A calibrated lumped-parameter interior-ballistics engine gives muzzle-velocity mean absolute error (MAE) of 7.551% on 30,327 graded rows from published manuals. This fitted-corpus result includes powders used for calibration. A separate default-physics arm reports 13.66% MAE; unmatched reconstruction inputs limit causal attribution of the difference. Four graded populations now bound the engine's velocity accuracy side by side. Fitted-corpus rows give 5.64% MAE. Held-out loads within fitted powders, on an edition-split partition, give 5.05% (16,082 rows), the best estimate for the dominant user case, a new load on a powder with a fitted model. Clean never-seen loads (8,966 rows) give 11.74%, the applicable figure for a user whose powder has no fitted model; the default partition gives 11.93%. Published-data lookup wins on 4,656 identical rows with enough printed charge points: 6.29% MAE against the engine's 11.66%, a paired gap of 5.370 points with a 95% bootstrap interval of [4.161, 6.504]. Lookup scores only 52% of the never-seen partition. The engine extends calculation outside printed support and estimates pressure, whose accuracy remains insufficient for load-development safety claims. On a separate fitted-powder battery, leave-one-out ignition anchors give 5.28% max-charge MAE. Three no-exit runs are formally excluded and disclosed; none entered the graded corpus. Western-source error (10.37% MAE) remains genuine model error, including sensitivity to Accurate lot data; the stratum's 1,637 imputed-14-inch rows are its worst cell (MAE 14.58%, 27.2% beyond 20% error). The pressure-independent burn law and selection on calibration holdouts limit generalization.
The combined record was measured on 2026-11-11. The following artifacts were retrieved on 2026-11-16; deployment updates can change their contents. The paper also reports the accompanying revision-analysis JSONs: