# RESULT — Row 057 diffusion-latents (EXT) — T3-exploratory

*Run 2026-08-03, tmux c100_057, launched 13:03:56 UTC, DONE 13:16 UTC (~13 min wall; smoke
passed first, artifacts deleted pre-launch). 0 errors, 8/8 batteries. Box CVD=0→phys0
hard-guarded; GPUs 35/15/15 MiB at claim AND exit; box tmux server empty both ends.*

## What ran
`diffusionfamily/diffugpt-s` (DiffuGPT-small 124M, Gong et al. 2024, arXiv 2410.17891 — GPT-2
adapted into a genuine masked discrete-diffusion LM; official HKUNLP/DiffuLLaMA inference code
vendored; pinned venv057: torch 2.5.1cu121 + transformers 4.44.2; battery stage in night8 venv,
read-only). Stimuli_v2 battery UNCHANGED (agent_patient 2000 + genitive 1200; linear+mlp; seeds
0,1,2; n_boot 1000) on 8 reps: denoiser final-layer mean-pooled states of the stimulus corrupted
at masking rate r ∈ {0, .25, .5, .75, .9, 1.0} (K=4 deterministic draws averaged for 0<r<1) +
measured anchors gpt2_h (the model's own AR parent, same arch/tokenizer/pooling) and emb_mean
(wte bag). Honest adaptation pre-logged: noise level = masking rate of the forward process
(masked diffusion has no Gaussian t); trajectory = conditioning on the stimulus at decreasing
corruption, not free-running generation.

## Gates
- G0 bidirectionality: PASS (late-token edit moves early-position logits Δ=2.74 in DiffuGPT;
  exactly 0.0 in gpt2 anchor; 148/148 hub tensors verified loaded pre-freeze).
- G1 functional denoiser: PASS — paragraph (~100-tok) denoise top-1 acc 0.559 @ r=.15 / 0.381
  @ r=.5 / 0.110 @ r=.9 (gate ≥.35/≥.15, monotone).
- G2 battery internals: z_bag = 0.500 everywhere; strict 0.9 gate → formal INSTRUMENT_FAILURE
  ×8 incl. both anchors (expected and pre-registered; measured-anchor interpretation per 052 A1).
- G3 probe power: PASS — dg_r000 ap within (mlp) 0.658 ≥ 0.60 → noise-arm nulls interpretable.

## Headline: SEVENTH NO — iterative denoising does not install role binding, and noise never
## induces abstraction

Commitment curve (agent_patient, mlp; within-AUC / primary cross-lexical pooled [CI]):

| rep | ap within | ap primary | genitive within |
|---|---|---|---|
| gpt2_h (AR parent) | **0.730** | 0.495 [0.475,0.519] | 0.741 |
| dg_r000 (clean) | 0.658 | 0.478 [0.460,0.495] | 0.704 |
| dg_r025 | 0.586 | 0.491 [0.474,0.510] | 0.632 |
| dg_r050 | 0.549 | 0.511 [0.487,0.534] | 0.557 |
| dg_r075 | 0.545 | 0.488 [0.461,0.517] | 0.536 |
| dg_r090 | 0.524 | 0.500 [0.465,0.534] | 0.517 |
| dg_r100 (all-mask floor) | 0.524 | 0.496 [0.464,0.543] | 0.508 |
| emb_mean (bag) | 0.645 | 0.493 [0.481,0.509] | 0.500 |

1. **Binding cell dead at every noise level**: primary 0.478–0.514 across all 32 rep×readout
   cells; max CI-lo anywhere = 0.497 (dg_r050 linear) — chance. The arc's 7th consecutive NO
   (051 contrastive, 052 MT, 053 instruction, 054 LCM, 055 latent CoT, 056 cross-attention,
   057 diffusion): no generation mechanism tested implements abstract role binding in pooled
   sentence states.
2. **No noise-induced abstraction — the row's genuinely open question answers NO**: the
   within-code decays monotonically with masking (0.658 → 0.549 → 0.524) with no mid-noise
   resurgence; the denoising posterior never has MORE role structure than the clean encoding.
   Nothing "commits" to who-did-what at any noise level (row 015's L22→L24 commitment has no
   masked-diffusion analogue in this model): information only leaves, it is never abstracted.
3. **Diffusion adaptation mildly ERODES the parent's code**: dg_r000 within 0.658 vs gpt2_h
   0.730 (mlp, −0.072; linear −0.066), genitive 0.704 vs 0.741. Converting GPT-2 to a
   bidirectional denoiser cost probe-accessible role/construction separability — the opposite
   sign from "diffusion pressure enriches intermediate representations".
4. **What does commit early is surface scaffold** (exploratory): at every rate the denoiser
   recovers function words far better than content words (r=.25: 0.230 vs 0.052; r=.9:
   0.054 vs 0.034) — the first thing the posterior locks in is the construction frame, not
   the proposition. Geometry: smooth collapse toward a degenerate all-mask attractor (norm
   18.8→26.9, meancos 0.873→0.9999, pc1_evr 0.094→0.900 at r=1.0) — a designed-input analogue
   of 055's OOD fixed-point collapse. emb_mean genitive within = 0.500 exactly (bag of GPT-2
   embeddings carries zero role info; replicates 055).

## Predictions scored (Brier = **0.134**)
- P1 (.75) r000 ap within (mlp) ≥ .60 → **T** (0.658). Brier 0.0625.
- P2 (.85) no cell primary CI-lo > .55 → **T** (max 0.497). Brier 0.0225.
- P3 (.70) within non-increasing r000≥r050≥r100 (tol .02) AND r100 ≤ .55 → **T**
  (0.658/0.549/0.524; floor 0.524). Brier 0.09.
- P4 (.60) |r000 − gpt2_h| ap within (mlp) ≤ .07 → **F by 0.002** (0.072; linear 0.066 would
  pass, prereg specified mlp — scored honestly as miss). Brier 0.36. Razor-thin; the
  directional content ("neither installs nor destroys") is half-wrong anyway: it destroys a
  little.

## What did NOT run / limitations
- Final-layer mean-pool only (no mid-layer sweep — a fold-in, not rerun-worthy).
- MDLM/SEDD rejected on flash-attn hard dependency; LLaDA/Dream rejected on 16 GB (BLOCKED
  branches recorded in PREREG; DiffuGPT is the best-provenance runnable diffusion LM here).
- Short battery sentences (5–14 tok) vs packed-1024 training (pre-logged OOD caveat); G1/G3
  bound the damage but clean-state within 0.658 < gpt2's 0.730 may partly reflect this.
- Masking rate ≠ continuous Gaussian noise; free-running generation trajectories not probed
  (would abandon controlled stimuli).
- Strict battery gate INSTRUMENT_FAILURE ×8 (incl. anchors) — all claims are measured-anchor
  relative (052 A1), T3 only.

## Follow-up worth funding? N as a row.
The 051–057 arc is now uniform across seven generation/encoding mechanisms; marginal value of
an eighth small-model NO is low. Cheap fold-ins if the arc resurfaces: layer sweep of dg
states; DiffuGPT-medium; LLaDA-8B int8 on a bigger card. The arc-level synthesis (what
training signal WOULD install binding) is the fundable question, not another mechanism cell.

## Artifacts
Repo `out/`: 8 battery JSONs, BATTERY_RESULTS_057.md, func_check_057.json, geometry_057.json,
run.log. Box `057-diffusion-latents/`: states npz ×8 (~75 MB) + venv057 (~6 GB, uv-cached
wheels) kept; HF cache +250 MB (diffugpt-s). tmux c100_057 self-ended; nothing killed;
GPUs 35/15/15 MiB at exit.