Reproducibility
Seeds everywhere
Every demo and every executed cell on this site fixes its RNG seed. SimulationConfig.seed and InferenceConfig.seed thread a single generator through observation noise and the particle filter respectively, so a given configuration is repeatable in the same numerical environment. Matching seeds alone do not guarantee identical trajectories or final numbers across platforms: NumPy draws, floating-point arithmetic, BLAS, and the host math library can differ. The frozen SMC replays therefore compare floating results with tolerances while checking counts and controls exactly. Use the locked dependencies and the displayed settings when reproducing a page.
Tests
The default suite — uv run pytest — covers the forward model’s physical domain, public data contracts, all resampling variants, adaptive tempering, conjugate evidence checks, MH invariance and rejection, observation atomicity, prior support, contrast-space likelihoods, the public API, wheel entry points, and visualization. Long-running fixed-seed acceptance scans are excluded by default (uv run pytest -m slow runs them). See tests/ for the current inventory; the README deliberately does not quote a test count, because hardcoded counts go stale.
CI
Three workflows run on GitHub Actions:
ci.ymlruns tests and lint on pushes tomainand on non-draft pull requests targetingmain, including when a draft is marked ready for review.site.ymlis triggered by each push tomainand re-renders and redeploys this site, so every executed-cell figure you see is regenerated from source at deploy time. The demo GIFs, the density PNG, and the echolocation range-study PNG are the exceptions: they are committed files, described in the next section.site-check.ymlrenders this site and runs the link and figure checks on non-draft pull requests targetingmainthat touch the site or the library, so a broken page is caught before it merges.
Where the committed figures come from
The story pages keep commentary on these figures’ origin to a short caveat. This section summarizes where each came from and what it does and does not establish; the development calibration report records the generating commit and the asset hashes.
Demo animations and the density figure. Each committed GIF, and the density PNG, is the packaged default output of its uv run demo-* command, generated explicitly from the corrected physics and SMC implementation after one-shot certification. The commands regenerate their scenarios, but image encoders are not claimed to be byte-deterministic across tool versions. Each figure depicts one named finite realization: a fixed scenario and seed. None establishes a population recovery rate. Where a story page also runs a smaller example in an executed cell, that is a separate finite run whose settings are shown in the cell.
The density demo samples a uniform box in \((\mu,\sigma,A)\) conditioned on that box and weak-field validity, and uses the same support for its prior density.
Echolocation animation. The committed GIF is the packaged default output from the corrected contrast-space SMC release, not a re-render of the current code. Its differential-rates panel draws the forward model evaluated at the posterior mean; the code now plots the posterior mean of the particles’ own predictions, so that panel will change when the assets are next regenerated. It places the mass at two circumradii along a fixed off-axis direction, and is one fixed-seed visualization at the closest, most favorable range, not a reliability measurement.
Model-comparison runs. The committed GIF on How Many Masses? is the packaged default run: 80 observations, 2,000 particles per candidate, and seed 42. With K=2 truth, its final estimate is K=2: 0.3990 and K=3: 0.6010. The executed cell on that page is a separate 25-observation, 400-particle calculation that ends near K=2: 0.740 and K=3: 0.260. A different finite data horizon changes the evidence supplied by the noisy realization, and a different finite-particle budget changes the Monte Carlo estimate of that evidence. The split between two and three masses also moves substantially with the inference seed at these budgets, so neither number is a claim about which model generated the data. Comparing models responsibly means examining data horizon, particle-count, and seed sensitivity rather than reading one run’s extra decimal places as certainty.
Echolocation range study. The range harness behind the figure on the echolocation page sweeps the declared circumradius grid along the same fixed off-axis direction as the animation, at one fixed mass, and records position error, mass error, posterior standard deviations, coverage, residuals, SMC controls, and forward-model evaluation counts. Development seeds 0–11 selected the controls by the declared ranking (uv run scripts/scan_echolocation_range.py --seed-block 0), and the controls and acceptance tolerances were frozen before the reserved block was run. The reserved 400–411 block was then run exactly once for the corrected implementation, using the scenario-default cell via --seed-block 400; protected blocks reject control overrides.
The displayed study was regenerated directly from the certified block-400 records. The frozen cell passed 46/72 cases overall; pass counts from close to far were 12, 12, 11, 8, 3, 0. The close range passed 12/12. The far/close median position-standard-deviation ratio was 66.524, above the frozen factor 20 gate, and the 12/12 far-range three-standard-deviation coverage is retained as a diagnostic rather than promoted after inspection into a new gate.
The completed protocol guards against tuning directly to the reported holdout, but its fixed seeds provide regression and calibration evidence only for those simulated cases. It does not estimate population reliability or justify a universal pass probability. Any claimed sensing horizon must therefore be attached to a stated noise model, prior, particle budget, SMC controls, pass rule, and finite seed set, not promoted to a property of the population.
Taken together: the protected fixed-seed results establish deterministic regression cases, not a population reliability estimate. SMC evidence is similarly a consistent Monte Carlo estimator with finite-particle sampling error, not a deterministic exact integral.
Locked Python dependencies
uv.lock is committed; CI enforces the Python package resolution with --locked in ci.yml and --frozen in the site build. The lock covers package versions for its supported Python-version markers. It does not pin the Python interpreter’s patch release; Quarto is pinned separately, in the site workflows. So the repository does not claim a fully bit-reproducible rendering toolchain.