EDITED DISCUSSION CONTEXT Gradient Dissent meeting, 2026-07-23 Prepared from the scientific portion of an automated transcript STATUS OF THIS ATTACHMENT This is a faithful thematic edit and paraphrase, not a verbatim transcript. Speaker labels in the source were unreliable, so attribution has intentionally been omitted. Unrelated café audio, personal information, and repetitive or obviously corrupted recognition output were excluded. The private source file was not copied into this repository or its Git history. DISCUSSION ARC The group began by questioning whether a domain-fine-tuned language-model paper offered enough mathematical novelty. The comparison appeared to mix model size, fine-tuning choices, and adaptation claims, making it hard to isolate what caused the reported effect. Interest moved toward models that participate directly in crystal discovery and toward the Matbench Discovery leaderboard. The leaderboard conversation produced a working hypothesis: rotationally aware graph attention might be central to the strongest current models. Because the meeting record came from noisy speech recognition and leaderboard state changes, the model identity and ranking needed an independent audit rather than being treated as a meeting fact. Motivating applications included superconductors, batteries, power transmission, motors, and magnets. The chromium dimer came up as a deliberately small but hard electronic-structure problem, prompting questions about what density functional theory approximates, why it remains expensive, and how a learned potential should be verified. Chromium is a conceptual adversarial example here; no toy simulation in the companion tutorial claims to reproduce Cr2 quantum chemistry. The group discussed several possible validation loops. A candidate might be sent back to DFT, or eventually to automated synthesis, but an energy-minimized crystal is not a synthesis recipe. Temperature and pressure schedules, metastability, and the path by which a phase is realized remain missing links. A browser-based atomistic demonstration was especially engaging. It showed a cutoff graph connecting nearby atoms and passing messages to approximate expensive electronic-structure calculations in real time. That led to the core technical questions for the present tutorial: * What mathematical structure is new? * Why should equivariance improve data efficiency? * What physics is lost through locality and a finite cutoff? * What physically impossible outputs remain possible? * How does regression error relate to actual discovery value? The group distinguished generating a merely stable candidate from generating one with a requested property such as conductivity, resistance, superconductivity, or magnetic response. Candidate conditioning might use property vectors, text, composition, partial structures, space groups, or Wyckoff positions. These choices change both the representation and what a benchmark should measure. The most animated mathematical discussion concerned the 17 wallpaper groups and 230 three-dimensional space groups. Reaction-diffusion animations suggested a runnable symmetry simulator. The same discussion exposed a limitation of static structure benchmarks: they say little directly about phase changes, transport, battery cycling, or long-time dynamics. The desired reading criterion became clear by the end of the discussion: choose a paper with a real mathematical object or representation, make its failure modes concrete, and build something the group can run rather than presenting only a literature review. Classical PDEs such as linear heat diffusion may already have excellent numerical solvers; DFT is a more compelling learned-surrogate target because each reference calculation remains expensive. HOW THIS SHAPED THE COMPANION The interactive field guide therefore includes: 1. a rotation/reflection equivariance unit test; 2. an angular-information and sample-efficiency toy fit; 3. conservative versus non-integrable force dynamics; 4. a convex-hull discovery and finite-validation-budget game; 5. a clickable genealogy and reweightable leaderboard score. It also keeps several questions explicitly open: long-range interactions, diffusion and kinetic barriers, active-learning disagreement, property-conditioned generation, and symmetry-constrained reaction-diffusion dynamics.