# 006 phase-transition (lexical-diversity sweep) — RESULT (harvested 2026-08-01)
**Tier T3-exploratory · ★ PROMOTION CANDIDATE — the binding block's first positive.**

**Verdict: generation-side role abstraction EMERGES with lexical diversity — a sharp
jump between N=117 and N=1000 fillers — while probe-transfer stays flat. The
lexical-lookup diagnosis (003) is confirmed along with its predicted resolution.**

## Numbers (fixed 003-objective; only filler-pool size N varies; budget 1.7e8/dose)
| cell | N=30 | N=117 | N=1000 |
|---|---|---|---|
| same-vocab swapped agent acc (lookup check) | 0.984 | 0.956 | 0.988 |
| **novel-filler agent acc** (never-trained fillers) | **0.000** | **0.050** | **0.727** |
| novel-filler swapped agent acc (first-noun baseline 0.0) | 0.000 | 0.020 | **0.807** |
| novel-filler tuple acc | 0.000 | 0.033 | 0.597 |
| battery primary (probe, cross-constr + lexical holdout) | 0.500 | 0.501 | 0.495 |

## Interpretation (scoped)
- At low diversity the organism memorizes a per-lexeme lookup (perfect in-vocab, zero
  out-of-vocab). Somewhere between 117 and 1000 fillers, memorization stops paying and a
  filler-general role mechanism forms: the decoder retrieves the agent for words never
  seen in any transitive frame, INCLUDING against surface order (0.807 vs 0.0 baseline).
  This is the abstraction 001-005 failed to produce — induced not by objective, format,
  data-decorrelation, or architecture, but by LEXICAL PRESSURE on a fixed objective.
- The battery's probe cell remains at chance even at N=1000: generation-readable role
  structure ≠ linear/MLP-probe-transferable role structure across constructions. (Two
  deltas between the cells — readout type and construction novelty — pinning which one
  blocks the probe is the obvious promotion follow-up.)
- Grain of salt: parse rate dipped at N=117 (0.87) — the 0.05 there is on parsed outputs;
  the qualitative 0→0.73 jump is robust to this.

## Brier: (a) 0.35→T, (b) 0.30→T, (c) 0.10→F, (d) 0.45→T. Mean **0.306** — miss was
UNDER-confidence in the transition (priors anchored too hard on 001-005 flats).

## Promotion sketch (if funded): 3 seeds × finer N grid {117, 250, 500, 1000, 2000} to
locate/characterize the threshold; add the middle cell (novel vocab × trained
constructions) to the battery to decouple readout-vs-construction; full prereg + breaker.
Also feeds 096 (toy theory now must predict the threshold's scaling) and mirrors LLM
grokking/diversity literature (cite in writeup).