The last edition I sent you, back on May 12, was titled "I Wrote 25 Rules to Fix My AI. Compliance Dropped to 7%."

That headline is wrong, and I'm the one who got it wrong. The experience was real. The math wasn't. And the specific way it's wrong turns out to explain both why AI rollouts stall and where this newsletter has been for seven weeks.

Here's the error. I conflated two different numbers.

Per-rule compliance never dropped to 7%. When I measured how often my AI followed any individual instruction under pressure, it held at roughly 90%. What collapsed was something else: the probability of a fully clean run. If each soft rule holds about 90% of the time, then a run that has to satisfy six rules comes out clean about 53% of the time (0.9 to the sixth power). At thirty rules, about 4%. Every rule is still holding at 90%. The chance that any single run satisfies all of them is what dies.

To be precise about what this is: a model, not a measurement. The 90% is an assumption, and a generous one. But the shape of the curve doesn't depend on the exact number, and the shape is the point.

Why the difference matters: those are two different diagnoses, and they have two different treatments. If compliance were actually 7%, the fix would be retraining — better prompts, better people, better models. But if per-rule compliance is fine and clean-run probability is what's collapsing, retraining does nothing. The fix is structural: fewer rules, and the rules that matter moved out of instructions and into constraints the system physically can't skip. You don't ask thirty things to go right. You build so that twenty-seven of them can't go wrong.

Most AI governance docs I've seen are thirty-rule systems waiting to be discovered.

Now the seven weeks.

I went quiet in mid-May because my own operating system hit the wall I write to you about. The system I run my business on — the knowledge base, the instruction files, the automation — had accumulated exactly the kind of debt I get hired to find: duplicated instructions, content in three places with three versions, files that a maintenance process had silently corrupted while reporting success. Nine of them, discovered in one sweep.

It looked fine in the demo. It fell apart at 4pm on a Thursday.

So I did the thing I tell clients to do and secretly hoped I'd never need: stopped adding features, audited everything, and rebuilt around one rule — one source of truth per concept. The full teardown, including what the corrupted files taught me about trusting "success" messages, is going in the Implementation Lab Notebook this month.

I teach this, and it still got me. That's not an embarrassing admission; it's the entire premise of this newsletter. The gap between buying tools and redesigning work doesn't care how much you know about it.

One thing to do this week: count the soft rules in your most important AI workflow. Every "always," every "never," every "make sure to" that lives in a prompt or a policy doc rather than in structure. If you count more than seven, pick the three that actually matter and turn them into constraints — a template the work has to pass through, a validation step, a field that won't accept the wrong thing. Demote the rest to guidance. At seven soft rules and 90% each, a clean run is roughly a coin flip. Your team deserves better odds than that.

One housekeeping note: The Implementation Lane is moving to every other Tuesday. Biweekly is the cadence I can hold while doing the implementation work these editions are drawn from, and you'd rather have that than a streak that ends in another seven quiet weeks.

The compliance-decay figures above are an illustrative model, not measured data. The Implementation Lane is a biweekly newsletter about making AI work inside real organizations — the actual work, not the theory. Written by Amanda Crawford, an AI Implementation Specialist building in the gap between buying AI and making it stick. If someone forwarded this to you, subscribe here.