Skip to main content
  • Status: accepted · Date: 2026-07-17 · Deciders: Davor Runje

​
Context

Given the Understanding principle (ADR-0004), how should examination be structured, invoked, and constrained — especially since the AI can’t know the ground truth of a novel claim. Contrasted with Matt Pocock’s grill-me (same mechanic, inverted epistemics: his lets the AI recommend answers; the human is ground truth).

​
Decision drivers

  • Author wants both self-invocation and an automatic guardrail against advancing work not understood.
  • Author wants teaching (explanations + references), not just testing.
  • Must respect agency: the AI can’t fabricate a novel claim’s answer or hard-block the human.

​
Considered options

  1. One cross-cutting defend skill; targets claim|cited-work|methodology; teach- and-reprobe loop; self-invoked + guardrail; guardrail = stop/surface/offer/ record with logged override (not a hard block); non-grading on substance.
  2. A mode baked into each pipeline skill (no standalone).
  3. A grader that passes/fails understanding.
  4. Hard-blocking gate at material decisions.

​
Decision

Option 1. Teaches the established (methodology, cited work) source-grounded; never asserts a novel claim’s answer; settled-vs-contested calibration; depth by stakes; anti-Goodhart. Personas per ADR-0016.

​
Consequences

  • Consistent stance across stages; one place to maintain.
  • The guardrail records logged overrides (accountability) but the human still drives; the thesis gate escalates to per-gap acknowledgement (ADR-0021).

​
Rejected alternatives

  • Per-skill mode — duplicates the stance; harder to keep invariant.
  • Grader — grades substance the AI can’t authoritatively judge; Goodhart-prone.
  • Hard block — violates agency / human-oversight-with-override.
meta-spec §3.7; sub-spec 1 §6; digest understanding-and-defense.md.