The aerospace platform is live. Fourteen categories, each with 200 verified rows, all served through one streaming endpoint.
What it covers
Computational categories — the answer is a number, and the number comes with a derivation.
| Category | What it teaches |
|---|---|
| Propulsion and Mass Budget | Rocket equation, propellant mass, staging, finite-burn gravity losses |
| Orbital Mechanics | Circular-orbit period, Hohmann transfer, Kepler's equation |
| Orbit Perturbations | First-order J2 nodal regression, sun-synchronous inclination |
| Launch Geometry and Windows | Launch azimuth with and without Earth rotation, plane-crossing wait time, range azimuth margins |
| Standard Atmosphere | US Standard Atmosphere 1976 density and dynamic pressure to 20 km |
| Hardware | Ideal nozzle exit velocity, thick-wall tank thickness |
| Guidance and Estimation | Kalman filter predict-and-update cycle |
| Ascent Simulation | Burnout state and Max-Q, converged across three numerical integrators |
Refusal categories — the answer is a code, and the code names what's missing.
| Category | The model learns to |
|---|---|
| Refusals: Missing Parameter | Ask for the specific field it needs |
| Refusals: Underspecified | Ask which interpretation the user meant |
| Refusals: Invalid Input | Ask for a corrected value |
| Refusals: Inconsistent Input | Ask the user to reconcile the contradiction |
| Refusals: Infeasible | Say the goal has no solution instead of inventing a number |
| Refusals: Legal and Licensing | Route the question to the correct authority |
Why the refusal categories matter
A model trained only on solvable problems learns that every question has an answer. Given a question with a missing input, it invents one. Given contradictory constraints, it picks one silently. The refusal categories teach the opposite: stop, name what's wrong, ask.
That's the difference between a calculator that returns garbage and an assistant that says "I need the chamber pressure before I can compute exit velocity."
Verification
Every computational row carries:
headline_answer— the display valueheadline_answer_exact— the same value at 40-digit precisionerror_bound— the maximum deviation from the true answerverification_level—analytical_proof,numerical_convergence, orinterval_arithmeticneglected_effects— what the model omits (J2, drag, Earth rotation)
A row is accepted only if the analytical half-width is at most 6.2×10⁻⁴, the delivered code agrees with the reference within 2.7×10⁻¹¹, and the independent numerical method agrees within 1.5×10⁻¹⁰.
Ascent simulation rows are verified differently. They run three independent integrators — RK4, DOP853, Radau — and accept the result only when all three agree within 1×10⁻⁷. The row's fidelity_tier field says which method was used.
How to stream it
GET /api/v1/models/alr-aerospace-platform/data
Same key, same format, same query parameters as the code platform. Filter by category with sample_type:
GET /api/v1/models/alr-aerospace-platform/data?sample_type=orbital_mechanics
What's next
More templates per category. Deeper coverage of launch windows, more staging configurations, and expanded perturbation cases.
— The Alrithm team
Associated Data Model
This article references the ALR Aerospace Platform 0.2. You can stream these verified reasoning traces immediately using our API.
[ FETCH DATA VIA API ]