Learned-policy deployment evaluation
Freedom within constraints.
Learn broadly. Deploy deliberately.
Know what changed before the robot does.
Evaluate candidate action traces against versioned controller constraints before physical deployment.
The deployment gap
Simulation ends before hardware validation begins.
Train
Did it learn?
Simulate
Does it work in the model?
Evaluate release
Does the trace satisfy the declared profile?
Controller
Can the target controller accept it?
Robot
What happens physically?
Aggregate scores hide release risk
A policy can score better and deploy worse.
Aggregate task performance can conceal release-specific constraint excursions and discontinuities.
Task success
What Synexis finds
Localise each finding to the trace.
Retain the run, timestamp, candidate version, and constraint for review.
Instability
Variation across comparable trace intervals
Discontinuity
Abrupt transition between adjacent commands
Constraint proximity
Action approaches a declared limit
Release regression
Candidate diverges from the trusted baseline
Trusted baseline → Candidate release
Isolate release-specific regressions.
Compare candidate and baseline action traces under the same versioned controller profile.
Trusted baseline
v0.8.3
Candidate
v0.8.4
Stability
0.91 → 0.84
Interventions
2 → 5
Flagged runs
1 → 4
Position in the stack
Between policy evaluation and physical execution.
Each candidate release creates a policy–controller compatibility decision with a recorded evidence trail.
Evaluation
Start with one release.
Submit one candidate action trace and a versioned robot/controller profile. Receive trace-linked compatibility findings for engineering review.