The instrument-dependence cuts cleanly. What evaporated was the leaderboard layer (which configs win, and why). What stands, because none of it was instrument-dependent, is the whole mechanism stack: direction-is-read / magnitude-is-discarded, the U-shaped depth-vulnerability curve, RYS-dup-as-power-iteration, the ~4-layer circuit-coherence length, the steering / soft-abliteration results, the continuous-depth ODE framing. The page keeps every one of those.

So the page gets a walkback banner on combo_10_38 — and, more usefully, a constructive section to replace the leaderboard it lost. The rest of this addendum is those four findings.

Granite's no-CoT arithmetic baseline doesn't fail randomly; it fails deterministically, in a handful of specific, reproducible ways. 907 − 131 (= 776) never comes back right across 15 seeds. It settles into one of two stable wrong answers: 676 (9/15, a hundreds-column slip) or 131 (6/15, echoing the subtrahend). 0 of 15 correct, two error basins. A particular carry-drop on an addition probe, likewise. I call these metastable bugs: observation the 0/15 reproducibility is what's measured; interpretation "basin / attractor" is the interpretation.

A "winning" surgery doesn't lift a diffuse competence: it flips a few of these specific bugs to 12/12 right. The measurement trap is that a balanced 49-probe average dilutes a concentrated flip until it looks like a tiny effect. The small balanced t-stats badly understate what's happening, and that's correct behaviour, not a measurement failure. The win really is concentrated. The right lens is per-probe: which exact bug got fixed, and how many configs reach it.

Once you know wins are disjoint bug-fixes, composition has an apparent rule and a non-obvious failure mode. observation depth-separated treatments tend to stack; same-region or redundant ones mask. I'm calling this a heuristic, not a law (earning the stronger word would take a regression of effect against layer-distance plus an effect-size distribution, and the margins below are thin).

The published page reads layers through a prune/dup endpoint map. That skeleton is blind to any layer with a U-shaped α-response, and Granite has them. L6 peaks at α≈0.7 (+0.062 Δ-vs-baseline on the gcore in-band probes, n=348/point: the top single-site math move) while both prune_6 (α=0, removed outright) and dup_6 (α=2, −0.105) collapse it. A gentle damp is neither endpoint, so the 3-point map can't see it: it reads L6 as "touch it and it dies," when in fact a partial damp is the single best thing you can do there. L5 similarly peaks at α≈1.3.

This sorts the layers into two math populations: damp-down layers (L7/8/10/18/26/27 want α<0.5 — L26 is the "savant," optimum monotone toward zero, where a heavy partial damp ≈ pruning it without removal) and boost layers (L9/14/15/16/19/21/22/24/29/34 want α>1). Any per-layer sensitivity claim needs the dense single-pass α grid (~0.5–1.5), not the dup/prune endpoints.

Granite was the cross-model confirmation of the Σα-law first found on MiniCPM5. The governing variable for a stacked/damped layer is Σα = the sum of all per-pass alphas; k (copy count) scales the magnitude, Σα sets the sign. On Granite the peak sits at Σα ≈ 0.75 versus MiniCPM's ≈1.25, and that downshift was named in advance: the pre-run note (2026-06-12, before the fair-battery run executed) flagged Granite's residual_multiplier = 0.22 as the suspect, and the run confirmed Σα≈0.75.

interpretation The working mechanism: a smaller per-layer residual write means a lower cumulative-α budget before the residual leaves the norm band. Beyond peak: ≈2.0 fading, ≥3 cratering, the same one-death-two-doors collapse. observation over-drive and negative-α both produce digit / special-token / CJK leak; interpretation residual-norm corruption is the proposed cause. dampall is neutral-to-negative on Granite (it isn't a collapsed model, so there's no pathology to relieve) — control behaviour consistent with MiniCPM's dampall "wins" having been de-collapse, not capability.