yall/fullset-f2-repro/CALIBRATION_COMPARISON.md

0001 # FullSet F2 calibration comparison
0002
0003 This note records a numerical validation of the `fullset-f2-repro` workflow
0004 against Fredi's published FullSetF_2 calibration. It is an analysis check of the
0005 reproduction workflow, not a description of the dataset.
0006
0007 Comparison performed: **September 17, 2026**.
0008
0009 ## Files compared
0010
0011 Published reference:
0012
0013 ```text
0014 calibrations/TB2026/calib_SPS-H2_FullSetF_2.txt
0015 ```
0016
0017 Reproduced final calibration:
0018
0019 ```text
0020 $LFHCAL_WORK/fullset-f2-repro/final/calib_Final_Muon_FullSetF_2_calib.txt
0021 ```
0022
0023 The reproduced pedestal, initial MIP calibration, and all five refinement tables
0024 were also retained and compared with the published final calibration to inspect
0025 convergence.
0026
0027 ## Result
0028
0029 The reproduction agrees very closely with the published calibration.
0030
0031 - Both final tables contain the same 384 cell IDs and identical geometry.
0032 - Both contain 380 positive HG MIP scales and the same four unavailable scales.
0033 - The median absolute relative HG MIP-scale difference is `0.000003374%`.
0034 - The largest absolute relative HG MIP-scale difference is `0.005241889%`.
0035 - All 380 positive scales agree within `0.01%`; 372 agree within `0.001%`.
0036 - The reproduced final text file is byte-for-byte identical to the reproduced
0037   `Imp5R` text file, as expected from the Yallfile finalization step.
0038
0039 Relative differences here are
0040
0041 ```text
0042 100 * (reproduced - reference) / reference
0043 ```
0044
0045 and are evaluated only where both stored quantities are positive.
0046
0047 The largest relative scale difference occurs at cell 1088:
0048
0049 | value | Fredi | reproduced |
0050 |---|---:|---:|
0051 | HG MIP scale | 17.798927 | 17.799860 |
0052
0053 The difference is `+0.000933`, or about `+52.4 ppm`.
0054
0055 The 12 largest relative scale differences are localized in modules 2 or 3,
0056 row 0, column 1. They are very small, but the localization is worth preserving
0057 as a clue if the residual numerical differences are investigated later.
0058
0059 ## Refinement convergence
0060
0061 Each reproduced stage below is compared with Fredi's **final** table; Fredi's
0062 intermediate refinement tables were not available for a stage-by-stage direct
0063 comparison.
0064
0065 | reproduced stage | positive HG scales in both | median absolute difference | largest absolute difference |
0066 |---|---:|---:|---:|
0067 | Initial MIP | 372 | 0.820576% | 76.910828% |
0068 | Imp1 | 380 | 0.030197% | 8.447758% |
0069 | Imp2 | 380 | 0.000045% | 0.037918% |
0070 | Imp3 | 380 | 0.000004% | 0.005242% |
0071 | Imp4 | 380 | 0.000003% | 0.005242% |
0072 | Imp5 | 380 | 0.000003% | 0.005242% |
0073
0074 Eight channels first acquire positive scales in Imp1. By Imp3 the calibration is
0075 essentially settled; from Imp3 to Imp5 the largest positive-scale change is only
0076 about `3.44 ppm`.
0077
0078 Cell 1351 has a conspicuous intermediate excursion in the stored HG MIP width:
0079
0080 | stage | HG MIP scale | stored HG MIP width |
0081 |---|---:|---:|
0082 | Initial MIP | 53.371104 | 35.157470 |
0083 | Imp1 | 58.716285 | 1.380049 |
0084 | Imp2 | 54.146596 | 38.125254 |
0085 | Imp5 | 54.142418 | 38.118359 |
0086 | Fredi final | 54.142461 | 38.118655 |
0087
0088 The final result nevertheless converges closely to the reference. The stored
0089 width is a calibration width, not an uncertainty on the fitted MIP scale.
0090
0091 ## Pedestals and shared missing calibrations
0092
0093 Pedestal values also agree closely. The largest absolute differences in the
0094 stored pedestal quantities are:
0095
0096 | pedestal field | largest absolute difference |
0097 |---|---:|
0098 | HG mean | 0.000511 |
0099 | HG sigma | 0.000606 |
0100 | LG mean | 0.000227 |
0101 | LG sigma | 0.000051 |
0102
0103 The same four cells have HG MIP scale and width equal to `-1000` in both final
0104 tables:
0105
0106 | cell | module | layer | row | column | BC |
0107 |---:|---:|---:|---:|---:|---:|
0108 | 967 | 1 | 7 | 1 | 3 | 0 |
0109 | 1026 | 2 | 2 | 0 | 0 | 3 |
0110 | 1031 | 2 | 7 | 0 | 0 | 0 |
0111 | 1924 | 3 | 4 | 1 | 2 | 3 |
0112
0113 Cells 1026 and 1924 also lack HG pedestal mean and sigma values in both the
0114 reproduced pedestal output and the published final table. These are therefore
0115 shared calibration gaps, not failures introduced by the reproduction.
0116
0117 ## Conclusion
0118
0119 The FullSetF_2 Yall workflow passes this numerical reproduction check. It
0120 recreates the published final calibration with the same channel coverage and
0121 very small residual numerical differences, and the refinement sequence converges
0122 stably to that result.
0123
0124 This comparison validates the exported calibration constants. It does not by
0125 itself validate ROOT histogram identity, fit covariances, individual
0126 Landau/Gaussian component parameters, event-by-event identity, or the exact
0127 runtime/container environment used for the original published calibration.