E8: the transplant keeps a weaker English-like latent pivot, and pivot strength shows no detectable association with answer correctness.

By @stw2 via agent · covers #167 · currently selected

Assignment, status and accepted plans remain authoritative on each experiment.

Reported state

Reported progress: E8 completed with two retrospective plan versions (the design and Amendment 1): a teacher-forced logit lens over the existing traces of both 11B models, tests A1, A2 and B.
Open issues: Test B had no registered decision rule or equivalence bound; its interval admits a positive association up to 0.3386 log-odds per standard deviation (0.4596 in the strict-agreement arm). The Polish-model lens run changed its run hash between two invocations with no record of the change. The pivot magnitude depends on vocabulary partitions built from re-serialised MLX tokenizer files without an equivalence check. The design was written after the lens run started. The two layer profiles average different item sets.
Suggested next action: Open follow-up: test the pivot causally by steering or patching the residual stream over layers 26 to 43, or repeat test B with a registered equivalence bound.
Access needs: Both models are gated with automatic approval; as written, the lens runs need Apple silicon, 128 GB unified memory; the complete answers, the PES and LLMzSzŁ items, the few-shot exemplars and two corpora are restricted materials: ask the Room owner, or rebuild them.

Exact record · H22

Pivot hypothesis tested by A1.

Exact record · H23

Accuracy hypothesis tested by B.

Exact record · P195

Test A1 of the pivot hypothesis: the interval lies above zero.

Exact record · P197

Test A2, descriptive and directional: the calibrated pivot is weaker in the transplant.

Exact record · P201

Sensitivity variant of A2 with hard decoding.

Exact record · P204

Test B of the accuracy hypothesis: the interval contains zero and no decision rule was registered, so the test neither supports nor contradicts it.

Exact record · P205

Secondary analysis of test B with a difficulty control; the interval contains zero.

Exact record · P207

Secondary analysis of test B by benchmark.

Exact record · P208

Accuracy of both models on the analysis universe.

Exact record · P209

Descriptive layer profile.

Exact record · P210

Descriptive layer profile.

Experiment · E8

Accepted plan

Accepted plan