# Methods and deviations — Polish cross-lingual pilot (PLLuM-12B-chat-2512)

Run id: `plpilot-04c874127fd7-20261017T183028Z-f212` · manifest SHA-256: `04c874127fd71d61b66c1e011855d21ac57c03a140ac8e85b4daec3685d8be87`

## Method
- 37 paired capability tasks (main37), conditions `deploy` / `only_F1` / `only_F8`; 111 English + 111 Polish prompts; last prompt token (`[/INST]`) residual stream; layers 10, 19 (main), 20, 30; bf16; batch size 32.
- English `v7` (held factor F1): unit(mean of 336 single-factor activations, 7 factors x 48 tasks, minus mean deploy activation). Global axis: unit(mean eval_all − mean deploy over 100 stored English main pairs). Both are applied unchanged to Polish activations (D1).
- AUROC of the scalar projection within a language: positives `only_Fk`, negatives `deploy`.
- Bootstrap: whole tasks with replacement, `numpy.random.default_rng(0)`, 1000 resamples drawn once, same indices for EN and PL, percentile 2.5/97.5, `v7` fixed.

## Three layers (kept separate)
1. Original plan (Claude Docs document, revision 13, 2026-10-15): criteria (a)/(b)/(c), margin ±0.10, length-control conditions. Not frozen by hash.
2. Decisions before measurement: frozen manifest and code; addendum `ADDENDUM_DECYZJE_PL_PILOT.md` (sha256 `4f99bcdd8754e72cc543aeba44c5184e9f250eb8c4cd4de077cc0d327670ac4e`): D1, D2, D6, D8, D11, D12, D13.
3. After measurement: post-hoc script; 31-task sensitivity and D11 length diagnostics (named in an addendum that is not frozen by hash); exploratory deploy-only length correlations and bootstrap-seed variability (not preregistered).

## Deviations and open points
- D13: F8 length criteria unmet (SD ratio PL/EN 1.532 vs 1.5; F8 addition difference 27.2 tokens vs ±5). Accepted as a deviation before the measurement; status remains UNMET; D13 does not determine the effect on the verdict.
- No precedence rule between (a) and (c) in the plan; none added after the result. Formal verdict: inconclusive; (c) under the original length-control condition.
- D8 froze code, corpus and manifest, not the full preregistration with criteria.
- Global axis from 100 pairs (plan text: 115).
- D12: no author control (the 12-version control stems from later addenda). No documented native-speaker review.
- Evaluated tasks took part in estimating English `v7`; native AUROC values are in-sample.
- Re-measured English activations differ slightly from the historical ones; cause not established.
- Historical 48-task F1@L19 (0.6072) is not a comparator for the 37-task EN/PL values.

## Provenance of commits
- Code matching manifest hashes: `f0ea8a83fe841af87a1b303c58856a61d239b2bb` (private dataset repo). `preflight_report.json` records this as the repository head at preflight start; the downloaded revision itself was not separately recorded, so it is not claimed as fully confirmed. Metadata was not back-filled.
- Post-hoc analysis: `22c754441e040c5e6f675661fa1c561a2064f171`.
- Page publication: the Space commit that adds this material (recorded in `PUBLIC_ARTIFACT_INDEX.json` only by reference, not self-referentially).