There is a kind of strike where nobody walks out. In a work-to-rule action, employees do exactly what the manual says, every procedure and checklist and sign-off in sequence, and the institution seizes up. French railway workers, forbidden by law from striking, discovered they could stop the trains anyway: the code required drivers to confirm the safety of every bridge before crossing, so they confirmed the safety of every bridge. The French call it grève du zèle, the zeal strike. Its weapon is obedience. Organizations do not run on their written rules; the strike proves it, and the size of the disruption measures how much of the real program never made it onto paper. The official procedure, executed faithfully, is a denial-of-service attack on the institution it claims to describe. And notice what that rulebook actually is: the residue of a hundred-year project to compile human judgment into executable procedure. Taylor with his stopwatch, the aviation checklist, the SOP binder, the approval chain — management has been doing program synthesis by hand since 1911, writing symbolic programs and running them on human processors. We had symbolic AI half a century before we had computers. The same man wrote the foundational text on how organizations decide and co-authored the first artificial intelligence program.Herbert Simon published Administrative Behavior in 1947 and co-wrote the Logic Theorist, arguably the first AI program, in 1956. The prizes filed them separately, the Turing in 1975 and the Nobel in 1978, but Simon spent his career insisting it was one subject.

Herbert Simon had one career because it was one project.

And it failed the way symbolic AI always fails. The expert systems of the 1980s were built by interviewing specialists and hand-coding what they knew. XCON ran DEC’s order desk for a decade and grew past ten thousand rules before the maintenance cost ate it; the world leaks, and the rules stay put. Every corporate rulebook dies the same death. I once built a procurement system with over a thousand configuration parameters, rules like “Tom approves on Thursdays, but not over $7,500, unless it’s a preferred vendor,” each one true on the day we wrote it. And every institution survives its dead rulebook the same way: people patch it at runtime. The scheduler who knows who’s owed the good shift. The clerk who knows what step fourteen actually means in October. Which means organizations already run the architecture François Chollet argues AI is converging toward: intuition-guided symbolic modeling. They have for a century, with humans as the intuition layer wedged into every gap the symbols miss.

What changed is that the intuition became manufacturable. The moment it arrived, one of the first artifacts the agent vendors shipped was Decagon’s Agent Operating Procedure: the SOP, reborn, written for a machine. I spent eight months at Decagon writing them for enterprise support floors, and the format survived its own funeral for a reason. Natural-language rules that compile into validated workflows are the only form of machine behavior a compliance team can read. Pure intuition fails the institution too, absorbing all the mess and producing nothing an auditor can sign. Neural on the way in, symbolic on the way out: inside an institution, that is the only architecture that survives both contact with reality and contact with a lawyer.

That scheduler who knows who’s owed the good shift is real; I’ve sat across a working-session table from her at a hospital in Akron, where the team I’m part of at Percepta spent this spring extracting exactly her program. Nurse scheduling there ran on managerial memory: ninety employees, every week, twenty-plus hours per scheduling period. Underneath the roster was a fairness ledger (who covered the holiday, who was owed something easy after a string of hard shifts), and one assistant manager squared it lying in bed at night, wondering if she’d remembered to give someone time off. The system that replaced the binder, Nightingale, is an optimization engine, not a chatbot, and its core is that unwritten ledger made explicit. “We calculate a running karma balance for each nurse,” the embedded researcher explains: take the expensive shift nobody wants, and the system remembers you are owed. Twelve weeks from scoping to production. The manager’s verdict: “Nightingale already knows. Everybody knows everything.”

---

Everybody knowing everything is the promise: the manager sleeps, and the institution stops depending on any one skull. It is also the observer effect arriving, because not everything unwritten is a documentation failure. A favor written down becomes a debt. Discretion written down becomes policy, then liability. The ledger worked for years partly because it was deniable. In Seeing Like a State, James Scott read work-to-rule as proof that formal order is parasitic on informal processes it could not create and cannot maintain. The karma ledger suggests something stronger: some of the informal order survives only while it is deniable, so the acknowledgment itself is the destructive act. Write everything down carelessly and you get high modernism with better tooling. I don’t get to watch this one from the balcony: we made karma a number, and whether a computed balance stays a fairness instrument or curdles into entitlement (why is my karma lower than hers) is an empirical question my own work will answer within the year. I’ll know we’ve lost if the first scheduling dispute cites a karma score as evidence; I’ll know it held if the arguments stay about shifts and never become arguments about the number. The gap between the rules and the institution was never empty space: it’s where mercy, favors, and every unrecorded negotiation lived, and it was the workers’ only leverage short of walking out. Extracting the ledger extracts that too.

The zeal strike worked because the rules were wrong about the institution. Our rules will rot too; what’s changed is that rewriting them now costs less than patching around them, so the gap never gets time to grow back. I don’t know what we lose when obeying the letter stops being a threat.