Every token, every dollar, every task, accounted for.

Most AI dashboards count tokens. Tokens aren’t the job. The Ledger accounts for work: tasks, verified outcomes, acceptance, reverts, and what each successful task actually cost.

The Waybill

Per task: the route, the cost per leg, the checks that ran, your decision, and, for a week after, whether it survived. The metric that matters rolls up from there: cost per successful task, where success means verification passed, you accepted, and no revert within seven days. A cheap model that fails often is expensive; the Ledger is how you see that instead of assuming its price tag.

The verification and seven-day survival fields arrive with the verification runner; the Waybill and the rollups do not wait for them.

The daily view

ledger — today

TODAY                                     (illustrative)

Tasks verified: 37
Tokens processed: 4.82M
Local: 3.71M 77%
Frontier: 1.11M 23%

Frontier spend: $6.90
Frontier-only baseline (est.): $31.84
Avoided: $24.94

“Avoided” is computed against a defined baseline — by default, routing every eligible call to a named frontier model — and the assumption is printed on the report, not buried in a footnote. We say “frontier tokens avoided,” not “tokens saved,” because local tokens still ran; they just ran on hardware you already own.

Haybox dashboard, illustrative: 37 tasks verified, 4.82M tokens at 77% local, $6.90 frontier spend, $24.94 avoided, and three recent Waybills.
Haybox dashboard, illustrative: 37 tasks verified, 4.82M tokens at 77% local, $6.90 frontier spend, $24.94 avoided, and three recent Waybills.

The dashboard, one day. Illustrative.

Audit-grade, on purpose

Savings claims get attacked, first by commenters and later by procurement. Fine. Sampled frontier comparisons put a confidence interval on the number, exports go out as CSV or PDF with the methodology attached, and the ledger itself is a SQLite file on your disk that you can query directly. If a savings figure can’t survive an adversarial reader, we’d rather not print it.

Questions

Where does the data live?

Locally, in SQLite, yours.

Can I export for finance?

CSV or PDF, with the methodology attached.

What baseline do you use?

Configurable; the default is a named frontier model for all eligible calls, and it is always shown next to the figure.

Local Model Manager · Rules files · What “success” means · Security · All features

See what your own traffic says.

Shadow Mode reports on the traffic you already have. Before that exists for you, the calculator will estimate it from your GPU and your monthly spend, and show you every assumption it used to get there.