# 007 two-slot-bottleneck — RESULT (harvested 2026-08-01, tier T3-exploratory)

**Verdict: NEGATIVE with an informative twist — given two separately attention-pooled
slots, the organism makes them REDUNDANT TWINS rather than factoring roles.**

## Numbers
- Slot specialization matrix (2k held-out items): s1 vs s2 near-identical — agent-identity
  0.833 / 0.834, patient 0.552 / 0.580, surface-first 0.682 / 0.679. No specialization by
  the registered criterion (|Δacc| > 0.15): **Δ ≈ 0.001-0.03**.
- Battery primary 0.495 [0.484, 0.506] linear (6th consecutive flat primary); within
  0.637/0.752; surface cross 0.675. Same-vocab swapped tuple check: agent acc **1.0**
  (the lookup is perfect). val_f1 0.6413 (5e8 budget).

## Interpretation (scoped)
Architectural slots alone don't induce factoring: with no pressure differentiating the
slots, training converges to two copies of the same pooled summary (a symmetric
solution — nothing breaks the permutation symmetry between q1 and q2). Notably each
128-d slot individually carries the full lexical-lookup role signal (0.833 agent
identity), i.e. the code is duplicated, not split. Any future slot experiment needs a
symmetry-breaking term (per-slot decoder roles, slot-dropout, or orthogonality pressure).

## Brier: (a) specialization 0.80 → **F (big miss — symmetric solution not priced in)**,
(b) 0.12 → F, (c) 0.80 → T, (d) P(val_f1≥0.60)=0.70 → T. Mean **0.216**.

## Follow-ups: slot-dropout or role-keyed decoding as the symmetry breaker (worth one
campaign slot if the binding block gets extended); feeds 096/097 (theory should explain
why the symmetric copy is the attractor).