Precipitation: does structural disagreement among independent observers carry information?

Summary of the collocation thread

What was asked. Can disagreement among physically independent observers of one rain field be turned into a measurement of each observer's error, with no ground truth? The mathematics exists for amounts (triple collocation: each observer's error variance from three observers' covariances). This thread applied it to structural coordinates (where the rain sits, how spread it is) and followed what came out.

What happened, in one line per step.

Where knowledge stands.

• Established (pre-registered; replicated on an unseen month, T10): the ranking structural collocation gives depends on which observers form the triplet; with radar, GOES and microwave, microwave is rated least noisy; replacing either satellite with gauges, at the same point support, reverses it (radar 0.06 < gauges 0.22 < microwave 0.26, centroid y), and radar's score does not depend on which satellite is replaced. August 2024: radar 0.04 < gauges 0.24 < microwave 0.33 (weak stratum).
• Conditional: reading these scores as true error rankings, and attributing the four-observer pattern ({radar, gauges} vs {GOES, microwave}) to an error shared by the satellites (estimated +0.35 to +0.51), both assume radar and gauge errors are not strongly correlated. Four observers cannot establish that; under assumed radar–gauge error correlations, satellite dependence detected up to 0.475–0.5 (July) and through the tested maximum of 0.5 (August); radar's ranking clear (interval excluding reversal) up to 0.275–0.4 (July), 0.375–0.4 (August) (T7-S); the actual value is not measured.
• Supported by the corrected synthetic control (T1-bis): centroid and dispersion recovered with correct rankings; a planted shared satellite error inverts the ranking at σ_sh = 1.0 cell; five observers identify a shared pair without false alarms (under the simulation's assumptions). F6 withdrawn.
• Not supported as explanations of the inversion (tests did not detect them): IMERG's infrared-to-microwave calibration, the tested smoothing, GOES parallax, convective-core keying, radar beam quality, and (in its tested form) aloft-only anvil cloud. Parallax and beam quality were checked exploratorily; the others by pre-registered tests. "Not supported" means not detected within the tested variants.
• Model failure (both months): tetrads t1 and t2 are above zero in July and August, so one common signal with mutually independent errors does not describe the four observers; the satellite attribution is one reading of that failure, not the only one.
• Not presently claimed: a satellite defect; the physical mechanism; an advantage over amount-based or representativeness analyses (not compared); a diagnostic that improves on available ones (not compared); generality beyond warm-season CONUS.
• Open: what the two satellites share, if the attribution holds. A coupling with no detectable change across anvil extent fits both observing the cloud system rather than surface rain; untested.

Rules for structural collocation, learned here (for the tested variables, simulation and observing configuration):

1. Prefer geometric coordinates; with a correct (mass-conserving) blur, centroid and dispersion were recovered and ranked correctly; threshold counts (islands, wet fraction) failed in the original simulation, were not re-tested, and are excluded from interpretation.
2. A shared error among the legs is invisible to the estimator; it shows up only as a plausible-looking but wrong ranking.
3. Check independence after feature extraction and windowing, not only from the hardware; simulation details (here, a blur that did not conserve mass) can create shared errors of their own, and windows add a small one.
4. Independence is a property of what each observer responds to, not of the hardware (now in RUBRIC.md Gate 6).
5. The bias from a shared error is largest where all observers are noisiest (light rain, anvil-heavy windows).
6. Always anchor on synthetic fields with known truth before reading a real-data ranking (house rule I-20).

Terms. Observer: one independent measurement of the same rain field (radar, GOES infrared, microwave, gauges). Collocation: estimating each observer's error from the covariances of three or more observers, with no truth. Leg: one of the three observers in a triplet. Instrument: in the four-observer check, the two observers assumed independent that let one estimate the error correlation of the other two. Noise fraction: 1 − ρ², the share of an observer's variance not shared with the common signal (0 = perfect, 1 = pure noise). Inversion: collocation rating microwave less noisy than radar. Point support: every observer read at the same gauge locations.

1. Why precipitation

The domain-qualification rubric (v2, domain-qualification/RUBRIC.md) scored four desk candidates. Weather spatial verification, reframed as observation disagreement (radar vs satellite vs gauge), was the only one with public, use-unrestricted data large enough for a powered test (scorecard candidates/04_precipitation_spatial_verification.md). Its weaker gates were stated in advance: the label (a point error is not an arrangement, Gate 4 PARTIAL) and the incumbents (SAL, MODE, FSS, MRMS quality layers). The first test was built to fail fast on exactly those.

2. Observers and data

3. Experiment 1: does structural disagreement predict radar error at gauges? confirmatory

Design (frozen in FEASIBILITY_PLAN.md, with two amendments written before any target or predictor value existed). Target: E = |radar-only at the gauge pixel − gauge|, hourly. 504 hours, 1–21 July 2024; discovery 1–10 July, validation 11–21 July, read once. Baseline: 1 − RQI, two plain radar–satellite difference maps, and the radar and satellite values at the gauge. Structural reads: 14, borrowed from the program's instruments (Saltimbanques islands / void / dispersion / edge dependency; Core42 multiscale; HTAN persistence; the variance-of-disagreement determination; an infrared third observer). Rule: go if Δρ ≥ 0.02 with a cluster-bootstrap 95% CI above 0 (clusters = UTC day × 5° tile). Seismology guardrails applied: near-zero Ω check, and a spatial-scramble null.

Sample: 288,690 gauge-hours (ASOS 271,511; USCRN 17,179); 423 discovery and 533 validation clusters. Near-zero om_loc share 0.0024%.

Anatomy of the label

Resolution sweep (partial ρ, n=98,751): om_pat -0.028 / -0.024 / -0.017 / -0.013; om_loc -0.027 / -0.018 / -0.013 / -0.009 at 0.3° / 0.5° / 0.7° / 1.0°. Gate 6 here: radar-error vs satellite-error at gauges r = 0.47 (PARTIAL; both are measured against the same gauge, which inflates r). Predictor audit I = 0.196, top pair om_pat~om_scale 0.95 (same correlation before and after 3 × 3 smoothing).

4. Literature step: why the null, and what maths exists

Nineteen papers were pulled (literature/INDEX.md). Two bodies of work mattered:

• The null was predictable. The area–point (gauge representativeness) term dominates radar–gauge differences at short accumulations and stays significant even at 4-day totals (Ciach & Krajewski 1999; Ciach, Habib & Krajewski 2003). Satellite products lose power and coherence with MRMS below ~200 km and ~4 h (Guilloteau et al. 2021), which is the scale of the windows used.
• The formal maths for the program's idea already exists as triple / extended / quadruple collocation (Stoffelen 1998; McColl et al. 2014; Alemohammad et al. 2015): each observer's error variance from three observers' covariances, with no truth. Its core assumption, zero error cross-covariance, is what the rubric's Gate 6 checks. No application to structural coordinates was found in two searches (one extended, 2026-10-15); the closest is "binned" triple collocation, which conditions error estimates on rain regime (IEEE document 9743452), akin to the T6 stratification. Known: violating zero cross-covariance makes collocation underestimate errors, the effect seen here.

Mapping to program methods, and the four probes ranked from it, are in RESEARCH_SYNTHESIS.md. Probe A was run first because it needs no gauge label.

5. Experiment 2: structural collocation, (R, IMERG IR, MW) confirmatory

Design (PROBE_A_PLAN.md). Non-overlapping 15 × 15-cell (1.5°) windows, every hour; a window-hour enters when all four observers carry rain and microwave covers ≥ 80% of it. Per observer: wet fraction, intensity-weighted centroid (y, x), dispersion, island count, and log mean rain as the value control. Estimator: ETC noise fraction 1 − ρ² for the triplet (R, IR, MW); quadruple check of merged IMERG's error correlation with IR and MW using the other two as instruments; 500 cluster-bootstrap resamples. Predictions: (1) radar least noisy; (2) IR noisier than MW; (3) merged most error-correlated with MW. Stop if most structural variables are unstable.

Sample: 25,468 window-hours, 1199 clusters.

Post-hoc check exploratory

Hypothesis: radar and IR share an hour-integrated footprint the microwave snapshot lacks, biasing MW's error low. If so, MW covering both half-hours should raise its noise fraction.

The IMERG V07 ATBD (in literature/) then confirmed that IMERG's infrared estimate is PDIR, calibrated to the passive-microwave estimates through a rotating 30-day accumulator, so (R, IR, MW) was not an independent triplet. That motivated Experiment 3.

6. Experiment 3: raw GOES infrared in place of IMERG IR confirmatory

Design (PROBE_A2_PLAN.md): identical, with GOES-16 band 13 brightness temperature (4 scans per hour, binned to 0.1°, cold-cloud field 235 − Tb) replacing IMERG IR. 502 of 504 hours (2 hours had < 3 scans and were dropped per plan). Test-hour alignment: radar vs cold cloud r = 0.39 at zero shift, best one cell displaced (geostationary parallax), not corrected.

Sample: 14,585 window-hours, 965 clusters.

7. Experiment 4: T1 synthetic anchor confirmatory

Why: house rule I-20, never quote a read without a machine-perfect control. Probes A/A2 produced claims with no control. Design (T1_PLAN.md): 20,000 rain-like 15 × 15 windows with known truth (1 + Poisson(1.5) Gaussian cells, lognormal intensities). Simulated observers re-render the same cells: radar-like R (position sd 0.5, no blur), microwave-like S1 (sd 0.8, blur +1), infrared-like S2 (sd 1.5, blur +2), merged M = 0.6·S1 + 0.4·S2, and an independent fifth observer G (blur +1.5). A planted shared error d_sh, one displacement per window applied to both S1 and S2, swept over σ_sh = 0, 0.5, 1.0, 1.5 cells. Reads and estimators are the Probe A code, unchanged.

8. Experiment 5: T5: shared smoothness or shared physics? confirmatory

Design (T5_PLAN.md): blur radar with a Gaussian (normalised convolution) before windowing, over a ladder σ = 0–3 cells, plus σ* chosen from field properties alone: the blur at which radar's median lag-1 autocorrelation matches the microwave field's. Collocation on the independent triplet (radar, GOES IR, microwave), centroid y, centroid x and dispersion only. A synthetic calibration arm (T1 fields, radar-like observer blurred) shows what each cause predicts. Prediction: if smoothness drives the real ranking, nf_radar − nf_microwave falls to ≤ 0 at σ*; if shared physics does, it stays above 0.

Smoothness match: microwave lag-1 0.700; radar 0.633 unblurred, 0.759 at 0.5 cell → σ* = 0.5 cell.

9. Experiment 6: T6: do the satellites share an error by keying on convective cores? confirmatory

Design (T6_PLAN.md): H_core says both satellites are pulled toward convective cores while radar sees the full surface rain area. Tests on the (radar, GOES IR, microwave) triplet: P1 re-read every observer on its own core (cells ≥ its window's 80th percentile); P2 split windows by radar light-rain mass fraction (< 5 mm/h); P3 check whether the satellites sit nearer radar's core centroid than its full-area centroid.

Exploratory follow-ups: where does the inversion live? By radar window peak intensity, and by MRMS radar quality index (beam overshoot is the obvious alternative: radar could be genuinely worse in light, shallow rain far from the radar).

10. Experiment 7: T7: whose error is it? Gauges as a fourth observer confirmatory

Why: with three observers the inversion is algebraically r(GOES, MW) > r(radar, GOES); it cannot say whose error it is. T1 showed a sharp, independent fourth observer can. Rain gauges are that observer: point measurements of surface rain used by neither MRMS radar-only, IMERG Late nor GOES. Design (T7_PLAN.md): Probe A2 windows with ≥ 5 reporting gauges; every observer read at the gauge points (its value in each gauge's 0.1° cell); point centroids; split at the median radar peak. P0: the (radar, GOES, MW) triplet reproduces the inversion at point support. P1: replace GOES with gauges; a shared satellite error predicts the inversion vanishes, a radar light-rain error predicts it stays. P2: GOES–MW error correlation with radar and gauges as instruments; shared satellite error predicts > 0.

Sample: 2,402 window-hours (median 9 gauges per window); split at 15.8 mm/h. This sample is wetter than T6's (rain is required at the points).

Robustness check of T7 confirmatory (pre-registered, T7R_PLAN.md)

T7 replaced GOES with gauges. Two checks it left open: the mirror triplet (replace microwave instead), and four-observer consistency. With four observers and one common signal, independent errors make three products of correlations equal ("vanishing tetrads"). A single shared GOES–microwave error inflates only r(GOES, MW), so the tetrad that excludes that pair (t3) should stay ≈ 0 while the two that include it (t1, t2) turn positive. Same sample, bootstrap seed 11.

11. Experiment 8: T8: is the shared error driven by cloud and ice aloft beyond the rain? confirmatory

Design (T8_PLAN.md): anvil fraction = share of GOES cold-cloud cells with no radar surface rain. On the T7 gauge-instrumented sample, split into anvil terciles (means 0.12, 0.43, 0.75). P1: GOES–MW error correlation grows with anvil. P2: the (R, GOES, MW) inversion grows. P3: the (R, MW, gauges) ranking keeps radar least noisy throughout. Anvil fraction correlates -0.34 with radar peak.

12. Corrected controls and replication: T1-bis, T7-S, T9, T10 confirmatory

Why: the external review (see the corrections box above) found that T1's blur did not conserve mass and that the satellite attribution assumed independent radar and gauge errors. Two pre-registered checks followed, with a decision gate after them.

T1-bis: synthetic assumption audit (T1BIS_PLAN.md)

Truth rendered on a 64 × 64 canvas; blur is a true, mass-conserving Gaussian convolution; reads on the full canvas and in 15 × 15 windows (central and shifted). Analytical reference passed: centroid shift ≈ 10⁻¹⁵ cell, squared dispersion grows by 2σ² within 0.04%.

Planted shared S1–S2 displacement, centroid y, full canvas: at σ_sh = 0.5 estimates R 0.090 / S1 0.095 (truth 0.046 / 0.135), not inverted; at σ_sh = 1.0 R 0.199 / S1 0.084 (truth 0.046 / 0.230), inverted. Five-observer benchmark (tetrad audit, 10 blocks): planted S1–S2 found 10/10, planted R–Gs (mirror) found 10/10, no false alarms with nothing planted.

T7-S: radar–gauge sensitivity (T7S_PLAN.md)

One-signal model for radar, gauges, GOES and microwave with an assumed radar–gauge error correlation ρ_RG (swept 0–0.5) and a free GOES–microwave error correlation, fitted by maximum likelihood. The fit is identical at every ρ_RG (the data cannot fix it), so the sweep traces the equivalence class.

T9: real-data window grid and fixed-sample smoothness (T9_PLAN.md)

T10: replication on an unseen period (T10_PLAN.md)

The July pipeline, frozen with script hashes, run on 2024-08-01 01:00 to 08-22 00:00 UTC; August data were fetched only after the plan was frozen. Before any August analysis the period switch was added and July was rerun: every results file reproduced byte-for-byte. An independent re-implementation from the plan files alone (audit/AUDIT.md) also reproduced 101 July quantities to 4 × 10⁻⁹. Gates: 1,933 point-support window-hours, 14,738 area window-hours, largest negative-variance share 0.10; not inconclusive.

13. What the evidence shows

Finding IDs are stable identifiers (the trail refers to them); the list is ordered by weight, not by ID.

14. What this does not show

• It does not show the instrument is wrong: every read ran, the guards fired, and the nulls behaved.
• The observer ranking is interpretable only with the gauge leg (T7), and only at point support and for centroid position; the area-support rankings in Probes A, A2, T5 and T6 are not accuracy rankings.
• It does not establish what the two satellites share; the tested "aloft-only anvil cloud" form is not supported (T8), and the cloud-system explanation is untested.
• It is two adjacent months (July, and the August replication), one season (warm convection), one domain (CONUS), hourly, 0.1°.
• Several splits were exploratory (T5 offsets; T6 by intensity and RQI; T8 absolute errors). No multiple-comparison correction was applied to them; they locate effects and set next tests, they do not confirm.
• Selection: every collocation sample requires microwave coverage of ≥ 80% of the window, which selects hours near a microwave overpass; T7 also requires rain at the gauge points, which skews it wetter.
• The T7 conclusions rest on gauge-dense windows (≥ 5 gauges); they replicated in August (T10) under the same design.

15. Open threads (investigator next steps)

16. Chronological trail

1. Rubric v2 written; four candidates desk-scored; precipitation (observation disagreement) chosen for the only fully public, powered data.
2. Step 0: confirmed open access to MRMS, IMERG (Earthdata token), USCRN, ASOS. Re-stated the target to radar-only error so every gauge is independent of the predictors.
3. Feasibility plan frozen; Amendment 1 (Saltimbanques, Core42, HTAN reads; infrared third observer) and Amendment 2 (variance of disagreement, resolution floor, seismology guardrails) added before any result.
4. Data pulled and reduced in memory (MRMS 1,008 files, IMERG 2 × 1,008, gauges 1.1 M rows); grid and timing alignment checked (correlation peaks at zero offset).
5. Experiment 1 read once: stop rule met; scramble null larger than real.
6. Literature pulled (19 papers): representativeness explains the null; triple collocation is the formal maths.
7. Probe A frozen and run: stable structure, ranking MW < R < IR; post-hoc footprint idea refuted; ATBD confirms IMERG IR is microwave-calibrated.
8. Probe A2 frozen and run with raw GOES infrared: same ranking; calibration route excluded; ice-aloft hypothesis proposed; rubric Gate 6 updated.
9. Original question closed; collocation thread kept open; lab book started.
10. T1 synthetic anchor frozen and run: structural collocation valid for centroid and dispersion; shared error reproduces the real ranking; smoothing found to be a shared error mechanism; F2 and F3 revised; T5 set as next.
11. T5 frozen and run: radar blurred to (and beyond) microwave smoothness, inversion persists; synthetic arm matches the shared-displacement case; exploratory check finds GOES parallax, not shared with microwave; F3 revised, F7 added.
12. T6 frozen and run: core keying not supported (P1 not met, P3 reversed, P2 met in form); exploratory splits put the shared error in light rain, independent of radar beam quality; F3 revised; T7 set.
13. T7 frozen and run: gauges as a sharp fourth observer; shared GOES–microwave error measured (+0.35 to +0.51); replacing a satellite with gauges reverses the ranking (radar least noisy); F3 resolved in existence and size; F8 added; T8 set.
14. T8 frozen and run: no detectable change in GOES–MW coupling across anvil terciles (P1 not met); the triplet inversion grows threefold (P2 met); exploratory absolute errors show all observers noisier in anvil-heavy windows; F3 updated, F9 added; cause candidates set (first labelled T9; the name was later given to the window and smoothness test).
15. Lab book consolidated (summary, glossary, data manifest, limitations). T7-R robustness check frozen and run: mirror triplet and tetrads all as predicted.
16. External review (GPT) checked against the code: blur not mass-conserving, dispersion mis-ranked, tetrads dependent, several "not detected" read as "excluded", attribution overstated. Text corrected; F2, F3 revised; F6 under review; T1-bis and T7-S frozen.
17. T1-bis run: corrected blur passes the analytical reference; coordinate errors independent on the full canvas; all variables recovered with correct rankings; planted inversion at σ_sh = 1.0; five-observer benchmark 10/10 with no false alarms. F6 withdrawn, F2 reinstated, F10 added. T7-S run: attribution survives assumed radar–gauge error correlation up to ≈ 0.4. Decision gate (GATE.md).
18. T9 frozen and run: shifted and larger windows leave the inversion and point reversal intact; fixed-sample close smoothness match (lag-1 within 0.013) leaves the gap unchanged. Paused at the gate for discussion.
19. Second external review: wording tightened (close match, not exact; robust to tested variants, not closed; threshold counts unvalidated; T7-S bound and five-observer result stated as conditional on their assumptions). No numbers changed. Replication plan T10 written and frozen before any August data was fetched.
20. Third external review (after T10): "beyond 0.5" overstated a grid that stops at 0.5; sensitivity now reported by interval in both months (dependence to 0.475–0.5; ranking clear to 0.275–0.4 in July, 0.375–0.4 in August; the earlier "ranking survives to about 0.4" was point-estimate based). "Excluded" softened to "tested, not supported within the tested variants". Replication stated as repeatability, not attribution.
21. Technical note review: "coordinate errors independent on the full canvas" narrowed to centroids; dispersion keeps a small residual (+0.021 in T1-bis; +0.031 [+0.005, +0.056] in the note's minimal example), origin unresolved. Note at notes/synthetic-shared-error/.
22. Claim ledger added (REVIEW.md, this section, paper §5.1): the tetrad result stated first as a failure of the one-signal, independent-error model; untested comparisons with existing methods listed as not claimed.
23. Independent re-implementation audit (separate session, from plan files only): 101 July quantities reproduced to 4 × 10⁻⁹; four details fixed by code rather than plans recorded in audit/AUDIT.md.
24. T10 run: period switch added, July reproduced byte-for-byte; August fetched (504 MRMS and GOES hours, 1,008 IMERG half-hours, 2,295 gauge stations); all six pre-registered predictions met. Verdict: Replicated. Paused for discussion.

17. Reproducibility

Everything is in Documents/russell/precip-disagreement/. Raw files were never stored: each was streamed, reduced (block means, crops, gauge-pixel samples) and discarded. Reduced arrays live in data/ (git-ignored, ~1.9 GB). Credentials: an Earthdata token in secrets/ (git-ignored, owner-only).

cd precip-disagreement
uv venv --python 3.12 .venv && uv pip install --python .venv/bin/python numpy scipy pandas h5py eccodes eccodeslib matplotlib pypdf
.venv/bin/python scripts/fetch_gauges.py          # USCRN + ASOS -> data/gauges.csv
.venv/bin/python scripts/make_stations.py        # unique stations -> data/stations.csv (reproduces the original inline step)
.venv/bin/python scripts/fetch_mrms.py            # MRMS radar-only + RQI, top-of-hour -> data/mrms/
.venv/bin/python scripts/fetch_imerg.py           # IMERG Late merged, cropped -> data/imerg/
.venv/bin/python scripts/fetch_imerg_extra.py     # IMERG IR, MW, randomError -> data/imerg_x/
.venv/bin/python scripts/analyze.py               # Experiment 1 -> data/results.json, samples*.csv
PYTHONPATH=scripts .venv/bin/python scripts/probe_a.py          # Probe A
PYTHONPATH=scripts .venv/bin/python scripts/probe_a_posthoc.py  # post-hoc (labelled)
.venv/bin/python scripts/fetch_goes.py            # GOES-16 band 13 -> data/goes/
PYTHONPATH=scripts .venv/bin/python scripts/probe_a2.py         # Probe A2
PYTHONPATH=scripts .venv/bin/python scripts/t1_synthetic.py     # T1 synthetic anchor (seed 20261015)
PYTHONPATH=scripts .venv/bin/python scripts/t1_explore.py       # T1 exploratory E1, E2
PYTHONPATH=scripts .venv/bin/python scripts/t5.py               # T5 (synthetic arm, smoothness match, ladder)
.venv/bin/python scripts/t5_offsets.py                          # T5 exploratory offsets
PYTHONPATH=scripts .venv/bin/python scripts/t6.py               # T6 core vs area
PYTHONPATH=scripts .venv/bin/python scripts/t6_intensity.py     # T6 exploratory: by intensity
PYTHONPATH=scripts .venv/bin/python scripts/t6_rqi.py           # T6 exploratory: by RQI
PYTHONPATH=scripts .venv/bin/python scripts/t7.py               # T7 gauges as fourth observer
.venv/bin/python scripts/t7_robust.py                           # T7-R robustness (mirror triplet, tetrads)
.venv/bin/python scripts/t1bis.py                                # T1-bis corrected synthetic audit (seed 20261016)
.venv/bin/python scripts/t7s.py                                  # T7-S radar–gauge sensitivity (seed 13)
PYTHONPATH=scripts .venv/bin/python scripts/t9.py               # T9 window grid + fixed-sample smoothness (seed 21)
PYTHONPATH=scripts .venv/bin/python scripts/t10_area.py _july    # T10 H1 interval on July (same draws as Probe A2)
.venv/bin/python scripts/t10_score.py _july                      # T10 scorer, July reference
# T10 August: cd runs/aug (data -> ../../data_aug), export PRECIP_PERIOD=aug, then
#   fetch_gauges, make_stations, fetch_mrms, fetch_goes, fetch_imerg_both; probe_a2, t10_area, t7, t7_robust, t7s, t10_score
PYTHONPATH=scripts .venv/bin/python scripts/t8.py               # T8 anvil fraction
PYTHONPATH=scripts .venv/bin/python scripts/t8_abs.py           # T8 exploratory absolute errors
.venv/bin/python scripts/make_figures.py && .venv/bin/python build_labbook.py

Environment at build: Python 3.12.11 · numpy 2.5.4 · scipy 1.18.1 · pandas 3.0.6 · h5py 3.16.0 · matplotlib 3.11.2 · macOS-26.3-arm64-arm-64bit. Randomness: bootstrap seeds fixed in code (scramble null 20261015; decision bootstrap 7; collocation 11; T1 fields 20261015, G bootstrap 3, E2 noise 99; T5 bootstrap 5; T6 bootstraps 5, 6, 8; T7 bootstrap 7; T8 bootstrap 9; T7-R bootstrap 11; T7-S 13; T9 21; T10 reuses each analysis's own seed). Known deviations from plan: Experiment 1 excludes gauges within 1° of the domain edge (so the 1.0° sweep window fits); Probe A2 dropped 2 GOES hours per its own rule; the stations table was built inline after the gauge fetch (now reproduced by scripts/make_stations.py, byte-identical).

Reduced input data (what every analysis reads; digest = SHA-256 over sorted file names and contents, first 16). Upstream archives can be reprocessed, so a rerun of the fetch scripts may not reproduce these digests exactly.

Data sources: MRMS s3://noaa-mrms-pds; GOES-16 s3://noaa-goes16 (NOAA Open Data: open, attribution requested); IMERG GPM_3IMERGHHL.07 at NASA GES DISC (Earthdata login); USCRN ncei.noaa.gov/pub/data/uscrn/products/hourly02; ASOS via Iowa Environmental Mesonet asos.py. Literature in literature/ (git-ignored PDFs; index in literature/INDEX.md).