Did It Actually Stop

India Today reported on 2026-07-10 that this role "can pay up to a million dollars a year" - in an article whose body would not fetch, so whether that is base pay or total compensation is unresolved, and The Economic Times reported up to $200,000 for the same role on 2026-06-03. That spread is the story. Nobody agrees what this job is, so nobody agrees what it pays. The checkable figure is narrower: Anthropic posts no title called Forward Deployed Engineer at all, names the phrase once in a requirements bullet on its Head of Applied AI Architecture posting, and states $315,000 to $380,000 there.

The Wall Street Journal's headline for the work, on 2026-07-19, was "The hottest job in tech isn't very glamorous." You fly to somebody else's office, log into a network you do not control, and make their mess produce something. AWS has put a billion dollars into embedding engineers like this inside its customers. TCS is building a team of 8,900.

Here is the part no job ad mentions. On day two, somebody with the authority to stop the engagement asks whether you can stop the agent you just installed, and how you would know it stopped. You will answer from memory, because that is all you have. You will say press Escape. You will say the process gets a signal. You will say the server honours a cancellation.

Two of those three sentences might be true in that customer's environment, and nobody in the room can tell which two.

Across twelve chapters you build one directory, cancelcheck, and leave it behind. A dated twenty-row table of the places a deployment can be told to stop, each naming its mechanism, the page behind it, and the version that page holds from. Probes that issue a stop and record a decision rather than a duration: process alive or dead, file still growing or not, exit code, cancelled request answered anyway or quiet. A stub MCP server written to keep working after it has been cancelled. A reaper that closes an input stream, waits, then force-terminates, whose POSIX path ran and whose Windows branch has never executed. A guard for the timeout a peer can keep resetting by reporting progress. And a verify.sh the customer can re-run next quarter without you.

Now the number the book turns on, printed here and not on page nine. cancelcheck card prints fifteen rows and the tally above them reads: reached 0, not reached 0, unproved 15. Nothing proved. Every record in these pages was taken against a stand-in, a stub, a fixture or a walked trace I wrote. Nothing here has booted against a real Claude Code install or spoken to an MCP server somebody else wrote. Four rows carry no record at all. So instead of fifteen cells reading UNPROVED, the card prints every reason each row cannot carry a verdict, thirty-one reason lines. On five surfaces the book's own records contradict each other about whether the work stopped.

That zero is the argument. A second program grades the same fourteen records the way most people would, and manufactures six green REACHED cells off no live run at all. Six rows you would have handed a reviewer. The book teaches the rule that refuses them. Read an observation of the work, never the receipt, and never print a verdict from a run that proved nothing.