Reproducibility · manifests
Reproduce this lab
The lab contract: every experiment, simulation, derivation, training run, scan, or analysis ships a manifest with sealed inputs (external data with links, or repo-pinned internal artifacts), exact entrypoints (the literal command that runs it), and a verification method — so a stranger, or Hubify, can reproduce it without asking us anything. The full-reproduction pass across every program is the final pre-publication test of this lab, not an afterthought.
Research programs
3
Experiment manifests
52
Runnable now
41
Est. total reproduction cost
$33.30
9 experiments need a data restore before they can run; 2 are superseded and kept for lineage only, never offered as a live reproduction target. Per-program cost/time estimates are rollups, not literal sums of the individual experiment estimates below — see each program's full_reproduction.order note for sequencing.
Research program · anomaly-discovery
DESI anomaly discovery
Question: What unusual spectra emerge from a full-scale DESI anomaly search, and which candidates survive scientific validation?
17 experiments · 12 runnable nowfull reproduction: $5 · Highly uncertain — the single most consequential open run (clean_rerun's run_scan.py full DESI DR1 iron scan, AUG-011) has not yet executed and has no wall-clock estimate beyond a qualitative 'download-bound, ~200GB volume' note. Of the legs that ARE runnable-now: most P3 legs (dedup, kfold gate, DP3-15 held-out, eROSITA control) are minutes-to-an-hour local-CPU; the NANOGrav PTA MCMC (192,000 samples) is hours-scale; the historical flagship legs (silver crossmatch, taxonomy, injection recovery, NEOWISE, gold z6 QSO, fnl tracer selection) are each minutes-to-an-hour on their preserved result JSONs.
Papers
anomaly-flagship
lead
Rebuilt DESI anomaly-science flagship (future primary paper)External data sources
Experiments — reproduction DAG order
| Experiment | Kind | Status | Venue | Est. wall-clock | Est. cost | Verification |
|---|---|---|---|---|---|---|
BigAE enhanced 18M/22.5M-row DESI inference (historical, unreconciled) anomaly-bigae-18m-inference-historical | inference-scan | needs data restore | not recommended for reproduction — superseded by the clean_rerun campaign (see anomaly-clean-rerun-scan) | n/a | free (local) | NOT independently verifiable as a reproduction target: the surviving catalog_summary.json (catalog_total=22,504,897; anomalies_score_gt5=249,905) cannot be checked against the 46 absent enhanced Parquets, and the enhanced checkpoint's own row counts (23,798,995 spectra / 22,748,720 rows) disagree with the summary by 243,823 rows with no batch manifest to reconcile them — there is no fresh-run tolerance test that can pass honestly for this historical artifact. |
Silver crossmatch (2,145-row SNR-filtered slice) anomaly-silver-crossmatch · depends on anomaly-bigae-18m-inference-historical | crossmatch | runnable now | local | minutes to low hours (depends on external crossmatch service throughput) | free (local) | Re-run and confirm the silver-tier crossmatch summary reproduces the 2,145-row SNR-filtered slice count (exact row-count equality expected; no numeric tolerance needed for a deterministic filter pass). |
Uncataloged taxonomy (1,127 objects, 10 families) anomaly-uncataloged-taxonomy · depends on anomaly-silver-crossmatch | analysis | runnable now | local | minutes | free (local) | Re-run and confirm the taxonomy classification reproduces 1,127 objects across 10 families (exact count equality expected on a deterministic classification pass over the same input catalog). |
Injection recovery test (per-class completeness / false-positive rates) anomaly-injection-recovery-test · depends on anomaly-bigae-18m-inference-historical | validation | runnable now | local | minutes | free (local) | Verification requires an independent re-derivation of the completeness/false-positive claim per the audit finding — do NOT treat the existing 0% FP / 10-1,377x enrichment figures in the committed JSON as verified facts; a reproduction is confirmed only when the re-derived per-class completeness and false-positive rates are recomputed from the injected-source recovery statistics and cross-checked, not by matching the existing headline numbers. |
NEOWISE crossmatch (IR variability, 16/283 meet variability rule) anomaly-neowise-crossmatch · depends on anomaly-silver-crossmatch | crossmatch | runnable now | local | minutes to low hours (depends on external IRSA/NEOWISE query throughput) | free (local) | Re-run and confirm 16/283 objects meet the IR-variability rule (exact count equality expected on a deterministic threshold pass over the same crossmatched sample). |
Gold anomalies — z6 QSO spectra (12 DESI Redrock z>6 QSO candidates) anomaly-gold-z6-qso-spectra · depends on anomaly-uncataloged-taxonomy | analysis | runnable now | local | minutes to low hours (DESI spectra download-bound) | free (local) | Re-run and confirm the same 12 DESI Redrock z>6 QSO candidates are recovered (exact object-id set equality expected); note these candidates have no independent redshift validation, so verification is limited to candidate-selection reproducibility, not redshift confirmation. |
Photo-z from latent vectors (supervised MLP, sigma_NMAD=0.0279) anomaly-photoz-latent-vectors · depends on anomaly-bigae-18m-inference-historical | training | needs data restore | local (blocked pending data restore) | n/a until parent latent features are restored | free (local) | Re-run and confirm sigma_NMAD=0.0279 on the same 800k train / 200k test split, within a small numeric tolerance (<=0.001 absolute) — currently blocked because the input latent-feature parent is absent locally. |
f_NL tracer selection / step4-6 bias validation + alpha empirical calibration (negative result) anomaly-fnl-tracer-selection | validation | runnable now | local | minutes | free (local) | The reproduction target is confirming alpha remains statistically consistent with 0 (jackknife alpha_internal_jk = 0.1936 +/- 0.650, i.e. within ~0.3-sigma of zero), not reproducing a positive detection — a re-run passes if the recomputed alpha_internal_jk stays within its own jackknife standard error of 0. |
clean_rerun campaign (AUG-011) — calibration stage, sealed 2026-08-05 anomaly-clean-rerun-scan | inference-scan | runnable now | runpod (A4000-class GPU or CPU-strong instance, ~200GB volume) | calibration stage: hours (bounded ~25GB two-stage PPS-cluster download over ~200 coadd groups). Full DESI DR1 scan (run_scan.py): NOT YET EXECUTED anywhere — no wall-clock evidence exists; download-bound across tens of thousands of healpix groups, so this is a multi-day estimate pending a real run, not a recorded fact | $15 | Re-run the calibration stage and confirm n_fit=20,000 / n_validation=20,000 over the same 200-group / 40,000-row two-stage PPS-cluster sample, with the stability check passing under the committed rule abs(validation_mse_mean - mse_mean) <= 5*(mse_std/sqrt(n_fit)) (observed deviation 0.00543 against a bound of 0.0481 in the sealed calibration.json). Full-scan verification (row counts, anomaly rates at DESI-DR1 scale) is not yet possible since run_scan.py has not been executed. |
DP3-15 held-out re-inference (structural-ceiling demonstration) p3-dp3-15-heldout | validation | runnable now | local | 30-60 minutes (dominated by the live SPARCL re-pull of ~20,000 held-out spectra) | free (local) | Re-run and confirm (A) the recoverable-fraction bound of ~1.31% of released rows (est_recoverable_released_rows / released_rows) and (B) the 5-seed BigAE ensemble held-out MSE median of 0.2327 (native-scale, matches the committed reconciliation reference of 0.233) plus the injection-recovery gate (broad emission-spike recall 98.8% @5-sigma, spectral-break recall 100% @5-sigma) within a small numeric tolerance (<=0.01 absolute on the MSE median, <=2 percentage points on recall). |
6-way / 7-way / 8-way positional dedup (275,151 -> 269,317 unique, 2.12% collapse) p3-positional-dedup | analysis | runnable now | local | minutes | free (local) | Re-run and confirm total_survey_level_detections=275,151 collapses to n_unique_physical_objects=269,317 (n_collapsed_detections=5,834, a 2.12% collapse rate) — exact integer-count equality expected on a deterministic 5.0 arcsec positional crossmatch over the same six-survey input. |
DESI 5-fold cross-validation reproducibility gate (mean pairwise Jaccard 0.862) p3-kfold-cv-gate | validation | runnable now | local | 1-3 hours (5-fold model training, 47,000-row pool) | free (local) | Re-run and confirm mean pairwise Jaccard >= 0.70 gate (committed value 0.8625) with 464/546 objects appearing in >=3 of 5 fold-model top-1% lists (consensus_fraction_of_union=0.8498), within a small numeric tolerance (<=0.02 absolute on mean Jaccard given training-seed variance). |
Planck held-out membership test + native re-inference (partial, 48/200 vs 30 expected) p3-planck-heldout-membership | validation | needs data restore | local (membership test only; full native re-inference blocked) | minutes for the membership test; full native re-inference is not schedulable until best_cmb_native.pt + cmb_native_patches.npy + cmb_native_all_scores.parquet are re-staged | free (local) | Re-run the membership-test half and confirm 48/200 top anomalies fall in the held-out split against an expected 30.0 under random assignment (1.60x over-representation, binomial one-sided p=5.49e-04), within exact integer-count equality on the same held-out split definition (val_frac=0.15, seed=42). Full native re-inference verification is not possible until the checkpoint/patch tensor are restored. |
eROSITA scaler-leakage bounded control (top-298 overlap 257/298, J=0.76) p3-erosita-scaler-leakage-control | validation | runnable now | runpod (pending generating-script recovery) | unknown until the generating script is recovered or rewritten from the committed result JSON's documented method | free (local) | Confirm the committed result: top-298 overlap between run A (full-catalog scaler) and run B (train-split-only scaler) = 257/298 (Jaccard 0.7581), top-1%-union overlap 7,279 (Jaccard 0.6427), Spearman rank correlation over the full 930,203-source catalog = 0.9413 — a bit-for-bit fresh-run check additionally requires recovering/rewriting the generating script against the eROSITA feature table; absent that, verification is against the committed JSON's own internally-consistent numbers. |
NANOGrav 15-yr free-spectrum PTA MCMC (real Zenodo KDE likelihood, emcee) p3-nanograv-pta-mcmc | mcmc | runnable now | local | under 1 minute (committed run reports production_seconds=24.97 for 32 walkers x 10,000 production + 2,500 burn-in) | free (local) | Re-run and confirm the gamma posterior mean/std (2.5665 +/- 0.3818, median 2.5913) on the 320,000-sample chain (32 walkers x 10,000 production steps), within a small numeric tolerance (<=0.05 absolute on mean gamma given emcee stochasticity with a fixed seed), and confirm the Savage-Dickey Bayes factor B_matter_bounce_vs_free reproduces near 3.23 (log10_B_matter_bounce_vs_smbhb near 3.85) from savage_dickey_2026-05-29.json. |
Multi-survey summary / crossmatch / spatial-clustering / score-distributions (8 surveys) p3-multisurvey-summary-crossmatch | crossmatch | needs data restore | local (script survives; per-survey raw output regeneration untested since the 2026-04-08 snapshot) | hours (8-survey crossmatch, external archive query throughput bound) | free (local) | Re-run bulk_cross_match_all.py and confirm per-survey detection counts and spatial-clustering/score-distribution summaries reproduce the 2026-04-08 snapshot outputs within the same tolerances used for the 6-way dedup headline (p3-positional-dedup): exact integer-count equality on deterministic crossmatch passes per survey. Not yet re-verified against a fresh run. |
UMAP multi-seed stability (Pod 1 production) p3-umap-multiseed-stability · depends on p3-multisurvey-summary-crossmatch | analysis | needs data restore | unknown — blocked until the UMAP pipeline script is restored | unknown | free (local) | Cannot currently define a fresh-run numeric tolerance test — no generating script survives to re-run. Verification is limited to inspecting the committed umap_stability.json for internal consistency until the pipeline script is restored. |
Research program · bounce-theory
Bounce theory
Question: Does matter-dominated contraction produce a distinctive, reproducible primordial non-Gaussian amplitude?
12 experiments · 10 runnable nowfull reproduction: $10.30 · ~1-2 days sequential; dominated by the P1A delta-Neff MCMC (~12-18h to R-1<0.01) and the P1B SN-overlap control chains (several hours per chain pair). All 6 P2 derivation/analysis scripts are minutes-scale and independently parallelizable; the 500-MC NaMaster birefringence recovery (~1.5-2h) and NaMaster window regen (~5-15 min) can run alongside the MCMC legs.
Papers
P1B
support
namaster-proof: exact pseudo-Cl window inference and tamper-evident provenance (JORS)view paperExternal data sources
Heinrich et al. 2023 SPHEREx forecast covariance (Cov_B) — NOT publicly released by original authors; DP2-26/DP2-29 gap
covariance-matrix
not publicly releasedExperiments — reproduction DAG order
| Experiment | Kind | Status | Venue | Est. wall-clock | Est. cost | Verification |
|---|---|---|---|---|---|---|
Four-vertex f_NL^local = -35/16 amplitude derivation (P2 headline) p2-vertex-check | derivation | runnable now | local | minutes | free (local) | Re-run and confirm symbolic output equals f_NL^local = -35/16, equilateral = -255/128 exactly (rational-number equality, not a numeric tolerance). |
G1 gradient-transmission scheme-dependence (Phase 1, T_c ~ 1/dcut) p2-g1-gradient-transmission · depends on p2-vertex-check | analysis | runnable now | local | minutes | free (local) | Re-run and diff g1_gradient_transmission_results.json against the committed copy; scheme-dependence coefficient c ~ 1/dcut should match to numeric tolerance (<1e-6 relative). |
G1 dressed-metric (Wilson-Ewing) transmission closure (T_c(k)=1, |delta f_NL| <= 6.8e-8) p2-g1-dressedmetric-transmission · depends on p2-g1-gradient-transmission | analysis | runnable now | local | minutes | free (local) | Re-run and confirm T_c(k)=1 and |delta f_NL| <= 6.8e-8 at k*eta_B=1e-2 (numeric tolerance <1e-9 absolute on delta f_NL). |
G3 torsion four-fermion bound (Einstein-Cartan estimate) p2-g3-torsion-fourfermion-bound · depends on p2-vertex-check | derivation | runnable now | local | minutes | free (local) | Re-run and confirm the bound coefficient matches paper Eq. 5 (v1.7.123, commit 275846c5) exactly. |
Honest-negative in-in bounce attempts (pathz / pathz2, superseded) p2-honest-negative-inin | derivation | superseded | local | minutes | free (local) | Re-run and confirm the negative-result structure is reproduced (no positive in-in bounce closure); this is a historical-negative artifact, not a headline claim. |
Channel-native Fisher surrogate (c15) + covariance chain (c8-c15) p2-channel-native-fisher · depends on p2-vertex-check | analysis | needs data restore | local | 2-6 hours (CAMB-bound) | free (local) | Re-run and confirm the nuisance ladder 3.5-sigma/3.1-sigma/2.3-sigma/0.4-sigma reproduces to numeric tolerance (<1% relative); true-covariance closure requires the gated Cov_B and is out of scope for this reproduction. |
MCMC full-tension / Planck+BAO+SN / third-combo chains (delta N_eff) p1a-mcmc-dneff | mcmc | runnable now | runpod (32-vCPU CPU-optimized instance) | 12-18 hours to R-1<0.01 convergence | $8 | Re-run and confirm sample counts (176,240 / 132,949 / ~114,992 = 424,181 total) and R-1<0.01 convergence; posterior means for delta N_eff within numeric tolerance (<0.02 absolute) of the committed chains. |
500-MC NaMaster EB birefringence recovery (beta = 0.238 deg, SNR 20.3) p1a-namaster-500mc-birefringence | mcmc | runnable now | runpod (CPU-bound instance, ~$0.17/hr class) | ~1.5-2h for 500 MC realizations at nside=512, lmax=1024 | $0.30 | Re-run 500 MC realizations and confirm recovered beta=0.238 deg (input 0.27 deg) with SNR within numeric tolerance (+/-1 SNR unit) of 20.3, at f_sky=0.3226. |
100,000-sample N_tot sensitivity Monte Carlo (Spearman |rho_s|=0.996) p1a-ntot-sensitivity-mc | analysis | runnable now | local | minutes (100,000-sample Monte Carlo is CPU-light) | free (local) | Re-run 100,000-sample Monte Carlo and confirm Spearman |rho_s|=0.996 on N_tot to numeric tolerance (+/-0.005), and viable-fraction 2.2% of parameter space. |
ALP prior-predictive / spectator-conditioned prior-predictive p1a-alp-prior-predictive | mcmc | runnable now | local | minutes | free (local) | Re-run and confirm prior-predictive coverage statistics match the committed receipt to numeric tolerance (<1% relative). |
SN-overlap control chains A (Pantheon+) / B (DES-SN5YR) p1b-sn-overlap-control-chains · depends on p1a-mcmc-dneff | mcmc | runnable now | runpod (A4000-class, $0.17/hr) | several hours per chain pair to R-1<0.01 | $2 | Re-run and confirm Control A w0=-0.874+/-0.059 / wa=-0.530+/-0.241 and Control B w0=-0.787+/-0.063 / wa=-0.785+/-0.263 within numeric tolerance (+/-0.02 on each parameter). |
NaMaster window regenerability check (pymaster 3.0) p1b-namaster-window-regen · depends on p1a-namaster-500mc-birefringence | validation | runnable now | local or runpod (CPU-only, no GPU needed) | ~5-15 minutes | free (local) | Re-run and confirm max|Delta| < 1e-10 against the regenerated NaMaster workspace (committed result: max|Delta|=9.926e-24, PASS). |
Research program · galaxy-chirality
Galaxy chirality
Question: Is there a large-scale observed-label chirality dipole in the released DESI imaging catalog?
23 experiments · 19 runnable nowfull reproduction: $18 · ~2-3 days sequential; dominated by the G1 ViT-Small retrain (~4h on RunPod A4000) and the e2e mirror-flip full-catalog inference (10.45h wall on RunPod A100, 16.9M inferences). The C1-C3 NaMaster null-test batch, G2/G3/G4 validations, and the A_95 dipole limit are each minutes-scale once the trained model and full-catalog inference exist and can run in parallel. P5's DESI crossmatch build (~hours, local CPU) gates its four downstream environment analyses plus the cosmic-web/DESIVAST and r-conf legs, which are independently parallelizable after that.
Papers
External data sources
Experiments — reproduction DAG order
| Experiment | Kind | Status | Venue | Est. wall-clock | Est. cost | Verification |
|---|---|---|---|---|---|---|
v2 ViT-Small production training (26,616-object historical realization) p4-v2-vit-production-training | training | superseded | runpod (A100/H200-class GPU, ViT-Small image classifier training) | 3-6 hours for a 26,616-object realization at typical ViT-Small batch throughput | $5 | No numeric target is defined for this historical realization since its labels/manifest were not retained; status is superseded. Use p4-g1-vit-retrain-manifest's best_val_acc=0.9931 (epoch 47) as the current verifiable production-training benchmark instead. |
G1 — regenerable ViT-Small retrain with manifest (supersedes historical v2 training) p4-g1-vit-retrain-manifest | training | runnable now | runpod A4000 16GB on-demand | ~1.5h smoke test + ~4-5h full retrain | $1 | Re-run and confirm best_val_acc=0.9931 at epoch 47 (early-stop epoch 62) within +/-0.001 tolerance; for a byte-identical reproduction, sha256 of g1_ckpt_best.pt should match aed109dc…, otherwise fall back to the accuracy tolerance for a fresh-seed retrain. |
G1 CE-included full composition (826-vs-846 adjudication) p4-g1-ce-composition-assembly · depends on p4-g1-vit-retrain-manifest | analysis | runnable now | local CPU | minutes to ~1 hour for the 26,609-object composition/adjudication pass | free (local) | Re-run and confirm 26,609 total objects with ce_not_spiral=819 adjudicated (the 826-vs-846 CE-inclusion adjudication resolved) — exact integer match, not a numeric tolerance. |
G2 — training-disjoint held-out GZ1 validation p4-g2-disjoint-validation · depends on p4-g1-vit-retrain-manifest | validation | runnable now | runpod A4000 or equivalent 16GB GPU (inference-only) | ~6-10 minutes | $0.05 | Re-run and confirm accuracy=0.9867, Cohen's kappa=0.9733 on the n=3000 disjoint GZ1-spiral held-out set, within +/-0.001 numeric tolerance. |
G3 — joint estimator covariance (local bootstrap leg + RunPod MASTER-leg refinement) p4-g3-joint-estimator-covariance · depends on p4-g1-vit-retrain-manifest | analysis | runnable now | local CPU for the base bootstrap leg; RunPod A4000 (or a local pymaster install) for the MASTER-leg refinement | ~10-15 min local-bootstrap leg + ~1h MASTER-leg | $0.36 | Re-run both legs and confirm the joint estimator covariance matrix entries match the committed JSONs within a relative-difference tolerance of <1%; cross-check against the HF backup p4_compute_phase2_2026-07-18/ (sha256-verified) as the primary hash check. |
e2e mirror-flip full-catalog inference (8.47M galaxies x 2 passes) p4-e2e-mirror-flip · depends on p4-g1-vit-retrain-manifest | inference-scan | runnable now | runpod A100 (or equivalent 40GB-class GPU) | ~10-11 hours for the full 192-shard, 16,949,062-inference run | $12.44 | Re-run and confirm T_raw=0.2303, T_eq=0.99974 (transfer function values) with all 192/192 shards completing 16,949,062 inferences, within numeric tolerance stated in the result file; md5 925649b7… for a byte-identical reproduction. |
G4 — per-pixel confusion + generative parity-null (monopole mechanism) p4-g4-monopole-mechanism-injection · depends on p4-e2e-mirror-flip | analysis | runnable now | local CPU or RunPod aggregation-only instance (no GPU inference required, reuses banked e2e mirror-pair record) | tens of minutes for per-pixel confusion + generative parity-null aggregation over the banked record | free (local) | Re-run against the banked e2e_mirror_pairs.parquet record and confirm the per-pixel confusion map (nside=64) and generative parity-null statistics match the committed JSON/NPZ within numeric tolerance stated in the result file. |
A_95^obs coverage-calibrated dipole injection upper limit p4-a95-dipole-injection-limit · depends on p4-e2e-mirror-flip | analysis | runnable now | local CPU | ~10 minutes | free (local) | Re-run and confirm A_95^obs = 0.98% (coverage-calibrated dipole injection upper limit) within the numeric tolerance stated in the result JSON. |
C1 — NaMaster fsky sweep (part of the C1/C2/C3 monopole/dipole-null batch) p4-c1-namaster-fsky-sweep | validation | runnable now | runpod CPU-strong instance (12+ vCPU, no GPU required) | ~1-1.5 hours for the 2x500 MC fsky sweep (fsky~0.85 and fsky~0.65) | $0.20 | Re-run and confirm the canonical fsky=0.32 reference (recovered_beta_deg=0.238, bias_deg=0.032, snr_se=20.32) and the fsky=0.85/0.65 sweep results (e.g. fsky_target=0.85: beta_recovered_deg=0.237, bias_deg=-0.033, snr_se=181.38) match the committed JSON within numeric tolerance. |
C2 — monopole/dipole N_all binomial null p4-c2-nall-binomial-null · depends on p4-e2e-mirror-flip | validation | runnable now | runpod CPU-strong instance or local CPU (job is CPU-bound) | ~6-10 minutes | $0.02 | Re-run and confirm the N_all binomial-null test statistics match the committed JSON within numeric tolerance stated in the result file. |
C3 — Wp (N_all vs N_spiral) fsky invariance null p4-c3-wp-invariance-fsky · depends on p4-e2e-mirror-flip | validation | runnable now | runpod CPU-strong instance or local CPU (job is CPU-bound) | ~6-10 minutes | $0.02 | Re-run and confirm the Wp (N_all vs N_spiral) fsky-invariance null test statistics match the committed JSON within numeric tolerance stated in the result file. |
GZ1-only classifier retrain + dipole null (pseudo-label independence check) p4-gz1only-retrain-dipole-null | validation | needs data restore | local CPU for the surviving dipole-null recomputation; RunPod A4000 16GB would be needed to redo the full GZ1-only classifier retrain from scratch | ~10 min local for the dipole-null recomputation on the surviving JSON (per COMPUTE_CAMPAIGN's own note); full end-to-end retrain time not estimable without the lost training script | free (local) | Re-run the surviving dipole-null analysis and confirm dipole z=-0.04 sigma (pseudo-label independence check) within +/-0.02 sigma numeric tolerance. |
Empirical b/a (axis-ratio) DR8 morphology cross-match p4-dr8-axis-ratio-crossmatch | crossmatch | needs data restore | local CPU or any RunPod instance (TAP pull + metric computation are CPU-bound, network-bound on the TAP query) | ~30-60 minutes depending on TAP query throughput for the full catalog cross-match | free (local) | Re-run the TAP pull + edge-on metric computation and confirm f_edge=15.8% (edge-on contamination fraction, 505,889 of 3,201,160 objects per the v1.0.240 regeneration) within +/-0.5 percentage-point numeric tolerance. |
Dipole analysis (8.47M full-catalog) p4-dipole-8m-fullcatalog | analysis | needs data restore | runpod A100/H200-class GPU for the full 8.47M-galaxy dipole computation | several hours, scaling with catalog size and TTA passes | $10 | Re-run the full 8.47M-galaxy dipole analysis and confirm 2.31 sigma raw dipole significance and 0.43 sigma post-TTA (test-time-augmentation-corrected) significance, within +/-0.1 sigma numeric tolerance. |
P4xDESI DR1 crossmatch + matched catalog build p5-desi-dr1-crossmatch-build · depends on p4-e2e-mirror-flip | crossmatch | runnable now | local CPU (implied local/CPU-bound crossmatch — no venue was explicitly logged for the original run) | ~1-3 hours for the DESI DR1 fetch + 2,232,212-row crossmatch, dominated by DESI DR1 download bandwidth | free (local) | Re-run and confirm the matched catalog contains 2,232,212 rows (1.3GB parquet) via p5_matched_chirality_desi_summary.json's row count, exact integer match. |
P5 redshift-dependence analysis of spiral chirality p5-redshift-analysis · depends on p5-desi-dr1-crossmatch-build | analysis | runnable now | local CPU | minutes to tens of minutes over the ~2.23M-row matched catalog | free (local) | Re-run and confirm the redshift-binned chirality fraction statistics in analysis_redshift/ match the committed outputs within the numeric tolerance stated in the result files. |
P5 local-density-dependence analysis of spiral chirality p5-density-analysis · depends on p5-desi-dr1-crossmatch-build | analysis | runnable now | local CPU | minutes to tens of minutes over the ~2.23M-row matched catalog | free (local) | Re-run and confirm the local-density-binned chirality fraction statistics in analysis_density/ match the committed outputs within the numeric tolerance stated in the result files. |
P5 HEALPix sky-map analysis of spiral chirality p5-healpix-analysis · depends on p5-desi-dr1-crossmatch-build | analysis | runnable now | local CPU | minutes to tens of minutes over the ~2.23M-row matched catalog | free (local) | Re-run and confirm the HEALPix-binned sky-map chirality statistics in analysis_healpix/ match the committed outputs within the numeric tolerance stated in the result files. |
P5 systematics analysis of spiral chirality measurement p5-systematics-analysis · depends on p5-desi-dr1-crossmatch-build | analysis | runnable now | local CPU | minutes to tens of minutes over the ~2.23M-row matched catalog | free (local) | Re-run and confirm the systematics-control statistics in analysis_systematics/ match the committed outputs within the numeric tolerance stated in the result files. |
Cosmic-web / DESIVAST void analysis (16, 27, 35-39 series) p5-cosmic-web-desivast-void · depends on p5-desi-dr1-crossmatch-build | analysis | runnable now | local CPU (no RunPod hit found for the p5 script family per the inventory's own grep) | ~1-2 hours across the cluster-bootstrap and RSD/void-reconstruction steps over DESIVAST VAC + the ~2.23M-row matched catalog | free (local) | Re-run and confirm the DP5-12 closure result in 27_rsd_void_recon_bound.json (RSD void-reconstruction bound, 2026-07-12) matches the committed JSON within the numeric tolerance stated in the result file. |
r23conf/r24conf/r27conf closure recomputes p5-rconf-closures · depends on p5-desi-dr1-crossmatch-build | analysis | runnable now | local CPU (the pod-implying filename is unconfirmed; a local or RunPod CPU-strong instance both work) | ~30 min - 2 hours across the four closure/batch scripts | free (local) | Re-run all four closure recomputes and confirm the r23conf/r24conf/r27conf statistics match the committed JSONs within the numeric tolerance stated in each result file. |
Focal cluster inference sensitivity + interaction clustering robustness p5-focal-cluster-robustness · depends on p5-rconf-closures | analysis | runnable now | local CPU | ~30 min - 1 hour across both robustness scripts | free (local) | Re-run both scripts and confirm the focal-cluster inference sensitivity and interaction-clustering robustness statistics match the committed JSONs within the numeric tolerance stated in each result file. |
Astra per-object crossmatch + HuggingFace mirror p5-astra-crossmatch-hf-mirror · depends on p5-desi-dr1-crossmatch-build | crossmatch | runnable now | local CPU | ~30 min - 1 hour for the per-object crossmatch + HF mirror push | free (local) | Re-run the crossmatch and confirm the per-object summary statistics in analysis_astra_per_object/summary.json match the committed output, and that the HF mirror bamfai/astra-desi-edr-mirror reflects the same row count. |