Six independent tools
plan_observation— Compare proposed observations using their predicted effect on position uncertainty.stress_test_observation— Search declared calibration and timing perturbations for a counterexample to a stability claim.assimilate_observation— Refit supplied tracks with new angular observations, keeping measured and synthetic origins explicit.improve_schedule— Compare affordable deadline and optional-task changes, then check one finalist with a separate simulation.forge_route_policy— Generate, mutate and cross bounded expressions on training cases, then lock one experimental candidate.run_route_policy— Run a bounded rule on another closed Euclidean tour without retraining or implying independent acceptance.
Choose where to observe
plan_observation fits the supplied angular tracks, then compares up to 64 proposed observation series at the same future time. It predicts the longest one-standard-deviation position axis from the full position/velocity covariance and new angular information. It checks cost, station identities, sample limits and geometric elevation. It never treats predicted angles as measurements. The best candidate is best among the supplied admissible alternatives, under the stated local linear model.
Generate a method within a declared grammar
forge_route_policy generates, mutates and crosses arithmetic route-construction expressions in Rust, using only supplied training cases. Its complete journal records every proposal, ancestry, rejection and reused result. One selected expression can be passed explicitly to run_route_policy on new points. The language is bounded: five features, seven constants, five operators, at most 31 nodes and four edges of depth. Tours close at depot 0 and use a fixed 2-opt improvement primitive. A new expression is not a new scientific model, an optimum guarantee or evidence of superiority. The laboratory never changes the standard optimizer automatically.
Open Method Lab: formula, 3D route and independent validation. Search and held-out validation are separate; the published study locks one training winner before examining 48 fresh synthetic cases, compares two competent reference methods and preserves failures. Live searches are experimental, regardless of the published study verdict.
A validated implementation can become reusable
The observation planner retains its independent MCP tool and scientific contract. A reviewed execution policy evaluates every supplied candidate and date while reusing station geometry and symmetric information updates. Small inputs keep the reference. The returned policy_trace identifies the version, evidence fingerprint and implementation that actually ran; a rejected baseline executes no ranking kernel. Inspect the sealed study, previous rejected attempts and live examples. This is an engineered numerical improvement with automatically generated evidence, not autonomous invention of a scientific model.
Add information and check the forecast
assimilate_observation refits every original and newly supplied angular pair once. The original fitted mean initializes the solver; the old covariance is not added again as a prior. A new series can come from a later station without overlapping the old track times. Its measured or synthetic origin remains caller-declared and unverified. A smaller conditional covariance does not prove smaller error against unknown truth.
Look for a counterexample
stress_test_observation examines up to 32 fixed hypothetical angular biases or clock-correction changes. It asks whether the reconstruction remains usable and within a declared position-shift limit. Invalid perturbations are inadmissible; solver non-convergence is inconclusive. No counterexample found in a finite list is not a global robustness guarantee.
Apply the same approach to a schedule
improve_schedule compares up to 32 approved combinations of deadline extensions and optional-task removals. Required tasks are preserved. Positive caller-defined cost weights make changes comparable. Fixed Monte Carlo search uses a simultaneous error budget; one finalist receives a separately seeded fixed confirmation. Failed confirmation does not trigger retries on other finalists using the same validation results.
Assumptions remain visible
The observation model assumes constant velocity in spherical Earth-fixed coordinates, exact station positions and corrected clocks, and independent angular errors. Its covariance is a local conditional approximation. The schedule model assumes independent uniform duration variation. Neither model is automatically calibrated to real-world observations.
Optional orchestration
The capability planner can return a direct binding to each complete request. Its structural graph cost does not measure scientific quality. Forecast receipts cannot satisfy an acquired-observation contract: the system will not invent missing measurements. Tools can always be called independently over the public MCP endpoint.
Inspect and replay
The laboratory preserves the exact submitted example, returned native result and browser timing in a downloadable receipt. Public examples use synthetic data. Private scientific inputs and outputs are not copied into the public activity journal.