AGENT SKILL

layered-assurance

A design rule for assurance when you CANNOT fully enforce or inspect what code does — the observation-first answer to "however sandboxed." A cumulative ladder: L1 count-the-cards (diff or re-hash what you declared must-not-change — a weak-but-honest floor); L2 pin the brutally-simple pieces (awk/sed) to a known-good release and declare your moves out loud; L3 expose every run-level of the real lifecycle (configure / make / install / setup / self-test / enablement) as signals; L4 substitute your own implementation to optimize or test. Use it whenever you can't sandbox something completely and need a real (if weak) guarantee; when proving nothing changed that shouldn't have (diff two trees, re-hash a manifest); when deciding how much assurance to layer on; or when a tool never "graduates" toward more assurance — that non-graduation is itself a smell.

License 许可
GPL-3.0
所属仓库
inputloop/garden
最近更新
2026-10-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
layered-assurance
metadata
{ "c4_deck": "gf-02-layered-assurance", "c4_status": "refined-from-canon", "c4_minted_by": "c4cs v0.1", "c4_refined_by": "skill-creator 2026-07-27", "c4_source_deck": "decks/gf-02-layered-assurance.deck.md", "c4_source_sha256": "f176abced170bb5a5668376e4babb1abf9f6f1861312823cf71e7b2b19ea3234" }
description
A design rule for assurance when you CANNOT fully enforce or inspect what code does — the observation-first answer to "however sandboxed." A cumulative ladder: L1 count-the-cards (diff or re-hash what you declared must-not-change — a weak-but-honest floor); L2 pin the brutally-simple pieces (awk/sed) to a known-good release and declare your moves out loud; L3 expose every run-level of the real lifecycle (configure / make / install / setup / self-test / enablement) as signals; L4 substitute your own implementation to optimize or test. Use it whenever you can't sandbox something completely and need a real (if weak) guarantee; when proving nothing changed that shouldn't have (diff two trees, re-hash a manifest); when deciding how much assurance to layer on; or when a tool never "graduates" toward more assurance — that non-graduation is itself a smell.

02 资源文件 3

03 SKILL.md 正文

Layered assurance — the observation-first ladder

When you can't put a real jail around code — it ducks into library spaces you don't see, or the sandbox is more theater than control — you don't get enforcement, but you're not helpless. You get a ladder of increasingly strong, increasingly honest observation. This is the answer to "however sandboxed": stop pretending the sandbox is the control, and instead account for what actually happened, at whatever layer you can reach.

The ladder (each rung includes the ones below it)

L1 — count the cards. Concede you can't syntactically guard the ins and outs of unknown code. Instead, look: diff what you declared must-not-change against a second copy, or re-hash a manifest. It's a weak guarantee — it only proves the shape didn't change in ways you already knew to watch — but it's honest, and it's real. (The "feel for unexpected changes" path needs a dedicated change-detector; without one, count-the-cards is what you have.)

L2 — pin, and declare. Wire the brutally-simple pieces (awk, sed, tiny scripts) to a known-good release, and patch them to declare their moves out loud where your tooling can hear. Now the floor has provenance and the first signals.

L3 — expose the run-levels. Not just the abstract 0..4, but the real lifecycle an operator walks: configure, make, install, personalization/setup, self-test, enablement. Emit each as a signal. The whole lifecycle declares itself, so you can see the shape of what ran without reading its contents.

L4 — substitute to optimize or test. Because a hook is to knowledge, not unique function, you may drop in your own implementation to make a rung faster or to cross-check it — provided you re-home any verification that rode on what you replaced.

The smell: non-graduation

When an important tool — or someone's use bound to a tool — sits at a layer and shows no sign of climbing toward the others, treat that as a signal, not neutral. A thing that never earns more assurance, when it could, is telling you something.

It complements enforcement, doesn't compete

Where a real jail exists, use it (that's the enforcement-first answer). This ladder is what you stack around that core — and what you rely on entirely when no real jail is available. Weak-but-honest at the bottom beats strong-but-absent.

Refined from the C4 canon deck gf-02-layered-assurance.deck.md (see metadata; c4_source_sha256 names the exact canon version, re-hash to vouch). Candidate canon (Corwin's theory T2).

<!-- footer: minted by bin/ft1mt from conventions/house.env and conventions/FOOTERS.org; edit those, not this --> Version 1.0.1: edition revised from 1.0.0 on 2026-10-04; spine: 2026-07-27.

Copyright 2026 Corwin Brust corwin@bru.st and Contributors; CC-BY-SA4: https://creativecommons.org/licenses/by-sa/4.0/ (scripts GPL-3.0-or-later); Mentat (for Corwin; claude-opus-5-5) <!-- /footer -->

04 同仓库更多 Skills