[Proposal] New Guide: Mechanical Guards for AI Agents

[YuhaoLin2005]

Proposed Section: Mechanical Guards for AI Agents

A new guide covering a category of prompt engineering currently missing from the guide: mechanical verification patterns that compensate for an LLM's structural inability to verify its own NL output.

The Gap

Current guides cover prompt design, reliability patterns, and adversarial robustness. But there's a distinct category not yet covered: deterministic, filesystem-level checks that bypass the LLM's reasoning layer entirely — because the LLM cannot reliably judge its own configuration integrity.

Core Concept: The Prose Barrier

Any NL-based AI agent shares a structural constraint: its verification loop and generation loop use the same channel. An agent cannot independently verify its own NL output — it can only produce more NL claiming the first NL was correct. This is not a bug — it's an architectural property.

What the Guide Would Cover

1. The Prose Barrier — why AI agents structurally cannot self-verify NL output
2. Mechanical Gate Patterns — filesystem checks (mtime, regex, exit codes, hook wiring) that verify config integrity without AI involvement
3. Dual-Layer Architecture — process-layer soft monitoring + output-layer hard blocking
4. Causal Evidence — config rules causally shape agent behavior (n=30 experiment, p=0.0092)
5. Implementation Patterns — SessionStart/PreToolUse/Stop hook designs for mechanical verification

Evidence

• 50+ session production deployment
• Causal experiment: odds ratio 11.0, p=0.0092
• Independent structural convergence with Anthropic J-space (July 2026)
• ECC 226K★ repo: delivery-gate merged (#2378)

Scope

~2000 words, fits as a new guide under guides/prompts-mechanical-guards.md. Complementary to existing prompts-reliability.md — focuses on the verification layer rather than prompt design patterns.

Happy to draft the full guide if maintainers are interested.

[YuhaoLin2005]

Update (July 11): expanded evidence and implementation

The proposed Mechanical Guards chapter now has:

1. Prose Barrier formalization: The structural constraint that makes mechanical guards necessary — NL-agent verification and generation share the same channel. This is not a bug; it's an architectural property. Filesystem-level checks (mtime, regex, exit codes, hook wiring) bypass this barrier.

2. Dual-layer architecture: Process layer (soft monitoring — config-health, DIVERGE signals) + Output layer (hard blocking — quality-gate, exit 2). The boundary: can this be fixed retroactively?

3. Causal experiment: n=30, p=0.0092, odds ratio 11.0. Config rules causally shape agent behavior. Mechanical enforcement is not aspirational — it measurably works.

4. J-space structural convergence: The self-model architecture independently converges on the same functional topology as Anthropic's J-space (July 2026) — compact, causally-placed representations. Convergent evolution at the functional level, not homology.

5. Independent adoption: Components merged in ECC (226K stars), submitted to anthropics/skills as 4 independent modules.

Happy to draft the full chapter (~2000 words) covering Prose Barrier theory, mechanical gate patterns, implementation examples, and evidence.