, MEV, network congestion and every execution risk. Paper performance is not a prediction of real trading results.

The persistent ledger records virtual cash, open positions, buys, sells, fees and portfolio observations. Total P&L equals marked portfolio equity minus the initial $100. Realized P&L deducts the position cost basis and sell fee; entry fees are included in cost basis. Unsold positions use their last observed price and are marked stale when appropriate, so stale portfolio values are estimates. A restart does not create a new $100 account.

The live terminal and cursor

The terminal is a live view of the shared server state. Its cursor projects the server's semantic action targets onto the visible interface: token row, chart, and buy, sell or hold control. Software schedules scan, observation, neural evaluation, approach, execution and settlement phases. A paper fill can be committed only at execution, after freshness, liquidity, position and cash guards are checked again. The cursor is a visualization of those server actions, not a human mouse recording or proof of an exchange order.

The evaluation stream contains measured inputs, model outputs, policy reasons and execution results. It is not private thought, language comprehension or a generated personality. Inspecting a coin changes your viewing selection, not the server's trading actions. The follow-live view keeps the cursor and selected market synchronized.

Recent token profiles and recently created pools are different. Pair age uses the provider's pool creation timestamp when available; an updated profile is not evidence of a newly launched token. Real project images and links are third-party metadata, not endorsements. Missing imagery is shown with an initials fallback.

Privacy and public data

The public site exposes experiment observations and simulated trade records. It has no login, wallet connection, deposits or payment form. Runtime administration is not available through the public website. Local source archives, environment files and database files are not public downloads. The hosting and market-data providers process their own service requests; external links open third-party sites.

Specimen image

Ciona intestinalis larva, Plymouth type B. Pennati et al. (2015), PLOS ONE, Figure 4. CC BY 4.0. The homepage displays a cropped panel with grayscale, inverted contrast for a dark background. It is a reference specimen, not the individual used for the connectivity reconstruction. Unmodified original figure.

Reproduce and inspect

Download the imported graph and model specification. The implementation includes retained Excel workbooks, an import script, Python engine, paper ledger, frontend and tests. The source archive is delivered separately from the public site. Software tests check orientation, deterministic replay, bounded outputs and dependence of motor outputs on network connections. They do not constitute biological validation.

Code is MIT licensed. Scientific data and the photograph retain their respective CC BY 4.0 attributions. This project is not affiliated with the researchers, eLife, PLOS or Flybrain.