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

​
Context

The defend guardrail is everywhere else a stop-and-confirm with a single logged override (ADR-0015), to honor agency. The thesis defense is the highest-stakes, least-reversible material decision; should its gate be special?

​
Decision drivers

  • Agency: the human must always be able to proceed; the AI must not grade a novel claim’s substance or decide a gap is “critical” (ADR-0003, ADR-0015).
  • Stakes: waving a thesis through with known gaps should be hard and on record.

​
Considered options

  1. Uniform stop-and-confirm (single blanket override), same as other gates.
  2. Hard block until no unaddressed gaps.
  3. Heavier confirmation: per-gap acknowledged sign-off — not a hard block, but the author must explicitly acknowledge each surfaced gap in writing.

​
Decision

Option 3. The thesis gate stays non-blocking (human decides) but escalates from a single override to a per-gap logged acknowledgement.

​
Consequences

  • Waving through a thesis is deliberate and fully on record, per gap.
  • The AI never adjudicates “critical” — it surfaces gaps; the human acknowledges.
  • Slightly more friction at exactly the highest-stakes gate (intended).

​
Rejected alternatives

  • Hard block — forces the AI to decide a gap is critical and to block the human, violating agency + the non-grading rule.
  • Single blanket override — too frictionless for the biggest, least-reversible decision.
sub-spec 1 §6, §10; meta-spec §2.1–§2.2; ADR-0015.