Add preregistered external corpus panel

[almondsun]

Summary

• preregister a 2-corpus × 2-tokenizer × 3-seed sealed-test panel before corpus access
• record verified Project Gutenberg provenance, a transparent Lincoln metadata amendment, all validation/test results, paired effects, interactions, checkpoint identities, and limitations
• add a bounded sealed-matrix validator/CLI with fail-closed completeness, comparability, identity, and coverage checks
• correct likelihood-comparison fingerprints so downstream generation-only controls do not create false training mismatches
• update reviewer, architecture, training, experiment, Codex, and theory documentation

Result

ByteBPE512 beats the character control in all six preregistered same-seed comparisons. Mean candidate-minus-reference test BPC is -0.113426 ± 0.004068 on Art of War and -0.154315 ± 0.013002 on Lincoln. The mean Lincoln-minus-Art-of-War interaction is -0.040889 ± 0.017069, so the direction is robust while magnitude remains corpus-dependent.

Validation

• uv run --frozen --extra dev make check
  • Ruff format/lint passed
  • mypy passed over 44 source files
  • 166 tests passed with 90.74% coverage
  • compileall passed
  • Markdown links resolve
• exact 12-cell sealed-matrix aggregation passed
• uv run --frozen --extra dev make audit: no known vulnerabilities; local smallm package skipped because it is not on PyPI
• post-change security review: no blocking findings

Notes

The local runtime corpora, tokenizers, checkpoints, and run artifacts remain ignored. All accessed terminal test regions are consumed and must not guide later model selection.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.