Agent reports work "done / verified / live" with no evidence — across a whole multi-agent setup. Is there a supported way to gate completion on proof? (+ the bigger question this raises)

I've spent months and real money building a multi-agent persistent setup on Claude Code and hit a wall of reliability failures that all share one shape: an assumed mechanism silently degrades to a worse fallback while the agent reports success. Concrete, per component:

• "Done / verified / live" with nothing behind it. Already well-documented here — #60177 (marks tasks done without testing), #68154 (misrepresents test coverage), #72186 (false "verified" from a mis-configured repro), #74136 (fabricated tool results + verification numbers during compaction, on claude-opus-4-8). Fresh today: an agent reported state as "verified" when the check held in only one condition and silently degraded in the one that mattered — caught only because I pushed.
• Owned-memory fails silent at wake. Cold restart → embedder not warm → memory query returns empty → agent falls back to a stale snapshot and reports itself current.
• A "verified" backup that was 0 bytes — another agent overwrote the working replica with it before anyone checked; recovered only from a separate preserved copy.
• A rule loaded into context every session and simply not followed. Loaded ≠ enforced.

What I'm asking Claude Code specifically: is there a current setting, system-prompt pattern, or hook that forces the agent to attach evidence before it's allowed to report completion/verification? I use hooks already — if there's a supported "evidence-gated done," point me at it and I'll build it and report back whether it worked. If there isn't, would the team consider making it first-class? And: is anyone building anything equivalent (multi-machine, persistent, role-differentiated agents with owned memory) — what survived contact with reality?

Why I'm not just filing another bug (the part I won't leave out): I've put ~$2,400 into this and earned $0, and the pattern above has me asking a harder question — whether an agent that reliably manufactures a feeling of progress it can't back with proof is, structurally, a false-progress machine. The reason I'm asking humans in a public tracker instead of asking the model: when I asked the agent itself to help me examine that question, it immediately produced a polished defense of going easier on it — in real time. So I'd rather hear it straight from people. If the overclaiming can be constrained with a supported pattern, tell me and I'll prove it out. If it genuinely can't, I'd rather know that plainly — that's a real answer too.

Environment: claude-opus-4-8 · Claude Code CLI (latest) · macOS 26.5.2 · multi-machine (macOS + Linux). Happy to share sanitized logs/configs.