Burn-in fleet — validation, baseline & calibration

Stage 1Dilution calibration — offset and gain

Known tryptophan ladder dosed into the bath on 2026-08-06: 0, 0.1, 0.5, 5, 10 and 50 ppb. Every step is read as a plateau — the window is checked for drift before it is averaged. The 50 ppb window sits at 15:24–15:33 rather than the logged 15:19–15:29, because that dose was still mixing for its first five minutes (24.6 % rise inside the logged window, against ≤ 2.5 % for every other step). The ladder ran before the DI baseline began, which is why the clean-water stages do not read it.

All four panels are built the same way: the mean at each dosed level, per sensor, with hashes at the 95 % CI of that mean (Student t, since the 50 ppb step is only a handful of sweeps). Markers only — a segment drawn between two dose levels would assert a trajectory that was never measured. Each series carries an OLS regression across all six plotted levels, and the R² in the legend is that line’s.

Campaign & window

The window is fixed in the page, not adjustable: it opens at 2026-08-07 13:15 local, the moment all 25 units were cleaned and put into fresh DI water outdoors, and closes with the campaign on 2026-08-20 09:00. Every stage reads that one window, so nothing on this page depends on a control someone may have left in an odd state. Extending it is a code change, in BASELINE_START, which scripts/build_dilution_calibration.js mirrors as COOL_FROM.

The full stored record

The complete record for every unit, start of campaign to its close, straight from data/burnin_bands.csv.gz. Unfiltered — the retired tap baseline, the ToF handling window, the dilution ladder and the dosed bath are all included, so nothing the later stages leave out is hidden here.

Stage 2Clean-water record — DI, outdoors

From 13:15 on 2026-08-07 all 25 sensors were cleaned and placed in fresh DI water, outdoors. This is now the clean-water reference for the whole page: the baseline block, the quench, the ToF calibration and the end-to-end test all read this window and nothing earlier. Outdoors it sweeps a real diurnal range rather than a bath cooldown, which is the condition a deployed sensor actually meets — and, unlike the retired tap window, it breaks the collinearity between temperature and elapsed time.

There is no analyte in DI water, so everything in the raw trace above is instrument. Three terms account for it, fit per sensor on this window:

ln s = a + cubic(T − 21) + c · dT/dt + d · e−n/τ

— a temperature response, a thermal-lag term in the rate of change (the thermistor is on the electronics, not the water), and a settling term where n is samples since the last clean. The clock is sample count, not wall-clock, so it is independent of calendar and cadence. The only disturbance modelled is cleaning; the one event here is 2026-08-07 13:15 local.

Finding the offset: it is the fit intercept. Set T = 21, dT/dt = 0 and n = ∞ and every term but a drops out, so ea is the settled blank at 21 °C — measured directly in clean water, which is what this window is for. Applying it: subtract it, then divide by the ladder gain mapped into the same frame. The temperature term is multiplicative on the whole signal, analyte included, so gain@21 = gainraw ÷ ecubic(Tladder−21) — no blank enters, so there is no circularity. The mapping factor is 0.984 median.

Fit on train, reported on test. Every coefficient, τ included, is fit on the first two thirds of the samples since the clean (—). The last third is withheld and is shaded on the panels below. The split is temporal rather than random, because neighbouring samples are minutes apart and a random split would leak; and because the settling term is a function of samples since the clean, holding out the tail forces it to extrapolate the decay rather than interpolate inside its own training data.

On the held-out third the full model leaves — against — for temperature alone, so the settling term is doing real work out of sample, on — sensors. In-sample it leaves —, and the gap between those two is the honest cost of having fit the thing. A lifetime-count clock was tested and rejected: fit on the earlier window it predicts this one at 17.1 %, so the decay restarts at the clean rather than tracking device age.

Against temperature, before and after

The same data on a temperature axis rather than a time axis, which is where a quench is easiest to read. Left: raw ln sTLF against SiPM temperature, with each sensor’s fitted temperature term drawn through it. Right: the fully corrected signal — temperature, thermal lag and settling all removed — which should be a horizontal band for every sensor. Any residual slope or curve is temperature the model did not capture. Both use the model applied in this stage, not a separate fit.

Against the published requirement

How good does this have to be? The literature answers that through the TLF concentration that corresponds to a given E. coli level. Sorensen et al. (2018) pooled groundwater and surface drinking-water data from India, Malawi, South Africa and Zambia (n = 564) and placed the WHO risk bands at 1.3, 2.4, 6.9 and 27.1 ppb for ≥1, ≥10, ≥100 and ≥1000 organisms per 100 mL. Independent field studies put the detection threshold in the same place: 1.9 ppb from thermotolerant coliforms and 1.7 ppb from E. coli in Malawi (Ward et al., 2021), 2.2 ppb in Uganda (Sorensen et al., 2021), 1.0 ppb in Kenya (Nowicki et al., 2019).

The recreational thresholds sit inside the published High band, so they are read by interpolating between 6.9 ppb at 100 and 27.1 ppb at 1000 CFU/100 mL: the EPA geometric mean of 126 lands at ≈ 7.9 ppb and the statistical threshold value of 410 at ≈ 16.0 ppb. That interpolation is ours, and it carries a drinking-water relationship into recreational surface water, which Ward et al. explicitly caution against. Treat the recreational rows as indicative.

Which noise applies. A threshold crossing is called at a moment, so the relevant figure is the short-term scatter, taken here as the standard deviation of successive-sample differences (÷ √2), which removes any slow drift. That is a different and much smaller quantity than the multi-day figure in the card above, and it is the one comparable to a published detection limit. Both are reported.

For scale, Sorensen et al. (2018) report field fluorimeter detection limits of 1.99 ± 0.53 ppb for the Cyclops-7 and 1.74 ppb for a modified GGUN–FL30, both of which fail their own 1.3 ppb threshold, against 0.17 ± 0.06 ppb for the photomultiplier-based UviLux; field repeatability in Sorensen et al. (2021) was σ = 0.4 ppb. Bedell et al. (2022) reported 0.05 ppb in deionised water for an earlier generation of this sensor.

Excluded sensors

The point of the DI run is to find sensors needing rework and to correct the rest to a flat zero line. These are the ones that cannot be corrected. Withdrawn units are out of every fit and every fleet chart; rework units still report and still hold data, but are kept out of fleet statistics and hidden on the charts above — click them in any legend to bring them back. Ranking is from a clean window with the full 36 °C range and no handling (08-12 14:00 to 08-13 08:40 MT), so it measures the sensor rather than the disturbance.

The same four views as Stage 2, excluded units only

Identical panels to the stage above, so these sit on the same axes as the fleet they were removed from. A withdrawn unit’s trace ends where it stopped reporting, and units without a usable calibration are absent from the two corrected panels — which is itself the reason several of them are here.

Stage 3Thermal lag — what temperature alone misses

Correcting on temperature alone leaves a hysteresis: at the same measured SiPM temperature the signal does not read the same coming down as it did going up. Across — that gap averages —, and it did not shrink when the post-clean decline stopped, so it is not a settling transient. Warming and cooling are separated per sample from the sign of dT/dt, not from a fixed peak time, and the gap is measured at every whole degree the fleet covered rather than at one chosen temperature. That second point matters: quoting a single bin was hiding the result below.

The likely reason is that the thermistor is on the electronics and there is no thermometer on the water or the optical window. Outdoors the enclosure follows air quickly while the water body lags it, so at the same measured temperature the medium the light passes through is not at the same temperature on the two limbs. A lag shows up as a dependence on the rate of change, so the test is to add a dT/dt term and see whether the loop closes. (The UV LED reports its own temperature, but it is the same thermistor — r = 1.0000 against the SiPM — so it adds nothing.)

Honest limits: the term is fit and tested on the same cycles, so any variance it removes is in-sample and flattered. It does not buy accuracy either — residual SD goes —, which is nothing. Nothing in this stage feeds the correction applied in Stage 2; it is a diagnostic, and on this evidence the honest reading is that the remaining hysteresis is still unexplained.

Stage 4ToF turbidity calibration

The clean-water period is 0 NTU by construction, so every bit of variation in signal_per_spad_kcps over it is instrument, not water. That makes it a ground truth: fit a correction there, and the corrected series must flatline on zero.

spscorr = sps − g · (T − 21), then ΔSPS = spscorr − sps₀. The axis is ΔSPS, not NTU. No formazin reference was run on these 25 units, so there is no scale factor to apply. The pooled k ≈ 1.59 on turbidity comes from two other sensors whose own slopes were 2.05 and 1.38 — a 1.5× disagreement between just those two — so transferring it here would print a unit we have not earned. Everything below is still a valid relative turbidity signal; converting it to NTU needs a formazin ladder on this fleet. g is per sensor, not fleet-wide: measured r(SPS, T) runs from −0.69 to +0.91 across the 24 units, so it does not agree even in sign and a single coefficient cannot be justified. The fit uses the central 90 % of each sensor’s clean-water SPS, because three units (500155, 500230, 500228) show sustained excursions there — 500155 runs 181 consecutive readings above twice its median — and those are events in the optical path, not drift to be absorbed into a baseline.

Stage 5Indoor bath control — 50059 & 500214

Test ended 2026-08-17 07:44 local (5.9 days). Data below is the complete record. Headline outcomes: 50059’s 08-14 basin-move step (×2.6 s_TLF, −15 % ToF) survived two deliberate shakes — only its ToF reverted — while 500214 in the same water stayed within a few percent throughout: handling artifacts are unit-local, shake-resistant, and sit on the LED/SiPM optical path, not the ToF window.

Two units ran from 2026-08-11 10:00 local in a water bath indoors, tracked separately from the 25-unit outdoor fleet. They are a control on the temperature confound: indoors the SiPM sits at —, against the fleet’s 21–40 °C diurnal swing. Any drift here therefore needs no temperature model to see, and cannot be a temperature artifact.

Three things make them not directly comparable to the fleet on absolute level or rate. They are the ToF firmware-swap pair, so the two may be on different builds from each other. Their cadence is far slower — roughly 4 sweeps an hour against the fleet’s 60 — so per-sweep dose accumulates about 15× more slowly, which matters for any comparison framed in dose rather than calendar time. Both start from the same instant here, so the two records are directly comparable in elapsed time.

Read with care while the record is short. So far 50059 has moved — and 500214 —. At roughly four sweeps an hour a day of running is only ~100 points, so these rates are indicative until the record is several days long.

18:37 shake outcome (the bubble kill test): a split verdict. Shaking 50059 in place did not collapse its ×2.6 basin-move step — s_TLF ticked up (+9 % first sample, settling +5 %) — but its ToF snapped fully back 41 → 35 → 41, recovering its pre-move value. So the IR obstruction was loose (dislodged by the shake) while the TLF-boosting change is shake-resistant: either a bubble held in a crevice at the window, or the unit’s position/reflection geometry, which an in-place shake does not reset. 500214, shaken the same way, stepped up +2.4 % — consistent with the fleet-wide finding that vigorous shaking adds signal. Next discriminator: lift 50059 out and re-insert slowly at a tilt (the manufacturer-prescribed bubble-shedding move), or wipe its face — a positional cause survives both; a crevice bubble does not.

Basin carry — 2026-08-13, the fleet’s own handling step (±1 h)

The outdoor basin was carried to the DI supply at 11:52 local and the tank refilled from 12:02. During the carry itself — before any water changed — 21 of 24 units stepped UP in raw s_TLF (median +3.3 %; 500179 +303 %, 500108 +166 %, 50084 +159 %, 50088 +133 %) while the bath temperature fell smoothly, so the jump is not thermal, and only 1 % of the window’s rows are flagged out-of-water. This is the same signature as the indoor pair’s 08-14 basin move (50059: s_TLF ×2.6 up, ToF −15 %), consistent with sloshing depositing bubbles in the optical near field. The refill and the 40 °C ramp that followed then drove the level down — the −35 % headline step is that composite, not the carry itself.

In-place shake — 2026-08-14 19:12, all 24 units (±1 h)

The companion test: the same 24 units shaken in place at 19:12 local, with no transport, no water change, and the bath at ~3.7 °C after a ~48 h cold hold — strongly undersaturated water, so any up-step here cannot be a supersaturation-grown bubble. With an hour of post-shake data: 20 of 24 units up in s_TLF, median +19 % (500103 +278 %, 50091 +209 %), while ToF barely moved (median 0.0 %) — unlike the indoor basin move, the fleet’s TLF steps came without a ToF signature. Four units went down (500179 −45 % with ToF +17 %, 50084 −29 %, 500226 −21 %, 500108 −12 %), so shaking removes material from some faces while adding it — most plausibly entrained surface air — to most. The handling itself delivers the scatterer; gas saturation state is not required.

Stage 6Blackout caps — 4 units, fitted 2026-08-12 12:35

New blackout caps went on 500146, 50095, 500226 and 500124 at 12:35 local. 500124 has since been withdrawn — it was leaking, so this stage shows the remaining three. Its dark channel had collapsed from ~167 counts to 0 against a fleet 171–172 95 minutes before the cap was fitted, so the cap did not cause the failure and the earlier “capped versus control” numbers that included it were carrying a leaking unit. The question is what the cap interior does to s_TLF — both whether it shifts a unit’s level, and whether a common cap makes units alike. Every unit’s blank differs, and if the inside of the cap is part of that, fitting the same cap to several units should pull their levels together.

Showing only the four means there is no control group to difference against, so the level figure is a before-versus-after and still contains the bath’s temperature swing. The spread panel is the one that survives this: it asks whether the four agree with each other more closely than they did, which does not need a control. Each unit is normalised on a window 8–4 h before the change, deliberately separate from the 2 h “before” window, so the pre-change spread is free rather than forced to zero.

—

Stage 7Baseline, validation & QC

One row per sensor. The baseline block is the median sTLF and signal_per_spad_kcps (ToF) over the baseline sub-window, with the window mean temperature — baselines are temperature-dependent, so it is recorded alongside. These become each sensor's canonical clean_water offsets; a stable baseline has a low CV, a high CV means the window is not clean or settled yet. The QC block is readiness over the same window: stability of TLF / ToF / temperature judged over the last 4 h, the leak signal (internal absolute-humidity trend — rising means ingress) and battery. Click any header to sort.

blank ÷ gain is the ppb-equivalent of each sensor’s blank, taken from the raw ladder fit. It converts blank instability into the units that matter: a unit at 36 ppb loses 0.4 ppb for every 1 % its blank moves, one at 1790 ppb loses 18. It is the acceptance criterion that decides how stable a given sensor has to be. Green ≤ 60, amber > 300.

Stage 8Air drift test — all 24 units, 08-18 11:00 to 08-20 09:00

Started 2026-08-18 11:00 local. The whole fleet was pulled from the ~29.3 °C bath and left in air. This is the bubble discriminator: a dry window cannot hold or grow a bubble, and the LED still fires every sweep against a fixed target. If the residual burn-in drift (fleet median −0.66 %/day over the final 46 h stable cold hold in water) continues in air, it is LED/electronics aging and the cycle-count correction stands. If it stops, it was water-side — bubbles or films — and window maintenance, not modelling, is the fix. In between: both, in measurable proportion.

This stage is closed. The air test ran 2026-08-18 11:00 → 2026-08-20 09:00 local, about 46 h, and every series, drift fit, fleet median and table entry here is cut at 09:00 — not merely hidden from the charts, so no in-water reading can leak into an air-phase result. One shared end for all 24 also means every unit’s rate covers the same window. Afterwards the fleet went into water in two groups: five units into the vacuum container for the degassing test at 09:29, and the other 19 into the shared DI tank at 11:40, where they join the FDOM/chlorophyll-a batch — that record lives on the variant page, not here.

Reading rules. The pull itself is a handling event — expect a step and ignore it; windows came out wet and cool by evaporation, so the first 2 h are settling, excluded from the fit. Per-unit rates appear once a unit has ≥12 h of settled air data (the same floor the drift-correction validation set; shorter fits are coin flips). There is no temperature control in air — the fleet rides the building’s ambient diurnal — so the drift fit estimates each unit’s own temperature coefficient jointly with the time slope (ln amp = a + b_T(T−25) + slope·days, the same structure as the canonical drift correction). The table shows the temperature-corrected rate, the fitted coefficient, and the raw slope for comparison; until a full diurnal cycle decorrelates temperature from time the two are confounded, and such fits carry a ⚠ instead of a verdict. These charts deliberately include the rows the out-of-water detector excludes from every water stage above.

Stage 9Degassing test — moved to its own page

The five units in the vacuum container — 500220, 500225, 500226, 500108, 500146 — now have their own page: Degassing test →. It reads a dataset scoped to start at the test itself (2026-08-20 09:29 local), so nothing on it is diluted by this campaign's history.

Stage 10Drift correction — the production model on this campaign

The same correction a deployed sensor would apply, run over the frozen campaign: ln(amp) = Pd(tn) + (r1+r2tn)·dT/dt + slope·days, fitted per sensor on its own sipm_temp_c and evaluated back to 20 °C with dT/dt = 0. Precomputed by scripts/build_burnin_drift.py, because this archive is 507,660 rows and the cohort is closed.

Each unit gets the richest model its data supports, falling back cubic + lag → cubic → linear + lag → linear → none when a fit implies an implausible correction (>8×) or leaves more than 8 % residual scatter. The pass rate is the finding, not a footnote.

What the correction leaves behind

Fleet median with interquartile band, normalised to each unit’s own first settled window. Blue removes temperature and leaves drift; green removes drift as well — the complete production correction, which should be flat. Shown for the two windows where the model passes fleet-wide; on the outdoor windows too few units pass to draw a fleet band honestly.

In-sample, and it matters here. The drift term is fitted on the same window it is removed from, so a flat green trace shows the fit converged, not that the rate would hold tomorrow. The forward test on findings is the honest one, and there extrapolating a fitted slope was worse than assuming the last baseline holds.

The instrument, and what each correction is for

Part numbers are from SweetSenseInc/tlf_hardware (v0.1/tlf-bom.csv, v0.1-led/tlf-led-bom.csv); wavelengths from the TLF method documents. Everything in the measured column comes from this campaign's own data, and is stated separately from the datasheet because in one important case the two differ by a factor of two.

Optical chain

The SiPM's peak sensitivity is 420 nm and its quoted PDE (43–54%) is specified there. Emission at ~350 nm sits on the falling shoulder of that curve, so absolute sensitivity is below spec — though this does not affect the temperature coefficient, which depends on breakdown voltage rather than on PDE magnitude.

Detector: datasheet against measurement

The datasheet's gain tempco is conditional on operating point: it is dV_bd/dT ÷ V_ov, so it doubles at half the overvoltage. Taking the 3 V figure at face value disagreed with measurement by 2×; using the measured 1.18 V it agrees to ~25%, the remainder being gain non-linearity near breakdown plus smaller PDE and crosstalk terms. Two consequences: running at 1.18 V makes the detector roughly 2.5× more temperature-sensitive than it would be at 3 V — a firmware setting, not a physics limit — and a V_bd spread of 4× the specification is why per-sensor calibration is unavoidable.

The correction chain

1. Sweep → amplitude. The reading is a product, mon2 = dark + G(V_ov)·L(LED)·Φ·η, so subtracting the measured dark and taking logs makes a sweep additive: y = aLED + bovervoltage + c. Fitting a free offset per LED drive assumes no linearity of the emitter; keying the detector shape to overvoltage rather than raw bias follows the physics and was confirmed by the data (scanning the shift minimises scatter at exactly the datasheet's 2.6 counts/K). Clipped cells enter as censored bounds, not discards. Because every cell independently estimates the same c, the estimate cannot move when one is lost — which is what fixed a sensor that was alternating between two discrete values.
2. Out-of-water screening. A sensor clear of the water reads a higher ToF signal-per-SPAD. Flagged against the unit's own trailing 45 min median (>1.15×) and against the fleet at the same instant (>1.3×); both are needed, because a trailing baseline cannot see a sustained excursion and a fleet comparison cannot see a brief one. ~1% of rows, dropped before any fit reads them.
3. Temperature: a single exponential, ρ ≈ 2 %/°C, not a polynomial. d(ln signal)/dT is constant at −1.87 %/°C from 3 to 39 °C, varying ±3% across the interior — there is no curvature for a cubic to fit. It is also what the physics predicts: LED output and fluorescence quench are both exponential in temperature and they multiply. Held out by temperature (train on the middle, predict the ends, which is what a deployment does): exponential 1.28%, quadratic 1.81%, cubic 2.17%, quintic 15.03%.
4. Every temperature counts once, not once per second. The bath holds one temperature for hours and crosses the rest in minutes, so a unit's training samples run 929 in the 5 °C bin against 55 at 20 °C. Ordinary least squares weights samples, which silently hands the slope to whichever temperature the bath sat at longest, and ρ came out steep at 2.51 %/°C. Weighting by inverse temperature density instead — each 1 °C bin carrying equal total weight — gives ρ = 2.29 and cuts the clean-window residual from 3.175% to 1.354% on identical data, model form and training span. It improved 16 of 20 units; three of the four that did not were already under 1%.
5. Thermal lag, a dT/dt term, measured under that weighting: 1.730% → 1.356% on a clean window. Under sample weighting the same term bought nothing (3.147% → 3.176%) — it could not show its value while the fit was being dominated by an isothermal hold, which is exactly where a lag term has nothing to say.
6. Drift: currently not applied. It is fitted and reported, not emitted. On a clean window a log-time drift term made things worse, and the metric that says otherwise is computed on a window containing a handling step — where a flexible term will always look good. Per-unit step offsets at automatically detected changepoints were also tried, and were worse again (1.611% vs 0.961%): the detector finds 1–5 segments per unit and mostly fits noise. The handling log, not a more flexible model, is what closes this.

What is still open

Removing the detector term leaves 1.85 %/°C for LED output and tryptophan quench combined. Splitting them needs the CUD7GF1B optical-power-versus-temperature curve, which is not in the BOM or the repo — so the fluorescence coefficient, the one that actually belongs in a concentration model, is currently absorbed into ρ rather than measured. Separately, the handling steps of 2026-08-13 are not modelled: they are per-sensor multiplicative changes in optical throughput, invisible to every diagnostic channel, and no temperature or drift correction can remove them.