Findings

We measure before we build. The mechanism's exchange rate, its verdict, its corpus and its cadence were each set by a measurement rather than a choice, and the measurements are kept with the owner's operating records, with the spend that produced them.

The measurements

Every constant in the mechanism comes from one of these. Each is our own measurement on real models with real costs, reproducible from the repository, and the negative results are published alongside the positive ones.

λ cannot be fitted to the dearest model; it can be fitted to the best fixed policy

The exchange rate between quality and dollars was first fitted between the two price extremes — always-cheapest and always-strongest. On the live pool that measures whether a stronger single model buys quality, and it does not: over three replicated draws of 124 tasks per benchmark, always-strongest minus always-cheapest was +0.010 ± 0.054 on LiveCodeBench and −0.051 ± 0.066 on HLE — both straddling zero, so the dominator flag was a coin flip and one benchmark or the other was excluded on every run. The cheap-then-escalate cascade beat the floor by +0.148 ± 0.051 and +0.139 ± 0.071 on the same tasks. Fitted to that pair — the floor and the best fixed policy — λ is 0.09 and 0.08, both benchmarks price, and the dominator flag fires in 0% of resamples. The mechanism's question is whether the best fixed policy beats the floor, not whether the dearest model does. ($2.6, 2026-09-05..07; the M3a record.)

Third-party workers are not deterministic at temperature 0

Six identical calls to one HLE task, one pinned endpoint, temperature = 0 — six distinct answers. Test-retest across draws: 12–17% of tasks score differently. Five of five candidate cheap rungs, MoE and dense, behaved the same. This is provider-side batch composition, not fixable from our side; the mechanism absorbs it by measuring the king once per slice and by ε, and a bootstrap over tasks cannot see it because it resamples tasks, not runs.

A verdict at two benchmarks has two working conditions, not four

Duelling a fixed policy against a second draw of itself cleared delta, lower bound and median, and was refused only by leave-one-out on an HLE delta of −0.015 whose sign is a coin flip. At two benchmarks the median is arithmetically the aggregate, so condition three tests nothing. Recorded as a known fragility of the N = 2 launch corpus and pinned in the test suite rather than hidden.

Corpus breadth is the binding constraint

Of the twelve agentic benchmarks the design names, two price today: LiveCodeBench and HLE. GDPval is refused (human-panel grading). R2E-Gym was removed — no between-model spread under the shipped scaffold, even after the read channel moved its apply rate from 0/162 to 7/162. SWE-bench Pro was measured on 2026-09-07: 21 tasks, three fixed policies, every episode 0 — the one-shot floor, with the gold patch grading 1.0 on the same images — so it is out of the launch corpus until a read channel against digest-pinned images exists. A third benchmark that prices is the next measurable win.

The first live window

One window of the real daemon on real tasks for real money (2026-09-07): the power gate separated the reference arms by +0.268 (lower bound +0.178, 60 tasks per arm); the king's arm scored quality 0.565 / final 0.369 over 80 tasks; one real challenger was admitted, served under its digest-derived name, duelled over 516 metered calls ($0.91, metered = recorded), and refused on all four conditions (Δ −0.213). Three arms were resumed from checkpoint at zero new calls after a wedged worker call was diagnosed and fixed — a read that trickles bytes outlives every timeout unless the deadline is enforced per chunk. Record: runs/m9-live-smoke/reveal.json.