# Claim 5 limitations and deviations

- No Table 5 model was trained. The exact head implementations, row split,
  schedule, seeds, and checkpoint-selection rule are not public enough for a
  faithful comparison.
- No Table 6 model was trained. The exact smaller training subset, evaluation
  row IDs, schedule, seeds, and RLM checkpoints are unidentified.
- The cited Qin et al. implementation establishes a defensible interpretation
  of normalized regression but is not substituted for the paper experiment.
- Public T5Gemma base metadata verifies rounded architecture sizes only; it
  cannot verify or falsify a correlation improvement.
- Gated base access is recorded as an access constraint, not as scientific
  evidence.
- The released 181.5M RLM discrepancy is not treated as a counterexample
  because that release is not identified as either exact Table 6 checkpoint.
- Public non-discovery cannot prove that private artifacts do not exist.
- The final scientific verdict is BLOCKED with LOW confidence and earns no
  forecast points.