model flight
correction branch
survivor beacon
no-fly zone
tick 0 / 0
Corrections on this rollout
None yet.
Fleet — live operations
Four staged robots flying different policies over simulated video feeds. Watch, take control when a flight goes wrong, and the takeover lands in the record store — the correction loop, entered from the ops floor.
MIRA 2v2 — four views, one world
Recorded 2v2 Rocket League from the kyutai/rocket-science corpus (General Intuition · Kyutai · Epic Games), played back the way MIRA §4.5 frames it: four synchronized perspectives of one shared world on a single master clock. Judge the play, trace an object across views, and export the segments as schema v0.2 records.
Judge which behavior you'd rather see deployed. Both bad routes the segment to the takeover queue.
queue:
A
B
Takeover queue (from both-bad)
Empty.
Browser Trainer — close the loop
Corrections become a new checkpoint; the new checkpoint gets evaluated on held-out seeds; the eval attributes improvement by failure category. This panel is the half of the loop nobody builds.
Before / after — held-out worlds, by failure category
Correction shelf life
Which corrections does the new checkpoint still visit? Stale corrections are the DAgger decay made visible — the reason this platform is a flow, not a dataset.
Training runs
Training runs from the Python / Prime Intellect pipeline appear here once the pipeline has produced artifacts.
Runs Explorer — the platform at scale
Scale — 1M clips/sec and the storage split
Lake — the whole record store, visualized
Every pane is a live query against the real Vortex lake of sim rollouts. Requires the local sidecar (train/lake_server.py) — the deployed demo shows this surface offline by design.
Operator corps — quality, calibration, economics
Five simulated operators judge the same gold-standard pairs. Calibration decides each operator's export weight — and the value table asks the guild-vs-crowd question: is one veteran worth more than the crowd?
Calibration (gold tasks)
| operator | archetype | gold accuracy | export weight |
|---|
Guild vs crowd
Value table
| operator | judgments | effective signal | cost units | signal / cost |
|---|
Organic play mining
Free-play runs from Review are auto-paired against model rollouts on the same world. Directed correction is performance under observation; organic play is the Medal thesis — capture people at their best because nobody's watching.
Learned verifier — online Bradley-Terry over trajectory features
Trains live on your Compare judgments. Human preferences are the only supervision signal — this is the verifiable-domain loop in miniature. Same math as MIRA's Bayesian-Elo human studies (App. I): Bradley-Terry is the Elo model, run online instead of as a one-off study.
Rollout ranking (by current verifier score)
| # | rollout | checkpoint | outcome | score |
|---|
Failure taxonomy by checkpoint
Auto events from the sim plus your annotations. This table is the seed of the eval loop: did corrections in a category reduce failures in that category on the next checkpoint?
Export — schema v0.2 records
Records are the source of truth; training formats are views. The bundle strips internal fields and carries a manifest.