Fruit Fly Lab — fruitfly.arnao.ai

Proposal v2 · 2026-09-26 · status: PROPOSAL (synthetic mockup live, nothing built) · decision needed: go / park

v2 = v1 + the Fable adversarial pass integrated (§10): P1 shrunk to one impeccable experiment, a coupling badge for the brain–body link, the missing nerve cord drawn as a ghost, the interpretability bridge page, the visitor-prediction study, honest re-estimate.

One line: a public, phone-first 3D lab where anyone can watch a simulated fruit fly, tap it, silence parts of its brain, and see what changes — built only on the real public science (Nature 2024–2026), with every element labeled by where it came from. The point is not the fly. The point is the sentence the site exists to test: a complete wiring diagram is not an understood mind. You only learn what a mind does by intervening on it. That argument is Jonas & Kording's (Could a Neuroscientist Understand a Microprocessor?, PLOS Comp Bio 2017); what this site adds is that the public does the intervening, and we measure how well the diagram lets them predict the outcome.

Live synthetic mockup: https://fruitfly.arnao.ai/ (procedural body + fake neuron cloud, labeled as such). Panel of eight reviewers: https://fruitfly.arnao.ai/personas.html · Research dossier with every source: https://fruitfly.arnao.ai/research.html · Fable critique: https://fruitfly.arnao.ai/critique.html

0. The decision in 60 seconds

• What happened: in two years the fruit fly became the first animal that is both fully mapped and runnable. The complete adult brain wiring diagram (FlyWire, 139,255 neurons, ~50M synapses, Dorkenwald et al., Nature, Oct 2024); a physics-accurate whole body with trained walking and flight controllers from Google DeepMind and HHMI Janelia (Vaxenburg et al., Nature, Apr 2025, code TuragaLab/flybody, Apache-2.0); a whole-brain spiking model that predicts behavior from the wiring (Shiu et al., Nature 2024, MIT); the male whole-nervous-system connectome (Berg et al., Cell, Sep 3 202600942-6), CC-BY); and in March 2026 a startup wired brain and body together (Eon Systems, no code released).
• Why now: this month it went viral. A "fly brain plays Beat Saber" clip hit ~22M views and a hobbyist wave produced 129 catalogued projects, ~35 playable in a browser (timeline, catalog). Almost all are "poke a neuron and watch dots blink" toys. None combines real body physics, real neuron shapes, real recorded behavior, and honest provenance in something a person on a phone can use without a login.
• The brand call (the reason to do it at all): on its own, a cute simulated bug dilutes Byron's RAI and agent-security positioning. Framed as interpretability by intervention, it becomes the third leg of an existing portfolio: J-Space (jlens.arnao.ai, interpretability of a model's inside) and OREO (oreo.arnao.ai, ablating defense rungs against attacks). Fly Lab is the same epistemic move on a biological mind: open weights are not understanding; silencing a circuit and watching behavior is. That is a talk for EU/EMEA RAI leaders, not a toy.
• Cost shape: static site, no runtime LLM, no login. Pre-baking runs on the Mini. Vercel Blob egress is the only recurring line (cents per thousand visits at the default under-30 MB payload). Build: P1 ≈ 2–3 sessions, P2 ≈ 3–4.
• Recommendation: go on P1 only, shrunk to the honest core (§8): one impeccable brain-resident experiment, the full provenance system including the brain–body coupling badge, the ghost nerve cord, plain-English captions, the interpretability bridge page, and the prediction study. Honest estimate: 4–6 sessions, not 2–3. P2/P3 are earned by an unsolicited researcher or educator citation, not by traffic.

1. What the visitor experiences (Watch → Tinker → Understand)

The panel was unanimous on one thing: the first touch must be the fly, not a button.

1. Watch (0–5 s, no UI). Camera low and behind the fly mid-stride, pushes in toward the head over ~2.5 s. The head turns translucent and the neurons inside light up. One line: "This fly's brain is simulated too. Tap it." (Designer's opening shot; museum designer's attract loop.)
2. Tinker (direct manipulation only). Tap the fly → air puff: flinch, wing buzz, a red ripple through the brain, the fly bolts, a one-line caption says what just happened. Tap a brain region → it dims, the body changes, the caption says "in the model, silencing X predicts Y" with a citation. Drag to orbit. Slow-motion scrub at 0.25× so the ripple is watchable. Sound (footstep patter, wing pitch, the hush when a motor circuit goes quiet) and phone haptics on every state change. A visible reset.
3. Understand (after 2–3 interactions). A dismissible strip offers guided experiments: five-minute stories, each a hypothesis, a run it button that sets the state and holds the camera, a plain-English result, a live citation. Sliders and camera presets live behind an advanced expander, never on first contact.
4. Every element wears a provenance badge: measured (real recorded fly data) · replayed (precomputed model output) · live (running in your browser now) · illustrative (animation, not a model). This is the honesty layer three of the eight reviewers independently put above every feature.
5. Share a state as a URL with a thumbnail. Classroom mode: seeded start, predict-before-run prompt, one-tap snapshot of the footprint trail (the AP Biology reviewer's three asks).
6. Accessible by design: keyboard camera, a live narration region that says in words what the scene is doing, a data-table view of the brain, reduced-motion respected, Fold outer/inner layouts by aspect ratio.

2. The experiments: living figures (each a replayable, DOI-linked reproduction of a published result)

Rule from the biologist, adopted and promoted to the public tagline of the site: no perturbation ships unless it maps to a cited experiment. The mockup's "silence left motor → circles" is invented and goes away.

† Fable's catch: three of the six experiments execute in the ventral nerve cord, which FlyWire does not contain. They ship only as behavioral replays with an explicit caption "the executing circuitry is unmapped in this dataset", or move to the MaleCNS phase where the cord is real. The site draws the missing cord as a ghost (dashed, labeled "not in this dataset"), which is the most honest pixel on the site and makes the argument for us: even the famous "complete" connectome is not complete.

3. What is real, what we would load

• Body: flybody MJCF (67 bodies, 102 DoF, ~20k faces) from TuragaLab/flybody / MuJoCo Menagerie, Apache-2.0. Raw meshes are 172 MB of OBJ; decimated + meshopt glTF lands under ~10 MB.
• Real behavior: Janelia's Figshare release: ~80 minutes of real fly walking as full-body joint angles (3 GB), flight trajectories (13 MB), trained policies (6.5 MB). Also Tuthill lab 3D leg kinematics on Dryad, CC0 (Anipose). This is what "measured" badges point at.
• Brain geometry: FlyWire v783 skeletons on Zenodo, CC-BY 4.0 (record 10877326); a curated 1,000-neuron subset as LOD'd glTF, lazy per neuron, plus a whole-brain soma cloud (~1.7 MB). The male connectome (MaleCNS, CC-BY, no login) is the commercial-safe alternative.
• Brain dynamics: Shiu's model (repo, MIT; 127,400 modeled neurons of FlyWire's 139,255 after filtering out fragments and neurons without complete synaptic data, 50M synapses). Full-brain runs are pre-baked on the Mini and replayed; a filtered subcircuit (a stimulus plus one or two hops, hundreds to low thousands of neurons) can run live in a Web Worker. Four independent projects already run the full model in a browser (fly-brain-bench, webgpu-fly), so live-full-brain is a P3 stretch, not a fantasy.
• Physics in the browser: DeepMind's official MuJoCo WASM bindings (@mujoco/mujoco 3.14, Apache-2.0). Walking (0.2 ms timestep) is real-time-feasible; flight (0.05 ms) runs sub-real-time with a speed indicator. The walking policy is a 3×256 MLP (~200k params, <1 MB), hand-ported to ~60 lines of JS, no ONNX runtime needed.

4. Architecture (static, honest, cheap)

• Site: static three.js (0.186) on Vercel, bundle under 10 MB including single-threaded MuJoCo WASM. WebGL2 everywhere; WebGPU as progressive enhancement (~87% global, caniuse).
• Assets: public Vercel Blob, every object under 512 MB (the cache cliff, docs), content-hashed names. Default first-visit payload under 30 MB; neuron meshes lazy.
• Compute: MuJoCo in one Worker, policy MLP in JS, LIF subcircuit in a second Worker, main thread renders. No live calls to FlyWire/CAVE/Codex (they require sign-in and forbid bulk API use); everything pre-baked by a Python pipeline on the Mini.
• Fallbacks: reduced neuron count and no physics on low-tier phones; FPS watchdog degrades edges → point count → pixel ratio → antialiasing; render loop pauses when the tab is hidden.
• Analytics: self-hosted Plausible with custom events (tap-fly, silence, run-experiment, share, story-complete). No login, no PII.

5. Honesty and licensing rules (non-negotiable)

• Provenance badge on every element (§1.4), plus a badge on the coupling: there is no published pipeline from Shiu's spikes to flybody's joint actuation (that is exactly what Eon built and did not release). Wherever the brain appears to drive the body, the link itself wears its own label: coupling: illustrative until a real spikes→motor mapping exists. Never "silencing X makes it circle"; always "in the model, silencing X predicts Y", with the paper.
• Say the neuromodulation point before a visitor does: the same wiring supports many circuits depending on modulatory state (Marder, Neuron 2012). The wiring underdetermines the behavior; that is the thesis, not a weakness.
• Neuron names: every displayed label (MDN, DNa02, sugar GRNs) is traced to a CC-BY paper, not to FlyWire's CC-BY-NC community annotation tables. Audit before P1.
• State plainly: FlyWire is one female fly; the model's synaptic weights are fitted assumptions, not measurements; the whole-brain model was validated on specific circuits (sugar, grooming), not as a general oracle.
• Drop pseudo-metrics: no "spikes/frame" counter unless it comes from a real simulation.
• Licenses: flybody code Apache-2.0 (no Google/DeepMind/Janelia logos or implied endorsement); FlyWire data CC-BY 4.0 with community annotations CC-BY-NC 4.0 (ToS); MaleCNS/MANC CC-BY; Shiu model and flyvis MIT; the Figshare dataset is tagged GPL-3.0+ while the paper says CC-BY, so email the Turaga lab before redistributing trajectories or checkpoints. The site carries no ads and no paid tier, which keeps CC-BY-NC clean.
• Every citation is a live URL at the point of claim (the standing rule).

6. The panel (eight reviewers, full text on the panel page)

Consensus ranking of the eight candidate features (votes across all reviewers): 1. Guided experiments · 2. Explainer layer with live citations · 3. Silence/stimulate tied to real (replayed) model output · 4. Real-vs-sim side-by-side · 5. Real connectome subset with click-to-inspect · 6. Real flybody physics · 7. Share-a-state · 8. Bring-your-own CSV (cut: off-thesis, upload risk, nobody wanted it).

7. What the mockup gets wrong (and what the build changes)

The live mockup is a deliberate stand-in and is labeled so. The panel's corrections, all adopted for the build: brain regions relabeled to real circuits (visual, olfactory, central complex, descending L/R) with the leg motor circuitry drawn as a second structure in the thorax (the ventral nerve cord); the tripod/tetrapod toggle replaced by a continuous speed→coordination mapping; wings fold at rest and beat at ~200 Hz only in flight; halteres idle during walking; the fixed 1.2 s "pulse" replaced by continuous noisy activity or replayed spikes; the spikes/frame counter removed; internal links (panel, critique) moved off the public header. Immediate mockup patches shipped with this proposal: tap-the-fly poke, plain-English "what just happened" caption, reduced-motion and hidden-tab pause, labeled controls, aria states.

8. Phases, cost, and what is earned vs committed

Recurring cost: Vercel Blob egress only (roughly $0.05/GB); at a 30 MB default payload that is about $1.50 per thousand visits. Pre-baking runs on the Mini (Shiu model ~5 min per simulated second per CPU thread; a handful of conditions overnight). No LLM in the runtime path.

9. Open decisions for Byron

1. Go / park. P1 only is the recommendation.
2. Female (FlyWire) or male (MaleCNS) first — RESOLVED by the Fable pass, and not for fame: the only published whole-brain dynamics model (Shiu) is fitted to the FlyWire female wiring; MaleCNS has none. MaleCNS geometry with Shiu dynamics would be a chimera; MaleCNS without dynamics is the poke-a-neuron dot cloud we are trying to bury. FlyWire for P1, cord drawn as a ghost; MaleCNS at P3 as its own launch.
3. Name. "Fruit Fly Lab" is the working title. Alternatives: "Known Mind", "Wiring Is Not Understanding". The brand strategist wants the thesis in the name; Fable wants "no uncited perturbation" as the public tagline.
4. The prediction study. Logging anonymous predictions and publishing the aggregate ("N thousand people given the complete wiring diagram predicted the outcome at X%") turns the site into a citable result. Yes/no, and whether to pre-register it.
5. Launch framing. One LinkedIn post leading with "a complete wiring diagram is not an understood mind", linking J-Space and OREO as the other two legs. Draft goes through Branding Central for approval; nothing auto-publishes.
6. Public header links. Strip panel/research/critique from the public page at P1 (keep them under an "about this build" link) or keep them visible as a transparency signal.

10. Self-critique — Fable adversarial pass (v1 → v2, integrated)

Fable 5 reviewed v1 as a ruthless critic (raw critique). Its verdicts and what changed:

• "The framing is post-hoc; the thesis lives in the LinkedIn post, not the product, and the thesis isn't yours." The argument is Jonas & Kording 2017, missing from the dossier. → Cited at the top; the site's addition stated plainly (the public does the intervening, and we measure it). The interpretability bridge page (fly intervention ↔ ML analog, linked to J-Space and OREO) moved into P1 as a first-class page. Without it: "senior guy played with three.js."
• "Badges label elements; a neuroscientist will attack the couplings." There is no published spikes→actuation pipeline; staging brain→body causality is the most sophisticated overclaim on the internet. → A coupling: illustrative badge on the brain–body link itself (§5).
• "Three of six experiments execute in tissue your connectome doesn't contain." → Experiments reclassified by where the circuit lives (§2); cord-resident ones ship only as captioned behavioral replays or wait for MaleCNS; the missing cord is drawn as a ghost.
• "Neuromodulation: someone will invoke Marder within the first hundred visitors." → Said first, cited, and folded into the thesis (§5).
• "Neuron names may import the CC-BY-NC annotation clause into the core of the product." → Name-provenance audit is a P1 gate (§5).
• "127k vs 139,255 in the same document reads as carelessness." → Reconciled with the filtering stated (§3).
• "P1 is 8–12 sessions in a 2–3 session costume; the least honest number in an honesty-themed proposal." → P1 cut to one impeccable experiment plus the provenance system; sound, haptics, classroom mode, Fold layouts, the table view and experiments 2–6 moved to P2; re-estimated at 4–6 sessions (§8).
• "≥1k uniques is a meaningless gate." → P2 gate is one unsolicited researcher or educator citation.
• "30 MB on a phone-first site is a UX problem before a cost problem." → First paint under 10 MB (§4, §8).
• "Living figures: researchers cite tools and data, not exhibits." → Each experiment reframed as a DOI-linked, replayable reproduction of a published result; "no uncited perturbation" promoted to the public tagline (§2).
• Definitive call adopted: FlyWire female first, because the only whole-brain dynamics model is fitted to it, not because it is famous (§9.2).
• Bold idea adopted: run the microprocessor test on the visitors. Show the real subgraph, force a prediction before the run, log it anonymously, publish the aggregate. The measured gap between having the diagram and understanding the mind is the product; the fly is the demo (§8 P1, §9.6).

Links