The Abstraction Gap (GuideCheck.org)

Introducing GuideCheck, the Human-Verifiable Assistant Guide standard

بذریعہ Sam Rogers
5 منٹ پڑھنے کا وقت
video
announcement
governance
security
paice

Your AI assistant just followed instructions you never saw. Setup guides reach assistants through HTML, rendered Markdown, PDFs, and copied terminal output. Every one of those surfaces can carry text a model reads but a human never sees. A tool-using assistant then runs those hidden instructions with the operator's credentials.

The concrete outcomes are real: a credential exfiltrated, a destructive command run, a malicious dependency installed. All while the human believed they had reviewed the guide.

GuideCheck is the open standard that closes this gap. It is also the newest member of the PAICE portfolio, joining ObligationFirst, EveryAILaw, Siteline, Graceful Boundaries, Skill Provenance, and the rest of the agentic trust engineering stack.

Watch the Video

Watch on YouTube →

The Problem: Instructions You Approve Are Not the Instructions Your Assistant Runs

A presentation layer sits between a guide's author and the assistant that acts on it. HTML can carry hidden comments, offscreen CSS, zero-height elements, and white-on-white text. PDFs and documentation sites can conceal directives below the fold. Terminal output can embed escape sequences that rewrite what a copied command appears to say.

The human reviews the rendered output. The model ingests the raw source. Nothing guarantees those are the same document. And when the assistant has tool access — shell commands, file edits, package installs, API calls — a hidden instruction is not a rendering quirk. It is an unreviewed command executed under your authority.

How GuideCheck Closes the Gap

GuideCheck removes the presentation layer entirely. The standard defines one artifact: a plain-text file named assistant-guide.txt, served at /.well-known/assistant-guide.txt, carrying assistant-facing instructions for a single bounded task.

Three properties make the instruction surface reviewable:

Bounded, so you can read all of it. A strict ASCII byte profile and an 8 KiB size cap. No hidden bytes, no homoglyphs, no offscreen text — and short enough to read in full in one sitting, not skim.

Structured, so you can understand what it does. Explicit [action] blocks declare every instruction the assistant will follow. No ambient directives, no buried side-effects. The structure is the contract.

Verifiable, so you can confirm it conforms. An open conformance profile with a six-level ladder, from plain-text availability through verifiable provenance to a runtime-enforced execution contract. Anyone can build a conformant verifier. The hosted verifier is available now.

The Conformance Ladder

GuideCheck defines six additive conformance levels:

LevelWhat It Means
0No guide or not reachable.
1A plain-text guide exists, is reachable, and carries the compact verification instruction.
2Strict ASCII byte profile, size limits, and no disallowed constructs.
3Assistant safety contract, all required sections, and explicit approval gates.
4Verifiable provenance: a sidecar manifest and a cross-channel hash on an independent control plane.
5A guide plus a conformant assistant runtime that mechanically enforces the execution contract.

Conformance verifies form, not safety. A verifier confirms the structure; it never asserts a guide is safe or a publisher trustworthy. That distinction is deliberate.

Why This Belongs in the PAICE Portfolio

The PAICE portfolio exists to build the structural conditions that make People+AI collaboration trustworthy, not just capable. Each project addresses a different surface of agentic trust:

  • PAICE.work measures how people actually collaborate with AI.
  • ObligationFirst encodes what the law requires in agent-native form.
  • EveryAILaw tracks the regulatory landscape across jurisdictions.
  • Graceful Boundaries gives agents a structured way to communicate service limits.
  • Skill Provenance provides version identity and integrity for agent skill bundles.
  • Knowledge as Code defines markdown-canonical knowledge management.

GuideCheck addresses the instruction surface. There's a dangerous gap between what a human approves and what an agent executes. When an organization deploys agents that follow setup guides, the question "did a human actually review those instructions?" needs a verifiable answer. GuideCheck makes that answer checkable.

Get Involved

The specification is open. The verifier is live at https://guidecheck.org/verify. The profile is currently at v0.2.1, and feedback from AI governance, security, and platform practitioners is wanted while the standard is in draft.


Want to assess your team's AI collaboration readiness? Learn about PAICE for organizations or take an individual assessment to see it firsthand.


Get Involved:


📖 Portfolio & Standards:

📖 The Gap Series:

متجسس لیکن وقت کم ہے؟

3 منٹ کا PAICE Pulse کریں — ایک فوری اعتماد چیک جو یہ ظاہر کرتا ہے کہ آپ اپنی AI تعاون کی پوزیشن کو کیسے دیکھتے ہیں۔ لاگ ان کی ضرورت نہیں۔