atelier: one-of-one card — cut, silhouette, GitHub Python. YAML preserved.

README.md

@@ -22,6 +22,38 @@ from szl_khipu.train import mini_embed
 
 emb = mini_embed.build(seed=20260721)
 vec = emb.embed("knot the run")
+
+
+<!-- SZL-ATELIER-CUT:v1:START -->
+## The cut
+
+Nobody ships an embedding where retrieval is provenance. MiniEmbed is not BGE, not NV-Embed. Mean-pool of L2 rows, seed 20260721, CPU NumPy.
+
+An embedding you can re-derive bit-exact from the seed. No SGD theater. Cosine is a claim about the table, not a vibe.
+
+### Silhouette → leave → SZL
+
+| Leader | Take, then tweak |
+|---|---|
+| Anthropic | Honest about what it is not — not a foundation embed. |
+| NVIDIA | Tiny table instead of NV-Embed / NeMo retrieval stacks. |
+| Unsloth | No LoRA. Construction is `build(seed)` — Unsloth is the wrong tool and we say so. |
+
+Nobody else ships this combination. That is the point of a one-of-one.
+
+## Intended use
+
+Silhouette of receipted retrieval. Teaching and tests.
+
+## Limitations
+
+- Not a neural embed.
+- Not comparable to BGE-base 768-d.
+- Hit@2 is on five doctrine pairs — SAMPLE.
+
+Canonical GitHub: [`szl-holdings/szl-khipu`](https://github.com/szl-holdings/szl-khipu/blob/main/szl_khipu/train/mini_embed.py)
+<!-- SZL-ATELIER-CUT:v1:END -->
+
 print(emb.V, emb.D, vec.shape)
 # 64 12 (12,)
 emb.save_npz("mini_embed.npz")