# Retriever Phase 1 evidence supplement

This supplement reports computational evaluation on published datasets. It contains
methods and recorded results; it does not distribute the Retriever application or UI.
Its manifest identifies the exact included files and their SHA256 digests.
The provenance index identifies the source run/file digest for each result; named-field
projections exclude private implementation and unrelated runtime details. Source hashes
provide identity, not independent proof of correctness.

## What readers can check

- Cohort eligibility, comparison denominators and stated exclusions. The dataset-flow
  table records an existing development snapshot, not a retrospective confirmation lock.
  Dataset inventory counts and evaluated-subset counts have different denominators.
- The numbers plotted in the human ranking and canine presentation figures.
- Declared sequence-construction, raw-read and mutant-RNA methods.
- Limits on what each experiment establishes, including missing patient DLA typing,
  training exposure, artificial peptide backgrounds and mixed human ranking results.

Published metrics and methods do not by themselves permit a complete rerun of the
private pipeline. Scientific implementation access for an independent rerun is a
separate, scoped review arrangement. No independent reviewer rerun or endorsement
is claimed here. Review of a blog is distinct from formal journal peer review.

The original publications remain the source for study data and their reuse conditions:

- Canine mammary sequencing: https://doi.org/10.1038/s41597-019-0149-8
- Canine presentation: https://doi.org/10.1016/j.isci.2026.115975
- Human TESLA: https://doi.org/10.1016/j.cell.2020.09.015
- Human IMPROVE: https://doi.org/10.3389/fimmu.2024.1360281

This supplement does not establish clinical mutation sensitivity, canine neoantigen
immunogenicity, vaccine efficacy or a partnership agreement. Candidate windows from
one mutation are correlated outputs, not independent biological observations.

## Review questions

1. Are source identities and exclusions justified without guessing unresolved aliases?
2. Do sequence checks cover the stated supported scope and disclose omitted variants?
3. Do the RNA rules distinguish lack of coverage from lack of mutant-read support?
4. Are comparison populations, model exposure and baseline denominators fair?
5. Which additional measurements would most strongly challenge the interpretation?

Please identify the supplement version, figure/table or method, concern and supporting
evidence. Any published reviewer quote or video requires the reviewer's permission;
feedback, our response and subsequent revisions should retain their dates and scope.