Robustness study / frozen before confirmation
Passed validation.
Then passed fresh tests.
FalkorDB selected the rehearsal checkpoint from complete validation evidence. A training-only collision penalty and 600,064 additional PPO steps produced a new saved policy. Its hash was frozen before one fresh, four-family confirmation.
Target confirmation: 8 collisions in 100 episodes. Limit: 10%. All four families passed their original success and collision requirements. This checkpoint is the current 2D demo policy.
Saved artifact / 180 validation episodes
The gate stayed fixed.
The new study required two episodes of collision clearance on every validation suite before selection stopped. This stricter stopping rule did not change promo-1.
One confirmation batch / 400 fresh episodes
A new test, opened once.
One hundred episodes per family. No confirmation outcomes entered the planner, and robustness-v2 stopped before opening them. The previous 13% target failure stays in the earlier study.
MuJoCo / new physics / no retraining
Physics still exposed a gap.
The new policy collided in 6/40 target episodes, above the 10% reference limit. Crossing collisions were 4/40. These fresh transfer seeds differ from the earlier study. The checkpoints were trained in 2D, with no MuJoCo adaptation. This demo is not a physical robot deployment.
Follow the separate mixed-domain physics adaptation study. The original unchanged-weight transfer results below remain frozen.
Watch the real physics recording · Every episode and checksum
Live FalkorDB / recorded decision
Why this parent and intervention?
The graph supplies exact-artifact validation and failures. An authored rule first ranks failed suites, then collision clearance and rate deficits. The Decision preserves those reads before training.
Ready to query FalkorDB.
Every eligible origin, ranked before training
These are the saved CONSIDERED relationships. Lower ranks win. Collision clearance asks for two fewer collisions than each suite allows; it is separate from promo-1. The graph revision saved on the Decision can differ from today's revision.
Cypher for the origin ranking
Selected parent's validation evidence
Exact query receipt
Download the recorded evidence index · Protocol and interpretation