| Qwen3.8-27B abliterated: integer attention with frozen PR 178 scales on 8 held-out prompts, range transfer and per-head fidelity | negative, unreplicated | Held-out check of PR 178's integer attention: per-layer exponents frozen from PR 178's in-sample fit, evaluated open loop on 8 new prompts (541 tokens) on the 16 full-attention layers of Qwen3.8-27B abliterated, nothing fitted on the new prompts. Two axes per arm: H_R = zero q/k/v clamps and zero guard failures; H_F = PR 88's A2 per-head bar (median rel_y <= 0.01, max <= 0.05 over 192 (prompt, head) cells per layer, every layer). Measured, unreplicated, one deterministic CPU run: C2047: H_R FAIL (clamped q/k/v 0/0/1 of 53,182,464/8,863,744/8,863,744, guard failures 0, range-fail layers [23]), H_F FAIL (13 of 16 layers meet the bar, failing [31, 35, 39]; per-layer median rel_y 0.00336 to 0.0127, per-layer max 0.00482 to 0.0345; pooled 3,072 cells median 0.00709, over 0.05: 0); C1665: H_R FAIL (clamped q/k/v 1/1/2 of 53,182,464/8,863,744/8,863,744, guard failures 0, range-fail layers [11, 59, 63]), H_F FAIL (10 of 16 layers meet the bar, failing [27, 31, 35, 39, 43, 59]; per-layer median rel_y 0.00525 to 0.018, per-layer max 0.00736 to 0.0325; pooled 3,072 cells median 0.00822, over 0.05: 0). Open loop per head: not token agreement, no logits. Eight prompts are not a deployment distribution. No proof | abliterated checkpoint (stock layers 0 to 16 byte-identical), 8 held-out prompts written for this experiment (prompts-heldout.json, split sha256 e5245bd6), no exact match with PR 124's 10 calibration prompts; open loop per attention layer on the float64 stream; lowered from rotated q, k, v to o only; E_q, E_k, E_v, sh copied from PR 178 runs/layers.jsonl (sha256 73404d47), int_attn.py imported unchanged; CPU, one deterministic run, determinism checked; shared loaded host; no proof | [experiments/20261114-204-int-attn-heldout/RESULTS.md](experiments/20261114-204-int-attn-heldout/RESULTS.md) |
0 commit comments