Astronomy and Geomagnetic Feature Alignment (2026-08-17)

Implements Findings Log item 100. Before this, astronomy was present in the multimodal table in name only: the audit in item 99 found two channels reaching the transformer, one of them a monthly constant repeated on all 11,020 rows and the other a count of files the collector happened to write. No astronomy ablation run before this date measured astronomy.

Nothing here is evidence that astronomical or geomagnetic data helps predict earthquakes. It made the question askable, and the answer came back negative: the held-out ablation in log item 102 shows these features add nothing at a 7-day horizon in the fixed-cell design. What this work establishes is that the null is now a real measurement rather than an artifact of a constant channel.

What the channels are now

Two kinds, with different missingness contracts.

Ephemeris — deterministic functions of UTC, computed in src/elfquake/features/ephemeris.py. Never missing, so their masks are honestly always clear.

Channel	Meaning
astro_moon_phase_angle_deg	Moon-Sun elongation, 0 at new moon, 180 at full
astro_moon_phase_sin, astro_moon_phase_cos	Wraparound-free encoding of the same angle
astro_moon_illuminated_fraction	Illuminated disc fraction
astro_moon_distance_km	Geocentric lunar distance
astro_tidal_potential	Degree-two lunar + solar tidal potential at the anchor
astro_tidal_potential_min, _max, _range	Same, summarized over the lookback window

Observed — normalized from the archives in data/derived/astronomy/, each with an age in hours and its own missing flag.

Channel	Source	Cadence
astro_kp, astro_ap, astro_kp_max, astro_ap_max	GFZ Kp_ap_since_1932	3-hourly
astro_dst_nt, astro_dst_min_nt	Kyoto WDC Dst	hourly
astro_f107	Spaceweather Canada daily flux table	3 readings per day

The named-phase category astro_usno_next_phase is gone. A next-event label is a sawtooth countdown over an arbitrary ordering, not a physical state.

The alignment rule

Anchors are the window end, which is the VLF capture time — a 30-minute grid against source cadences of 3 hours, 1 hour and 1 day.

1. Zero-order hold, never interpolation. Interpolating between the readings either side of an anchor mixes a future reading into a feature that is supposed to be causal. Each channel takes the most recent reading whose observation interval has closed at or before the anchor. A Kp bin covering 09:00–12:00 is not usable at 09:57.
2. Staleness is a field, not a footnote. astro_*_age_hours is the gap from the end of that interval to the anchor. A daily F10.7 value presenting as a 30-minute observation was the failure this replaces.
3. Held values expire. Past 6 h (Kp, Dst) or 72 h (F10.7) the hold is abandoned: the value is blank and quality_missing_* is 1. Holding a three-day-old Kp reading forward would manufacture exactly the constant this work removes.
4. quality_missing_astro reports observation, not computability. It is set only when Kp, Dst and F10.7 are all missing. Ephemeris channels are deliberately excluded from the test — they are always computable, so letting them satisfy it would pin the flag to 0 forever, which is what the old implementation did.

Window aggregates cover only readings that closed inside the window. astro_dst_min_nt is the storm depth, since Dst goes negative during a storm.

The tidal potential

The degree-two potential at a reference site, in units of the lunar term at mean distance with the Moon overhead:

V = (r̄_moon/d_moon)³ · P₂(cos ψ_moon) + 0.4599 · (AU/d_sun)³ · P₂(cos ψ_sun)

The reference site is the centre of the project's Italy bounding box (41.4°N, 12.5°E). Italy spans about 13 degrees of longitude, so the semi-diurnal phase at the extremes differs from the centre by under an hour. Positions are geocentric; topocentric parallax moves the lunar zenith angle by up to a degree, well under a percent of the term's range.

The 0.4599 solar coefficient is (M_sun/M_moon)·(r̄_moon/AU)³, which reproduces the textbook ratio of the solar to lunar tide.

Accuracy

ephemeris.py uses the truncated periodic series from Meeus, Astronomical Algorithms, chapters 12, 25 and 47. Checked against the worked examples there and against published 2026 lunation times:

Quantity	Reference	Computed	Error
Moon longitude, 1992-04-12	133.162655°	133.1354°	1.6'
Moon latitude, 1992-04-12	-3.229126°	-3.2292°	<1"
Moon distance, 1992-04-12	368409.7 km	368399.3 km	10 km
Sun longitude, 1992-10-13	199.90988°	199.90987°	<0.1"
GMST, 1987-04-10 00:00 UT	197.693195°	197.6932°	<0.1"
New moon	2026-08-12 17:37Z	within 1° of elongation 0	—
Full moon	2026-08-28 04:18Z	within 1° of elongation 180	—

Meeus's examples are in dynamical time; UTC differs by about a minute, which is itself roughly half an arcminute of lunar motion. Two arcminutes is far more than a tidal proxy needs.

Observed range in the rebuilt table

Over the 761 anchors of all_italy.prospective_vlf_windows.csv, every channel now varies:

Channel	Range	Distinct values
astro_kp	0 – 7.333	20
astro_ap	0 – 154	20
astro_dst_nt	-150 – +53 nT	101
astro_f107	94.8 – 258.1	65
astro_moon_phase_angle_deg	0.19 – 359.9	760
astro_tidal_potential	-0.804 – +1.278	761

Compare the audited state: astro_noaa_solar_cycle_f107_value constant at 125.69, astro_capture_count nonzero on 3 of 17 days. The window covers a real geomagnetic storm (Dst to -150 nT, ap to 154), so the geomagnetic channels have something to be tested against.

The transformer path

The tabular fixture and the transformer are fed by different artifacts, and until 2026-08-22 only the tabular one carried this work.

build-common-transformer-fixture.sh writes data/derived/models/common_transformer_fixture.csv; materialize-common-transformer-sequences.sh turns that into the per-modality sequence tensors under data/derived/models/common_transformer_fixture_sequences/, which is what evaluate-self-supervised-transformer actually reads. The second step is manual and unguarded, so the sequences sat 13 days behind the fixture. The fixture had the 19 aligned channels from 2026-08-17; the astronomy sequence still had astro_capture_count and astro_noaa_solar_cycle_f107_value, the two dead channels item 99 found. Every transformer run in between trained on the dead pair.

After rematerializing, italy_all_astronomy_sequence/manifest.json reports:

Field	Before	After
channel_count	2	19
row_count	9,082	15,960
time_count	478	840
masks that ever read present=0	0 of 2	Kp/ap family, 152 rows

("After" is the 2026-08-22 rebuild of the whole chain, from the labeled spatial table forward; an intermediate rematerialization from the 2026-08-17 fixture gave 14,478 rows over 762 steps with the same 19 channels.)

The mask column matters. Item 99's complaint was that the missing-modality mask was unconditionally true; the per-channel masks now fire where the observation is absent — the five Kp/ap channels blank on 152 rows at the end of the record, where the GFZ file has not yet closed its 3-hour bin. Dst, F10.7 and the ephemeris channels are genuinely present throughout, so their masks staying clear is correct rather than broken.

What the Transformer does with them

The full chain was run on 2026-08-22 — labeled spatial table, fixture, sequences, then run-cross-region-generative-smoke.sh — so this is the first result in which the model saw the aligned channels rather than the dead pair. Synthetic masked pretraining, Japan self-supervised continuation, then chronological Italy fine-tuning on seismic + VLF + astronomy; 11,004 training rows and 2,752 test rows with 678 test positives; one seed, CPU.

Evaluation	Balanced accuracy
all three modalities	0.513960
astronomy masked	0.511239
seismic masked	0.511085
Italy VLF masked	0.493117
linear probe on the frozen representation	0.498439

Read that block as indistinguishable from 0.5. The spread is two to twenty thousandths on a single seed, and the masking test measures how much a model trained on everything leans on a channel at inference — not whether the modality adds value, which would need retraining without it. What the run establishes is that the interface works end to end with 19 real channels; it establishes nothing about astronomy.

run-cross-region-generative-smoke.sh now checks both hops with the shared freshness guard (see Input freshness) and refuses to read sequences older than the fixture they claim to come from.

These channels encode the calendar

A caution that applies to any model reading them, and the reason a positive astronomy result would need more than an ablation to believe.

Correlation of each channel with the anchor index, over the 799 distinct anchors of the labeled spatial table (2026-06-28 to 2026-08-19):

Channel	r with time
astro_f107	-0.880
astro_moon_phase_cos	+0.838
astro_moon_illuminated_fraction	-0.838
astro_moon_distance_km	-0.730
astro_tidal_potential_min	-0.604
astro_kp, astro_ap, astro_dst_nt, ...	|r| < 0.26

Five channels are near-monotone in time over this record. That is not a leak — no future information enters any row — but under the time-based split the training rows are early and the test rows late, so those five are close to an indicator for which side of the split a row is on. A model can use them to recover the era's base rate rather than any physical state, and a lift built that way would not survive a different split or a longer record.

Two consequences. The lunar channels are periodic and only look monotone because 52 days is under two lunations; they will decorrelate as the record grows, so this is a property of the current window, not of the features. astro_f107 is different — solar flux trends over months, and it will keep behaving like a date stamp for as long as the record is short compared with the solar rotation. Read any astronomy result against the shift control in Within-cell null control, which destroys feature-label alignment while preserving the time structure, before treating it as physical.

The channel gate

src/elfquake/models/channel_gate.py runs inside the fixture builder and raises rather than warns on three defects:

• constant_channel — every value identical across at least 8 rows. This is the item-94 failure generalized from the standardizer to the builder, where it would have been caught five weeks earlier. A channel that is constant by design has to be named with --allow-constant-channel, so the decision is recorded rather than inferred.
• unmasked_missing_channel — blank on some rows with no quality_missing_* flag set on all of them. Such a channel is imputed silently downstream and the model cannot tell a measured zero from no measurement.
• empty_channel — never populated at all.

--no-strict-channels downgrades the gate to a report entry for exploratory runs. The defect list is written into the fixture report either way.

Reproducing

./scripts/refresh-space-weather.sh              # fetch + normalize the archives
./scripts/refresh-prospective-labels.sh         # rebuild the window tables
./scripts/build-italy-spatial-vlf-targets.sh    # rebuild the labeled spatial table
./scripts/build-common-transformer-fixture.sh   # tabular fixture
./scripts/materialize-common-transformer-sequences.sh   # sequence tensors

The last line is not optional and used to be missing from this list. The fixture and the sequences are separate artifacts, nothing rebuilds the sequences when the fixture changes, and the transformer reads the sequences. See The transformer path.

refresh-space-weather.sh is deliberately on its own daily timer (deploy/systemd/elfquake-space-weather.timer) rather than the prospective job's 30-minute cadence: the Kp/ap and F10.7 sources are whole-history archives of roughly 16 MB and 2 MB, and both publish at most once a day. It skips a refetch when a copy under 20 hours old is already on disk.

Source notes

• GFZ Kp/ap — CC BY 4.0. Whole history from 1932; the file ends at the last completed 3-hour bin, so the newest anchors legitimately report Kp missing.
• Kyoto Dst — non-commercial use. Recent months are served only by the realtime tier: probed on 2026-08-14, final and provisional both 404 for 2026-06 through 2026-08 while realtime returns 200, and provisional serves 2025-12. Realtime values are provisional and get revised, which is why dst_tier is carried through normalization and the previous month is refetched on every run.
• Spaceweather Canada F10.7 — three Penticton observations per day at 17, 20 and 23 UT. The 20 UT fluxadjflux reading is the conventional daily F10.7, so the observation time is preserved rather than collapsed to a date.

Known gaps

• The ablation has been run twice, and it is negative both times. See log items 102 and 104. On the item-104 design (1-day horizon, M>=2.0, cell-stratified, thresholds calibrated on training rows only) the era_3 held-out result is seismic_only 0.552836 stratified against seismic_astronomy 0.500000 — and that 0.500000 is exact because the 21-feature model predicts the negative class on all 1,976 test rows, having reached 0.613508 balanced accuracy on its training rows. Twenty-one features against 19 label transitions is the whole story. full_multimodal is 0.431988 and all_features 0.404832: every family added to seismic history lowers the score. These features are correct, varying, and aligned; they do not improve prediction at a 7-day or a 1-day horizon in the fixed-cell design.
• src/elfquake/features/astronomy.py and multimodal_smoke.py still emit the old columns for the older build-multimodal-smoke path. They are marked superseded; the channel gate will reject any fixture built from them.
• The tidal potential is computed at one reference point, not per target cell. For a country the size of Italy that is a sub-hour phase approximation, but it is an approximation.