Add provenance correction: not a surgical child of the declared base

README.md

@@ -21,7 +21,31 @@ model_type: qwen3_5_text
 
 **The first auto-tuned model from jBlaze.**
 
-Seven stacked behavioral modifications. Automatically calibrated. No conventional fine-tuning required.
+Seven stacked behavioral modifications. Automatically calibrated.
+
+> **Provenance correction.** An independent weight-level audit of this
+> checkpoint against its declared base, `Qwen/Qwen3.5-27B`, found that it is
+> NOT a surgical child of that base. Gradient training lies somewhere
+> between the two:
+>
+> - 0 of 14 normalisation and bias tensors survive unchanged (weight surgery
+>   leaves these untouched); they moved 5.1e-3 to 1.6e-1.
+> - Cosine to the base is 0.775-0.942 at essentially unchanged scale
+>   (norm ratio 1.014-1.031) -- related, same magnitude, substantially
+>   rotated. That rules out both rescaling and a wrong base.
+> - `q_proj` (rel. Frobenius 0.46-0.52) and `embed_tokens` (0.47) are both
+>   modified. Neither is a weight-surgery target.
+> - The deltas are spectrally dense throughout (leading singular direction
+>   carries 0.05-1.2% of the energy; 64 directions carry 3-10%), which is
+>   the signature of training rather than of a constructed low-rank edit.
+>
+> The audit CANNOT say which step is responsible. `Sharona_Q27B-R_CodeSecurity`
+> (v1) predates this model and sits closer to it than the declared base does,
+> so the real chain is likely base -> v1 -> v2, and an undisclosed
+> intermediate is indistinguishable from training in the final step. What is
+> established is that the path from `Qwen/Qwen3.5-27B` to these weights
+> includes gradient training, so the `weight-surgery` tag does not describe
+> this checkpoint on its own. The full chain should be documented here.
 
 This is a 27-billion parameter code security scanner built on [Qwen 3.5-27B](https://huggingface.co/Qwen/Qwen3.5-27B) with 7 stacked behavioral modifications, each automatically calibrated by **jTuner** -- a new capability in [jBlaze](https://jblaze.dev) that we are announcing with this release.