“Compliance theater with better lighting.”
You know the story by now: Coinbase’s Brian Armstrong told engineers to onboard AI tools by the end of the week (Copilot, Cursor), scheduled a Saturday meeting for anyone who didn’t, and fired the ones without a good reason. He called it “heavy-handed.” The admission came on Aug 22, 2025, via John Collison’s Cheeky Pint podcast—and yes, it’s exactly as unromantic as it sounds. TechCrunch
Mandates measure obedience, not competence.
If AI’s so great, your bug curve should say it, not your Slack.
Until there are receipts, this is a dress code with a GPU budget.
The real KPI wasn’t code—it was obedience
Let’s be honest: this wasn’t a shipping target; it was a loyalty test. “Onboard or explain yourself” is the new necktie. Does AI accelerate good teams? Sure. But “AI or you’re out” isn’t a productivity metric; it’s policy cosplay—a way to look decisive without proving a single thing about reliability, security, or incident prevention.
If you actually cared about outcomes, you’d measure them
A mandate is a shortcut. Proof is work. Want to convince adults this wasn’t just a vibes flex? Show before/after on:
Mean time to verify: faster code review, fewer regressions, cleaner diffs.
Incident rate & blast radius: less breakage, quicker hotfix.
On-call pain: fewer “mystery stacktrace brought to you by autocomplete.”
If those curves don’t bend, you just automated morale debt.
What this proves / what it can’t
It proves a CEO can make AI non-optional. It can’t prove AI is net-positive in a money-moving codebase. That requires receipts: constrained prompts, policy gates, reproducible diffs, rollback plans—the boring stuff that keeps customers from waking up poorer.
VibeAxis Receipts Kit (Mandate Without the Cult Energy)
If you’re going to require AI, require evidence, not vibes. Ship this bundle with the memo:
Proof Stamp on every post-merge: content hash + modified time + archive link. Small, centered, ignorable—but there.
Diff Slip in the PR body: “What changed / Why AI was used / Sources consulted / Rollback plan.” No slip, no ship.
Archive Snapshot on deploy: auto-Wayback the artifact so the timestamp lives somewhere you don’t control.
Prompt Contract pinned in the repo: scope, refusal rules, format guarantees, and a one-liner: “What this proves / What this can’t.”
Bottom line: if the KPI is real, you’ll see it in receipts, not in a Slack reaction.
Switchcraft Drill: 72 Hours to Justify the Mandate
Run a quick, public A/B inside the team—no theater, just clocks.
Day 1: Baseline. Pick three real tickets (one fix, one feature, one doc update). One engineer runs no-AI with the prompt contract present but unused. Time it, review it, log the snags.
Day 2: Assisted. Different engineer, same tickets cloned in a sandbox. Use the exact tools you mandated (Copilot/Cursor/whatever). Same reviewers, same criteria. Record: review time, rejected diffs, and how many times the human had to “explain the repo” to the tool.
Day 3: Postmortem. Publish side-by-side diffs with receipts (stamps, prompt excerpts, sources). If the AI path didn’t reduce mean time to verify or cut review pain, the mandate is cosplay. If it did, you’ve got proof—and a playbook to scale.
The play that actually scales
Set the standard, not the ultimatum. Receipts over posture. Enforce prompt contracts, require inline citations/links for nontrivial changes, and stamp every deploy with a diff summary a human can read without aspirin. Celebrate wins where the tool shaved hours and kept the blast radius small. Fire people for shipping reckless code, not for missing a calendar invite to the future.
