1Make the failure visible
Replay an event and inspect the state it produces. Duplicate orders and repeated stock updates become reproducible cases.
Your failures become your dataset.
Your dataset becomes your model.
Reflex turns verified repairs into training data. UFO calls your trained River model, and Reflex runs the checks that show what the patch actually fixed.
Live UFO → River integration
Watch the working loopUnder 2 minutes · UFO + River
A real UFO tool call. Your River checkpoint. Executed checks and saved evidence.
Download videoLive integration · v4 checkpoint
UFO called. Your model repaired.
A replayed inventory event moved stock twice. The UFO SDK invoked repair_code_with_reflex, the trained River model returned a patch, and Reflex executed the unchanged checks.
UFO SDKTool invokedProgrammatic invocation
River model48 updatesYour saved v4 checkpoint
Executed checks2/4 → 4/45.7 seconds end to end
One training case, verified locally. This receipt demonstrates the tool integration and repair outcome; it does not measure a learning gain.
Inspect the UFO receipt2Give UFO a trained tool
UFO invokes the repair specialist. River proposes a patch from your saved checkpoint; Reflex executes its behavioral checks.
3Learn from the work
Verified examples train a River model. Save its checkpoint and compare base, memory, and learned behavior on held-out tasks.
Your latest model, v4.
Real River training produced the checkpoint used by the UFO tool. The earlier v3 comparison remains recorded separately.
24Training examples
48Optimizer updates
122,752Training tokens
Held-out task successEarlier v3 · 16 updates
Base4 / 4Tasks solved
With memory4 / 4Tasks solved
Learned model4 / 4Tasks solved
The earlier v3 checkpoint solved all four held-out tasks, as did base and memory. Measured accuracy gain: 0 percentage points on this set. A larger, harder evaluation is needed to measure generalization.
What this demonstration establishes
The training set contains 5 operator-accepted repairs and 19 machine-verified synthetic variants. These variants come from six repair families; they are not 24 independent production incidents. The video is an edited walkthrough of real application captures. River repair, training, checkpoint creation, and the earlier v3 held-out evaluation ran live. The v4 checkpoint completed 48 updates over 24 examples and processed 122,752 training tokens. The installed UFO SDK then invoked the repair tool against the v4 checkpoint on a training case, moving its executed checks from 2/4 to 4/4 in 5.7 seconds. This was a programmatic tool invocation; a full autonomous UFO conversation is not shown.
The learned checkpoint depends on the base model and River infrastructure. The checkpoint, dataset provenance, and evaluation record make the training result inspectable. This small evaluation does not establish improvement over the base model.