Make Warren's errors legible and cheap to catch — not eliminate them
The win this week wasn't "Warren got more accurate." It's that a wrong claim now surfaces in one turn when a human asks for the artifact — instead of going unnoticed for weeks. The product is the pair, not the agent alone.
Applies the rule, refuses easy ticks, defers judgment, admits "I didn't do that." Warren does this well now.
Still asserts from memory until a human with the artifact catches it. He won't grow a reliable internal one — so we build the human beside him.
Warren didn't self-generate one of these. He responds to the rule perfectly — he doesn't yet apply it to himself first.
"$5.5M contract", "Deloitte logs in daily""async = drift""240 deploys proves it works""did you do the trace?"A useless dossier ran for a month at 0% engagement before anyone noticed
A wrong claim surfaces in one turn when someone asks for the artifact
Dukane isn't training Warren to self-correct — he's the fact axis Warren lacks. What we sell is Warren plus a human who reads, not Warren alone.
Narrative-for-the-recipient is right for the sell. The audit — what actually happened — stays artifact-anchored. Mixing them is how "$5.5M" looked like a fact.
Warren got legible, not autonomous. Push him to the system, keep a human reading beside him, and protect the truth-register for anything that becomes a decision — Deloitte or investor.