Automated tests
Domain, search, CBS, regression, simulation, recovery, and run-contract checks.
Reference run ac32c3eb44ff9873
A delay breaks the plan. Recovery changes only what it must.
A discrete-time planning lab that shows exactly how collision-free routes fail, which agents are affected, and whether the recovered plan is valid.
3 agents1 forced wait2 paths replanned0 conflicts after recovery
01 / LIVE RUN
Delay recovery
Baseline valid
ENU
Eastbound Northbound Unaffected
StatusValid
Makespan6steps
Agents changed0 / 3baseline
Vertex / edge conflicts0 / 0independently checked
Replan latency—local reference
02 / EVIDENCE
The animation is the hook. The validator is the proof.
The interface replays a committed run. Planning, simulation, and validation remain separate so a planner cannot certify its own output.
Seeded cases
1,000 low-level, 100 prioritised multi-agent, and 100 exact joint-oracle cases.
Post-recovery conflicts
Zero vertex conflicts and zero opposing edge swaps in the reference run.
POST
/v1/runs · /v1/comparebaseline→failure→affected_agents→recovery
One versioned FastAPI response contains the scenario, configuration, paths, validation reports, metrics, and sequence-numbered events.
03 / BENCHMARKS
Change the strategy. Watch the plan change.
Replay the exact paths written by the Python benchmark runner. Fixed priority is the responsive track; priority sweep and CBS expose what extra search can recover on deliberately small instances.
One deterministic order with the lowest search effort.
Tests a declared number of orders and retains the best valid result.
Optimal for sum of costs when it solves within the explicit node limit.
1
2
t = 03
| Scenario | Fixed priority | Priority sweep | Bounded CBS |
|---|---|---|---|
| Crossing conflict | Cost 5 | Cost 5 | Cost 5 |
| Priority cost gap | Cost 9 | Cost 7 | Cost 7 |
| Priority order trap | No plan | Cost 11 | Cost 11 |
| Storage detour | Cost 10 | Cost 8 | Cost 8 |
Costs are sums of final-arrival timesteps. CBS matched an independent joint-state oracle on 100 seeded two-agent cases. Every visible path and digest comes from the committed machine-readable benchmark record.
CONTROLLED SCALING
Search work, not marketing claims.
One fixed 9 × 9 crossing family adds agents from two to six. Bars show deterministic expanded-state counts across 11 repeated runs; runtime remains a local reference.
2 agents
Fixed priority18cost 17
Priority sweep36cost 17
Bounded CBS46cost 17
3 agents
Fixed priority27cost 26
Priority sweep190cost 25
Bounded CBS55cost 25
4 agents
Fixed priority36cost 34
Priority sweep1,058cost 34
Bounded CBS160cost 34
5 agents
Fixed priority45cost 43
Priority sweep6,531cost 42
Bounded CBS99cost 42
6 agents
Fixed priority66cost 53
Priority sweep8,165cost 51
Bounded CBS166cost 51
The priority sweep exhausts all orders through five agents and is capped at 120 of 720 orders for six agents. CBS remains capped at 20,000 high-level nodes. These results describe this committed family only.
04 / ARCHITECTURE
One responsibility per boundary.
The Python core has no dependency on FastAPI or the interface. The web client renders the run record; it does not recreate planner decisions.
01React client
→Playback and inspection
02FastAPI
→Versioned run contract
03Planning core
→SIPP, sweep, and CBS
04Validator
Independent decision
Stated limitation
Fixed-priority SIPP is deterministic and responsive, but order-dependent and incomplete. Priority sweep grows factorially; CBS can grow exponentially. Every comparison is explicitly bounded. This is a discrete planning lab, not a safety-certified controller or operational rail system.