ALP-ttbar Toy MC v2.7 — Preregistration: Tag Formula Diagnostics in Displaced Light ALP → e⁺e⁻

Preregistration for v2.7 of the ALP-ttbar Toy MC series: Tag Formula Diagnostics. The v2.6 multi-seed ensemble (N_seeds = 50, E3 configuration with η = 2.0) revealed that the Tag discriminator exhibits unexpectedly large stochastic variation at m_a = 17 MeV: σ/mean = 7.4%, compared to only 0.16% for the HV discriminator — a 46× ratio. Both discriminators share the same vertex populations, so this disparity points to a mechanism specific to the tag formula. We hypothesise that extreme-λ vertices drive Tag variability: the tag formula's conversion probability density p_conv(z) has steep gradients near material band edges and drops to a numerical floor in inter-band gaps, producing extreme λ values that disproportionately influence the Tag degradation ratio. This preregistration specifies three complementary diagnostic approaches to test this hypothesis: Approach A — Frozen-Seed Variance Decomposition: Three seed groups isolate the contribution of each vertex population. Group F (Full) varies both signal and conversion seeds (SEED_CONV = 100i+5, SEED_SIG = 100i+11, reproducing v2.6). Group C (Fixed-conv) fixes SEED_CONV = 5 while varying SEED_SIG, isolating signal-vertex randomness. Group S (Fixed-signal) fixes SEED_SIG = 11 while varying SEED_CONV, isolating conversion randomness. Comparing σ_Tag across groups identifies which population dominates the variance. Approach B — Per-Vertex Weight Analysis: For each signal vertex, we record the per-vertex λ(z) value and classify vertices into four mutually exclusive groups based on their proximity to material band edges and their p_conv magnitude. "Near-band-edge" vertices satisfy min(|z−(Z_d−HW_d)|, |z−(Z_d+HW_d)|) < D for distance thresholds D ∈ {1, 2, 5, 10, 20, 50} mm. "Inter-band" vertices have p_conv(z, σ_z) < ε = 10⁻¹². These groups can overlap — for example, z = 310 mm is 7 mm from the nearest band edge and has p_conv ≈ 1.6 × 10⁻¹⁴ — so we use a 4-way classification: near-edge only, inter-band only, both, or neither. The numerical floor p_conv ≤ 10⁻³⁰⁰ is recorded separately. Fractions f_near and f_gap use N_used (actual vertices passing basic cuts) as the denominator. We measure correlations R(f_near, Tag) and R(f_gap, Tag) separately, and compute λ distribution moments for all four groups. Approach C — Robust Estimator Comparison: We define trimmed and winsorised variants of the Tag to test whether removing or clipping extreme-λ vertices stabilises the discriminator. Trimmed Tag (k%) sorts the signed λ values for each population and removes n_rem = ⌊k × N / 100⌋ vertices from each tail (the n_rem smallest and n_rem largest), totalling approximately 2k% removal. A boolean removal mask ensures these vertices never contribute to S or B in any (t, mass-window) combination within scan_tag_endcap. Winsorised Tag (k%) selects the identical ⌊k × N / 100⌋ vertices from each tail but replaces their λ with the nearest retained boundary value instead of removing them. This shared vertex selection ensures the R2 comparison isolates the effect of removal versus clipping with no difference in which vertices are targeted. We scan k ∈ {0.5, 1, 2, 5, 10}%. The generation pipeline follows v2.6: endcap_signal generates 300,000 signal vertices and endcap_conversions generates 1,500,000 conversion-photon candidates per seed. After basic-cut selection (fiducial z ∈ [200, 2500] mm, minimum p_T), subsample() retains the first N_sig = 50,000 signal and N_conv = 100,000 conversion vertices passing cuts (or fewer when the base-passing yield is below the cap); vertex weights are rescaled to preserve the total cross-section. Twelve predictions with quantitative thresholds are preregistered across the three approaches: four variance decomposition predictions (F1–F4), four per-vertex weight predictions (N1–N4), and four robust estimator predictions (R1–R4). Strong diagnostic evidence requires all four of {F1, N1, N3, R1} CONFIRMED; moderate evidence requires exactly three; two or fewer indicates weak or no evidence for the extreme-λ hypothesis. These four criteria confirm that extreme λ values drive variability but do not by themselves distinguish whether band-edge gradients or inter-band p_conv floor is the dominant mechanism — the Approach B breakdown addresses that distinction. Configuration: E3 (f = 0.25, z precision from 9 sensor disks, η = 2.0), 17-band geometry (9 sensor + 8 service disks), z ∈ [200, 2500] mm. Focus masses: 17 MeV and 200 MeV. Confidence intervals use 1.96 × SE (normal approximation for N_seeds = 50 ≥ 30), consistent with v2.6. Frozen components: mtag_v24.py (tag formula, 17-band geometry, frozen since v2.4), mtag_v22.py (9-band baseline, frozen since v2.2), endcap_sim.py (simulation engine, frozen since v2.5). Independent personal toy study, not affiliated with the authors of arXiv:2610.38932 or arXiv:2306.08686, nor with the CMS, ATLAS, LHCb or ALICE Collaborations or the FCC study. Detector numbers are assumed; only qualitative trends and ratios are meaningful. Builds on v2.6 post-registration results (DOI: 10.5281/zenodo.24373115). Full DOI chain from v2.0 through v2.7 is listed in the README.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24388227
Primary Topic
Particle physics theoretical and experimental studies
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

ALP-ttbar Toy MC v2.7 — Preregistration: Tag Formula Diagnostics in Displaced Light ALP → e⁺e⁻

Won Shik Paik
10 links
Zenodo (CERN European Organization for Nuclear Research)
Particle physics theoretical and experimental studies
preprint

ALP-ttbar Toy MC v2.7 — Preregistration: Tag Formula Diagnostics in Displaced Light ALP → e⁺e⁻

Won Shik Paik
preprint en

Abstract

Preregistration for v2.7 of the ALP-ttbar Toy MC series: Tag Formula Diagnostics. The v2.6 multi-seed ensemble (N_seeds = 50, E3 configuration with η = 2.0) revealed that the Tag discriminator exhibits unexpectedly large stochastic variation at m_a = 17 MeV: σ/mean = 7.4%, compared to only 0.16% for the HV discriminator — a 46× ratio. Both discriminators share the same vertex populations, so this disparity points to a mechanism specific to the tag formula. We hypothesise that extreme-λ vertices drive Tag variability: the tag formula's conversion probability density p_conv(z) has steep gradients near material band edges and drops to a numerical floor in inter-band gaps, producing extreme λ values that disproportionately influence the Tag degradation ratio. This preregistration specifies three complementary diagnostic approaches to test this hypothesis: Approach A — Frozen-Seed Variance Decomposition: Three seed groups isolate the contribution of each vertex population. Group F (Full) varies both signal and conversion seeds (SEED_CONV = 100i+5, SEED_SIG = 100i+11, reproducing v2.6). Group C (Fixed-conv) fixes SEED_CONV = 5 while varying SEED_SIG, isolating signal-vertex randomness. Group S (Fixed-signal) fixes SEED_SIG = 11 while varying SEED_CONV, isolating conversion randomness. Comparing σ_Tag across groups identifies which population dominates the variance. Approach B — Per-Vertex Weight Analysis: For each signal vertex, we record the per-vertex λ(z) value and classify vertices into four mutually exclusive groups based on their proximity to material band edges and their p_conv magnitude. "Near-band-edge" vertices satisfy min(|z−(Z_d−HW_d)|, |z−(Z_d+HW_d)|) < D for distance thresholds D ∈ {1, 2, 5, 10, 20, 50} mm. "Inter-band" vertices have p_conv(z, σ_z) < ε = 10⁻¹². These groups can overlap — for example, z = 310 mm is 7 mm from the nearest band edge and has p_conv ≈ 1.6 × 10⁻¹⁴ — so we use a 4-way classification: near-edge only, inter-band only, both, or neither. The numerical floor p_conv ≤ 10⁻³⁰⁰ is recorded separately. Fractions f_near and f_gap use N_used (actual vertices passing basic cuts) as the denominator. We measure correlations R(f_near, Tag) and R(f_gap, Tag) separately, and compute λ distribution moments for all four groups. Approach C — Robust Estimator Comparison: We define trimmed and winsorised variants of the Tag to test whether removing or clipping extreme-λ vertices stabilises the discriminator. Trimmed Tag (k%) sorts the signed λ values for each population and removes n_rem = ⌊k × N / 100⌋ vertices from each tail (the n_rem smallest and n_rem largest), totalling approximately 2k% removal. A boolean removal mask ensures these vertices never contribute to S or B in any (t, mass-window) combination within scan_tag_endcap. Winsorised Tag (k%) selects the identical ⌊k × N / 100⌋ vertices from each tail but replaces their λ with the nearest retained boundary value instead of removing them. This shared vertex selection ensures the R2 comparison isolates the effect of removal versus clipping with no difference in which vertices are targeted. We scan k ∈ {0.5, 1, 2, 5, 10}%. The generation pipeline follows v2.6: endcap_signal generates 300,000 signal vertices and endcap_conversions generates 1,500,000 conversion-photon candidates per seed. After basic-cut selection (fiducial z ∈ [200, 2500] mm, minimum p_T), subsample() retains the first N_sig = 50,000 signal and N_conv = 100,000 conversion vertices passing cuts (or fewer when the base-passing yield is below the cap); vertex weights are rescaled to preserve the total cross-section. Twelve predictions with quantitative thresholds are preregistered across the three approaches: four variance decomposition predictions (F1–F4), four per-vertex weight predictions (N1–N4), and four robust estimator predictions (R1–R4). Strong diagnostic evidence requires all four of {F1, N1, N3, R1} CONFIRMED; moderate evidence requires exactly three; two or fewer indicates weak or no evidence for the extreme-λ hypothesis. These four criteria confirm that extreme λ values drive variability but do not by themselves distinguish whether band-edge gradients or inter-band p_conv floor is the dominant mechanism — the Approach B breakdown addresses that distinction. Configuration: E3 (f = 0.25, z precision from 9 sensor disks, η = 2.0), 17-band geometry (9 sensor + 8 service disks), z ∈ [200, 2500] mm. Focus masses: 17 MeV and 200 MeV. Confidence intervals use 1.96 × SE (normal approximation for N_seeds = 50 ≥ 30), consistent with v2.6. Frozen components: mtag_v24.py (tag formula, 17-band geometry, frozen since v2.4), mtag_v22.py (9-band baseline, frozen since v2.2), endcap_sim.py (simulation engine, frozen since v2.5). Independent personal toy study, not affiliated with the authors of arXiv:2610.38932 or arXiv:2306.08686, nor with the CMS, ATLAS, LHCb or ALICE Collaborations or the FCC study. Detector numbers are assumed; only qualitative trends and ratios are meaningful. Builds on v2.6 post-registration results (DOI: 10.5281/zenodo.24373115). Full DOI chain from v2.0 through v2.7 is listed in the README.

Zenodo (CERN European Organization for Nuclear Research)
Particle physics theoretical and experimental studies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon