Three changes to ForecastHead and the conformal calibrator, each measured
walk-forward on real data over seven train splits.
Hedge normalised its losses by the current sample's worst loss, so on a quiet
step where every member agreed to within 0.01 C whichever happened to be worst
still took the full exp(-eta) penalty, exactly as if it had been wrong by 5 C.
The regret bound assumes a fixed loss range, not a per-sample one, and the
symptom was weights that jumped around with no relation to horizon. Normalising
by the running member MAE instead is worth 1.81% of MAE, better on 106 of 126
heads, coverage unchanged.
The residual handed to the conformal calibrator was computed after this
sample's loss had already moved the weights, so it was better than anything the
forecaster could produce and the intervals were calibrated about 2% too narrow.
Coverage survived only because ACI notices the extra misses and reopens the
band, a correction that should never have been needed. Scoring the blend with
the pre-update weights leaves MAE untouched, as it must, and widens the
intervals 2% to the honest width.
The conformal quantile refused to produce a band below 20 scores. That number
is arbitrary: the (1-alpha) empirical quantile is the ceil((k+1)(1-alpha))-th
of k order statistics, so alpha = 0.10 needs 9. The 20 became actively harmful
in the previous commit but one, because striding pairs by the horizon leaves a
long-horizon head about 13 scores per refit. Twelve of eighteen heads therefore
fell through to 1.645*sigma with sigma from an unconstrained x'Px, giving bands
of +/- 45 C and +/- 115% relative humidity on a young station. Those cover, by
being absurd, which is why the backtest never flagged them: a long walk-forward
passes 20 scores early and never looks back. After the change all eighteen
heads have a band from the first fit, +/- 3.1 C and +/- 7.7% in the same place.
Also measured and deliberately not done: adding the Kalman level variance to
the predictive spread. It moves sigma by 0.06% at the shortest horizon and
0.00% everywhere else, so the plumbing to carry it through three files buys
nothing.
The companion to the Kalman fix. RecursiveLeastSquares.from_dict and
AdaptiveConformal.from_dict restore lambda, delta, alpha, gamma and the
conformal window alongside their data, and load_dict replaces the
config-built heads with those, so every one of those knobs was immutable on
any station that already had state. Editing config.yaml and restarting looks
exactly like a change with no effect, which is the failure mode that cost real
time on the Kalman side of this before it was found.
Only the estimate is state now. Weights, covariances and conformal scores are
restored; tuning is re-applied from config on every load. The conformal deques
are rebuilt when the configured window changes, preserving their contents.
load_dict also skips heads for a target or horizon this build no longer has,
rather than resurrecting them from a stale file.
Found while implementing a damped-trend ensemble member, which was then
abandoned: see the following note.
q and r were written into station_state.json and restored over the configured
values, so tuning was effectively immutable in the field. This was found the
expensive way: the retune in the previous commit was deployed, the service
restarted cleanly, and the filters carried on with q = 2e-6 because that is
what the state file said. Measured median rate afterwards was 14.4 C/h against
12.4 before, which is to say nothing happened.
Only the estimate is state. x, P and initialised are restored; q and r now
come from config every time. P may be momentarily inconsistent with a changed
q, which costs a few hundred samples of reconvergence and is far cheaper than
a configuration change that appears to work and does not.
load_dict also now skips filters this build no longer has, rather than
resurrecting them from an older state file.
Two changes to the same signal path, one large and one small.
The large one: all three filters were tuned to track one to three decades
faster than their signals move. In a still room the temperature filter
reported a median rate of 12.4 C/h while the air moved 0.4 C/h, and it
overshot a real -36 C/h event by 77%. Sweeping q against the RMSE of the
reported rate versus the true rate, using noise measured on the board
(temperature 0.088 C, pressure 0.022 hPa, humidity 0.40 %):
temperature 6.45 -> 0.37 C/h RMSE 2e-6 -> 1e-9
pressure 2.15 -> 0.24 hPa/h RMSE 1e-5 -> 1e-8
humidity 27.94 -> 3.55 %/h RMSE 5e-5 -> 2e-8
Tracking does not suffer. Lag against a genuine 2 C/h ramp is 0.003 C at both
the old and new values, and the peak response to a five-minute event moves
closer to the truth rather than further from it, because the overshoot goes
away. What is given up is response to sub-minute transients, which for a
station forecasting fifteen minutes to a day ahead is noise to reject.
This matters most for pressure, whose tendency drives the precipitation
forecast, and which was the worst tuned of the three.
config.yaml shadowed kalman_q_temp, so editing the dataclass alone changed
nothing. All six values are now listed there with that hazard spelled out,
because a silent shadow cost real time here.
The small one: temp_raw was the plain average of two thermometers whose
white-noise sds differ by 7x (LPS25HB 0.007 C, HTS221 0.049 C), which throws
the quiet one away. Inverse-variance weighting cuts the raw noise 3.5x.
The trap is that the chips do not agree. They sit at different distances from
the SoC and stand about 1.3 C apart, so weighting by variance alone drags
temp_raw 0.48 C onto the LPS25HB, which after the 1.55x gain of the inverse
compensator is 0.75 C of silent bias on every reading, since k was fitted
against the mean of the two. The gradient is therefore tracked and removed
before weighting and only the deviations are fused: measured mean shift
0.0001 C, noise still 3.5x lower. The tracked gradient is retained because it
is a second observation of self-heating.
Also corrected: the earlier claim that the HTS221 was the quieter channel was
wrong, taken from twelve samples at a cadence slow enough that real drift
dominated. At 0.5 s over 120 samples the LPS25HB is quieter by 7x and takes
98% of the weight.
fit() trained every head on every consecutive grid row. At the 1 d horizon on
a 5-minute grid adjacent pairs share 287 of their 288 samples, so the filter
was handed the same outcome 288 times and RLS with forgetting read each one as
fresh evidence:
horizon steps overlap independent events in a 400-score window
15m 3 66.7% 133.3
1h 12 91.7% 33.3
3h 36 97.2% 11.1
6h 72 98.6% 5.6
12h 144 99.3% 2.8
1d 288 99.7% 1.4
The day-ahead head was therefore fitted on roughly two independent outcomes by
a filter carrying 667 updates of memory, and its interval was a 90th percentile
of a sample of size one.
This is not a compute shortcut that trades accuracy for speed. Measured
walk-forward on four days of real station data and averaged over five train
splits, striding improves every horizon past fifteen minutes:
15m +0.6% 1h -12.2% 3h -31.7% 6h -33.3% 12h -39.5% 1d -14.4%
with coverage unchanged at 87 to 92%, and the fit 11.6x faster. The redundancy
was not merely wasted work, it was collapsing P onto the one direction the
repeated sample excited.
The stride phase rotates each refit and is persisted, so a long-lived station
eventually trains on every offset rather than seeing one sample in 288 forever,
and a restart does not pin it to phase 0. A floor relaxes the stride when a
long horizon on a short record would otherwise yield one or two pairs; 12 was
chosen by sweeping it across five splits rather than picked.
Single-split runs showed 10 to 17% regressions at the 1 d horizon that moved
with the parameter. Averaging over five splits removed them, which is the
expected result for a head fitted and scored on under two independent
outcomes. That horizon cannot be evaluated on a four-day record and was not
tuned against.
Incidentally, this also retires the parallel-retrain idea: the Pi's 42 s
retrain becomes a few seconds, and multiprocessing inside a 280 MB cap buys
nothing for a job that short.
The board carries two independent thermometers and the code averaged them
into temp_raw without ever recording either. Measured over 12 samples on a
real station: HTS221 30.973 C at sd 0.060, LPS25HB 29.810 C at sd 0.443, a
standing gradient of 1.163 C with the SoC at 44.55 C.
Two things follow from that and neither is possible without the raw channels.
A plain average of a quiet sensor and one seven times noisier lands at sd
0.223 where inverse-variance weighting reaches 0.060, and the gradient between
two chips at different distances from the SoC is a second observation of
self-heating that could identify the compensator's k with no reference
thermometer. Both need history, and history cannot be backfilled, so the
columns land on their own ahead of the work that consumes them.
CREATE TABLE IF NOT EXISTS is a no-op against a table that already exists, so
adding to COLUMNS would have reached a fresh install and silently missed every
station already running, then surfaced as an OperationalError inside
insert_telemetry. That sits on the sample loop, so it takes a station down
rather than leaving a gap. Store now reconciles the table against COLUMNS on
open, which makes every future column addition safe rather than just this one.
The simulator gains the same two channels, with couplings solved so their
forward models average to exactly the k = 0.55 the compensator is tuned
against. Aggregate behaviour is unchanged; only the per-channel detail is new.
Simulated temp_raw noise does rise from 0.05 to 0.223, which is not a
regression but the end of an over-optimistic figure: it was modelling the
quiet sensor and calling it the average.
Reported from a real station after 1.5 days: a six hour temperature forecast of
53 C in a 24 C room, and 9 C at one day, both carrying a plus or minus of 0.43.
Confidently wrong is the one failure this project is supposed to refuse.
Root cause. fit() replayed history into the live RLS on every retrain tick and
never reset, so 453 grid rows had produced 64,676 updates in a day and a half.
RLS with forgetting reads every update as fresh evidence, so the model believed
it had a hundred times the data it had: P collapsed, in-sample error looked
excellent, and the weights drifted without bound in directions the data never
excited. Measured: cond(P) 3.1e9 and ||theta|| 1680 against a median |theta| of
1.67. A refit now starts from the prior, which makes retraining idempotent.
Across 25 refits on the real data ||theta|| holds at 11.35, drifting 0.03, where
before it grew without limit.
The two largest weights were sin_doy and cos_doy at +1174 and +1191. Annual
harmonics were in the design matrix from the first sample, where they are
near-constant, near-collinear with each other and with the bias, and a
rank-deficient regressor is what RLS answers with enormous cancelling weights.
They are now held at zero until the record spans the same 120 days the
climatology fit already requires, because a day and a half of data says nothing
whatsoever about the season.
Also raised the standardiser's variance floor from 1e-8, which only caught a
bit-exactly constant column, to 1e-3. A feature that merely barely moves was
being divided by its own noise.
The conformal calibrators and Hedge weights are deliberately not reset by a
refit: those are earned from scored forecasts, not from this regression.
Backtest unchanged within noise, coverage still 89 to 91 across all 18 heads.
Four regression tests added, including that refitting the same history twice
must give the same model.
I claimed 20/20 clean and was over-stating it: that was a fixed list of
viewports, and a fixed list cannot cover a user's zoom level, a larger default
font, or simply more accumulated history than a panel was designed around. All
three change how much space a card needs. I could not reproduce the reported
clipping at any width or height I tried, so rather than keep guessing at
viewports I made the failure impossible by construction.
The page now measures itself after every layout change and on resize: if any
card's content escapes it, density escalates to compact, and if that still is
not enough the height lock is released and the page scrolls. Content is never
silently clipped, whatever the viewport. Scrolling beats hiding.
The escalation is exercised, not dead code: across 34 viewports it used compact
on 3 and scroll on 2, and at 125% zoom it used both.
Also added Cache-Control: no-cache to the dashboard route. The page is generated
from live config and changes every deploy, but carried no cache headers, so a
browser could hold an old copy indefinitely and show layout bugs that were
already fixed. That is a plausible reason a fix can look like it did not land.
Verified: 34 viewports from 1920x1080 to 320x568, six tabs each, 204 tab
renders, zero card overflow and zero horizontal overflow; plus five zoom levels
from 100% to 200%.
Reported on a 13 inch screen: the Detectors column and the Learner bank spilled
past their cards. My earlier sweep called those tabs clean because it only
checked content escaping the viewport, never content escaping its own card, and
the cards are overflow-hidden so it was invisible rather than obviously broken.
The worst was not either of those. The Station log was discarding up to 299 px
of entries behind overflow-hidden with no indication, at every viewport. It now
binary-searches the number of rows that fit and says how many are not shown.
The count is searched rather than divided out of an assumed row height, because
a long detail line wraps to two lines and no constant knows that. It runs from
a ResizeObserver rather than once: measuring at call time reads a stale
clientHeight, since the pane has only just become visible and Chart.js resizes
its siblings 40 ms later, which fitted 13 rows into a box that holds 10.
Also: denser learner and attribution rows on short screens, three precipitation
coefficients instead of four, and a lower floor on the tendency chart.
Two mistakes of mine on the way, both worth recording. I declared a second
const rows inside loadModels, which is a parse error that killed the entire
dashboard script; the audit reported it as a flood of sparkline layout faults
because it was not listening for page errors, and it now fails loudly on them.
And the min-height override did nothing at first because the Tailwind CDN
injects its sheet after this style block, so min-h-[42px] won at equal
specificity.
Verified: 20 theme x viewport combinations, 120 tab renders, from 1920x1080 to
375x667 in both themes. Zero card overflow, zero horizontal overflow, zero
console errors.
Auto (follow the system), light, or dark. Auto is a real preference rather than
the absence of one, so it tracks prefers-color-scheme live rather than only at
load. Stored in localStorage, not the station's settings overlay: a theme
belongs to the screen you are reading on, not to the weather station. Control in
the header for one click and in Settings for the explicit three-way choice.
The markup is dark-first Tailwind utilities. Adding a dark: variant to several
hundred class attributes would have been a large diff with a large blast radius,
so light is an overlay that remaps the slate scale and the accent hues under
[data-theme=light]. The dark path is byte-identical: nothing is re-specified
unless the attribute is set.
Chart.js keeps its own copy of every colour and cannot follow a CSS class
change, so the charts read the same tokens from the computed style and are
re-rendered on switch.
Three things the screenshots and measurements caught that reading would not:
- Accent text is the 200-400 shades, picked to glow on black. On white they
wash out. Remapped to the 600-700 shade of the same hue so the colour coding
survives.
- My first remap listed exact class names and silently missed every opacity
variant, because text-amber-200/60 is a different class. Measured at 1.02:1.
Now matched on the hue prefix.
- The value-changed flash is a pale indigo that dropped a headline reading to
2.06:1 on white for half a second, which is precisely when you are looking at
it. Now theme-aware.
Contrast measured with alpha properly composited to the page ground: light 212
nodes below WCAG AA against dark's 208, so light is no worse than the dark
theme it mirrors. The remainder is the design's deliberately quiet tertiary
text, present in both. Layout re-verified: six tabs across three viewports in
both themes, twelve combinations, zero clipping and zero console errors.
Three real faults, all measured rather than guessed.
The forecast chart was unreadable on a 13 inch laptop. A 1280x800 screen has
desktop width but 280 fewer vertical pixels, and the fixed rows (stat cards,
outlook, diagnostics) took that entirely out of the one row meant to flex.
Measured: 404 px tall at 1920x1080, 124 px at 1280x800, 92 px at 1024x768.
Height-aware media queries now compact the furniture instead, and the chart is
201 px at 1280x800.
Mobile navigation did not exist. At 375 px all six tabs were clipped with no
affordance they were there, so five of six sections were unreachable. Below
1024 px the tab row is now a native select on its own full-width line, showing
the current section rather than a bare chevron.
Horizontal overflow on phones: the shell measured 452 px inside a 375 px
viewport. Flex and grid children default to min-width:auto, so a long subtitle
refused to shrink and dragged the layout wider than the screen.
Also restructured the conditions column. It was a fixed stack that simply
overflowed its own card on short screens, and shrinking each piece by media
query chased the symptom; it is now a flex column with exactly one flexible
child, so the tendency chart absorbs the slack and the column fits at any
height. And the no-scroll contract gained a height floor: below 700 px tall the
page scrolls, because at 1024x600 the chart had collapsed to 1 px and a
scrollbar is the better answer.
Verified across ten viewports from 1920x1080 to 375x667, six tabs each: 60
combinations, zero clipping, zero overflow, zero console errors.
A room held at a setpoint is a different process from one left to drift. It is
a closed loop, and persistence, the baseline everything here is scored against,
is the wrong statement about it: the truth is not that it stays where it is, it
is that it returns to the setpoint.
So site.heating adds a fourth ensemble member, first order because that is what
a controlled system is:
dT_set(h) = (T_set - T_now) * (1 - exp(-h / tau))
Humidity follows and is the part that is easy to get wrong. Heating adds no
moisture, so vapour pressure is conserved and not relative humidity:
RH(h) = RH_now * es(T_now) / es(T_now + dT_set(h))
Warm the air and RH falls although nothing was dried, which is why a heated
house in winter is dry. The test asserts the dew point is unchanged to 1e-6.
Pressure gets zero: a thermostat cannot move the synoptic field.
Offered, not imposed. Hedge scores this member on realised error like any
other, so a wrong tau or a stale setpoint costs accuracy and gets down-weighted
rather than quietly biasing every forecast. Verified: on history with no
heating the ensemble assigned it weight 0.000. With heating off it returns zero
and is identical to persistence.
Going from three members to four means old saved heads must migrate.
from_dict reinitialises weights and member_mae. I missed member_mae first time
and it did not fail on load, it failed later inside learn() on a broadcast
error, which is a much worse place to find out; the migration test now covers
both and calls learn() to prove it.
Settings tab gains the toggle, setpoint and time constant. Turning heating on
or off is treated as a regime change like a door: discontinuity marker plus a
queued retrain.
Everything that was previously a curl command now has a surface: surroundings,
site geometry, the matrix, the psychrometric flag, and the maintenance actions.
Changes persist to data/state/settings.json, not config.yaml. That file is
hand-annotated and hand-edited per station, and rewriting it from an API would
destroy the comments and risk clobbering something the owner set. The overlay is
applied last in load_config, so a dashboard change beats both the file and the
environment, and deleting the overlay reverts everything. Written atomically via
a temp file so a crash cannot truncate it.
Every field applies live. A settings page that needs a restart is one people
stop trusting, so site geometry is re-read per sample, the compensator flag is
set on the live object, and the display picks up its rate the next frame.
Two deliberate frictions. Selecting a surroundings pill only stages it: nothing
is recorded until you press the button, because that writes a discontinuity
marker and queues a retrain. And changing altitude or the psychrometric flag
says outright that the stored history is now inconsistent and offers the
re-derive, rather than leaving a silent mismatch.
Verified in a browser: pills stage and apply, the toggle round-trips, re-derive
ran 6201 rows in 0.29 s from the button, all six tabs report zero scrollbars and
zero clipping, zero console errors.
recompute replayed the Kalman over stored rows at their own spacing while q
stays tuned for the live 2 s cadence. Q scales with dt^3, so at the 30 s
persist interval the process noise was 3375x too large and the filter tracked
noise instead of smoothing: it wrote indoor temperature rates of +/-20 C/h into
the history. This is the exact trap DESIGN.md section 2 documents for
simulate.py, which does scale q, and I walked into it anyway. Now rescaled per
step, because tiering means the stored cadence is not constant. Mean |rate| on
the real board dropped to 2.73 C/h; what remains above 10 is the filter's
warm-up transient in the first four samples, which is honest.
Readout scene puts the actual numbers between the animations: temperature,
humidity, sea-level pressure and the signed three hour forecast, each in its
channel colour, scrolling. Text is drawn whole-pixel on purpose. Everything
else here is sub-pixel and that is what makes it look good, but splitting a
3 px glyph across two columns halves its peak and smears it illegible. Crisp
beats smooth when the thing has to be read.
site.environment and site.enclosure record where the sensor lives and what has
changed around it, with POST /api/environment to change them at runtime. This
is not cosmetic: closing a door changes how strongly the sensor couples to
outside, which is a regime change in the process the heads are fitting, and at
lambda 0.9985 they carry about 55 hours of memory. Left alone they keep
predicting the old room for two days. Page-Hinkley would notice eventually but
needs matured forecasts to do it, which at the long horizons is the same two
days. So the endpoint marks a discontinuity and queues a retrain.
Two bugs, both found by checking the real board rather than the test harness.
The glyphs never showed. _pick_glyph gated each one behind narrow conditions
and returned None otherwise, so on the Pi's actual state (27.3 C, rain
probability 0.024, condition settled, solar elevation -20.3) nothing qualified
and the panel silently fell back to the ambient scenes every time. A forecast
symbol is the default, not an exception, so it now always returns one of the
three: cold wins, then wet, then fair.
Day and night were inverted. night = _smoothstep(2.0, -8.0, elev) passes a
descending range, and _smoothstep treated edge1 <= edge0 as a degenerate step
returning the opposite of the intent, so the panel drew a moon at midday and a
sun at midnight. Caught by rendering it and looking, not by reading it.
_smoothstep now handles descending ranges, and only the degenerate equal-edge
case takes the step branch.
The fair-weather glyph also needed to survive after dark or it vanishes for
half of every day, which is how it went missing in the first place. Same
geometry, cool palette, rays drawn in to a halo.
Three references were requested as 8x8 animations. Copying their frames does
not work and I measured it rather than asserting it: at 8x8 the sun is a
2025:1 area reduction and its rays vanish, the umbrella loses canopy and
handle, and the snowflake averages into the background. Downsampled they move
0.0037, 0.0175 and 0.0027 per frame against 0.0177 for the aurora already on
the panel, so frame-copying would have been a downgrade. The sun source is
only 3 frames and the umbrella 4. These are hand-drawn at 8x8 instead, taking
the palette and subject from the references, which also keeps three artists'
frames out of an Apache-2.0 repo.
Transitions are now the data. _pick_glyph reads rain probability, Kalman
temperature, solar elevation and cloud index and selects sun, umbrella or
snowflake; a change preempts whatever is on screen and crossfades immediately,
so the panel dissolves because the weather moved, not because a timer expired.
Between changes the informational scenes still rotate. Verified switching live:
sunny -> sun, rain forecast -> umbrella, temperature to 0.4 C -> snowflake,
clearing -> sun.
Getting them to read took two failed passes, both recorded in comments. First
version blew the canopy to white and fused the snowflake into a blob, because
seventeen arc samples over ten pixels overlap 1.7 deep. Dropping alpha made
them muddy instead. The fix was sampling density, not brightness.
Profiled again since these share the board: the glyphs first cost 11 to 13% of
a core. Making plot() write scalar components rather than a 3-vector slice, and
expressing the sun's eight-fold rays as one angular field instead of 56 splats,
took the sun from 275 to 43 us and the worst scene overall from 13.2% to 8.0%.
24 fps costs about 11% of one core on a Zero 2 W, measured on the board. That
is a reasonable default for something you look at, but it is a decorative load
sharing a 512 MB machine with the forecaster, so it should be the owner's
choice. server.led_fps is clamped to 4..30.
Particle fall speed now divides by the configured rate rather than the module
constant, so rain falls at the same real-world speed whatever the frame rate,
instead of slowing down when you turn the frame rate down.
The old display drew static glyphs, held them, and cut to the next, which looks
like a microwave clock. This is a continuous 24 fps renderer.
Three things do most of the work. Gamma, because LED duty cycle is linear and
perception is not, so ungamma'd gradients band and dim colours vanish.
Sub-pixel rendering, so a dot at x=3.4 lights two pixels and motion glides
rather than steps. Crossfades, so scenes dissolve over 1.3 s and nothing ever
cuts.
Added temporal dithering after finding the framebuffer is RGB565: 32 levels of
red and blue, which after gamma leaves very few steps exactly where an aurora
and a star field live. A Bayer pattern rotated each frame alternates between
adjacent hardware levels, measured landing on 1.75, 4.31 and 8.06 where the
panel can only display integers. The panel is also dimmed by measured lux on a
log curve, so at night it is a glow rather than a searchlight.
Five scenes, each a reading rather than decoration. Aurora: hue is temperature,
curtain drift direction is pressure tendency, contrast is humidity. Solar sky:
sun at its true azimuth and elevation over a dawn/day/dusk gradient, becoming a
twinkling star field and moon after sunset. Precipitation: drop count from rain
probability, snow below 1.5 C with sideways sway, lightning with exponential
afterglow when stormy. Forecast ribbon: six horizons scrolling, height is the
predicted delta, pale caps are the conformal half-width so uncertainty is
visible. Barometer: a breathing ring whose period is the tendency.
Profiled because it shares a 512 MB board with the station. The first ribbon
cost 330 us a frame, about 16% of a core scaled to a Zero 2 W; vectorising it
into fields rather than 84 sub-pixel splats brought the worst scene to 7.6%.
Verified 23.6 fps sustained with zero malformed frames.
The tab bar was its own grid row costing about 70 px of vertical space on
every tab to hold five buttons, which is a poor trade on a layout that refuses
to scroll. The tablist now sits in the header between the title and the status
block, so the shell drops from three rows to two.
That space goes to the pressure tendency chart, which was 56 px and had no
room for Chart.js to lay out its tick row: measured, the caption sat 2 px below
its own card. The chart is now 80 px with explicit layout padding, and the
column has 11 px of slack instead of overflowing.
Also made the tendency x-axis adaptive. Hour-only labels collapsed to three
identical ticks on a short window, which is what a young station always has.
Below a six hour span the label now carries minutes.
Verified at 1600x900: Live, History, Models and Nerd all zero scrollbars, zero
clipping, zero console errors. Tabs stay reachable and the header stays one row
at 1280 and 1024 wide. Methods improved incidentally, from 68 px of overflow
to 4.
1. POST /api/recompute re-derives every compensated column from the untouched
raw values, removing the step a calibration otherwise leaves through the
history. Possible because temp_raw, cpu_temp and hum are never overwritten.
Idempotent by construction and tested per row: 0 of 6051 rows change on a
second run. 6069 rows in 0.25 s here, so a few seconds on the Pi.
2. Calibration now emits a 'discontinuity' event alongside the calibration log,
so downstream views can find the boundary without parsing prose.
3. Vendored Tailwind, Chart.js, hammer, the zoom plugin, KaTeX with its 20
woff2 faces, and both Google fonts into ashvale/static, served by the
station. 1.4 MB. Verified with every non-localhost request aborted in the
browser: zero external requests, equations still render, fonts still load.
The dashboard no longer needs internet.
4. 54 pytest cases over the pure numerics: physics closed forms and round
trips, both compensator inverse properties, the Kalman covariance
invariants and NIS consistency, the RLS trace cap under a deliberately
unexcited regressor, conformal coverage, and the Zambretti ordering. Wired
into CI after the seed step so the recompute cases have history. Writing
them caught my own sign error on the conformal update: a hit raises alpha
and narrows the band, which reads backwards until you follow it through.
5. Stats for Nerds gains the condition number of each head's covariance, a
standardised innovation histogram per Kalman filter from a bounded 600
sample ring buffer, and a reliability strip of realised against nominal
coverage. All arithmetic on data already in memory.
6. OutdoorProbe reads a DS18B20 over the kernel 1-Wire driver, no new
dependency. Polled on its own slower cadence because the sensor blocks for
up to 750 ms during conversion, which would eat a third of the 2 s sample
budget. Rejects the 85000 power-on sentinel and out-of-range values, and
reports age so a dead probe cannot masquerade as fresh.
Shutdown hang, the cause of every restart taking systemd's full 90 s timeout
and ending in SIGKILL: /api/stream looped forever with no disconnect or
shutdown check, so an open dashboard was an in-flight request that never
completed and uvicorn's graceful shutdown waited on it. Reproduced cleanly:
with no stream client the process stopped in 2 s, with one open client it was
still running after 15 s. Fixed by bounding timeout_graceful_shutdown, and by
having the generator exit on client disconnect and on a shutdown event. Now
7 s with a client attached.
Equations rendered as blank boxes whenever the KaTeX CDN was unreachable,
which is a real case for a Pi on wifi. The elements carried the TeX only in a
data attribute, so with no KaTeX there was nothing to display. The source is
now written into the element as text first and KaTeX replaces it, so it
degrades to readable TeX. Verified by aborting the katex request. A window
load handler re-runs typesetting for a slow CDN. The code comment claiming
this fallback already existed was wrong and is corrected.
New Stats for Nerds tab over a new read-only /api/nerd endpoint: Kalman NIS
and covariance per signal, both compensators, all 18 RLS heads with trace(P)
against the cap, |theta|, EWMA RMSE, conformal alpha against target, realised
coverage and ensemble weights, plus per-head feature attribution over the 33
standardised weights, the Mahalanobis and Page-Hinkley detector state,
climatology harmonics and precipitation coefficients.
Methods overhaul: KaTeX now renders the equations. They were previously passed
through .replace(/[{}\\]/g,' '), which stripped every brace and backslash and
turned real mathematics into mush. Stages 2, 3, 5, 6 and 7 gained full
derivations (RLS normal equations and the trace cap, Joseph-form Kalman with
NIS, adaptive conformal with its coverage limit, ridge harmonic regression with
anomaly decay) and a per-symbol legend rendered inline.
Conditions ahead gains weather icons chosen from measured cloud index, solar
elevation and temperature rather than the barometric class alone, so a fine
barometer under overcast draws a cloud and after sunset draws a moon. Snow is
selected on temperature.
Seven day outlook moves to the top of Live, directly under the nav. Tab renamed
Models and Calibration.
Verified in Chromium at 1600x900: Live, History, Models and Nerd all report
zero scrollbars, zero clipping, no page scroll, zero console errors. Methods
keeps its documented prose scroller. Backtest numerically unchanged.
Seven day outlook moves from History to Live, which now runs four rows.
Conditions ahead tightened so the Live column no longer needs a scroller.
Adds HumidityCompensator: an additive RH offset estimated by one-step RLS from
a trusted hygrometer, clamped to +/-35%, persisted, exposed at
POST /api/calibrate/humidity and on the renamed Models and calibration tab.
It also implements the psychrometric term (RH moved from element temperature
onto air temperature via conserved vapour pressure) but leaves it OFF by
default. The thermal argument predicts a hot element reads low; measured
against a reference hygrometer this board read 75.4% where the truth was
50.4%, so it reads HIGH and that correction would push it the wrong way. When
the flag is enabled, simulate.py applies the exact inverse, per the
simulator/compensator trap in DESIGN.md section 2.
Models pane rebuilt: the scorecard is one column per target so all 18 heads
are visible, and no panel on the tab uses an internal scroller. Verified in
Chromium at 1600x900: Live, History and Models all report zero scrollbars,
zero clipping, no page scroll, zero console errors. Backtest is numerically
identical to the previous commit, confirming the humidity work is a no-op
while the flag is off.
The Live tab now carries the observed-and-forecast chart (a superset of the
old rolling window, which plotted the same observed series without the
prediction) and the precipitation panel. Estimator internals moves to Models,
beside the calibration input that sets the coefficient it reports. The seven
day outlook moves to History. Verified in Chromium at 1600x900: all four tabs
scrollHeight 900 against innerHeight 900, zero clipped elements outside
internal scrollers, zero console errors.
CONTRIBUTING.md now states plainly that this is a solo project: bug reports
welcome, pull requests unlikely to be merged, fork it instead. No other
developer was ever named anywhere in the repository.
kemal