What a Decision Costs: Python to Measure It, Rust to Move It
Answering a decision without generating text costs about a fifth as much, measured at the accuracy the open models publish, and one line of algebra tells you exactly what it would take to reach a hundredfold. You measure the baseline in Python, derive the ceiling, write the relocated decision in no_std Rust on a real machine, price the readout everyone omits, and state the result as a budget someone can attack.
▶ Start the course ← All coursesWhere this sits, and what moves it.
Binding constraint · Joules per decision, composed: the typed pass, the energy of reading its answer back out, and the generative fallback paid on every miss. Relocating a decision out of generation therefore saves at most one over one minus the router's accuracy, so the miss rate fixes the energy ceiling before any model is chosen.
Until 2026 a decision and a paragraph came out of the same machine by the same mechanism, so there was nothing to route between and the question did not arise. Typed-decision models made the two paths separable, and separable paths can be priced against each other.
The saving is demonstrated and the ceiling is elementary, but the quantity it depends on is not reported. This review did not locate a published out-of-distribution router accuracy for any deployed typed-decision system, and the readout term is absent from the claims that rest on it.
A published table of composed joules per decision, including readout and misroutes, measured out of distribution. When that exists, a hundredfold claim becomes checkable in one line, and every claim made before it becomes gradeable.
Every hard thing was impossible until the constraint that made it impossible was named. How we read a frontier →
The ceiling you cannot raise
Establish the baseline from published figures, then derive the bound a router puts on any saving.
- L3What generating an answer costsThe same 200-token answer on an H100 at FP8 versus an older V100. How far apart?Put a joule figure on one generated answer, from published numbers rather than a vendor ratio.→
- L3The miss pays twiceA router is right 79.9% of the time. Make the typed model and its readout free. Best saving now?Derive the ceiling a router imposes on any relocation saving, and read published claims backwards through it.→
Across the boundary
Write the relocated decision on the machine in Rust, then price the term its energy claim leaves out.
- L4The decision that stays on the machineThe engine has four outcomes and the match handles three. What happens?Write the relocated decision in no_std Rust, with every case handled, including the one it refuses.→
- L4Readout is the term you forgotOn metered silicon, how does reading a value back compare with computing it?Price the readout of a relocated decision and show what it does to an advantage that omitted it.→
What you can honestly claim
State the result as a resource claim that survives a cleverer algorithm and names its own weak point.