xlab-budget(7)
NAME
xlab-budget — know the declared spend before starting a run.
SYNOPSIS
xlab plan experiments/S1-demo.toml
xlab queueDESCRIPTION
The product brief is a lab that plans, prices, and pairs before spending tokens. The current implementation provides estimates and rolling-window admission. Automatic agent token pricing and a hard mid-run dollar stop are still planned.
An adapter estimates resource use for a plan. A spec's [spend] values override the corresponding adapter estimates per sample. Planned samples include warm-up; retry attempts can add work beyond that estimate. If nothing is declared, “no declared spend” means unknown, not free.
[budgets] in lab.toml caps declared resources over a rolling window, rather than a calendar day. The runner picks the first queued run that fits every cap. It charges the estimate once on starting, including across resumes, and settles the charge by the fraction of samples actually taken when the run ends.
Admission is not a hard spending ceiling
A run whose estimate exceeds an entire resource allowance may start once that resource has an empty window. Running samples are not stopped at a dollar limit. Use estimates conservatively; this mechanism cannot guarantee a provider bill cap.
EXAMPLES
These are illustrative configuration values, not measured costs:
# lab.toml: resource allowances, including standing runs
[budgets]
window_h = 24
usd = 5.0
seconds = 28800
bytes = "200GiB"# experiment spec: declared estimate for each sample
[spend]
usd = 0.02
seconds = 1.5Standing runs share the same allowances as the rest of the queue. Account for their scheduled spend when deciding how many experiment blocks fit.
CALIBRATION DATA
The agent A/A blocks-needed table and exported noise.json do not exist yet. The DS-05 calculator is deferred until measured noise and pinned prices can supply its inputs. No invented noise floor or token price is substituted here.