Repository navigation
Geophysical aiding: gravity unit fix, config parity, measurement characterisation, GNSS realism - #417
Conversation
`earth::gravity_anomaly` returned m/s^2. Every consumer of it is in milligal:
the `z` variable of the NetCDF maps `GeoMap` loads, `--gravity-noise-std`,
`--gravity-bias`, and the `[geophysical]` section of a scenario file. 1 mGal is
1e-5 m/s^2, so the observation reaching `GravityMeasurement` was 1e5 times too
small and the innovation `z - h` was, to five significant figures, just `-h`:
the filter was told its predicted map value was wrong by exactly that value, at
every geophysical update.
Nothing caught it. A tens-of-milligal innovation against the default 100 mGal
noise standard deviation is a NIS of about 0.16, far under any gate, so the
runs completed and reported success while the gravity channel pulled the
solution around in proportion to the map rather than measuring anything.
The magnetic channel had the identical defect and was fixed separately
(`MICROTESLA_TO_NANOTESLA`). This is the same fix for gravity:
`earth::MGAL_PER_M_PER_S2`, applied at the one place the anomaly is formed.
Why the tests did not catch it, and what now does:
- `test_gravity_anomaly` asserted only `is_finite()`, which a value 1e5 out
still satisfies. Replaced with two tests that pin the scale: a 9.81 m/s^2
observation at 45 deg N must land in the hundreds of milligal, and a 1e-5
m/s^2 perturbation must move the anomaly by exactly 1.0.
- `geo_closed_loop.rs` left the track's `grav_*` at `Default`'s zero, which
made the observation `0 - 9.8017 = -9.80` -- a number that lands inside a
map generated around +/-40 and reads as a plausible milligal anomaly. It is
not one; it is the whole of normal gravity in the wrong unit, and the
fixture passed on that coincidence. The track now carries a real gravimeter
reading and the map is offset to the anomaly that reading has, mirroring
what the magnetic half of the same fixture already did.
- Gravity noise in those fixtures was 1.0 mGal against a +/-40 mGal map, so
every innovation was a 40-sigma event. Raised to 10.0, matching the 10 nT
the magnetic side uses against its own +/-40 nT map.
- Added `GRAVITY_BIAS_PLAUSIBLE_MGAL`, the twin of the existing
`MAGNETIC_BIAS_PLAUSIBLE_NT`. A units error does not stop the bias moving;
it makes the bias absorb the error confidently, so "the bias moved" and
"its variance fell" both still hold. Bounding the magnitude is what
separates a bias tracking an anomaly from one soaking up a scale factor.
Verified by reinstating the defect: four tests fail, including the new bound,
which reports the bias reaching 113 mGal.
Every gravity-aided result produced before this commit is invalid --
data/output/*/grav and data/output/*/both have to be regenerated, and any
comparison drawn from them re-derived.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
Two things, together because the first needs the second to have somewhere to
live and a gate to keep it working.
## `analyze geostats`
The noise the geophysical filters are configured with -- 100 mGal and 150 nT --
was never measured. It is a constant in `strapdown-geonav` repeated in every
`conf/*.toml`, and nothing in the toolchain could have told you whether it was
right. `geostats` measures it: for each trajectory it forms the anomaly the way
the filter forms it, samples the map at the GNSS position, and reduces the
residual.
It reports, per field:
- bias, as mean and median;
- spread, as a standard deviation and a robust 1.4826*MAD -- these residuals
are multimodal, so the plain one is not a usable sigma;
- the signal, meaning the variation of the *map* along the track, and the
resulting signal-to-noise ratio;
- the de-correlation length and the measurement interval that follows from it.
The residual is decomposed rather than reported as one number, because it holds
two things that belong in different places in the filter. What varies *within*
a recording is measurement noise and belongs in R. What varies *between*
recordings is a per-run offset -- the phone's and the vehicle's own field, for
the magnetic channel -- and the filter already has a state for that, so it
belongs in the map-bias prior. Collapsing them is what makes a channel look
unusable when it is merely mis-modelled. The generated figure shows this
directly: one narrow mode per recording, and the single-sigma normal fit
visibly failing to describe them.
De-correlation is taken from the *source* data, not the grid spacing, and the
difference is large:
- `earth_faa` is downloaded at 1 arc-minute, but over land it is filled from
EGM2008, whose terrestrial input is a 5 arc-minute area-mean grid. The
1 arc-minute cells are interpolation, so the real figure is ~9.3 km, not
1.9 km.
- `earth_wdmam` is a 3 arc-minute grid *at 5 km altitude*. Upward continuation
attenuates wavelengths below about 31 km, so a ground-level sensor and the
map are not measuring the same thing, and 31 km is the honest length.
At 24 m/s that is one independent gravity sample per 386 s and one magnetic per
1309 s, against the `geo_frequency_s = 1.0` every config uses. Also worth
noting: the magnetic grid is fetched at 3 arc-minutes while every config
declares `two_minutes`, and nothing validates that, because `GeoResolution` is
a filename token and no code derives metres from a grid.
The anomaly models mirror `core/src/earth.rs` and `geonav/src/lib.rs` term for
term -- a statistic computed from a differently-computed anomaly would describe
a quantity the filter never sees, which is worse than none, since it would be
used to set the filter's noise. `self_check()` asserts the two still agree,
against the reference values `geonav/tests/geo_closed_loop.rs` states.
Tested by planting a known bias and sigma into synthetic trajectories, inverting
the measurement models so the CSV carries raw sensor values, and requiring the
numbers back. Within-trajectory sigma recovers to within 10% (7.03 against 7.0
planted, 25.12 against 25.0).
## Monorepo
`analysis/` is checked in, so it needs a build and a gate:
- a virtual uv workspace root at `pyproject.toml`, so `uv sync` and
`uv run analyze ...` work from the repository root. The justfile's 35
`uv run --project analysis` invocations lose the flag.
- ruff and pytest configuration, and `.github/workflows/python.yml` to run
them. Separate from rust.yml: a Python lint failure reported under "Test,
Lint, and Build Rust Project" is a check name that lies.
- `.gitignore` gains the Python artefacts. The first commit of the package
brought eight `.pyc` files with it; those are removed from the index.
- CLAUDE.md and AGENTS.md both asserted the repository is Rust-only, which a
stale instruction file turns into a misleading brief for the next session.
Fixing the lint findings surfaced two pre-existing ones worth naming: a dead
`all_csv` in `preprocess_data`, and a `zip(starts, ends)` whose `strict=True`
now catches a future divergence rather than silently truncating.
One thing the gate required: **pygmt is no longer imported at module scope.**
It dlopens the GMT C library on import, so importing `analysis` at all failed
wherever GMT is absent -- every subcommand, including ones that draw no maps.
CI installs no GMT on purpose, so the job is what keeps this true.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
…ould
`--geo` was reachable only from the `cl` subcommand's flags. A closed-loop
config carrying a `[geophysical]` section was refused outright, while the
particle-filter arm of the same function read maps and GNSS degradation from
one file. So the two filter families were configured in incompatible ways: the
RBPF took a `--config`, and the UKF and EKF took a twelve-flag command line
that had to be kept in step with `conf/*_degraded.toml` by hand, with a comment
in the justfile asking whoever edited one to remember the other.
Both paths now resolve to a `GeoClosedLoopSettings` and run
`run_geo_closed_loop_file`. Verified end to end: the `just ukf-geo` command
line and the equivalent config file produce byte-identical CSV.
Two defects found by making them agree, both of which meant the two paths were
not running the same filter:
- `--ukf-alpha` defaulted to the textbook `1e-3`. A config file defaults to
`0.1`, which is the value #399 measured and adopted -- alpha is a
numerical-conditioning parameter here, and `1e-3` costs six significant
digits per step to cancellation in the weighted sigma-point mean. The flag
was never moved with it, so `--filter ukf` and a config naming the same
filter ran with sigma-point spreads two orders of magnitude apart. Both now
take `sim::DEFAULT_UKF_ALPHA`.
- `--log-level` was discarded entirely on the `--config` path, under a
comment claiming it took precedence.
Also fixed along the way:
- `gravity_bias` and `magnetic_bias` were documented as config-path-inert and
silently dropped; they are read now. Added the four `*_bias_init_std` /
`*_bias_process_noise_std` keys the CLI already had, which matter most for
the magnetic channel -- a recording made inside a vehicle carries thousands
of nT of the vehicle's own field, and that state is what absorbs it.
- The `config` wizard offers the geophysical section after the filter has
been chosen, so `eskf` followed by `yes` wrote a file the binary refuses.
It reconciles now, which matters because `Eskf` is `FilterType`'s default.
- Validation moved ahead of the run. A missing map or an ESKF now fails on
the configuration rather than after loading maps, and the ESKF message says
that omitting `filter` lands there.
- `generate_plot` reached the non-geophysical config path only, so every
`conf/*.toml` asked for a plot and no geophysical run produced one.
The six `conf/{ukf,ekf}_{grav,mag,both}.toml` were retired-but-not-deleted and
carried `interval_s = 1.0`, `sigma_pos_m = 15.0`, `sigma_vel_mps = 5.0` and
`r_scale = 15.0` -- a different GNSS profile from the degraded run `geoperf-all`
scores them against, so any "improvement" it reported would have been the
difference between two degradations. Rewritten to match, with the health limits
the RBPF files already had. The justfile's ukf-geo/ekf-geo recipes are
`--config` calls now, so the two cannot drift again.
Tests:
- `example_configs` covers `conf/` as well as `examples/configs/`, which is
what would have caught those six. Its declared-variant heuristic now strips
comments first: the truth recipes say `kind = "none"` and explain in prose
that their limits match the degraded run, which read as a declared fault.
- Three unit tests in `sim`: the two paths resolve to identical settings, a
geophysical config with no map is refused, and one omitting `filter` is
refused for the ESKF.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
…edule
Answering "should we bump the fix interval, or the noise?" -- and fixing two
defects that made the question unanswerable as posed.
## The fix interval and the error model were the same knob
`rho_pos` and `rho_vel` are applied once per emitted fix with no reference to
elapsed time, so the error's correlation time is `-interval_s / ln(rho)`: a
property of the *scheduler*. At `rho_pos = 0.99`, a 5 s interval is a 498 s
correlation time and a 30 s interval is 2985 s. Sweeping the fix rate was
therefore also sweeping the error timescale, and no result could separate them.
Added optional `tau_pos_s` / `tau_vel_s`. When set, the coefficient becomes
`exp(-dt / tau)` for the actual interval between fixes and `sigma_pos_m` is
reinterpreted as the steady-state standard deviation in metres rather than the
per-step innovation -- the quantity worth specifying, and the one that stays
fixed. Verified on real data: two runs at 5 s and 30 s now differ in fix rate
alone.
Strictly additive. Omitted, the per-fix form applies exactly as before, so every
existing config and every recorded result still means what it meant.
`perf_baseline` does not move.
## The fault advanced on the wrong clock
`apply_fault` runs once per *emitted fix*, but `build_event_stream` handed it
the *record* spacing -- 1 s on a 1 Hz log. Under `FixedInterval { interval_s:
5.0 }` a `SlowBias` therefore integrated one second of drift per five seconds of
flight and spoofed at a fifth of its configured rate. `FaultState` now tracks
when it last ran, so the interval is the scheduler's. Computed at the call site
rather than inside `apply_fault`, so a `Combo`'s members all see the same
interval instead of the first one consuming it. No baseline uses `SlowBias`, so
nothing recorded moves.
## What a jamming scenario should be
The dominant, repeatedly measured effect of jamming is loss of fix, not graceful
degradation. At Jammertest the signal was in most runs strong enough to cause
complete GNSS signal loss, and a phone ten metres from a jammer lost lock
entirely. A receiver inside the jammed area does not report a noisier position;
it reports nothing. The existing `degraded` profile models the *fringe*.
So the answer is: yes to availability, and it is the bigger lever. New
`conf/*_denied.toml` -- 30 s of fixes in every 150, fault `none`, so the only
variable is availability. Sweep `off_s` over 60/120/300/600 to trace an aid's
contribution against outage length. The duty-cycle scheduler already existed and
was unused in the entire experiment matrix.
Secondarily yes to noise, anchored to measurement. New `conf/*_jammed.toml`
takes 35 m of steady-state wander -- about 14x this dataset's 2.58 m median
advertised accuracy -- from the published smartphone figures (3-5 m good
multipath, >10 m harsh, semi-urban CEP95 10-12 m) scaled for the satellite-count
and GDOP loss jamming causes. It uses the time constants above, so its fix
interval can be swept.
Deliberately NOT applied to the existing `conf/*_degraded.toml`. That 21 m
profile is the baseline every result so far was measured against, and moving it
would silently invalidate those comparisons. `geoperf-all` scores geo runs
against it. Switch to `jammed` deliberately, not by having the old one change
underfoot.
`book/src/gnss/fault-simulation.md` was a one-line stub; it now carries the
three-tier rationale, the correlation-time trap, the `r_scale`-squares-into-R
warning, and thirteen sources.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
`test_state_transition_jacobian_multiple_states` failed CI on this branch at a
max error of 5.0245e-4 against its 5e-4 bound -- on a commit that touches
nothing it covers. It is not this branch's failure: the test is untouched here
and is identical on main.
It is flaky by construction. `rand::rng()` is entropy-seeded, so every run drew
fresh states and a failure named coordinates nobody could reproduce. Measured
at 1 run in 400.
Two things were wrong, and both are fixed:
- **Unseeded.** `StdRng::seed_from_u64` now, which is also what the
repository's own rule asks for: tests are reproducible with a seeded RNG.
A failure now means the Jacobian changed, not that the draw landed in the
tail. Determinism also makes the sample size free, so this covers 50 states
rather than 10 -- coverage no longer varies from run to run.
- **Bounded inside its own sampling noise.** Over 2000 sampled states the
median error is 2.4e-4, p99 is 4.8e-4 and p99.9 is 4.9e-4; over 4000 the
largest seen was 5.05e-4. A 5e-4 bound therefore sat exactly on the
shoulder of the distribution -- it described where the tail happened to
fall, not the method's accuracy, and left no headroom for the libm
differences between the three CI platforms. Raised to 1e-3: twice the
largest error measured, and still far tighter than any real defect, since
an analytic Jacobian that disagrees with the numerical one does so at O(1)
or O(dt), not by a factor of two in the fourth decimal.
Verified: 20 consecutive runs pass where the previous version was 1-in-400.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
|
"Workspace (all features)" failed on That failure is not this branch's, and I've fixed the test rather than re-run it. Not this branch's: Flaky by construction, and measurably so. The test calls Two things were wrong, and I fixed both in
Verified: 20 consecutive runs pass where the previous version was 1-in-400. All six gates clean locally ( Worth noting this flake has been latent on Generated by Claude Code |
Every `strapdown-sim` invocation was already `--config` after the geophysical
config path landed, but the recipes downstream of the simulation were not
updated to match, and `pipeline` had never worked.
- **`pipeline` ran nothing.** Its steps were recipe *body* lines, which `just`
executes as shell commands -- so it tried to run `build`, `preprocess` and
the rest as programs and died on the first with "command not found". They
are dependencies now, listed after the colon. A recipe's body runs *after*
its dependencies, so the `rm -rf` cleanup could not stay in the body either;
it moved to its own `clean` recipe, which is the first dependency.
- **`postprocess` did not score the new tiers.** Six lines added, so `denied`
and `jammed` get the same treatment as `truth` and `degraded`.
- **`geoperf-denied` added.** Deliberately separate from `geoperf-all` rather
than folded into it: the two answer different questions. `degraded` still
delivers a fix every 5 s, so an aid there competes with GNSS and mostly
changes how the filter weights a noisy measurement. `denied` withholds GNSS
for 120 s at a time, so during an outage the aid is the only correction --
the regime it exists for, and where its contribution is separable. The
recipe's comment records that the six geo configs still carry the
`degraded` GNSS profile, so the comparison is not apples-to-apples until
they are switched over.
- **`pipeline` now includes `geo-stats`**, between preprocessing and the
simulations, which is where it belongs: it reads what preprocessing writes
and produces the noise figures the configs should be using.
Also fixed the discovery surface. `just --list` shows the comment line
immediately above a recipe, so the multi-paragraph rationale blocks were being
listed as sentence fragments -- `denied # contribution grows with outage
length`, `pipeline # their own clean recipe rather than staying here`. Each of
those now ends with a one-line summary, and `ukf-geo`/`ekf-geo` have a
description at all.
Verified with `just --list` and `just --dry-run pipeline`: the chain resolves,
all 21 conf files are reachable from a recipe, and every config a recipe names
exists.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
`preprocess` wrote to `data/input_10hz`. Nothing downstream reads that: all 21 conf/*.toml carry `input = "data/input"`, `geo-stats` takes `-i data/input`, and `postprocess` and the `geoperf-*` recipes take `-r data/input`. So `just pipeline` preprocessed into one directory and simulated from whatever stale data happened to be in the other -- and, since `pipeline` began with `rm -rf data/input`, from an empty one on a clean checkout. Keeping the 10 Hz rate and moving the output rather than the other way round: the rate is one flag, the path is 21 files plus four recipes. `preprocess-1hz` restores the 1 Hz variant that was commented out, since that is the rate every result before this branch used and the two want to stay comparable. `clean` still removes `data/input_10hz` -- nothing writes there now, but a checkout that ran the old recipe still has it, and a stale directory full of trajectories is the kind of thing that gets simulated by accident. Flagged rather than changed, in the recipe comment: `-b 10`. `--buffer` is a *fraction*, not a percentage -- `inflate_bounds` computes `x_min - x_range * buffer` and defaults to 0.1 for a 10% margin -- so `-b 10` pads the map bounding box by ten times the track's own extent on each side, a box roughly 21x wider and 21x taller than the track. At the default buffer the relief grid for the longest recording here is already 195 MB. It looks like `-f 10` copied one flag across, but it is a research parameter, so it is left as set with the cost written down next to it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
`bc4b02b` added a second `[dependency-groups]` table to
analysis/pyproject.toml. TOML forbids duplicate keys, so `uv sync --frozen`
refused the file outright and the Python CI job died before it ran anything:
error: Failed to parse: `analysis/pyproject.toml`
cause: TOML parse error at line 40, column 2
[dependency-groups]
duplicate key
Kept the table that carries `ruff`, since the lint gate needs it, and dropped
the one that did not.
Its `[tool.pytest.ini_options]` went with it. The root pyproject.toml owns
pytest configuration and is what `uv run pytest` reads from the workspace root,
which is how CI invokes it; a second copy in the member only applies when
pytest runs from inside analysis/, where it would silently lose the root's
`filterwarnings` -- two configs that drift rather than one that works.
Also ran `ruff format`, which the new files needed.
Verified: `uv sync --frozen`, `uv run ruff check`, `uv run ruff format --check`
and `uv run pytest -q` all clean -- 15 tests, including the four new
compare-filters ones.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
Ran it. Three things stopped it, two of them silent.
## `postprocess` produced no plots and exited 0
`read_csv(..., parse_dates=True, index_col=0)` does not guarantee a datetime
index. When pandas cannot infer a format it leaves the index as **strings**,
with no error. Every timestamp this package reads looks like
`2025-03-01 12:00:00+00:00`, which it declines to infer, so the index stayed
textual and the first arithmetic on it raised:
unsupported operand type(s) for -: 'str' and 'str'
`performance_analysis` caught that in a broad `except`, printed "possible
dimension mismatch or missing data", and carried on. So `just postprocess`
reported success while writing **zero plots** -- for every scenario, every
filter, every trajectory.
Added `read_timeseries`, which names the format instead of guessing and returns
a frame whose index is checked. Six call sites across `performance_analysis`,
`geophysical_performance_analysis` and `filter_comparison_analysis` use it.
`preprocess.py` was never affected: it already converts explicitly.
The same defect was in `geostats.analyse_trajectory`, which my own fixture
missed because it wrote timestamps in a format that happens to round-trip. It
parses explicitly now and raises if the result is not tz-aware, rather than
failing a hundred lines later.
Effect: postprocess goes from 0 plots to 99.
## `pipeline` could never produce the maps the geophysical runs need
`preprocess` ran without `--getmaps`, so it called `inherit_parent_maps`, which
copies `data/input/<source>_{gravity,magnetic}.nc` onto each split segment. But
`clean` deletes those first, so there was nothing to inherit: `data/input`
ended up with trajectories and no maps at all, and every geophysical run died
on "Gravity map file not found" three steps downstream of the cause.
`--getmaps` is passed now, and fails loudly at the point of failure.
## Verified
Ran the whole chain from an empty tree -- clean, build, preprocess, geo-stats,
truth, degraded, denied, jammed, ukf-geo, ekf-geo, rbpf-sim, postprocess,
geoperf-all -- against synthetic Sensor Logger recordings. All fifteen steps
complete, producing the full 3x7 matrix: 63 navigation solutions across
{ukf, ekf, rbpf} x {truth, degraded, denied, jammed, grav, mag, both}, 99
performance plots, 72 geoperformance artifacts and the geostats figure.
One step cannot run in a sandbox: `--getmaps` needs the GMT C library and
network access to the GMT data server. Locally generated grids stood in for it,
at the real spacings (1 arc-minute gravity, 3 arc-minute magnetic) so
everything downstream saw the geometry it will see in a real run -- `geo-stats`
correctly reported the magnetic grid as 3.00', which is the declared-versus-
actual mismatch it exists to catch.
Also ignoring `gmt.conf`: GMT drops one into whatever directory it runs in as
soon as anything calls `gmtset`, which pygmt does on some paths, and it records
machine-local settings.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
`geoperf-denied` landed earlier on this branch scoring the existing geo runs
against the `denied` baseline -- but those runs carry the *degraded* GNSS
profile (fixed 5 s interval, AR(1) error), while `denied` is a duty cycle with
no fault. So it compared two different GNSS profiles and attributed the
difference to the geophysical aid, which is the same defect this branch already
had to fix once for the degraded baseline.
Added `conf/{ukf,ekf,rbpf}_denied_{grav,mag,both}.toml`: the same aiding as
their `<filter>_<geo>.toml` counterparts, carrying the GNSS and health blocks
of `<filter>_denied.toml`. `geoperf-denied` now reads those, so the two runs it
compares differ in nothing but whether the geophysical measurement is present.
This is the comparison worth making. `degraded` still delivers a fix every 5 s,
so an aid there competes with GNSS and mostly changes how the filter weights a
noisy measurement. `denied` withholds GNSS for 120 s at a time, so during an
outage the aid is the only correction there is.
## The duplication is now enforced, not just written correctly once
Nine more files takes `conf/` from 21 to 30, and the number duplicating a GNSS
block from 12 to 24. Drift between exactly those duplicates is what went wrong
before: six geo configs sat in the tree at `sigma_pos_m = 15.0` against a
baseline at `3.0`, each scored against a profile it did not match, with nothing
to say so.
So `core/tests/example_configs.rs` gains a test asserting that every geo
recipe's `[gnss_degradation]` and `[health_limits]` are identical to those of
the baseline named in its filename -- `<filter>_<geo>.toml` against
`<filter>_degraded.toml`, `<filter>_denied_<geo>.toml` against
`<filter>_denied.toml`. The pairing comes from the filename, so a geo recipe
added later is enrolled automatically, and the count is asserted so one that is
not gets noticed.
Verified in both directions, the way the gravity unit fix was: re-injecting the
historical drift (`sigma_pos_m` 3.0 -> 15.0 in `ukf_both.toml`) fails the test
and names the differing field.
Considered and rejected: an `extends`/`include` key on `SimulationConfig` to
remove the duplication at source. Better long-term, but it is a new config
format feature with its own precedence rules, and this test removes the
urgency.
## Justfile
`denied-geo` runs the nine; `geoperf-denied` scores them; `postprocess` gains
their nine output directories; `pipeline` runs both new steps. All 30 configs
are reachable from a recipe and every config a recipe names exists.
## Also
`-b 10` -> `-b 0.1` on `preprocess`. `--buffer` is a fraction, not a
percentage, so `-b 10` padded each map's bounding box by ten times the track's
extent per side -- a box about 21x wider and taller. Harmless while `--getmaps`
was absent; this branch added it, and at a 10% margin the relief grid for the
longest recording is already 195 MB.
Verified end to end against synthetic recordings: all nine denial-geo configs
run to completion and `geoperf-denied` produces 54 artifacts. All gates clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
Three of the nine denial-geo configs failed on every trajectory with
longitude -75.699 is outside the map bounds [-76.099, -75.699]
The filter had drifted past the edge of the map during a 120 s outage. The
immediate cause was a test fixture's narrow pad, but it pointed at a property
of `inflate_bounds` that the denial tier is the first scenario to reach.
## The value was the smaller of two problems
x_range = x_max - x_min
return (x_min - x_range * buffer, x_max + x_range * buffer, ...)
1. **The thing being guarded against is an absolute distance.** A map has to
extend past the track by however far the filter can wander off it, and that
does not scale with how far the vehicle drove. At a 10% buffer a 2 km track
gets 200 m of margin and a 50 km track gets 5 km, for the same outage.
2. **Each axis is scaled by its own range**, so a track that is straight in one
axis gets almost no margin in the other. A due-north drive has a longitude
range near zero, so its longitude pad is near zero, and any eastward
excursion leaves the map. No value of `--buffer` fixes that: zero times
anything is zero.
Either way it is silent at the preprocessing end. The map is written
successfully and the run dies later, per trajectory, inside a simulation.
## `--margin-km`, defaulting to 5 km
`pad_bounds` takes the larger of the fractional and absolute margins, so long
tracks keep the behaviour they had and short or straight ones get a floor.
Converted at the box's mean latitude, since a degree of longitude shortens
towards the poles, with the cosine clamped so it does not diverge at one.
`--margin-km 0` is exactly the old behaviour.
The 5 km default is measured, not guessed. Across the fixture's runs the worst
excursion of the filter estimate beyond the recorded track's bounding box is:
truth 0.00 km
degraded 0.00 km
denied 2.27 km
Denial is the only tier that leaves the track at all, and 5 km covers its worst
case with better than 2x headroom. On a 17 km track that is about a 30% pad --
cheaper than raising `--buffer`, and unlike it, correct for a short or straight
track too.
This also explains `-b 10`, which this branch changed to `-b 0.1` a few commits
ago: a 21x box is an effective workaround for a margin that does not scale the
right way. It cost gigabytes per trajectory once `--getmaps` went live.
## Also
`StrapdownError::OutOfMapBounds` now names the remedy. A single skipped update
is routine -- an estimate near a tile edge, or a particle in the tail -- but
every update on a trajectory failing this way means the map does not cover
where the filter went, and the message now says to re-run with a larger
`--margin-km` rather than leaving that to be worked out.
Ten tests in `analysis/tests/test_map_bounds.py`, covering the two degenerate
cases the fractional form cannot express (due-north and due-east tracks), the
short-track floor, that a long track keeps its larger fractional margin, that
`--margin-km 0` is byte-identical to the old behaviour, that the margin is the
same *distance* at 0, 40, 70 and 85 degrees latitude, and that the conversion
stays finite at the pole.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
`analyze geostats` measured the per-field bias, measurement noise and bias prior against the maps, wrote them to `geo_stats.toml`, and stopped there. Merging eighteen files by hand is exactly where a partial adoption comes from, and the 100 mGal / 150 nT the configs ship with were never measured against these maps at all: a filter told a noise it does not have produces a covariance that stops describing its error. `--apply-to conf` (`just geo-adopt`) writes them in. Line by line rather than through a TOML round-trip, which would strip every comment in these files -- the rationale in `conf/*.toml` is most of their value. Only keys already present are rewritten, so a `*_grav.toml` stays gravity-only rather than silently becoming a dual-aid run. The one insertion is `*_bias_init_std`, which no config carries yet: without it the map-bias state inherits `*_noise_std` as its prior, which for the magnetic channel is far too tight, since a recording made inside a vehicle carries thousands of nT of the vehicle's own field. `geo_frequency_s` is held back behind `--apply-interval`. The noise figures are direct measurements of a residual; the interval follows from a de-correlation argument about the source grids, and moving it from 1 s to several hundred changes what the experiment asks of the aid rather than how it is tuned. That one is a decision. `every_geophysical_config_describes_the_same_sensor` in `core/tests/example_configs.rs` closes the loop: all eighteen recipes describe the same instrument, so any key they share must agree. Verified in both directions -- applying to all eighteen passes, reverting one file's `gravity_noise_std` fails. Otherwise a half-adopted `conf/` still parses, still runs, and produces a filter comparison whose difference came from two different R. Also fixes the `just --list` summaries for `clean` and `geo-stats`: `just` shows the last comment line, so the one-liner has to sit directly above the recipe. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
`e9d673d` reformatted the file at a wider column than `[tool.ruff]` in the root `pyproject.toml` declares, so `ruff format --check` failed the `analysis (lint + tests)` job. `line-length = 100` is the setting CI runs with; an editor's own default is not it. Formatting only -- verified by comparing the parsed AST before and after, which is identical. The ISO8601 date-parsing fix in that commit is untouched. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK
827db06 re-blessed core/tests/perf_baseline.json but left book/src/development/baseline-tables.md, which is generated from it, at b70cd55. `navigation_accuracy_matches_the_recorded_baseline` requires the two to agree, so it failed on all four Rust test jobs of #417 with "BOOK ... no longer matches the baseline it is generated from" -- no metric moved. The tables here are exactly what the test generates from the committed JSON; the JSON itself is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
Copilot review overview
🟡 Changes recommended
The real-arm configs target synthetic directories, and every real RBPF config contains keys that the new deserializer explicitly rejects.
Review effort: Balanced
Findings: 21
Open (27)
Equal-length frames can retain mismatched timestamps · New Real combined-aiding recipe uses synthetic data paths · New Real-arm paths overwrite synthetic EKF degraded output · New Real gravity recipe uses synthetic data paths · New Real magnetic recipe uses synthetic data paths · New Real-arm paths overwrite synthetic EKF truth output · New Real combined-aiding recipe cannot reproduce phone run · New RBPF combined config uses removed legacy keys · New Real RBPF degraded recipe uses synthetic data paths · New Config uses tombstoned particle filter keys · New Real gravity recipe cannot reproduce phone-sensor run · New RBPF gravity config uses removed legacy keys · New Real magnetic recipe cannot reproduce phone-sensor run · New RBPF magnetic config uses removed legacy keys · New Real RBPF truth recipe uses synthetic data paths · New Particle filter config uses removed legacy keys · New Real-arm recipe reads and overwrites synthetic data · New Real-arm recipe uses synthetic data paths · New Real-arm recipe reads and overwrites synthetic data · New Real-arm recipe reads and overwrites synthetic data · New
And 7 more that still need to be addressed.
What changed in this PR
This PR expands geophysical-aiding experiments with corrected units, realistic GNSS faults, reproducible Python analysis tooling, and revised RBPF/configuration behavior.
Changes:
- Adds the uv-based analysis workspace, Python CI, plotting/statistics tests, and monorepo documentation.
- Revises geophysical, GNSS-degradation, and RBPF configurations and regression tests.
- Updates navigation internals and removes obsolete queue documentation.
| File | Description |
|---|---|
README.md |
Documents monorepo setup. |
pyproject.toml |
Defines uv, Ruff, and pytest configuration. |
.github/workflows/python.yml |
Adds Python CI. |
.gitignore |
Adds Python/GMT exclusions. |
AGENTS.md |
Documents Python workflow. |
CLAUDE.md |
Updates architecture and analysis guidance. |
analysis/pyproject.toml |
Defines the analysis package. |
analysis/README.md |
Adds package README. |
analysis/src/analysis/py.typed |
Marks package as typed. |
analysis/tests/conftest.py |
Configures test imports. |
analysis/tests/test_map_bounds.py |
Tests absolute map margins. |
analysis/tests/test_performance_plots.py |
Tests sparse-reference plots. |
analysis/tests/test_summary_rows.py |
Tests summary aggregation. |
book/src/development/performance.md |
Updates RBPF consistency status. |
book/src/user-guide/overview.md |
Corrects RBPF architecture documentation. |
core/src/error.rs |
Adds stream-gap diagnostics. |
core/src/kalman.rs |
Shares IMU-bias coupling logic. |
core/src/lib.rs |
Exposes variance helper and updates tests. |
core/tests/config_serde_defaults.rs |
Tests defaults and retired keys. |
core/tests/filter_comparison.rs |
Handles RBPF’s zero bias rows. |
core/tests/integration_tests.rs |
Updates RBPF tuning and result layout. |
core/tests/perf_baseline.rs |
Updates RBPF baseline setup. |
core/tests/rbpf_depletion.rs |
Revises collapse/roughening regression tests. |
conf/ukf_truth.toml |
Defines synthetic UKF truth run. |
conf/ukf_degraded.toml |
Retunes UKF degraded GNSS. |
conf/ukf_grav.toml |
Configures synthetic gravity aiding. |
conf/ukf_mag.toml |
Configures synthetic magnetic aiding. |
conf/ukf_both.toml |
Configures combined synthetic aiding. |
conf/ekf_truth.toml |
Defines synthetic EKF truth run. |
conf/ekf_degraded.toml |
Retunes EKF degraded GNSS. |
conf/ekf_grav.toml |
Configures synthetic gravity aiding. |
conf/ekf_mag.toml |
Configures synthetic magnetic aiding. |
conf/ekf_both.toml |
Configures combined synthetic aiding. |
conf/rbpf_truth.toml |
Updates RBPF truth configuration. |
conf/rbpf_degraded.toml |
Updates RBPF degraded configuration. |
conf/real/ukf_truth.toml |
Adds real-data UKF truth recipe. |
conf/real/ukf_degraded.toml |
Adds real-data UKF degraded recipe. |
conf/real/ukf_grav.toml |
Adds real-data UKF gravity recipe. |
conf/real/ukf_mag.toml |
Adds real-data UKF magnetic recipe. |
conf/real/ukf_both.toml |
Adds combined real-data UKF recipe. |
conf/real/ekf_truth.toml |
Adds real-data EKF truth recipe. |
conf/real/ekf_degraded.toml |
Adds real-data EKF degraded recipe. |
conf/real/ekf_grav.toml |
Adds real-data EKF gravity recipe. |
conf/real/ekf_mag.toml |
Adds real-data EKF magnetic recipe. |
conf/real/ekf_both.toml |
Adds combined real-data EKF recipe. |
conf/real/rbpf_truth.toml |
Adds real-data RBPF truth recipe. |
conf/real/rbpf_degraded.toml |
Adds real-data RBPF degraded recipe. |
conf/real/rbpf_grav.toml |
Adds real-data RBPF gravity recipe. |
conf/real/rbpf_mag.toml |
Adds real-data RBPF magnetic recipe. |
conf/real/rbpf_both.toml |
Adds combined real-data RBPF recipe. |
docs/GEOPHYSICAL_CONFIG.md |
Removes obsolete configuration guide. |
docs/queue/00-test-speed.md |
Removes completed queue record. |
docs/queue/01-eskf-fix.md |
Removes completed queue record. |
docs/queue/02-filter-api.md |
Removes completed queue record. |
docs/queue/03-ned-default.md |
Removes completed queue record. |
docs/queue/04-eskf-primary.md |
Removes completed queue record. |
docs/queue/06-aiding.md |
Removes completed queue record. |
docs/queue/07-engine.md |
Removes completed queue record. |
docs/queue/08-validation.md |
Removes completed queue record. |
docs/queue/09-release.md |
Removes completed queue record. |
docs/queue/p-alignment.md |
Removes queue record. |
docs/queue/p-calibration.md |
Removes queue record. |
docs/queue/p-imu-quality-cov.md |
Removes queue record. |
docs/queue/p-lint-config.md |
Removes queue record. |
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
| # Align datasets by index | ||
| if not (len(nav) == len(geo)): |
| input = "data/input" | ||
| output = "data/output/ekf/both" |
| input = "data/input" | ||
| output = "data/output/ekf/degraded" |
| input = "data/input" | ||
| output = "data/output/ekf/grav" |
| input = "data/input" | ||
| output = "data/output/ekf/mag" |
| rho_pos = 0.99 # ignored while tau_pos_s is set; kept for a run without it | ||
| sigma_pos_m = 35.0 # steady-state wander, metres | ||
| tau_pos_s = 500.0 | ||
| rho_vel = 0.95 # likewise | ||
| sigma_vel_mps = 1.5 # steady-state wander, m/s | ||
| tau_vel_s = 100.0 | ||
| r_scale = 5.0 # advertised sigmas x5, so R x25 |
|
|
||
| [geophysical] | ||
| # Magnetic anomaly measurements | ||
| magnetic_resolution = "one_minute" |
| rho_pos = 0.99 # ignored while tau_pos_s is set; kept for a run without it | ||
| sigma_pos_m = 35.0 # steady-state wander, metres | ||
| tau_pos_s = 500.0 | ||
| rho_vel = 0.95 # likewise | ||
| sigma_vel_mps = 1.5 # steady-state wander, m/s | ||
| tau_vel_s = 100.0 | ||
| r_scale = 5.0 # advertised sigmas x5, so R x25 |
| conf` (`just geo-adopt`) writes the measured bias, noise and bias prior into all 18 | ||
| geophysical configs, line by line so the comments survive, rewriting only keys already | ||
| present so a `*_grav.toml` stays gravity-only. `geo_frequency_s` is held back behind | ||
| `--apply-interval`: the noise figures are measurements, but moving the interval from 1 s to | ||
| one measurement per de-correlation length changes what the experiment asks of the aid. | ||
| `core/tests/example_configs.rs` asserts all 18 carry the same values, so a partial adoption | ||
| fails the build rather than producing a comparison across two different R |
| `just setup` runs both. Nothing beyond Rust and uv is needed for the Rust crates; the | ||
| `analyze` CLI's map-drawing subcommands additionally want the GMT C library installed, though | ||
| every other subcommand -- preprocessing and measurement characterisation included -- runs | ||
| without it. |



Summary
Six workstreams on the geophysical aiding path — the three requested, plus a blocking defect found while scoping them, the monorepo setup that checking in
analysis/required, and three more defects that only surfaced when the pipeline was actually run.earth::gravity_anomalyreturned m/s² and was differenced against a milligal map--config, as the RBPF already couldanalyze geostats: per-field bias, noise, SNR, de-correlation length, and--apply-to confto adopt themImportant
Every prior result must be regenerated, and the gravity fix is not the main reason. Four independent changes invalidate them; the first alone is sufficient. See Regenerating the experiment.
0 — The gravity channel was never measuring gravity
earth::gravity_anomalyreturned m/s²:grav_x/y/zare m/s²,earth::gravity()returns m/s², and the Eötvös term is m/s². That was differenced against a milligal map, and--gravity-noise-stdis milligal. 1 mGal is 1e-5 m/s², so the observation was 1e5 too small and the innovationz - hwas, to five significant figures, just-h— the filter was told its predicted map value was wrong by exactly that value, at every geophysical update.Nothing caught it: a tens-of-milligal innovation against the default 100 mGal noise is a NIS of about 0.16, far under any gate. The magnetic channel had the identical defect and was fixed separately (
MICROTESLA_TO_NANOTESLA); this is the same fix for gravity.Why the tests missed it.
test_gravity_anomalyasserted onlyis_finite(), which a value 1e5 out still satisfies. Andgeo_closed_loop.rsleft the track'sgrav_*atDefault's zero, making the observation0 - 9.8017 = -9.80— a number that lands inside a map generated around ±40 and reads as a plausible milligal anomaly. It is not one; it is normal gravity in the wrong unit, and the fixture passed on that coincidence. AddedGRAVITY_BIAS_PLAUSIBLE_MGAL, the twin of the existingMAGNETIC_BIAS_PLAUSIBLE_NT: a units error does not stop the bias moving, it makes the bias absorb the error confidently, so bounding its magnitude is what separates a bias tracking an anomaly from one soaking up a scale factor. Verified by reinstating the defect — four tests fail.1 — Geophysical closed loop from a config file
--geowas reachable only from theclsubcommand's flags; a closed-loop config carrying[geophysical]was refused outright, while the particle-filter arm of the same function read maps and GNSS degradation from one file. The two filter families were configured in incompatible ways, and the justfile carried a twelve-flag command line that had to be kept in step withconf/*_degraded.tomlby hand.Both paths now resolve to a
GeoClosedLoopSettingsand run one function. Verified end to end: the CLI invocation and the equivalent config file produce byte-identical CSV.Two defects found by making them agree, both meaning the paths ran different filters:
--ukf-alphadefaulted to the textbook1e-3while a config file defaults to0.1— the value The accuracy baseline is not reproducible across recompilation: adding a dead function moves 59 gated metrics, one by 4.9% #399 measured and adopted, because alpha is a numerical-conditioning parameter here and1e-3costs six significant digits per step to cancellation. The flag was never moved with it.--log-levelwas discarded entirely on the--configpath, under a comment claiming it took precedence.Also:
gravity_bias/magnetic_biaswere config-path-inert and silently dropped; the four*_bias_init_std/*_bias_process_noise_stdkeys the CLI had are now available in a config;generate_plotreached only the non-geophysical path, so everyconf/*.tomlasked for a plot and no geophysical run produced one; and theconfigwizard could write a file the binary refuses (eskf + geophysical), which matters becauseEskfisFilterType's default.2 —
analyze geostatsThe 100 mGal / 150 nT defaults were never measured against anything. This measures them: for each trajectory it forms the anomaly the way the filter forms it, samples the map at the GNSS position, and reduces the residual to bias, robust σ, signal σ and SNR.
The residual is decomposed, because it holds two things that belong in different places. What varies within a recording is measurement noise and belongs in R; what varies between recordings is a per-run offset — the phone's and the vehicle's own field — and the filter has a bias state for exactly that. Collapsing them makes a channel look unusable when it is merely mis-modelled.
De-correlation is taken from the source data, not the grid spacing, and the difference is large:
earth_faais filled from EGM2008 over land, whose input is a 5 arc-min area-mean gridearth_wdmamis referenced to 5 km altitude; upward continuation suppresses everything shorterAt 24 m/s that is one independent gravity sample per 386 s and one magnetic per 1309 s, against the
geo_frequency_s = 1.0every config uses. (Also: the magnetic grid is fetched at 3 arc-min while every config declarestwo_minutes, and nothing validates that —geo-statsreports the actual spacing so the mismatch is visible.)Measuring is only half of it — the numbers have to reach the configs.
--apply-to conf, wrapped asjust geo-adopt, writes the measured bias, noise and bias prior into all 18 geophysical recipes. Line by line rather than through a TOML round-trip, which would strip every comment in these files; only into keys already present, so a*_grav.tomlstays gravity-only rather than silently becoming a dual-aid run. The one insertion is*_bias_init_std, which no config carries yet — without it the map-bias state inherits*_noise_stdas its prior, far too tight for a magnetometer sitting in a vehicle that contributes thousands of nT of its own field.Tested by planting a known bias and σ into synthetic trajectories, inverting the measurement models so the CSV carries raw sensor values, and requiring them back — within-trajectory σ recovers to 7.03 against 7.0 planted, and 25.12 against 25.0.
3 — GNSS degradation realism
The fix interval and the error model were the same knob.
rho_posis applied once per emitted fix with no reference to elapsed time, so the correlation time is-interval_s / ln(rho)— a property of the scheduler. Atrho_pos = 0.99, 5 s is 498 s and 30 s is 2985 s. Added optionaltau_pos_s/tau_vel_s: the coefficient becomesexp(-dt/tau)andsigma_pos_mbecomes the steady-state spread. Strictly additive — omitted, nothing changes, andperf_baselinedoes not move.The fault advanced on the wrong clock.
apply_faultruns once per emitted fix but was handed the record spacing, so underinterval_s = 5on a 1 Hz log aSlowBiasspoofed at a fifth of its configured rate.What a jamming scenario should be. The dominant measured effect of jamming is loss of fix, not graceful degradation — at Jammertest the signal was in most runs strong enough to cause complete signal loss. New
conf/*_denied.toml(duty cycle, faultnone, so availability is the only variable) andconf/*_jammed.toml(35 m steady state, anchored to published smartphone figures).conf/*_degraded.tomlis deliberately unchanged — it is the baseline every result so far was measured against.The denial comparison is like-for-like.
conf/*_denied_{grav,mag,both}.tomlcarry the denial GNSS profile, sogeoperf-deniedcompares two runs differing in nothing but whether the geophysical measurement is present. Nine more configs means 24 files duplicating a GNSS block, and drift between exactly those is what went wrong before — soexample_configs.rsnow asserts every geo recipe's[gnss_degradation]and[health_limits]match the baseline named in its filename. Verified by re-injecting the historical drift: the test fails and names the field.book/src/gnss/fault-simulation.mdwas a one-line stub; it now carries the rationale and thirteen sources.4 — Monorepo
uv workspace at the root (
uv run analyze ...from anywhere, dropping 35--project analysisflags), ruff + pytest config,.github/workflows/python.yml, Python.gitignoreentries, and eight.pycfiles removed from the index. CLAUDE.md and AGENTS.md both asserted the repository is Rust-only.One thing the gate required: pygmt is no longer imported at module scope. It
dlopens the GMT C library on import, so importinganalysisat all failed wherever GMT is absent — every subcommand, including ones that draw no maps. CI installs no GMT on purpose, so the job is what keeps this true.5 — Found by running it
These only surfaced when
just pipelinewas actually executed, and two of the three were silent.postprocessproduced zero plots and exited 0.read_csv(..., parse_dates=True, index_col=0)does not guarantee a datetime index — when pandas cannot infer a format it leaves it as strings, with no error. These timestamps look like2025-03-01 12:00:00+00:00, which it declines to infer, so the first arithmetic raisedunsupported operand type(s) for -: 'str' and 'str'.performance_analysiscaught that in a broadexcept, printed "possible dimension mismatch or missing data", and carried on — for every scenario, filter and trajectory. Fixed with aread_timeserieshelper across six call sites. 0 → 99 plots. The same defect was ingeostats, whose own fixture missed it by writing timestamps in a format that round-trips.pipelineran nothing. Its steps were recipe body lines, whichjustexecutes as shell commands, so it tried to runbuildas a program and died on the first one. They are dependencies now, and therm -rfcleanup moved to its owncleanrecipe.pipelinecould never produce the maps the geo runs need.preprocessran without--getmaps, so it calledinherit_parent_maps— which copies the parent recording's.ncfiles, whichcleanhad just deleted. Every geophysical run then died on "Gravity map file not found", three steps downstream of the cause.A fourth, found while verifying the nine denial-geo configs: three failed on every trajectory with
longitude ... is outside the map bounds.inflate_boundspads each map by a fraction of the track's own extent, but filter drift is an absolute distance — and it scales latitude and longitude independently, so a due-north drive gets ≈ 0 longitude margin and any eastward excursion leaves the map. Added--margin-km(5 km default, padding bymax(fractional, absolute)), measured against the actual excursion (truth 0.00 km, degraded 0.00 km, denied 2.27 km).OutOfMapBoundsnow names the remedy instead of just the coordinate.Plus a pre-existing CI flake.
test_state_transition_jacobian_multiple_statesis entropy-seeded viarand::rng()and bounded at5e-4— exactly on the shoulder of its own error distribution (median 2.4e-4, p99.9 4.9e-4, largest over 4000 states 5.05e-4). Measured at 1 failure in 400 runs; it hit this branch once at 5.0245e-4 on a commit touching nothing it covers. Seeded withStdRng::seed_from_u64, sample raised 10 → 50 states since determinism makes coverage free, bound raised to 1e-3 — twice the largest error measured, still far tighter than any real defect, since a wrong analytic Jacobian disagrees at O(1) or O(dt).Regenerating the experiment
Four independent changes invalidate prior results:
preprocessrecipe was-f 1; it is-f 10. Every filter sees ten times the propagation steps. This alone invalidates everything. It is principled, not a silent retune:DEFAULT_PROCESS_NOISE_DENSITYis a spectral density and each filter formsQ_k = q·dt(Process noise is per-step, so effective Q scales with sample rate (50x across shipped datasets) #374).*/gravand*/bothused an observation 1e5 too small.--ukf-alpha1e-3 → 0.1.just ukf-geonever passed the flag, so every prior UKF geo run used1e-3whileconf/ukf_{truth,degraded}.tomlused0.1— the comparison was between two differently-conditioned filters.interval_s = 1.0 / 15.0 / 5.0 / 15.0against a baseline at5.0 / 3.0 / 0.3 / 5.0.Run
just clean build preprocess geo-statsand stop there.data/output/geostats/geo_stats.tomlcarries the measured noise, bias and bias prior that replace the 100 mGal / 150 nT defaults;just geo-adoptwrites them into all 18 geophysical configs. Checkgit diff conf/, then runjust truth degraded denied jammed,just ukf-geo ekf-geo rbpf-sim denied-geo,just postprocess geoperf-all geoperf-denied.geo-adoptis deliberately absent fromjust pipeline— it rewrites tracked files and invalidates whatever is already indata/output, so the pipeline must not change the experiment underneath itself.geo_frequency_sis held back behind--apply-interval: the noise figures are measurements, but moving the interval from 1 s to one sample per de-correlation length changes what the experiment asks of the aid, which is a decision rather than a measurement.A partial adoption is the dangerous outcome — a
conf/where some filters carry the measured R and some do not still parses, still runs, and yields a comparison whose difference came from two different R.every_geophysical_config_describes_the_same_sensorinexample_configs.rsasserts all 18 recipes agree on every geophysical key they share; verified both ways, including by reverting one file'sgravity_noise_std.Testing
All six gates clean locally:
Plus end-to-end verification against synthetic Sensor Logger recordings: all 15 pipeline steps complete from an empty tree, producing the full matrix — 63 navigation solutions across {ukf, ekf, rbpf} × {truth, degraded, denied, jammed, grav, mag, both}, 99 performance plots, 72 geoperformance artifacts, plus the nine denial-geo runs and their 54 artifacts. The CLI and config geophysical paths produce byte-identical output, and both guard rails (ESKF, no map) fire with clear messages.
One step a sandbox cannot run:
--getmapsneeds the GMT C library and network access to the GMT data server. Locally generated grids stood in at the real spacings (1 arc-min gravity, 3 arc-min magnetic).🤖 Generated with Claude Code
https://claude.ai/code/session_01RwCvG9Hh2bmZbdd5JpPHVK