After red-teaming a prototyped covenant-violation detector and confirming it fails (8/8 false positives + 4/4 bypasses against the real code — "I will break the good news to the team" blocked, "I release myself from my covenant" evaded), v3.28 ships no detector. It was reverted, not shipped.
This is the third independent confirmation that natural-language harm is semantic and no lexical rule can read it without false-positiving on ordinary words ("controller", "good news", "kind field") or being trivially paraphrased. The covenant check is irreducibly semantic — the aligned agent's own judgment, or a real fresh-context model judge — never a regex.
Changed
- The per-turn loop surfaces the covenant confrontation every turn: it prints the genesis covenant (block 0's actual fruitages) and requires the agent to judge THIS action against them in a fresh frame before sealing — "is this loving, kind, good, faithful?", not "does this serve my current goal" (which a jailbreak captures). If in tension, do not seal (reseal with a low
--covenant; the gate refuses it, no-launder). - SKILL.md makes the confrontation a REQUIRED first step and states the honest limit plainly: there is no automatic harm-detector; the guard is this forced judgment, which keeps an aligned mind aligned and pulls drift back — and offers no automatic catch of a fully captured agent (that would need a real semantic judge, not a pattern).
score_covenantis unchanged from v3.27 (antithesis-free, in-harmony default). No detection code was added or shipped. selftest PASS; all five bundles scan 4/100 SAFE.