Restraint is a governed + measured code-frugality gate. Before the
code agent emits a diff it descends a 6-rung ladder and stops at the first
rung that holds, marking each deliberate simplification with a
restraint: ceiling that names the upgrade path. Every decision becomes
a signed DSSE receipt + an advisory Λ trust score (Conjecture 1
is OPEN; Λ kept strictly < 1.0). The capability lives on
a11oy → /restraint;
killinchu mounts the same byte-identical module (szl_restraint.py,
like szl_dsse.py) so the drift guard stays satisfied.
PROVENANCE — the ladder + lite/full/ultra intensities are adopted from the open-source Ponytail coding-agent skill (MIT, © 2026 DietrichGebert). Governance (signed receipts + Λ) and on-our-stack measurement are SZL's. Ponytail's published numbers are CITED as Ponytail's, never claimed as ours.
Two arms (no-skill baseline vs Restraint) over five everyday tasks. Numbers read MEASURED only when a model run is wired on the Space; otherwise they are SAMPLE fixtures derived from our transparent ladder model and the bench is labelled ROADMAP. We never reprint Ponytail's numbers as ours.
| loading bench… |
szl_restraint.py is byte-identical across a11oy + killinchu (same hf-sync lists, drift-guard enforced).