a11oy Restraint
governed · measured · frugal
/
LIVE
Code-frugality ladder adopted from the open-source
Ponytail skill (MIT)
— governed (signed receipts + Λ) and
measured on our stack. Ponytail's published numbers are cited as theirs; ours as ours.
Less code —
—
median LOC reduction (ours)
Cheaper —
—
median token-cost proxy reduction
Faster —
—
median latency reduction
Joules saved —
—
cumulative · tokens-saved × J/token
Two-arm benchmark …
No-skill baseline vs a11oy-restraint over the same five everyday tasks.
Frugality → energy less code = fewer tokens = fewer joules
Every restraint decision saves lines → tokens → joules on our sovereign GPU. Joules are priced at the live on-box MEASURED J/token when the GPU probe is live, else our honest SAMPLE constant.
Frugality rate
—
decisions that cut code
Lines saved
—
cumulative (MODELED)
Tokens saved
—
≈ lines × tokens/LOC
The 6-rung ladder stop at the first rung that holds
Loading ladder…
Intensities: lite (name the lazier alternative) · full (stdlib/native first, default) · ultra (YAGNI-extremist). Every decision → a signed DSSE receipt + advisory Λ (Conjecture 1, <1.0).
Reproduce flip SAMPLE → MEASURED on your stack
Run the harness with a wired model client; it counts emitted LOC deterministically and records tokens + latency from the API, then writes benchmarks/restraint/results.json. This dashboard flips to MEASURED for the run you actually executed.
…
Or via promptfoo (Ponytail's methodology):
npx promptfoo@latest eval -c benchmarks/restraint/promptfooconfig.yaml --repeat 10
Ponytail's published numbers CITED · theirs
These are Ponytail's results, cited honestly — never claimed as ours. Our numbers above are measured/sampled on our own stack.