Correction corpus
For data buyers: failed runs paired with the human fix, priced as a range.
ILLUSTRATIVE live DB · demo corpus
Rent is cash flow; this is the asset. How a human takes over and corrects a failing policy can't be boughti, and a hosted farm is the only place that produces that moment systematically.
Three ways to capture a correction
Shadow divergence is captured live in Shadow mode and reviewed per session in the Shadow report.
DAgger rounds
β = probability the expert takes over in a round; round 0 (β=1) is pure demonstration, then it decays.i
| Policy | Round | β | Corrections | What this round does |
|---|
Only some hours are sellable
Total hours are not sellable hours. Two gates: the customer chose the private tier, or time sync missed speci.
Total → trainable hours
Shadow divergence · value lies right of the threshold
Export buckets
| Bucket | Segments / points | Hours |
|---|
Label fewer frames, not cheaper ones
Versus Scale-style labeling the gap is not unit price but how many frames get labeled at all.
Work queue · by type
Priced as a range, never a point
The corpus has zero sales to date, so any unit price is derived: the upper edge of prices buyers actually paid × a labelled multiple, reported as a range.i
Comparables · buy and sell sides apart
Resulting range
Rights: in the contract on day one
Shared by default, private at a premium. The order can't be reversed.i
| Tier | Premium | Ref. unit price | Meaning |
|---|
Invariants
duration_s, so "intervention segment exceeds run duration" reports "not checkable", not OK.