The "interpretability by intervention" framing is post-hoc. You can tell because the brand strategist had to invent it, and because nothing in the product embodies it. The site as specced is a museum exhibit with excellent hygiene. The thesis lives entirely in the launch LinkedIn post — which means the brand value evaporates the moment someone clicks through and finds a fly you poke.
Two harder problems:
The thesis isn't yours. "A complete wiring diagram is not an understood mind" is Jonas & Kording, "Could a Neuroscientist Understand a Microprocessor?" (PLOS Comp Bio, 2017) — one of the most-cited methodological critiques of the last decade, and it's nowhere in your dossier. Any neuroscientist or interpretability researcher who sees your framing will think of that paper within thirty seconds. If you don't cite it, you look like you don't know the literature. If you do cite it, you must answer: what does the site add beyond the paper? The only honest answer is "it lets the public do the intervening" — which is genuinely new, but the proposal buries it under attract loops and haptics.
Timing works against seniority. Riding a meme wave (22M-view clip, 129 projects this month) is what hobbyists do. By the time P1 ships — realistically Q4 — the wave is dead. A 61-year-old positioning for RAI leadership doesn't win by being the 130th fly project; he wins by being the person who used the moment to make a point nobody else made. That point must be in the product (see change #2 below), or this is a charming detour that a 2027 hiring committee reads as "senior guy played with three.js."
Verdict: on-brand only if the ML-interpretability parallel is a first-class page, not a coat of paint. As written: rationalization.
The badge scheme labels elements. A working neuroscientist will attack the couplings:
Be honest about who cites things. Researchers cite tools and data, not exhibits. Teachers cite exhibits. Your realistic citation channels are (a) AP Bio / outreach and (b) the provenance system itself — a norm-setting artifact in a field currently drowning in "poke a neuron" slop. The one genuinely uncopied thing in this plan is the rule "no perturbation ships without a cited experiment" plus per-element provenance. That's the identity. Everything else — opening shot, sound design, share thumbnails, haptics — is a better-produced version of the other 128 projects and will be indistinguishable in a screenshot.
The framing that could earn a researcher citation: position each experiment as a living figure — a replayable, DOI-linked reproduction of a published perturbation result. Six papers, six interactive figures. That's a sentence a scientist can put in an outreach section.
Count the P1 line items: a 172MB→10MB mesh pipeline, ingesting a 3GB trajectory corpus (behind an unanswered license email — and the fallback, "render our own," is itself multi-session), running and choreographing Shiu-model replays for three experiments, the entire badge system, a cinematic opening, guided stories, captions, sound, haptics, keyboard camera, live narration region, a data-table view of the brain, Fold layouts, Plausible instrumentation, and classroom mode with snapshot export. That's 8–12 sessions wearing a 2–3 session costume. The estimate is the least honest number in an honesty-themed proposal.
Cut from P1: sound, haptics, classroom mode, Fold-specific layouts, data-table brain view (keep reduced-motion and captions — those are cheap and core). Cut experiments 2 and 3; ship one experiment — sugar→proboscis — impeccably. It's the best-validated, it's brain-resident (SEZ, actually in FlyWire), and one bulletproof experiment beats three assailable ones. Also: the P2 gate of "≥1k uniques" is meaningless — trivial during any residual meme traffic, unreachable after. Gate P2 on one unsolicited researcher or educator citation instead.
The 30MB default payload on a phone-first site is a UX problem before it's a cost problem. Target sub-10MB first paint; lazy-load the rest.
The proposal picks FlyWire for fame. Fame is the weakest reason. The decisive fact: Shiu et al. is fitted to the FlyWire female connectome, and MaleCNS has no published whole-brain dynamics model. Your entire scientific core — "replayed model output," the sugar experiment, the >90% validation claim — exists only on the female wiring. Ship MaleCNS geometry with Shiu dynamics and you've built a chimera worse than anything the mockup gets wrong; ship MaleCNS without Shiu and you've built exactly the poke-a-neuron dot cloud you're trying to bury. MaleCNS-first guts the thesis.
Mandatory consequences of the call:
Run the microprocessor test on your visitors. Before any perturbation runs, show the visitor the actual wiring — the real connectome subgraph — and force a prediction: what will silencing this do? Log predictions (no PII, you already have Plausible). After a few months you publish the aggregate: "N thousand people, given the complete wiring diagram, predicted the behavioral outcome at X% — barely above chance." That's a novel, citable, on-thesis empirical result generated by the site itself — a public replication of Jonas & Kording, a workshop-paper candidate, and the single most powerful slide a Responsible-AI leader could own in 2027: here is what "we released the weights, therefore it's transparent" is actually worth, measured.
The fly is not the product. The measured gap between having the diagram and understanding the mind — that's the product. Build that.