# 010 role-critical-mt — RESULT (harvested 2026-08-01, tier T3-exploratory)

**Verdict: TASK-LEARNING FAILURE, and it's the most informative kind — with a
scrambled-order case-marked target, the organism NEVER LEARNS role attachment at all:
case accuracy 0.503 on held-out SAME-VOCAB items (baselines 0.5025), vs ~0.99 for the
fixed-slot tuple objective at identical scale/budget (003/006).**

## Numbers
- Format learned perfectly: parse rate 1.0, verb marker acc 1.0, case_recon loss 0.248
  (lemma content + template mastered). Role assignment at chance everywhere:
  general 0.503, swapped subset 0.528, novel fillers 0.490. agent_acc == patient_acc ==
  full_acc exactly (one coin-flip per sentence, consistently applied to both markers).
- Battery primary 0.501 [0.494, 0.507] — flat (unsurprising given the objective never
  encoded role). val_f1 0.6468 (OWT unharmed).

## Interpretation (scoped)
Contrast with 003/006 isolates WHERE the role work was being done: with a fixed slot
order, the decoder's positional scaffold shoulders the role computation (read slot 1 →
emit agent) and same-vocab accuracy is ~0.99; remove the positional scaffold (random
output order, case as the only carrier) and rung-A at 5e8 tokens/20% mix cannot learn
the mapping AT ALL — not abstraction, not even the lexical lookup. Order-free role
marking appears to require a per-lemma role TAG readable at arbitrary output positions,
and nothing in this architecture/budget forms one. This reframes 006's positive: the
diversity transition produced filler-general role retrieval WITHIN a positional output
scaffold; the scaffold itself was load-bearing all along. (Caveat: single seed/budget/λ;
a longer run or higher mix might eventually learn it — "unlearnable" here means at the
matched budget where fixed-order was easy.)

## Brier: (a) novel-filler >0.6 (0.72) → F — biggest campaign miss so far, and for an
unpredicted reason (same-vocab failure, which no prediction priced); (b) 0.22 → F;
(c) 0.38 → F; (d) val_f1≥0.60 (0.80) → T. Mean **0.238**.

## Follow-ups: the binding block's closing synthesis should present 003/006/010 as a
triptych (scaffolded lookup → scaffolded abstraction → no scaffold, no learning).
Curriculum idea if ever extended: fixed-order pretraining THEN scramble (does the
abstraction transfer into the case system once formed?). Feeds 096 theory directly.