# 020 z-read-lens — RESULT (harvested tick 21, 2026-08-02, tier T3-exploratory)

**Verdict: ★-adjacent ARCHITECTURAL CHARACTERIZATION — z enters the decoder as a
CONSTANT per-layer bias (cross-attn read vector identical at every step: per-step
std/norm = 0.000 at all 24 layers, gate ≡ 1.0). There is no stepwise "reading" of z;
all sequencing/selectivity is internal to the decoder's autoregressive dynamics.**

## Numbers
- Constancy: 0.000 at every layer (the manager detected this during setup and
  preregistered it at P=0.97 as the arch prediction — confirmed).
- Teacher-forced read-ablation (per-class token acc, baseline → ablated):
  ablate-at-filler-steps: fillerA 0.985→0.851, fillerB 0.982→0.812 (targeted, modest);
  ablate-at-function-steps: function 0.953→0.570 with collateral filler damage
  (0.98→0.90); ablate-ALL: content acc collapses to 0.18-0.31 (z is everything).
- Free decode: baseline preserved 0.980 (anchor matches 013/019); fillerB-step ablation
  0.856 preserved (12% drop, below the 20% prereg bar); function-step ablation 0.532
  preserved / 0.422 dropped, chrF 60; ablate-all: chrF 17, 100% dropped.
- Effect sizes track HOW MANY steps lose the constant bias more than WHICH class —
  consistent with constant-broadcast + internal sequencing.

## Interpretation (scoped)
The TAE decode interface is: 24 fixed z-derived vectors injected at every position;
the decoder is a conditional LM whose entire output sequencing derives from internal
dynamics given that static conditioning. This closes the decoder block's mechanism
question and dovetails with 019 (no stepwise recovery point; order decays) and the arc
(roles/order reconstructed by surface reproduction under constant conditioning).
"Which z-directions does step t read?" — all of them, identically; selectivity is
downstream mixing.

## Brier: (arch) 0.97→T; (a) 0.30→F; (b) 0.30→F; (c) 0.30→F (12% < 20% bar);
(d) unscored (trivial under constancy). Mean (scored 4) **0.068**.

## Incidents: first run FAILED on a BOS-offset ablation-mask off-by-one (target-indexed
[B,34] vs BOS-prefixed [B,35]); tick-patched (right-pad convention: emission of target j
= decoder position j) and relaunched. Manager stopped for watcher violation (3rd);
harvest by tick.

## Follow-ups: the 24 constant read-vectors ARE z's functional decomposition at the
decoder interface — dictionary/geometry experiments (041-050 block) should analyze THEM
(W_v z per layer) rather than raw z; noted for the SAE block briefs.