Cost per outcome as the unit of account

Not throughput, not cache hit rate, not tokens consumed: dollars and latency per completed outcome. The cache-hit-rate case is the cleanest demonstration that the intermediate metrics mislead — a hit rate that fell alongside a 3.2x improvement in time to first token is proof the rate was never the thing being optimized. Annual compute budgets get blown in weeks because one agentic user costs 10–100x another, and per-seat capacity planning has no way to see that coming.

The unit does triple duty. It is a quality axis — best customer outcome per dollar rather than peak capability, quality per unit of latency and dollar rather than peak quality. It is what makes outcome-based pricing possible, priced per resolved ticket or per proposal and pegged to what the human job costs rather than to token cost. And it is the capacity-planning primitive that per-seat models can no longer provide.

The same reframing runs down into the hardware layer as tokens per second, per watt and per dollar, with joules per task proposed as a benchmark dimension alongside parameters and accuracy. And it runs up into architecture: cost, not capability, is what stops an agent from retrying harder — search at roughly 20 cents per dollar of generation holds agents to about one query per 5,000 generated tokens — which makes caching, model routing, compression and cheaper retrieval design decisions rather than optimizations.

Assay has none. The recurring reports count activity — pull requests, briefs, sessions — and neither dollars nor tokens are attached to any unit of work, so nothing can answer what a brief cost to land or which role is burning budget. A multi-role loop with a standing pool of parallel workers burns compute fast, and today that spend is invisible to the board (how it runs).

Three pieces, in dependency order. A per-brief ledger of cost per passed verify row — Assay’s version of cost per completed task, and what would make routing decisions evidence-based and catch a runaway loop. Counting briefs shipped is the paper-volume metric; counting verified claims per dollar is the knowledge metric. Reporting records tagged by brief type and model, a single schema addition that is cheaper than the metric it enables. And an inference budget in brief metadata, sitting alongside effort: effort is already estimated, cost is not, and the two stopped being proxies for each other some time ago (briefs).

The customization ladder — context engineering first, then LoRA, then fine-tuning — is exactly the order Assay bet on. Briefs, skills and memory are context engineering, chosen deliberately before any model perturbation; that is a shipped position rather than an aspiration. The contractive-versus-expansive distinction is the sharpest framing of why accumulated artifacts matter economically: a methodology that re-derives its judgments every session re-pays the search cost every time, while findings, skills and briefs are the shortcuts that shrink it. Assay has the artifacts. It has never measured whether they shrink anything.

• Whether cost-per-outcome numbers get published with quality attached, the way the $0.10-at-99.9% figure is — most cost claims still omit the quality half.
• Whether outcome-based pricing spreads past the ticket-deflection vertical into engineering work, which would give an external reference price for a landed brief.
• Energy per task appearing on a real leaderboard rather than as a proposal.