Rhodamine sensitivity screen — 2 September

ResultAll four channel types at 20.0 mg/L

Two statistics, because these channels do not behave alike. Amplitude is the median tlf_amp relative to each sensor’s own pre-dose level — the right measure for a channel that produces a signal. Lit fraction is the share of sweeps whose peak clears the pedestal by 20 counts — the right measure for the chlorophyll-a channel, whose sweeps are bimodal at either 1 count or ~3000, so a median flips between the two and summarises nothing.

ControlThe dose is confirmed without the sonde

ReferenceWhat the In-Situ sondes recorded

The two instruments look through effectively the same window. The Lume chlorophyll-a channel excites at 440 nm and detects through an Edmund 45786 filter, 676 nm CWL with 29 nm FWHM and OD 6 blocking, so its passband is 661.5–690.5 nm. The AquaTroll excites at 430 nm and detects 675–750 nm. They overlap across 675–690.5 nm, and both contain the 685 nm chlorophyll-a emission peak. Whatever the sonde saw at 46.6 and 121.3 sigma was therefore reaching a band the Lume channel also covers, so the difference between them is not a choice of filter. An earlier revision of this page put the Lume detection at 650 nm and argued the band sat closer to rhodamine than the sonde’s; that figure was wrong and the argument with it.

BenchDoes the sensing path still work?

Unit 500185 was lifted out of the barrel after the screen closed and taken through three conditions on the bench. Everything is the same unit, so nothing needs normalising between phases.

The cap-off phase is what separates the two hypotheses. Ambient room light reaches the detector without the excitation LED being involved at all, so it tests the sensing path on its own. Net signal is the wrong measure for this and is shown only to demonstrate that it stays flat: net = val − dark, and the dark subtraction cancels ambient by design.

How far the fault goes

We read this channel at the wrong operating point

Measured from the archived sweeps, at the calibrated combination the fleet normally uses (LED 512, SiPM bias ~3000) the chlorophyll-a channel returns about 110 to 155 counts, against 1,009 to 1,148 on FDOM and 475 to 483 on TLF at the same combination. It is 7 to 10 times weaker than its neighbours there. The signal lives at the top of the bias range instead: 500151 goes from about 150 counts at bias 3000 to about 2,000 at bias 3800, and at the very top both surviving units reach 3,082 and 3,104 counts, which is the same rail the working FDOM and TLF controls hit.

The screen the canonical pipeline applies is LOW = 200 on raw mon2, which rejects the sub-breakdown pedestal. It was set for channels that rail at 3,000 counts. Measured across the archives, FDOM and TLF units clear it on 79 to 86% of sweep cells and the canonical fitSlope returns a value on 100% of their sweeps. A healthy chlorophyll-a unit clears it on 21 to 66% of cells and yields a slope on 48 to 90% of sweeps. The channel is operating at the screen even in good health.

That is why neither experiment can be re-read to settle the dose question. On 500151, the one unit still producing signal, supportUnstable is true and chooseSupport finds no bias bin that stays unclipped across its sweeps, so the fixed-support amplitude declines to return a value. The ordinary matched filter does return one, and it reads +1,563% at rhodamine, but it is built from 14 of 65 sweeps in that window and 4 to 9 in the others. That is the rebasing artifact the estimator documents, not a measurement. Answering this needs a run designed for it, and probably its own screening constants: PED = 170 with LOW = 200 leaves a channel whose whole working range is roughly 170 to 350 counts just 30 counts of headroom.

DetailEvery sensor

LimitsWhat this does and does not establish