Cumiana Colour-Scale Change (2026-08-14)

Diagnostic of Findings Log item 95(a): decide from raw Cumiana spectrograms, not derived features, whether the level change across the July collector outage is instrumental or physical. Item 95(a) set the operational consequence in advance — an instrument change would mean era_0 and era_3 image features are not on a common scale and must not be pooled.

Result: the receiver's colour scale was changed during the outage. era_0 and era_3 pixel-derived image features are not on a common scale and must not be pooled.

Reproduce with ./scripts/diagnose-vlf-palette-shift.sh.

The colourbar moved; the dB ruler did not

Every last_E_VLF capture embeds its own 96-step colourbar above the spectrogram, with a fixed -100 dB … 0 dB tick ruler printed beneath it. The ruler is a plotting constant; the ramp inside the bar is a receiver-software setting. They can move independently, and here exactly one of them did.

Across all 699 captures from 2026-06-29 to 2026-08-14:

Variant	Solid-red onset	Displayed dB window	Captures	Span
`palette_0`	step `59`	`-80.0` … `-37.9`	277	`2026-06-29T09:45Z` – `2026-07-06T21:45Z`
`palette_1`	step `48`	`-91.6` … `-49.5`	422	`2026-07-11T06:00Z` – `2026-08-14T09:45Z`

The change is a single step, with no intermediate values: every capture on or before 2026-07-06T21:45Z reads 59, every capture from 2026-07-11T06:00Z onward reads 48. The ramp moved 11 steps = 11.58 dB, downward — the later palette assigns hot colours to lower absolute levels.

Everything else in the image is unchanged. The frequency tick rows on the right-hand axis are pixel-identical across the two variants (90, 112, 135, 157, 179, 201, 223, 246, 268, 290, 312, 334, 357, 379 for 14000 Hz down to 1000 Hz), as is the image geometry (842x573) and the dB ruler. This is a colour-scale setting, not a re-plot and not a layout change.

The change point falls inside the collector outage: captures stop on 2026-07-06 and the next one, on 2026-07-11, already carries the new palette. It is therefore invisible to any diagnostic that only compares dense era_0 against dense era_3.

Why this settles the pooling question on its own

A palette change is sufficient for the item-95(a) conclusion regardless of what prompted it. vlf_intensity_mean, vlf_hot_color_ratio, vlf_high_intensity_ratio, and the vlf_band_*_mean family are all functions of pixel colour. The same colour denotes a level 11.58 dB lower after the change than before it. Those features are readings on two different rulers, and pooling them across 2026-07-06 measures the ruler.

This is the mechanism behind the item-93 observation that image-content features dominate the era shift while cadence-derived features move far less. It also explains the reported band asymmetry: the earlier "bands 0-3 fell, bands 4-5 unchanged" split does not reflect frequency-selective physics, since bands 4-5 read the separate sub-1500 Hz zoom panel and sit deep in the palette's saturated region in both eras.

Decoding to absolute dB

Because each image carries its own colourbar, pixels can be inverted back to absolute dB by nearest-colour lookup against that image's ramp, which makes the two eras comparable again. The diagnostic does this over the upper panel's rightmost 45 columns — the most recent sweep, so successive captures are close to independent rather than re-reads of the same scrollback — restricted to 11:00–13:00 UTC, because the VLF record has a strong diurnal cycle and an unmatched hour comparison would mostly measure time of day.

Two limits are reported rather than hidden:

Saturation. The palette clips to black below its floor and to solid red above its top. Only the window resolvable under both variants supports a comparison: -80.0 … -49.5 dB. Levels outside it are marked censored, not compared.
Inversion residual. JPEG compression, gridlines, and overlaid annotations push pixels off the palette curve. Pixels further than 30 in RGB distance from any ramp colour are masked. This drops 20.5% of pixels in era_0 and 20.1% in era_1, with median residuals 16.4 and 15.3 — a symmetric loss, so it does not bias the comparison in either direction.

Hour-matched, 21 early captures against 51 late ones:

Band	`palette_0` median	`palette_1` median	Delta
12–15 kHz	`-57.9`	`-73.7`	`-15.8`
9–12 kHz	`-55.8`	`-71.6`	`-15.8`
6–9 kHz	`-54.7`	`-72.6`	`-17.9`
4–6 kHz	`-55.8`	`-80.5`	`-24.7`	censored
2.5–4 kHz	`-69.5`	`-83.2`	`-13.7`	censored
1.5–2.5 kHz	`-70.5`	`-83.2`	`-12.6`	censored
0.7–1.5 kHz	`-63.2`	`-82.1`	`-19.0`	censored
0.2–0.7 kHz	`-55.8`	`-79.0`	`-23.2`

Four of eight bands are resolvable under both palettes. Their deltas span -15.8 to -23.2 dB, a spread of 7.4 dB across the whole 200 Hz–15 kHz range.

What this does and does not settle

Settled:

The colour scale changed once, during the outage, by 11.58 dB. The dB ruler and image geometry did not change.
era_0 and era_3 image features are not on a common scale. Do not pool them, and do not read any cross-era feature comparison built from raw pixel statistics as a physical result. This is the item-95(a) condition, met.
The change is invisible to era-boundary diagnostics, because it happened between the two dense eras rather than at either edge of one.
A recoverable path exists. Each image embeds the colourbar it was drawn with, so palette-inverted dB features would be era-invariant by construction and would not need the eras kept apart.

Not settled:

Whether the underlying level change is instrumental or atmospheric. The palette move proves the features are incomparable; it does not prove what the receiver was seeing. After decoding to absolute dB the late era still reads 16–23 dB lower, and the images alone cannot separate a front-end gain reduction from a genuinely quieter period. The evidence leans instrumental — the shift is broadband and roughly uniform across two decades of frequency, is step-like rather than gradual, and coincides with an operator changing a display setting, which is what an operator does after a gain change. That is circumstantial, not conclusive. Station metadata or operator contact would settle it; pixels will not.
Whether palette-inverted features carry signal. Nothing here tests that. It removes a known scale artifact; it does not create evidence.

This diagnostic does not support any claim about earthquake-related VLF signals. It identifies and quantifies an instrument-side artifact in the capture record.

Absolute-dB features, and the pooling verdict (2026-08-21)

Answers Findings Log item 97(a), and settles item 105(b).

The recoverable path this document identified is now implemented: src/elfquake/features/vlf_image_db.py, run by ./scripts/extract-vlf-image-db-features.sh. Every capture is inverted through its own embedded colourbar, so the features are invariant to the palette setting by construction. All 824 captures decode, with a mean censored fraction of 0.229 and 7.65 of 8 bands scoreable per capture.

The answer to the question that motivated it is no. Palette-inverted features do not make era_0 and era_3 poolable, so the item-96 restriction stands and the transition budget in Target Design does not improve.

Censoring is carried, not smoothed over

Item 97(a) required this and it is the whole difficulty. The two variants resolve different dB windows, so a band below the shared floor is missing, not quiet, and a median over the pixels that escaped clipping would report the palette floor as a measurement. Each band therefore carries vlfdb_band_<i>_censored_fraction, and its level is withheld entirely — left empty, never filled with a number — once censoring passes 50%. Each capture also carries its own vlfdb_black_end_px / vlfdb_red_start_px, so a third palette variant would show up in the features rather than only in this diagnostic.

On the real record the censoring is almost entirely inversion-residual masking from JPEG artifacts, not clipping: hot- and black-clipped fractions are 0.000 on nearly every day. The decoded levels are not floor or ceiling artifacts.

A level step survives the inversion

Reproduce with ./scripts/diagnose-vlf-db-era-step.sh. Hour-matched to 11:00–13:00 UTC, and reduced to one median per day before comparing — captures 30 minutes apart are highly correlated, so days are the independent unit:

Band	Early median	Late median	Delta	Step ÷ within-era spread
12–15 kHz	`-56.6`	`-73.2`	`-16.6`	`2.93`
9–12 kHz	`-53.9`	`-70.5`	`-16.6`	`3.06`
6–9 kHz	`-52.4`	`-71.1`	`-18.7`	`2.92`
4–6 kHz	`-54.7`	`-76.3`	`-21.6`	`2.83`
2.5–4 kHz	`-66.1`	`-83.2`	`-17.1`	`2.12`
1.5–2.5 kHz	`-65.8`	`-82.6`	`-16.8`	`2.26`
0.7–1.5 kHz	`-62.1`	`-81.1`	`-18.9`	`2.29`
0.2–0.7 kHz	`-55.8`	`-76.3`	`-20.5`	`2.70`

Median step -17.9 dB, range -16.6 to -21.6 across two decades of frequency, at 2.8 times the day-to-day spread within either era. That is a discontinuity, not ordinary variability, and it is broadband and roughly uniform — which is what a front-end gain change looks like and what frequency-selective propagation physics does not.

Why inversion could never have removed it

The hope recorded in item 97(a) — that dB features would be "era-invariant by construction and would remove the need to keep the eras apart" — was only ever true of the display component. Decoding through the colourbar recovers dB as the receiver reported it, not dB at the antenna. A palette change alters the plot; a gain change alters the quantity being plotted. Inversion undoes the first and leaves the second exactly as it was.

So the two changes at the outage separate cleanly:

The 11.58 dB palette move is a display setting. It is now removed, and the pixel-feature incomparability it caused is fixed.
The ~17.9 dB level step is in the data the receiver reported. It is untouched by inversion, and pooling across it would hand a model an era indicator measured in dB instead of in colour.

That the two numbers are close is itself evidence: an operator who reduces front-end gain by roughly 12–18 dB and then moves the colour ramp down by 11.58 dB to keep the display readable is doing one job in two steps. This is consistent circumstantial evidence, not proof, and it does not change the operational answer either way — the eras stay separate.

What this does and does not settle

Settled:

Absolute-dB features exist, decode on the whole record, and carry their censoring explicitly.
The palette artifact is removed. Cross-era comparisons of vlfdb_band_*_db_median are on one ruler, unlike the raw pixel features.
The eras still must not be pooled. A 17.9 dB broadband step at 2.8 times the within-era spread remains, so item 96 stands and item 105(b)'s premise — that this would roughly double the usable label transitions — is refuted. The blocker recorded in item 104(d) is unchanged: calendar time.

Not settled:

Instrumental versus atmospheric, still. The evidence has moved further toward instrumental — the step is now measured on a common ruler, is uniform across 200 Hz–15 kHz, and is nearly twice as large as the display change that accompanied it — but pixels cannot separate a gain reduction from a quiet period. Operator contact (item 105(c)) remains the only route.
Whether absolute-dB features carry any signal. Nothing here tests that. They remove a known artifact within an era; they create no evidence.

Nothing in this section supports any claim about earthquake-related VLF signals.