Design rationale

DEC-MAPF separates simulation semantics, experiment execution, independent validation and presentation. The following boundaries make the system inspectable and extensible without treating a software check as a scientific result.

Boundary	Design	Verification and limits
Independent validation	Validate roster, geometry and trajectories separately from solver success.	Reject missing/unknown agents, incorrect starts, illegal movement and goal departure. An unfinished legal prefix is not a solution.
Joint movement	Preserve legal greedy moves; bound repair while enforcing waiting permission, live commitments and future parked occupancy.	Refuse an unauthorized joint step atomically. Physical validation and protocol obligations remain separate.
Solver geometry	Use instance geometry for omitted defaults; reject explicitly conflicting configuration.	Entry-point tests check consistent behavior across built-in solvers.
Solver identity	Display SimplifiedFocalCBS for the Python focal implementation, including its EECBS aliases.	Registry/capability tests do not confer a published EECBS bound.
Commitments	Store per-contract reservations with explicit ownership and absolute time.	SC/DC/ZC tests cover finite expiry and simultaneous obligations to different partners.
Tentative decisions	Snapshot owned participant state; validate tentative proposals and apply settlement atomically.	Exception and token-conservation tests check rollback. External I/O inside strategy callbacks cannot be rolled back.
Information disclosure	Count actual delivered broadcasts and distinguish spatial disclosure from space-time/all-message metrics.	Missing receipt evidence remains unavailable; see metrics.
Identity and durability	Separate condition digests from job/run/attempt identities; use a SQLite journal, atomic artifact writes, directory ownership and a process watchdog.	Failure, cancellation, restart and idempotency tests exercise recovery boundaries.
Portable JSON	Canonicalize with RFC8785 across Python and JavaScript.	Round-trip and tampering checks cover numeric representation and checksum consistency.
Costs and comparison	Derive first-goal action costs from paths, retain solver-reported values and pair on declared identities.	Common-solved cost cohorts differ from all-planned success denominators.
Replay	Start at t=0 and retain observation timestamps, current plans and commitments.	Local received information and global hindsight use distinct visual layers; absent information is unavailable.
Renderer	Use React/TypeScript/Vite, Canvas 2D and native controls with a bounded frame cache.	Coordinate, workflow and accessibility checks qualify finite interactions. A retained renderer spike supports performance investigation.
Packaging	Lock dependencies and bundle the built GUI with Python source.	Installed-wheel checks run outside the checkout; Node is unnecessary at wheel runtime.
Research evidence	Keep synthetic correctness fixtures separate from empirical datasets and published results.	Acceptance tests neither supply the article's full experiment matrix nor establish numerical replication.

For implementation responsibilities, read architecture. For behavior, read algorithm semantics. A passing test supports its declared stimulus and expectation, not universal collision freedom or equivalence to the complete published study.