Supplementary source for the complete FB50 native/reference/calibration evidence.

These are exact workspace Python sources named by the frozen scientific plans,
plus their local Python import dependencies. The main twelve evidence parts
contain native-plan.json, pilot-plan.json, query-z-plan.json, fact-split.json,
per-chunk execution commands and timing, raw scores and sufficient statistics.
Use official CATT commit de4de13a705de88cc67c0b0e8f888b9803bf3397 with the pinned
model/data revisions and hashes in those plans. The experiment used Python3.12,
Torch2.6.0+cu124, Transformers5.4.0, PEFT0.18.1, BF16 held native gradients and
FP32 token contractions. CPU thread/affinity limit was10; GPU stage time limit
10800seconds. The original absolute paths and historical prerequisite hashes
are provenance, not portable paths. Regeneration requires preparing those
upstream corpus/query/native-validation artifacts in a fresh workspace and
recording an explicit path-mapped plan. This is not claimed as a turnkey replay
of GPU extraction. The main package's CPU evidence verifiers run independently.

GPU extraction entry: experiments/collect_unprojected_expandable.py --plan
native-plan.json --cell INDEX, for the20 FB indices16..35. It delegates to the
original collector, with allocator-only fragmentation repairs. The supplied
scheduler examples describe this node's execution history; use your own
approved allocation environment for a new run. No credentials or wrappers are
included. Never copy this run's private identities or ledger state.

The main package preserves the failed original packaging attempt and a JSON
whitespace-only re-encoding audit. That packaging change does not change
scientific values. Collection covers tracked LoRA factors; FB trained embedding
and output-head parameters are outside that inventory. Code MIT.

Additional CPU check: python replay_fb50_native_contrasts.py --root . --out NEW_REPORT.json . It independently reconstructs native hits from raw scores and the three descriptive headline intervals from integer per-function counts. Comparator counts are retained measurements, not a rerun of their raw score matrices.
