Mission

Objective

Operationalize total cost per verified useful outcome across compute, retries, latency, review, incidents, intervention, and recovery.

Why this matters

Token cost alone can reverse deployment decisions when human and failure costs are material.

Scope

Included: prospective resource accounting, failed-attempt costs, task-level distributions, and sensitivity analysis. Excluded: universal monetary weights unsupported by observed use cases.

Existing context

research/frontier/records/RFR-006.md and evaluation Workstream D.

Initial hypotheses

All-in cost changes at least some system or policy rankings relative to token cost. The competing result is ranking stability across plausible weights.

Required evidence

Versioned cost schema, complete run/intervention logs, valuation assumptions, sensitivity ranges, and reconciliation records.

Constraints

Separate observations from assigned weights and preserve failed runs.

Execution instructions

Instrument costs before collecting the target runs and reconcile every run.

Deliverables

Cost schema, observed dataset, sensitivity report, and decision guidance.

Success criteria

Every included run reconciles and conclusions remain explicit across plausible cost weights.

Stop conditions

Stop if missing telemetry prevents all-in reconciliation.

Handoff requirements

Provide reusable cost fields and identify the weights that materially change decisions.