Validation

Validation

lob is numerical software: correctness is established by evidence, not by assertion. Each claim below cites a test or measurement; no universal accuracy promise is made.

Strategy

• No invented accuracy claims. Each assertion below cites a test or a measurement; the phrasing mirrors the tests' tolerances rather than “lob is accurate to X everywhere.”
• Evidence before synthesis: tests are the primary contract; technical pages describe what the code does and the tests bound how well.

Reference data

• BRL drag tables: 87-point Mach 0–5 tables for G1…G8 from JBM (source/tables.hpp). The spline reproduces them within 5e-3 max absolute error over a 5000-point truth grid (test/source/splines_test.cpp, SplineOptimization::BaselineAccuracyBudget).
• Atmosphere & ballistics: reference trajectories in test/source/lob_env_test.cpp etc. for ICAO, altitude (+4500 ft), hot/low-pressure, barometer-offset and weather-station cases, each with expected velocity ±1 fps, energy ±5 ft·lbf, 0.1 MOA equivalent and ±0.01 s out to 1000 yd — matching McCoy's ISA treatment and Huang's saturation formula.
• Spin/jump: Boatright and Litz against their papers; the builder test forces the Boatright supersonic integration to timeout (kLobErrorInternalError, test/source/lob_builder_test.cpp) to prove the guard fires.

Numerical accuracy

• Default 1-yard Heun step meets the reference tolerances above. Halving the step reduces error monotonically (Builder::StepSize); the benchmark benchmark/ode.cpp shows linear cost.
• The PCHIP splines preserve monotonicity and never overshoot by construction (source/splines.hpp, test/source/splines_test.cpp).
• Angle solver tolerance is 0.01 MOA with at most 10 iterations (source/solve_angle.hpp); tightening to 0.001 MOA is tested for convergence to ±0.01 in (test/source/solve_angle_test.cpp).
• Inverse vs forward consistency: iterative SolveInverse matches FastInverse within 0.1 MOA on the cross-check trajectories (test/source/lob_inverse_test.cpp). Forward vs inverse sign flip is asserted (test/source/lob_inverse_test.cpp).

Invariants

• CurveView clamps below 0 and above 5 Mach to the edge drag (test/source/splines_test.cpp); seeks from start to end and back are O(1) (test/source/splines_test.cpp). See also Errors for the full Build()-vs-solve error split and Physical Constants for the ISA values behind the tolerances.
• FindInterval is a binary search; its edge cases are unit-tested (test/source/splines_test.cpp).
• Custom-table extrapolation that would yield negative Cd at 0 or 5 is rejected (test/source/lob_builder_test.cpp, source/lob_builder.cpp).
• BC bands are clamped flat outside the user Machs (test/source/lob_builder_test.cpp).
• Zero/inverse angle is bounded to ±45° and returns NaN if exceeded; unreachable targets return NaN/kLobErrorZeroUnreachable (test/source/solve_angle_test.cpp).
• Inputs are strictly increasing where required (ranges, Machs, fps); non- monotonic returns 0 or the corresponding *NotMonotonic error.

Coverage and checks

— coverage from cmake --preset=ci-coverage && cmake --build build/coverage -j && ctest --test-dir build/coverage && cmake --build build/coverage -t coverage (see .github/workflows/ci.yml coverage job).

• CI runs lint, sanitize (ASan+UBSan), coverage and cross-OS test (macos, ubuntu, windows, shared/static).
• Unit tests live under test/source/ and mirror each production file: calc, cartesian, eng_units, helpers, ode, splines, solve_angle, solve_step, boatright, litz, plus integration tests lob_env, lob_coriolis, lob_wind, lob_spin_drift, lob_cwaj, lob_builder, lob_inverse, lob_api.

Gaps and follow-ups

• Per-step lapse is gated (Fast* firing-site vs DsDx/SolveStep/SolveAngle when forward drop>100ft), not universal — see Atmospheric Model and source/lob_solve.cpp kDynamicDropThreshold. Long-range >100ft inverse accuracy vs field data would be a stronger anchor than builder-time ISA alone.
• No field-data validation beyond the BRL tables and borrowed reference trajectories; a “field vs radar” data set would be a stronger anchor.
• Performance is benchmarked (ci-benchmark cachegrind) but no regression gate is enforced in CI beyond CG_MAX_REGRESSION_BP.
• The spline knot optimization (5e-3 budget) is validated but its derivation is not documented beyond the exhaustive test's 5000-point grid — worth a short note if the knot set ever changes.