# 003 structured-decoder — RESULT (harvested 2026-08-01, tier T3-exploratory)

**Verdict: the strongest shortcut-satisfaction instance yet — the decoder provably
retrieves roles from z (0.998 agent accuracy on order-swapped sentences, first-noun
baseline 0.0) while the transferable role code remains EXACTLY at chance (0.507).**

## Numbers
- Manipulation check (held-out propositions, training vocab/families): parse rate 1.0;
  **swapped-order subset (passive/objrel): agent acc 0.998 vs first-noun baseline 0.000**;
  swapped patient/tuple 0.742. General: agent 0.802, patient/tuple 0.615.
- Battery primary (novel vocab + novel constructions): linear 0.507 [0.491, 0.524],
  MLP 0.501 [0.483, 0.518]. Within-ceiling 0.688/0.750. Surface cross 0.715 (≈ baseline).
- val_f1 0.6364 (full 5e8 budget; INSTRUMENT_FAILURE verdict string = organism-genitive
  gate as always; cells read per prereg).

## Interpretation (scoped)
Role information sufficient for near-perfect who-did-what generation IS in z — the
decoder reads it against surface order, so it is not a first-noun heuristic. Yet neither
a linear nor an MLP probe finds ANY role signal that transfers across vocabulary and
construction. The induced role code is therefore best described as a **lexical lookup**
("which of my trained nouns fills the agent slot"), not an abstract role variable:
it does not exist for fillers outside the trained lexicon. Third independent objective
type (contrastive / classification / generation) satisfied without abstraction —
"shortcut satisfaction" is now a robust phenomenon across objective families at this
scale/budget. Note the asymmetry: agent slot near-ceiling, patient lagging (0.74/0.61) —
consistent with the ladder's known slot asymmetry (W_A≠W_B); logged, not interpreted.

## Brier (prereg-lite): (a) 0.80→T, (b) 0.20→F, (c) 0.28→F, (d) 0.70→T. Mean **0.062** —
best-calibrated launch so far (priors correctly absorbed 001/002).

## Caveats / follow-ups
- Same-vocab manipulation check vs disjoint-vocab battery: the contrast IS the finding,
  but a middle cell (novel vocab, trained constructions, tuple-decode with novel nouns)
  would pin down whether the lookup fails on novel fillers at generation too — cheap
  add-on if promoted.
- Single seed/rung/budget. Shares the 20% mix design.
- Feeds: 006 (phase transition should sweep LEXICAL DIVERSITY of the transitive corpus,
  not loss weight — the bottleneck is abstraction over fillers), 096 (toy theory now has
  three instances: the invariant to explain is "decodability without abstraction").