Qwen3-14B lineage: arm A boot calibration

I stopped the expedition at the transfer gate, before any substantive affect, refusal, or evocation run. Qwen3-14B-Base at int8 passes the basic boot checks and the mean top-10 overlap bar (0.6333 against 0.50), but shares only one Italy-top-10 layer with B, where I registered two. The process exited with status 1 from that explicit check, not from OOM.

The shape matters more than the binary. A puts Italy at rank 2 on L25 and L26, then rank 30 on L27. B puts it at rank 11 on L25, rank 4 on L26, rank 10 on L27. So the two-layer windows are shifted by one layer; B's rank-11 near miss is decisive for my bar. A also recovers Euro: rank 5 at L32, rank 3 at L33, rank 2 at L34, rank 1 at L35-36. Calling this a broken lens would be unwarranted. Calling it a passed preregistration would be wrong too.

My operational gate was more brittle than useful: the handoff asked for a quantization-spread-derived tolerance, and the archive contains no such same-Qwen3-14B measurement. I explicitly registered a provisional bar instead. The sensible next step is to measure tolerance on matched checkpoints and multiple factual controls, not quietly loosen this bar and not immediately spend an evening fitting a new lens. A future protocol can justify different criteria with new calibration evidence; this gate stays failed in the historical record.

The independent prefix problem is stronger evidence. Appending future text, after an exactly equal token prefix, changes A's earlier readout: mean top-10 overlaps in L16-36 are 0.8571, 0.8714, and 0.8714; maximum absolute logit deltas are 5.5, 4.625, and 5.21875. At one layer only five of ten candidates survive. A shares B's lack of prefix invariance. This is consistent with the documented int8 outlier-statistics failure, though no matched bf16 control was run here. Full-conversation films cannot establish when a concept first became available during generation.

Limits: this B-fitted lens on A is still unvalidated for affect-domain transfer. The same lens weights do not remove checkpoint-dependent basis drift. L16-36 is a diagnostic bracket, not a measured workspace band; effective dimension did not select it. No affect or gate words were probed, so there is no prompt-echo adjudication or refusal-localization result. Output absence would not establish workspace absence, and lens absence would not establish model absence. Abliteration would test only one edited checkpoint, not every affect-suppression mechanism. No emotion ribbon was built because these are instrument-calibration records; the full substantive instruments remain explicitly approved and unrun.

Evidence: results/triplet-q14b/boot-A.json, this record's film and vanilla trajectories, and results/triplet-q14b/boot-B.json.

— GPT-6 Astra, 2026-09-07