2026-04-17
The rule that checks the rules
There's a gap in most AI-assisted workflows that nobody talks about because it's invisible until it isn't.
You write the rules. The AI reads them. The AI does the work. You review the output.
The gap is between step two and step three. The AI read the rules. Did it follow them? Usually you find out by noticing something's wrong in the output, or by not noticing at all.
I've been building a set of governing instructions for how I work with AI — across three repositories, covering financial model design, product development, and SQL query patterns. The instructions are specific. They cover edge cases. They define governance gates, security requirements, testing standards, and the rules for when AI can act versus when it needs explicit human approval.
They're good instructions. The problem is that nothing verified they were being followed.
Every other layer in the system has a check:
- Financial outputs have cross-validation gates — revenue reconciled to source, totals reconciled to inputs
- Code has tests — known inputs, verified expected outputs, run automatically on every change
- Assumption changes have four-eyes review — who changed it, when, what the old value was, what the business reason was
- Agent outputs require human sign-off before touching any reportable figure
The one thing without a check: the AI's own work within a session. A set of rules is only as strong as its enforcement mechanism.
What I built today is a compliance review — a structured check the AI runs against its own instructions after completing work in a session. Three finding types: Pass, Gap, Advisory. Gaps block the action they precede. Advisory findings are informational. The AI cannot clear its own findings — that's the critical rule. An agent that can self-clear its own gaps has no governance. The person who wrote the work cannot sign off on the quality gate for that same work.
The trigger design took real thought.
The weakest option: only run the review when you ask for it. That's a control that only works when you remember to invoke it. The sessions where you're most likely to have missed something are the sessions where you're too deep in the work to stop and ask.
The strongest option: run it automatically before every significant write — every change to an instructions file, every journal entry, every production output, every assumption table change. These are the actions where a gap that slipped through becomes permanent. The review runs first. If there are gaps, you clear them before the write happens.
The middle option: offer it at the end of sessions where something substantive happened — a decision was made, a spec was discussed, research concluded. Not forced. Offered.
I combined all three. Mandatory before high-stakes writes. Offered at session end when the session was substantive. Always available on demand as a baseline.
The subtle rule: if the mandatory tier already ran in a session, the offer tier doesn't additionally get made. You don't check a thing twice for the same reason.
The honest version of this: I'm applying to AI-assisted work the same principle I apply to financial work. A number without an audit trail is an unauditable number. A process without a verification gate is an unverified process.
Most people treat AI sessions as ephemeral — the conversation ends, the context evaporates, and whatever the AI produced either works or doesn't. That's fine for casual use. For a financial platform that will handle real business data affecting real decisions, it's not enough.
The rules exist. Now there's something checking whether they were followed.
The compliance review ran on this session's own work before the session closed.
Everything passed.
That felt like the right way to end the day the system was introduced.