v0.2.0 — reproduction under gmat-sweep / gmat-run 0.6

Reproducibility refresh of v0.1.0 under the gmat-sweep / gmat-run 0.6 toolchain. Per-row Δr against the v0.1.0 outputs is at floating-point-roundoff scale across every aggregate, summary, and diagnostic — no manuscript reissue, no arXiv v2, no science changes.

Changed

• Sweep, CdA sensitivity arm, maneuver-threshold sensitivity arm, and sweep/manifest.jsonl regenerated end-to-end against gmat-sweep / gmat-run 0.6 (#68). The refreshed Zenodo deposit is published as version 0.2.0 under the existing concept DOI 10.5281/zenodo.20277028 (version DOI 10.5281/zenodo.20370326).
• Pipeline upgraded to gmat-sweep / gmat-run 0.6 (#62), dropping the 0.4-era script-templating workaround.
• Phase-3 post-processing folded into the sweep via gmat-sweep's postprocess hook (#63); the hook is now re-applied on Sweep resume so partial-restart runs match cold-start runs bit-for-bit (#66).

Added

• arXiv preprint badge and citation metadata pointing at arXiv:2605.19850 (#58).
• Regression test pinning run_id preservation when aggregating the CdA sensitivity arm (#67).

Removed

• make figures Makefile target (#71); figures are produced by make build via showyourwork's Snakemake DAG.

---

Reproducing on a fresh checkout (no GMAT required):

git clone https://github.com/astro-tools/paper-tle-divergence-atlas
cd paper-tle-divergence-atlas
make env && conda activate paper-tle-divergence-atlas
make build

Full CHANGELOG: CHANGELOG.md#020--2026-05-24

v0.1.0 — initial release

v0.1.0 — initial release

First citeable release of How long can you trust a Starlink TLE? An empirical comparison of SGP4 and high-fidelity propagation against operator-updated truth across a megaconstellation.

Headline findings

• H1 — staleness curves. Every populated (altitude shell × pooled generation) cell of the 24,641-pair corpus carries a statistically distinguishable power-law staleness curve $\lVert\Delta\mathbf{r}(\Delta t)\rVert \approx A,\Delta t^{k}$, with fitted exponents in $(1, 2)$ on every v2-mini cell and on the high-fidelity v1.x cells at 540 and 560 km. SGP4 v1.x and the high-fidelity v1.x cell at 550 km sit at or below $k=1$ — the boundary between mean-motion-bias and constant-along-track-acceleration error modes.
• H2 — hi-fid loses on the bulk. The high-fidelity propagator initialised from public TLEs does not improve median 3D position error over SGP4 at any of the four shell-pooled staleness horizons (6 h, 1 d, 3 d, 7 d), losing on 65–75% of pairs depending on horizon. The negative result decomposes into three mechanisms — at-epoch operator-OD residual dominance, SGP4-vs-SGP4 truth-construction kernel alignment, and asymmetric spacecraft-property bias amplification on the hi-fid arm.
• H2 cohort exception (v2-mini at long Δt). On the v2-mini cohort alone, the hi-fid arm wins 72.6 % [69.9, 75.4] of 7-day pairs at 550 km and 55.1 % [49.9, 60.4] at 560 km — both above the 50 % null. The cohort-uniform majority-wins regime at long Δt is invisible in shell-pooled win fractions because the 550-km cell is 78 % v2-mini and the 560-km cell is only 34 % v2-mini; the per-(shell × Δt × generation) decomposition in Table 8 surfaces it directly.
• H3 — solar modulation, preliminary. Per-shell ANCOVA of the per-satellite SGP4 staleness coefficient against daily-observed F10.7 returns a positive slope direction-consistent with the LEO density-gradient expectation at 540 and 560 km. The 560-km slope clears conventional significance ($\hat\beta = +0.014$ sfu⁻¹, $p = 0.034$); 540 and 550 km are slope-ambiguous on the present 30-day, ~17 sfu corpus. Reported as preliminary rather than a calibrated F10.7-modulation measurement.
• Practitioner-facing atlas. The per-cell $(A, k)$ table is the input form a downstream consumer can plug into their own error budget, and the corpus is in a form usable as a benchmark target by the enhanced-propagator line of work (SGP4-XP, differentiable SGP4, ML-residual correctors).

Reproducibility

• Code: this repository (MIT licensed).
• Sweep outputs: deposited on Zenodo at the version DOI 10.5281/zenodo.20277029. Concept DOI (version-stable, cite this): 10.5281/zenodo.20277028.
• make build renders the PDF from the Zenodo-cached bundle on a clean checkout without a local GMAT installation.
• make sweep reproduces the full sweep from scratch given GMAT R2026a, Space-Track credentials, and the project-installed EGM2008 coefficients.

Release artefacts

• ms.pdf — final compiled manuscript.
• arxiv.tar.gz — arXiv-ready submission tarball (make arxiv-tarball).