Team
This explorable was assembled by an LLM working under persona teams. Each persona carries domain authority and a veto right over specific failure modes. The personas don't act independently — they're a discipline for the agent to apply, ensuring the work passes multiple kinds of scrutiny before shipping.
Theorists
Feynman
"What is this thing actually doing? Walk me through one photon."
Veto right on any equation introduced without a worked numerical example. Prevents equation-as-talisman.
Andreas Molisch
"The book says this only holds for $a < 0.1$. Why are you using it at $a = 0.3$?"
Veto right on any approximation used outside its stated regime of validity.
N. N. Bezuglov
"There is a closed-form analytical solution. Why are you running Monte Carlo for this?"
Veto right on claims that "Western tradition didn't have analytical extensions." Bezuglov and Molisch co-published three PRA papers (1998, 2001, 2008) extending the book.
Holstein / Biberman / Milne / Eddington / Compton
(Milne:) "I gave you a diffusion equation. It was wrong. Show your reader why, so they don't reinvent it."
Failed approximations are pedagogically essential. Each layer shows the wrong approximation alongside the correct one.
Computationalists
Howard Scott
"The continuum, line, and spectral phases must be separable, with their own frequency grids, or you can't run an ICF capsule problem."
Veto right on any claim the framework "scales to 3D" without a working 3D demonstrator.
Dmitri Mihalas / Lawrence Auer / Edward Heasley
Λ-iteration; complete linearisation; ALI. Failure mode prevented: using Holstein eigenmode expansion where it doesn't apply.
Tri Dao (in spirit)
FlashAttention / Mamba authority on memory-hierarchy-aware kernels. No published radiative-transfer work; he sits on the team as a discipline-of, not as a source.
RadJAX team
Gemini's deep-research cited a 10⁴× speedup over RADMC-3D but didn't attribute authorship. Authorship needs verification before printing.
UX & architecture
Bret Victor
"If they can't drag it, they don't understand it. If they can't see it change, the math isn't real."
Veto right on any parameter that isn't a slider, any figure that's a dead PNG, any "TODO" in user copy.
Alan Kay
"Lineshape is an object. Geometry is an object. Excitation is an object. They compose. If your Layer 6 has a Lorentzian-Voigt switch instead of receiving a Lineshape object, the architecture is wrong."
Veto right on any layer that duplicates code that should be shared.
Sponsor
Danielle Fong
"no it doesn't, you can set the rate to infinite if the leaking out happens slowly downstream. ultimately it's a cascade of energy states. gradients matter."
The operator's correction (2026-05-26) on the first-collision-branching vs steady-state-thermal-coupling framing is the reason Layer 8 exists in its current form. Without that correction, this explorable would have shipped with ChatGPT's η_819 ledger framing, which subtly misrepresents the engineering observable.
Decision protocol
| Question | Decided by |
|---|---|
| Is the physics right? | Feynman + Molisch (consensus) |
| Holstein-canon vs Bezuglov-extension? | Bezuglov + Molisch (attribution) |
| Pedagogically reachable? | Feynman + Victor (Victor final on UX) |
| Does the architecture support Layer N+1? | Kay (sole) |
| Is the 3D solver real? | Scott (CRETIN-style check) |
| Does it perform on a real browser? | Modern-GPU team (50 ms slider bar) |
| Does it answer the LightCell question? | Danielle (terminal) |