PV-04 — How We Falsified Our Own Espresso Headline (static text equivalent) Generated by puckworks.public.analysis_autopsy.build_payload from puckworks.harness.result1_magnitude_comparison, source commit d4e131fa064c3394ac1dc32aaa55e6c483e14e14. Global badge: OBSERVED. Evidence strength: independent (this refers to the corrected observed replicate result only; see the per-panel partitions). Finding: an apparently dramatic interior 'fine-grind' extraction-yield maximum was superseded after the project rebuilt the target on a coherent TDS basis and checked the raw replicates. The raw cell means are ordered across dial with the middle below the coarse one; the model's interior feature is smaller than the shot-to-shot spread. Model capacity is not parameter identification. Key numbers (each produced by result1_magnitude_comparison): - observed extraction-yield CELL MEANS by dial: 18.27 -> 19.38 -> 19.62 % EY (middle BELOW coarse) - observed replicate points per dial cell (n = 3 / 6 / 3): dial 1.4 (finer): [17.743, 18.215, 18.843] % EY (cell SD 0.451) dial 1.7 (middle): [19.48, 19.439, 19.562, 19.284, 19.404, 19.117] % EY (cell SD 0.145) dial 2.0 (coarser): [19.675, 19.649, 19.542] % EY (cell SD 0.058) - middle-minus-coarse (dial 1.7 - dial 2.0) contrast: -0.241 EY-points, Welch 95% CI [-0.422, -0.06] (excludes zero) - mean within-cell sample SD across the three dial cells (uncorrected, ddof=0): 0.218 EY-points — a DESCRIPTIVE reference only, not each cell's error bar, not a noise floor, not a minimum-detectable-effect - model interior prominence: 0.154 (5 bar) / 0.0 (9 bar) EY-points; smaller than the descriptive spread: True - central-condition cup-mass value: a literal evaluation of the rounded published coefficients gives 6.723 g, the project refit gives 3.918 g, and the observed central mean is 3.876 g; limited printed precision prevents reconstructing the absolute level (rounding-limited: True) Evidence partitions (the panels do NOT share one global class): - corrected_replicate_data: observed / OBSERVED / independent — One machine, one coffee, one campaign, and one fixed nominal central condition; the within-cell spread is run-to-run variance at a fixed NOMINAL dial (the experimental unit), with no replication across machines/coffees/campaigns. A grinder dial is not a particle size and is non-portable to another grinder without a calibrated adapter. - rounded_coefficient_and_refit_audit: derived / RECONSTRUCTED / verification — A computational audit: the literal rounded-coefficient central value vs a least-squares refit vs the observed central mean. Verification against the committed data, NOT independent experimental validation. - model_interior_feature: simulated / EXPLORATORY_SIMULATION / qualitative — Model capacity to draw an interior feature at two pressures, compared with the descriptive replicate spread. This is capacity, NOT identification of channeling as the cause; 0.218 EY-points is a descriptive reference, not a formal minimum-detectable-effect or noise floor. - superseded_headline: superseded / (no current badge) / superseded/revised — The interior 'fine-grind maximum' headline read from the literal rounded-coefficient central value and the mixed-unit target. Retained visibly as a revised record; it carries NO current evidence badge. What this does NOT claim: that self-correction proves every current result is right; that a universal 'too-fine optimum' exists; that channeling is disproved or was identified as the cause (only that the data do not identify it); that the fine-grind reversal is 'fake'; any confidence claim beyond the reported Welch interval; that a grinder dial is a portable particle-size scale. Scope: Schmieder E65S grinder DoE at one fixed central flow/temperature condition; TDS-derived EY observable; grinder-dial positions (not particle sizes). Caveat: A dial number is NOT a particle size and is non-portable without a calibrated adapter; do not read a universal 'too-fine optimum'. Self-correction here downgrades ONE result — it does not prove every current result is correct. Static channeling remains a viable GENERATOR (model capacity); the data just do not identify it. The 1.4-vs-1.7 interval includes zero, so means are 'ordered', not 'statistically monotone'. No confidence claim beyond the reported Welch interval. Fidelity ceiling: Per-partition, NOT one global class: the corrected replicate correction is OBSERVED / independent; the rounded-coefficient/refit audit is RECONSTRUCTED / verification (a computational audit, not independent experimental validation); the model interior feature is EXPLORATORY_SIMULATION / qualitative and is model capacity, never mechanism identification; the superseded interior-maximum headline carries no current evidence badge. The documented revision record: - Superseded internal headline: an interior 'fine-grind' extraction-yield maximum, read from a literal evaluation of the rounded published RSM coefficients at the central condition (~6.723 g, treated as if it were the real absolute level) and a target that mixed mg solute masses with g TDS. - Problem 1 — unit/aggregation: mg-scale solute masses and g-scale TDS/cup-mass quantities were aggregated before conversion to one coherent observable; the target must first be rebuilt as a single-basis TDS-derived extraction yield. - Problem 2 — printed precision: limited printed coefficient precision makes the literal absolute reconstruction unreliable; the project's least-squares refit central-condition prediction (~3.918 g) is near the observed central mean (~3.876 g). The 6.723 g figure is a literal evaluation of rounded coefficients, NOT itself a coefficient and NOT a real 1.7x over-prediction. - Correction — coherent target: extraction yield rebuilt from TDS on one consistent basis, so the three dial cells can be compared as one observable. - Corrected replicate-level finding: the raw cell means are ordered across dial with the middle below the coarse one (18.27 -> 19.38 -> 19.62 % EY); the Welch 95% interval on the middle-minus-coarse (dial 1.7 - dial 2.0) contrast excludes zero -> no observed interior maximum in these cells. - Remaining model-capacity statement: a channeling model can still GENERATE a small interior feature, but the tested feature is smaller than the descriptive within-cell replicate spread -- model capacity, not identification of channeling as the cause. - Resulting downgrade: the interior 'fine-grind maximum' interpretation is superseded and revised down to a model-capacity statement; self-correction here downgrades ONE result and does not prove every current result is correct. Reproduce: python -m puckworks.public.analysis_autopsy verify