Before the a11oy Code agent emits a diff, it descends a 6-rung ladder and stops at the
first rung that holds — (1) YAGNI, (2) stdlib, (3) native platform feature,
(4) already-installed dependency, (5) one line, then (6) minimum viable code. Deliberate
simplifications are marked with restraint: ceiling comments that name the upgrade path.
Every decision becomes a signed DSSE receipt and an advisory Λ trust score; the
benchmark reproduces code-reduction / cost / speed numbers on our stack, labelled
MEASURED only when actually run, else SAMPLE/ROADMAP.
Type what the code agent is about to build. Restraint runs the pre-write ladder, names the ceiling, scores Λ, estimates lines/tokens/joules saved, and signs the decision.
Run the ladder to see which rung stopped it and why.
The governed differentiator: every restraint decision is a tamper-evident DSSE receipt. Verify in-browser against /cosign.pub.
Less code → fewer tokens → fewer joules on our GPU. Estimates are MODELED; the joules label defers to the on-box NVML meter (reads sample unless a fresh exporter reading is present).
Ported from Ponytail's promptfoo methodology (MIT): the same five everyday tasks, two arms, median reported. Numbers are SAMPLE/ROADMAP on this Space (no model run wired here) and flip to MEASURED only when a model run is supplied. We measure our own; we never reprint Ponytail's numbers as ours.
| Task | Rung | Baseline LOC | Restraint LOC | LOC −% | Cost-proxy −% | Latency −% | Label |
|---|---|---|---|---|---|---|---|
| Run benchmark to populate. | |||||||
/api/a11oy/v1/restraint/{evaluate,bench,info}.