# 005 passive-curriculum — RESULT (harvested 2026-08-01, tier T3-exploratory)

**Verdict: NEGATIVE — data-side decorrelation alone (P(agent-first)=0.500 hard-asserted,
40% mix, no role objective) induces no transferable role code.**

## Numbers
Battery primary 0.504 [0.488, 0.519] linear / 0.505 MLP — chance, 5th consecutive flat
primary in the binding block. Within-ceiling 0.632/0.725; surface cross 0.704 (≈baseline);
flipped-parity 0.540 [0.466, 0.613] (CI spans chance; not interpreted). val_f1 0.6409
(full 5e8 budget). INSTRUMENT_FAILURE verdict string = organism-genitive gate, as always.

## Interpretation (scoped)
Making surface order uninformative about role in 40% of the data gives plain DAE no
reason — and evidently no tendency — to develop role abstraction. Together with 001-004:
neither objective pressure (contrastive/classification/generation) nor data pressure
(decorrelation, forced order) produces a transferable role code at rung-A scale/budget.
The binding block's remaining hopes: 006 (lexical-diversity threshold, running), 007
(architectural slots), 008-010.

## Brier: (a) 0.08→F, (b) 0.15→F, (c) 0.40→F, (d) 0.65→T. Mean **0.078**.

## Caveats: single seed; mix fraction 40% differs from 001-003's 20% (pre-logged as the
design intent — data-hypothesis arm anchored to A_D_s0 baseline, not to 001-003).