Under the fixed custom-wrapper evaluation, the streamed endpoint improves TextVQA and DocVQA over its starting checkpoint. This is a single-run result with paired image-cluster confidence intervals. Full denominators and stored paired sufficient counts are in the result package.
It is not a frontier model, a SmolVLM-parity result, a pure data-quantity ablation or a population-level scaling law. Source mixture, replay and multiple-choice label formatting changed together. Higher-resolution or trainable-vision variants remain hypotheses for controlled tests.
Retention failed. Old natural-image QA fell from 56.25% to 50.86%; two consecutive checks exceeded the 5-percentage-point tolerance. Other retention checks passing does not erase that failure. The final model is not promoted.
AI2D's baseline default-filter score is 7.32; the same original predictions score 25.62 under a separate prefix-tolerant diagnostic. That diagnostic is not a replacement for the primary metric. The 26.26 after-score therefore must not be presented as a 19-point gain in visual reasoning alone.
ChartQA's 95% paired interval crosses zero. Two known exact-overlap questions from prior training remain in the reported full split; all questions are kept and the earlier sensitivity result is separately recorded.
The four full published splits contain 15,937 questions. They were evaluated with a custom wrapper aligned to pinned scoring/task definitions, not the full lmms-eval CLI. No leaderboard submission is claimed. Public baseline results were inspected before the new run and informed format handling, so they are not a blind confirmatory endpoint. Intervals describe evaluation-image sampling only, not training seeds, search choices or four-task multiplicity.
Exact hashes, pixel hashes and registered near-image checks are useful but do not certify semantic or original-document independence. Prepared source counts are not completed optimizer exposures. Final runtime in the saved training summary covers the resumed process only; it is not the entire run duration.
Published evidence uses machine-path aliases. Both original-source and published-file hashes are recorded; transformed JSON is not falsely claimed byte-identical to the remote files. Paired score exports contain image hashes, question counts and score sums, not images, prompts or reference answers.
The text base is not instruction- or safety-tuned. VLM weights and optimizer states remain private. The public repository is not a self-contained current VLM release. Code MIT licensing does not relicense external models or datasets.