Post-release system status
Updated 21 September 2026. The router v5 files, checkpoints and original evaluation remain unchanged.
Later Generator evaluations
| Evaluation | Result | Adoption |
|---|---|---|
| Strict exact-coordinate outline set | 31 / 32 | Failed the frozen 32 / 32 gate |
| Existing regression set | 127 / 128 | Exceeded the 124 / 128 floor |
| Action-only strict JSON contract | 0 / 32 | Failed |
The exact-coordinate failure produced a pitch outside the predeclared tolerance. The action-only responses all included forbidden arrays, coordinates, angles, dwell values, explanations or other structure. Outputs were not rounded, repaired, stripped or replaced. Neither candidate was adopted or flown.
These later results use different contracts and evaluation packs from router v5, so they do not revise the original v5 counts. The later exact-coordinate checkpoint is not distributed in this repository.
Current fixed-scene reference path
One continuous PX4/Gazebo mission completed a partially occluded observation, Hosted Reasoner inspect, one deterministically compiled reobservation flight, Hosted no_action under the same goal, and return. The completion audit passed all eight gates. Maximum estimated tracking error was 0.127640243 m, maximum ground-truth axis error was 0.266943785 m, and return error was 0.041820814 m.
The reference path made zero Generator calls and accepted zero native model proposals. Its success is therefore not performance evidence for router v5 or for the later failed candidates. It qualifies one known fixed scene in simulation only; real aircraft and broad scene generalization remain unvalidated.
Planned successor contract
The successor design does not ask the model to emit exact coordinates or choose an arbitrary route-candidate ID. One fine-tuned Cosmos 3 Edge derivative checkpoint will use two strict roles: the Reasoner names the unanswered mission question, and the Generator turns a validated question into the required view relationships, coverage and stopping conditions. Trusted deterministic code will compile that observation program into exact viewpoints and a safe route. The initial mission profiles are slope failure, access clearance, flood and river inspection, rescue targets, building and earthquake damage, and OrbitPod search and recovery staging. Future wildfire, mountain search-and-rescue and wildlife profiles remain untrained until suitable sensors and independent data exist.
No complete internal run archive, raw Hosted response, source footage, client material or new weights were added in this update.