⚖️ Scientific Claims Register

This page centralizes every major scientific claim made across the three research projects. Each claim is assigned an evidence level, linked to its primary figure and documentation page, and annotated with known limitations or caveats.

Purpose

This register exists to prevent overstating results — in conversations, in documentation, and especially in any future manuscript submissions. If a claim is not listed here with a supporting figure and run ID, it should not be stated as a finding.

---

Evidence levels

Level	Meaning
✅ Solid	Reproduced across multiple runs, consistent with physical expectations, no known confounders
🟡 Suggestive	Present in primary results but not yet replicated across independent datasets or model families
🔶 Preliminary	Single run or dataset, treated as hypothesis-generating, not conclusive
⚠️ Caveat	Result is real but requires explicit qualification before public communication

---

🌟 AstroSpectro

Classification performance

⚠v6 · includes 4 Gaia/photometric features — see Data Leakage Audit

Claim	Evidence	Run ID	Figure	Page
Balanced accuracy of 89.2% on 5-class AFGKM classification (N=60,947 test)	✅ Solid	20260522T161236Z	confusion_matrix_xgboost_20260522T161236Z.png	Classification Performance
ROC-AUC macro = 0.986 across all 5 classes	✅ Solid	20260522T161236Z	roc_xgboost_20260522T161236Z.png	Classification Performance
Class M achieves highest F1 (0.946); class F is the weakest (0.822)	✅ Solid	20260522T161236Z	confusion_matrix_xgboost_20260522T161236Z.png	F/G Confusion
F↔G confusion is the dominant off-diagonal confusion (F→G: 12%, G→F: 10%)	✅ Solid	20260522T161236Z	confusion_matrix_xgboost_20260522T161236Z.png	F/G Confusion

SHAP interpretability

Claim	Evidence	Run ID	Figure	Page
Mg b equivalent width is the top-ranked feature globally (mean SHAP = 0.42)	🟡 Suggestive	20260522T161236Z	FIGURE_2_shap_composite.png	SHAP Interpretability
96.8% of total SHAP feature importance comes from physically interpretable spectroscopic features	✅ Solid	20260522T161236Z	physical_vs_observational.png	SHAP Interpretability
Metallicity-sensitive features (Mg b, Ca II K, CaH2) dominate over Balmer temperature indicators globally	🟡 Suggestive	20260522T161236Z	global_importance.png, FIGURE_2_shap_composite.png	SHAP Interpretability
Class A is primarily distinguished by Ca II K eq_width (mean SHAP = 1.364) and G4300 index (0.773)	🟡 Suggestive	20260522T161236Z	class_specific_importance.png, beeswarm_per_class.png	SHAP Interpretability
Class M is primarily distinguished by index_CaH2 (mean SHAP = 1.589) and slope_deep_red (0.666)	🟡 Suggestive	20260522T161236Z	class_specific_importance.png	SHAP Interpretability
Mg b × CaH2 SHAP interaction (0.039) is the strongest pairwise interaction in top-10 features	🔶 Preliminary	20260522T161236Z	feature_interactions.png	SHAP Interpretability

Caveats & limitations

Claim	Evidence	Note
SHAP metallicity dominance has not been replicated on an independent survey (e.g. APOGEE, Galah)	⚠️ Caveat	Future work; stated explicitly in SHAP page and limitations
3.2% of SHAP importance comes from Gaia photometric features (M_BP, M_G, M_RP, ag_gspphot) — the model is not purely spectroscopic	⚠️ Caveat	Requires explicit statement in spectro-only-mode.mdx before RNAAS submission
F↔G confusion is partially irreducible due to spectral sequence overlap near 6000 K	✅ Solid	Physical, not a model failure
Dataset covers LAMOST DR5 only — results may differ on higher-resolution spectra (APOGEE R22,500 vs LAMOST R1800)	⚠️ Caveat	Scope limitation

---

🔭 AstroVision

Classification performance

Claim	Evidence	Figure	Page
A late-fusion multimodal model reaches 86.3% balanced accuracy on Galaxy10 DECaLS in one 50-epoch training on its own test split; the same configuration scores 85.7% in the 30-epoch ablation. The backbone features come from the fine-tuned DINOv2, so a leak from the training split cannot be excluded	🔶 Preliminary	results_final_comparison.png	Benchmarking
Late fusion scores above fine-tuned DINOv2 alone (86.3% vs 83.9%), but the two use different test splits and the ablation shows no measurable gain from the added modalities (86.0% image only vs 85.7% with morphometry and photometry). A modality gain is not claimed	⚠️ Caveat	results_final_comparison.png	Benchmarking
Disturbed class recall is capped at ~0.53 — a physically irreducible ceiling due to intra-class morphological heterogeneity	✅ Solid	results_metrics_by_class.png	Limitations

Calibration & uncertainty

Claim	Evidence	Figure	Page
Fusion model is natively well-calibrated (ECE = 0.049)	✅ Solid	—	Uncertainty
Temperature Scaling degrades rather than improves calibration for the fusion model	✅ Solid	—	Uncertainty
Conformal Prediction achieves 90% coverage with 91% singleton rate	✅ Solid	—	Uncertainty

Physical interpretation

Claim	Evidence	Figure	Page
DINOv2 features encode galaxy colour information in part: a cross-validated Lasso on the 768 zero-shot features predicts g−r with R² = 0.443 ± 0.103 (the earlier 0.536 is not found in any output and is withdrawn)	🟡 Suggestive	xm_feature_correlation.png	Physical Interpretation
Morphometric features (CAS, Gini, M20) contribute more than SDSS photometry to the fusion model because DINOv2 already encodes colour (ablation differences are below 0.4 point, within run-to-run variation)	🔶 Preliminary	xm_feature_correlation.png	Physical Interpretation

---

ξ Dark Energy

MCMC results

Claim	Evidence	Figure	Page
Ξosc model shows a marginal preference (~2.26σ) for non-zero oscillation amplitude ε	🔶 Preliminary	Corner plots	Known Limitations
ΛCDM and CPL posteriors are consistent with published Planck + DESI constraints (unsourced: _facts/xi holds no Planck reference values and no per-measurement DESI values; do not state publicly until a source is added)	⚠️ Caveat	Corner plots	Parameter Constraints

Caveats (v0.1.0-exploratory)

Claim	Evidence	Note
The 2.26σ preference for ε > 0 is fragile — the boundary prior ε ∈ [0, 0.3] artificially inflates significance	⚠️ Caveat	Fixed in v0.2 (symmetric prior)
Fixed r_d = 147.09 Mpc imports an implicit CMB constraint — results are not purely data-driven	⚠️ Caveat	Fixed in v0.2 (free r_d)
Self-consistency bias in t(z) mapping affects the distance calculation	⚠️ Caveat	Fixed in v0.2
This project is explicitly exploratory — no claim of new physics is made	✅ Solid	Core design principle of the project

---

How to update this register

When a new result is produced:

1. Add a row to the appropriate section above
2. Assign an evidence level honestly
3. Link the figure filename and the documentation page
4. Add any associated caveat immediately

This register is the single source of truth for what is claimed publicly. It takes precedence over any other documentation page in case of conflict.