# Study 15 — Skala DFT shear

**Status: CITED · court DARK — GMTKN55 2.72 / 2.8 kcal/mol fetched 2026-08-24**  
**Program:** [Language-games study board](Language-Games-Study-Board.md) · [Study 14 — Protein lattice](Study-14-Protein-Lattice-Manifold.md) · [Zero Float · Zero Shear](Zero-Float-Zero-Shear-Paradigm.md) · [Fourier Phantom](Impact-Study-Fourier-Phantom.md)

| Surface | URL / key | Cited vs sealed |
|---|---|---|
| Look | https://affine.earth/language-game/#researcher · [families GET](https://affine.earth/language-game/discoveries/families.json) | GET only. credentials omit. |
| IDE | https://affine.earth/language-game/ide.html#skala-cited | LOOK template. No kcal POST. |
| Court | none — no kcal / XC / GMTKN55 key | **DARK** |
| Study 11 `geometry` (contrast) | `polytope_id=unit_square` `dilation=12` | LIVE Ehrhart: count **169** vol `1/1` · `dilation=1.5` → HTTP 400 `REFUSED_FLOAT` |
| Humanity | A moving XC baseline treated as the molecule | **Using something that is not accurate (because it shears) may not be useless — it may be dangerous.** |

This page **contrasts methods**. It does not write a 2D verdict that Skala “is useless” or that Affine “solved” thermochemistry. The win is the verified bond. Edge word on an unbound GMTKN55 integer court: ***not known***.

---

## For every reader — one paragraph

Microsoft Skala 1.1 is a deep-learning exchange-correlation functional for DFT. Their own pages publish a GMTKN55 weighted error and a “continuously improving” release rule: each version supersedes the last. Affine.Earth OS freezes integer law. A float on the wire is refused. Nine cells replay the same Swift bytes. Study 11 at dilation 12 still returns count **169** volume `1/1`.

**Using something that is not accurate (because it shears) may not be useless — it may be dangerous.** A kcal/mol miss treated as the molecule is a spatial collision: wrong pocket, wrong bond, wrong dose. GMTKN55 “2.8 kcal/mol average error” is that danger, **cited from their page**, not a dunk.

Same continuum court as Anima’s FNO DFT (Aug 2026 — **their** 82,500-atom Mg / B300 claim). Skala shears the kcal. Anima shears the density via Fourier. Affine’s replacement is not a second XC functional: [Study 11](Study-11-Ehrhart-Volume-Shear.md) counts, [Study 12](Study-12-Quantum-Parallel-Repetition-Shear.md) locks state as rationals, [Study 13](Study-13-Connes-Rigidity-Shear.md) locks the Jordan link. Addendum: [Fourier Phantom](Impact-Study-Fourier-Phantom.md). Affine has not sealed an 82,500-atom Mg density.

---

## Fetches — 2026-08-24 (VERIFIED vs REPORTED)

| Claim | Number / fact | Grade | Source · date |
|---|---|---|---|
| GMTKN55 WTMAD-2 | **2.72 kcal/mol** | **VERIFIED** | [Skala 1.1 model card](https://microsoft.github.io/skala/model-card/skala-1.1.html) · fetched 2026-08-24 |
| GMTKN55 weighted average error | **2.8 kcal/mol** | **VERIFIED** (blog rounding of the same suite; **not** the same spelling as 2.72) | [Microsoft Research blog — Broadening access to Skala](https://www.microsoft.com/en-us/research/blog/broadening-access-to-skala-creates-a-faster-path-to-predictive-dft/) · fetched 2026-08-24 |
| W4-17 atomization | ~1 kcal/mol MAE (“chemical accuracy”) | **REPORTED** on the model card | same model card · 2026-08-24 |
| Implementation | PySCF SCF (CPU); traced functional for GPU / other SCF | **VERIFIED** | model card + [github.com/microsoft/skala](https://github.com/microsoft/skala) · 2026-08-24 |
| CP2K | Native path via GauXC; blog says available in CP2K | **VERIFIED** | blog + [CP2K / GauXC paper page](https://www.microsoft.com/en-us/research/publication/molecular-implementation-of-the-machine-learned-skalaexchange-correlation-functional-in-cp2k-through-gauxc/) · 2026-08-24 |
| CP2K vs PySCF | agree to within **0.1 kcal/mol** MAD on a GMTKN55 subset | **VERIFIED** (their figure caption) | same blog · 2026-08-24 |
| Release rule | “each release is designed to supersede the previous one”; “continuously improving” | **VERIFIED** | same blog · 2026-08-24 |
| Training vs prior | 2.5× more data than predecessor | **VERIFIED** | same blog · 2026-08-24 |
| Parameters | ~385k trainable | **VERIFIED** | model card · 2026-08-24 |
| Production claim | “this is not a production model” | **VERIFIED** | model card · 2026-08-24 |
| Paper | Luise et al., arXiv:2506.14665 (2025) | **VERIFIED** | model card citation · 2026-08-24 |

Do not collapse 2.72 and 2.8 into one invented number. Cite both spellings. The blog’s **2.8 kcal/mol** is the danger figure this charter names in the humanity clause.

---

## Comparison — Affine.Earth OS vs Microsoft Skala 1.1

**Using something that is not accurate (because it shears) may not be useless — it may be dangerous.**

| Axis | Skala (measured from their page) | Affine (measured from our membrane) |
|---|---|---|
| **Baseline** | Deep-learning XC on DFT. Releases supersede. Blog: continuously improving functional. | Frozen integer law. Study 11 dilation 12 still count **169** vol `1/1`. |
| **Error** | GMTKN55 WTMAD-2 **2.72 kcal/mol** (model card) / weighted average **2.8 kcal/mol** (blog). Both fetched 2026-08-24. | `REFUSED_FLOAT` or exact residue 0. No kcal/mol court today. |
| **Replay** | GPU FMA / non-associative SCF (PySCF, GPU4PySCF, CP2K/GauXC). Two codes agree to 0.1 kcal/mol MAD on their subset — still kcal. | Nine cells, same Swift law, same bytes. |
| **Consensus** | Shifting kcal/mol as the next release replaces the last. | V234 / court receipt. `source` + `role` required. Empty source → `REFUSED_UNATTRIBUTED`. |
| **Method** | Neural XC energy from density / meta-GGA features; SCF until a float energy. | Occupancy appointment (Study 14 grammar) or Ehrhart count (Study 11 LIVE). No second DFT law. |
| **Danger** | A 2.8 kcal/mol average miss treated as the molecule is a spatial collision: wrong pocket, wrong bond, wrong dose. | Refuse the float or seal the integer. Health court remains PK dose (`mass_milli`/`vol_milli`), not a pocket. |

Adversary: the “continuously improving” functional — a baseline that moves after the claim. Edge: we do **not** claim a GMTKN55 integer court today. That court is DARK. Honest word: ***not known***.

---

## What Affine does on this shear

1. **Refuse the float.** `dilation=1.5` on geometry is HTTP 400 `REFUSED_FLOAT` (measured LIVE on Study 11). A kcal field would be the same refuse — when a chemistry key exists.
2. **Seal the integer.** Study 11 `unit_square` + `dilation=12` → count 169, volume `1/1`. That is Ehrhart, not a fold, and not GMTKN55.
3. **Replay on nine cells.** Same Swift membrane. No GPU FMA as the law.
4. **Name the health court.** PK is `mass_milli` / `vol_milli` (dose). It is not a binding pocket and it is not a Skala energy.

Study 14 states the protein / ligand grammar. The public discovery index is [Known molecular discoveries](Known-Molecular-Discoveries.md) — Look LIVE, Claim DARK. This page states why a sheared XC baseline is the **adversary**, not a dunk target.

---

## What this page does not do

- Does not call Skala **useless**
- Does not claim Affine has a GMTKN55 integer court
- Does not claim docking, folding, or thermochemistry is solved
- Does not land `ProteinManifold.evaluateDocking` or an O(1) dock stub
- Does not treat Study 11 geometry as a protein fold

---

**Status: CITED · court DARK.** Their GMTKN55 numbers are on this page. No kcal game_id. **Not OPEN.**