PV-05 — More Physics Made It Worse (static text equivalent) Generated by puckworks.public.model_composition.build_payload from the named producers, source commit 7dd257e0b339f174c5da090a4a031e6e2c1b8a6a. Badge: EXPLORATORY_SIMULATION. Evidence strength: qualitative. Finding: a time-varying extraction-only model follows the rising 9-bar flow far better than a constant baseline. But adding an imported swelling branch through ONE shared porosity state OVER-closes that state, so the composite prediction decouples from the rise and scores worse than the constant. The result rejects THIS composition, not swelling physics in general. Key numbers (each produced by a named function, over the 15-95 s window): - constant baseline RMSE (1 free param): 0.573 g/s at a level of 1.584 g/s - extraction-only RMSE (0 free params): 0.116 g/s (cross-checked by degeneracy_rmse = 0.116 g/s; agreement 0.0 g/s) - extraction beats the constant by 4.94x - composite (extraction + swelling) RMSE: 0.648 g/s = 1.131x the constant and 5.586x the extraction-only model - shared porosity over-closes to 0.0998 (below the eps0 reference 0.17); swelling closes the shared state: True; clamp hit: False What this does NOT claim: that physics is universally harmful; that swelling is absent from real pucks; that the extraction-only mechanism is proven; or any connection to the Guided Espresso Pull primary chain. Scope: One dataset/campaign (waszkiewicz2025 9-bar trace), one observation operator, one tested shared-porosity composition, one shared-state definition; the swelling parameters are mo2023_2's, applied to a saturated pre-wet rig. Caveat: This rejects THIS composition, not swelling in general and not a universal rule that added physics reduces accuracy. It does NOT prove the extraction-only mechanism is correct or uniquely identified. The composition is not coupled into the Guided Espresso Pull. Components are fitted/calibrated; a lower residual is verification against one trace, not independent validation. Fidelity ceiling: coupled kappa(t) FRAMEWORK; branch fidelity inherited from donors (mo2023_2 swelling is fixed-dP and unvalidated on a saturated rig). Exploratory model-composition evidence, never a validated composition. Before combining components, a composition must answer: - Does each component use the same state definition? - Does the state share one reference volume and the same bounds? - Are the time origins and the pressure/flow nodes compatible? - Does a fitted parameter in one component already absorb what the other adds? - Are mass and energy balances preserved across the composition? - Do both components operate in overlapping evidence domains? - Is the combined observation operator still physically meaningful? - Does the composite beat a simple null benchmark? Reproduce: python -m puckworks.public.model_composition verify