We don’t count lines.
We measure expert-hours.
Every agent that works a session — the lead, and any worker it dispatches — resolves to exactly one of five kinds of work, read from structure, never content. A session’s total is the sum of what each of its agents earned, so the same hours can come from new code, deletions, debugging, exploring, or coordinating a worker.
Five kinds of work.
Pick one. The same expert-hours can come from any of them — each read from structure, each sized its own way.
Creation
Net-new capability that survives — after the model discounts what got written, rewritten and thrown away.
The dispatched worker's new retry handler: 1,600 lines added -> 400 retained (25% retention) -> 8.00h.
From agents to a number.
Each agent’s kind of work is sized into expert-hours, the session total is capped by a leverage ceiling, scaled by whether the work landed, and priced at your rate. Every step is a transparent line of working.
Shown with a confidence — 82% here — so the number is honest about its uncertainty, never false precision. The 5.05–11.45h band sweeps the model’s two most uncertain knobs — retention and exploration pace. A leverage ceiling (1.8h active × 10) caps the total at 18h; this session is well under it.
Merge-conditioned only for the code-producing types — creation, reduction, repair: 1.0 merged, 0.85 linked but unmerged, 0.6 otherwise. This session’s own type is orchestration, which holds at 1.0 regardless — a PR isn’t the coordinator’s own deliverable.
The one number you set. It never leaves your workspace.
12.00× the $82.50 of AI spend — expert work returned per dollar.
A line count
is not value.
Diff size rewards verbosity and treats every change as equal. Expert-hours measure the work, not the output.
Verbosity wins — more lines reads as more work, even when the result is worse.
Churn and copy-paste inflate the count without adding any value.
Deletions and refactors look like nothing, yet are often the hardest work.
It assumes a speed-up the evidence does not support.
Sized by outcome — the hours the work would have taken an engineer.
Deletions, debugging and orchestration each carry their own hours.
Content-free and deterministic — identical inputs, identical figure.
Always a number with a confidence — honest about its uncertainty, never a false point.
We’d love your feedback.
Each kind is sized by its own transparent formula over structural signal — line counts, tokens, turns, tool calls, subagents — never the content of your code.
We attach a confidence to every number, not false precision — and we tighten it as we reconcile estimates against real outcomes. Every figure is an inspectable line of working, never a black box. If a number looks wrong for your team, tell us; that feedback sharpens the model.
The research we lean on.
Run the pipeline on
your own work.
Connect a repo and step through the real pipeline behind every figure — stage by stage, inputs to dollars.