BTA — R1.8 counterfactual cohort (5 seeds)

5-seed cohort of the structural counterfactual adapter that recovers nonlinear-decodable Φ above the K_T text-only floor.

Architecture: identical to R0 (84M MLP-2 + conv adapter) Encoder: WavLM-Large layer 16 (frozen) LLM: Qwen3-8B-Instruct (frozen, enable_thinking=False) Loss: $\mathcal{L}{\mathrm{BLSP}} + 5 \mathcal{L}{\mathrm{cf}} + \mathcal{L}{\mathrm{artifact}} + 0.5 \mathcal{L}{\mathrm{cond}}$ (λ_NCE=0) Schedule: 400 steps + warmup 300, cosine to 1e-6, max_norm=1.0 Aug: aggressive_stage3_config (apply_prob=1.0, bandlimit ≤5kHz, SNR floor 0dB, codec mu_law/mp3/opus) Code: https://github.com/Nurgali-Kadyrbek/frozen-speech-llm-stress

Files

File	Size
`A_R1p8_seed1234.pt`	357 MB (dereferences from Stage 3.7; equivalent under λ_NCE-orthogonal-to-update)
`A_R1p8_seed1235.pt`	357 MB
`A_R1p8_seed1236.pt`	357 MB
`A_R1p8_seed1237.pt`	357 MB
`A_R1p8_seed1238.pt`	357 MB
`grad_norm_clip_log_seed*.json`	per-step λ + cosine-to-update + clip-fire diagnostics

Reported metrics (5-seed cohort)

Metric	Mean ± σ	Δ vs R0
Probe-G total	0.6326 ± 0.0061	+0.0020
Probe-G$_{\mathrm{neutral}}$	0.5122 ± 0.0039	+0.0000 (F3 — at consumer baseline)
Probe-G$_{\mathrm{explicit}}$	0.7531 ± 0.0138	+0.0040
Probe-K linear eval_full	0.2256 ± 0.0118	+0.0151
Probe-K MLP-2 eval_full	0.3059 ± 0.0436	+0.0613 (clears K_T floor 0.290)
C.1 PASS	5/5 seeds	(R0 was 0/5)
C.2 ratio	0.910	(R0 ~0.7)

The MLP-2 cohort > K_T finding is validated by Stage 4 controls: Control A (text-only adapter via projection P): MLP-2 = 0.2122; gap = -0.094. Control B (audio-decorrelated cf-pairs): MLP-2 = 0.2033; gap = -0.103. Shortcut probes (tts_vs_real, domain): 5/5 seeds clear the per-seed 0.10-absolute-drop threshold vs raw H baselines.