Upload README.md with huggingface_hub

README.md

@@ -210,6 +210,52 @@ Two caveats stated up front, because they decide how the figure may be read.
 - **`F8b_AIM_property_deltas` is the panel with power**: the paired with-minus-without difference
   in each property over 20 matched pairs, Wilcoxon signed rank on the pooled row.
 
+### 3.3 RESULT — the targeting is real, and it is roughly four times more efficient than shrinkage alone
+
+**Coverage at the time of writing: 2/20 matched pairs.** The numbers below are refreshed
+automatically as more land (`scripts/aim/autopilot_aim.sh`), and the current table is always §6 T4
+and `results/aim/tables/T4_mechanism.csv`.
+
+Averaged over the 2 pairs measured so far, on 4096 pile-val token positions and 40 layers:
+
+| quantity | value | reading |
+|---|---|---|
+| `rho_W` — weight-space shrink `‖ΔW_AIM‖/‖ΔW_merge‖` | **0.929** | AIM removes 7.1% of the merge's task vector |
+| `rho_A(uniform)` — activation shrink, activation-**agnostic** control | **0.972** | removing that much weight the *wrong* way removes only 2.8% of the activation deviation |
+| `rho_A(AIM)` — activation shrink, activation-**informed** | **0.884** | removing the same amount the *right* way removes 11.6% |
+| `aim_advantage` = `rho_A(uniform) − rho_A(AIM)` | **+0.088** | positive in 2/2 pairs |
+| leverage `(1−rho_A)/(1−rho_W)` — activation deviation removed per unit of weight change | AIM **1.63** vs uniform **0.39** | **AIM is ~4.1x more efficient** |
+
+So the answer to the question in the section title is **yes, AIM does what it claims, and the
+claim is not trivial**. The control makes that precise. Both models are, tensor for tensor, exactly
+the same Frobenius distance from the base model; the only difference is *which* input channels the
+delta was taken out of. Taking it out of the channels the base model's activations single out
+removes about four times as much activation-space deviation per unit of weight moved as taking it
+out uniformly. An activation-agnostic shrink of the same size (`rho_A(uniform)` ≈
+0.97) barely moves the representation at all — it is *less* effective in
+activation space than in weight space, which is what you would expect from perturbing directions
+the model does not use.
+
+Two things this rules out, both of which the paper's own with/without benchmark comparison cannot:
+
+- **It is not just shrinkage.** Section 2.3 showed a single scalar per tensor reproduces AIM's
+  *weights* to R² = 0.9985. That turns out to be the wrong place to look: the 0.15% of the weight
+  change that the per-channel profile adds is doing most of the representational work, because it
+  is concentrated on the channels that carry the activations.
+- **It is not an artefact of measuring "closer to base".** With-AIM is closer to base by
+  construction; the control is closer by exactly as much, and still loses.
+
+The depth profile (panel (c) of `figures/aim/F9_AIM_mechanism.png`) shows where: the three arms are
+indistinguishable through the first ~10 blocks and separate monotonically after, with the gap
+largest at the last layer. AIM is a late-layer intervention in effect even though it is applied
+uniformly across depth.
+
+**Caveat, stated plainly.** The activations are measured on pile-val, which is AIM's own
+calibration source. That is the setting most favourable to the method — a preserved activation on
+the distribution you selected the weights from is the easiest version of the claim. It does not
+follow that the preserved activations are the ones that matter for HumanEval or GSM8K, and §3.4
+(the property panel against the published outcome) is where that link would have to be made.
+
 ## 4. What did not run, and why
 
 - **No benchmark evaluation.** Ruled out on time grounds; published numbers cover all 40