The Gold Watch

There’s a monologue in Pulp Fiction where a soldier delivers a dead man’s gold watch to his son. You know the one. The watch was bought before the First World War, worn through it, carried across the Pacific in the Second, and smuggled through five years in a Vietnamese prison camp at — let’s say considerable personal cost — by two men in succession, one of whom died holding it. “He gimme the watch.” Four words, and a small boy inherits the only object in the story that crossed every war intact.

Everything else those men owned is gone. Uniforms, rifles, letters, the men themselves. The wars consumed all of it. The watch crossed because it wasn’t of the war — it was the thing worth carrying through the war, and somebody decided, at cost, to carry it.

I think about this more than is probably healthy, because it’s the cleanest picture I have of what survives in an AI stack — and what doesn’t.

The models are the wars. Every generation of them. The one you built on two years ago is deprecated. The one you’re building on now has a sunset date nobody’s announced yet. The prompt library lovingly tuned to one model’s quirks, the fine-tune, the workflow that exploits a specific model’s specific behavior — that’s materiel. It serves in this war and it dies in this war. Companies keep discovering this the expensive way: everything they built was of the substrate, so when the substrate rotated, they inherited nothing.

The watch is the record. The thing whose entire value is its lineage. In our shop it’s the Standards Ledger: every rule in it exists because something went wrong once, got caught by a person, and got written down — dated, attributed, versioned. A rule like that doesn’t care which model is on duty. It’s plain text. It crosses substrates the way the watch crossed wars, because it was never made of the same stuff the wars consume. When the next model generation arrives, you don’t restart. You hand the new recruit the ledger and say: here’s everything the people before you learned. He gimme the watch.

This is what “substrate independence” means when you take the suit off it. Not a compliance posture. An inheritance strategy.

And underneath it sits an older discipline: memento mori. Everything in the stack dies. The models die. The vendors die. The guidance dies — this April, the regulators rescinded the model-risk framework banks had leaned on for fifteen years, and explicitly left generative AI outside the replacement’s scope.^† Expert consensus dies. Common sense dies. My own judgment has a shelf life, and so do I — which is precisely why the deliverable is designed to be left behind. A consultant who builds dependency is selling you something that dies with him. A consultant who builds a record is leaving you the watch.

So the discipline is simple and a little morbid: put a death date on everything. Every claim carries a review-by date. Every superseded document gets a banner, never a silent edit. Every signed record states what was known, by whom, on what day — because the only honest way to be right is to be right as of a date. The dangerous document isn’t the wrong one; it’s the confident one that doesn’t know it can die.

Generation is free now. The machines will produce more than any war ever consumed — drafts, decks, code, claims, most of it materiel, destined for the mud. The judgment act, the only act left that’s worth a fee, is deciding what gets carried. What goes in the ledger. What you’d smuggle through five years of captivity because your kid is going to need it on the other side.

Filter the gold watches from the waste. Carry them at cost. Write down why.

Everything else, you leave on the field.

^† OCC Bulletin 2026-13 (April 17, 2026), issued with the Federal Reserve's parallel SR 26-2: the revised interagency model-risk guidance supersedes SR 11-7 (2011) and places generative and agentic AI explicitly outside its scope, with an interagency request for information to follow. Details in Briefing #1; this note inherits that briefing's review-by date.