card: rich prose (about + journey + ablation + stage notes)

README.md

@@ -64,6 +64,33 @@ license: apache-2.0
 
 ---
 
+## About this model
+
+Methodology validation artifact for the v2 forge pipeline + KL-distillation compensation LoRA. Demonstrates that aggressive head pruning + activation-metric importance + pad-mode defrag, when paired with output-distribution distillation against the unmodified teacher, recovers near-base HumanEval capability (61.0 vs 62.2 base, within calibration tolerance). This is the empirical anchor for PLASTICITY-COMPACTION §4.1.3.3 and the loss-function ablation that closes the §4.1.3.2 PPL/HumanEval disconnect. NOT a Pareto improvement over the unmodified base 7B at any single VRAM tier — published as proof that the methodology stack works end-to-end, in preparation for the Qwen3.5-35B-A3B and 397B-A17B forges where the pruning dimension actually wins.
+
+## The Journey
+
+This artifact is the punchline of a four-run experimental sequence on the same base model. The first run scored **50.0**; the final run scored **61.0**. Each run between them isolated a single variable, and each result narrowed the design space to the structural fix that recovered near-base capability.
+
+| Run | Configuration | HumanEval pass@1 |
+|---|---|---|
+| 1 | broken global-flat L2-weight | **50.0** |
+| 2 | layer-normalized activation, 1-cycle 500-step | **54.9** |
+| 3 | layer-normalized activation, 3-cycle (ablation) | **46.3** |
+| 4 | 1-cycle + KL compensation LoRA | **61.0** |
+
+## Loss Function Ablation
+
+The compensation LoRA was run twice with identical configuration, varying only the distillation loss. The result is a substantive methodology finding in its own right:
+
+| Distillation loss | HumanEval | HumanEval+ | Outcome |
+|---|---|---|---|
+| `mse_hidden` | **0.0** | **0.0** | degenerate fixed point — model collapsed to outputting '0' |
+| `kl_logits` | **61.0** | **53.0** | near-base recovery within calibration tolerance |
+
+MSE-on-hidden-states has a degenerate fixed point: the student can satisfy the loss by collapsing some downstream computation, regardless of whether the hidden states encode useful information. KL-on-output-logits has none, because matching the teacher's output distribution directly constrains task-level behavior. **For autoregressive language models, distillation must operate at the output layer, not at intermediate residual streams.**
+
+
 ## Benchmarks
 
 | Benchmark | Score | Base | Δ | Verified |
@@ -110,9 +137,14 @@ print(tokenizer.decode(output[0], skip_special_tokens=True))
 prune → lora → lora → eval (1 cycles)
 ```
 
-- **Pruning**: 12% heads via activation-magnitude
-- **LoRA**: rank ?
-- **LoRA**: rank ?
+- **Pruning**: 12% heads via `activation-magnitude`, layer-normalized, pad-mode defrag
+  > Layer-normalized activation-magnitude head importance (PLASTICITY-COMPACTION §4.1.3.1 fix). Pad-mode defrag preserves the q_proj invariant num_q_heads*head_dim==hidden_size so the artifact loads in llama.cpp (Finding 6 fix from VALIDATED-TENSOR-SURGERY).
+- **lora**: rank ?, 500 steps
+  > Single-cycle code-domain LoRA fine-tuning on the pruned student. 1-cycle ablation chosen because the 3-cycle multi-cycle test surfaced the §4.1.3.2 PPL/HumanEval disconnect (54.9 → 46.3 across cycles).
+- **compensation-lora**: rank 16, 500 steps, `kl_logits` distillation against `Qwen/Qwen2.5-Coder-7B`
+  > PLASTICITY-COMPACTION §4.1.3.3. KL divergence on output logits is the structural fix for the §4.1.3.2 disconnect. Loss-function ablation: MSE-on-hidden-states collapsed the model to 0.0 (degenerate fixed point); KL-on-logits recovered to 61.0. LoRA adapter merged into student weights at save time so inference-time VRAM and tokens/sec are unchanged from the un-compensated student.
+- **Calibrated evaluation**: anchored against `Qwen2.5-Coder-7B` (published 61.6, measured 62.2, ±3.0pt tolerance)
+  > All HumanEval numbers are anchor-calibrated against the unmodified Qwen2.5-Coder-7B base measured on the same hardware/pipeline in the same run. Hard-fail tolerance: ±3.0 points. Anchor delta: +0.6/+0.7 vs Qwen-published 61.6/53.0, deterministic across 6+ independent runs.
 - **Hardware**: NVIDIA GeForce RTX 5090
 - **Forge tool**: [Continuum](https://github.com/CambrianTech/continuum) Factory + [sentinel-ai](https://github.com/CambrianTech/sentinel-ai)