Learned Hand's 1932 opinion turned on six words most readers skip: new and available devices. The tugs that lost their barges could have carried a weather radio; almost none did; and Hand refused to let the industry's own custom set the standard of care, because a whole calling may have unduly lagged in the adoption of new and available devices. The sentence has been a sword against custom for ninety years. It also has a hinge that has never turned for software, and the hinge is the word available. A calling cannot lag behind a device that does not exist.
For oversight of what a program means, the device did not exist, and the reason was not a shortfall of engineering effort. In 1953 Rice proved that no machine can decide a non-trivial property of what another program computes โ and a software monitor watching an agent for misbehaviour is exactly that machine, asking exactly that question, sharing the exact failure domain of the thing it watches. Its certificate is a story about a story. So for ninety years the deployer held Hand's own clause as a shield: you cannot be faulted for declining a monitoring device when the device is foreclosed by a theorem. Non-use was not a lag. It was the honest report that there was nothing to adopt.
The shield holds exactly until the device exists โ and the device is not the monitor Rice foreclosed. That monitor asked the forbidden question, was the output good, the non-trivial property of meaning no machine can settle for every program. The new device asks a different question and signs only that: not whether the work was good in the world, but where it landed against the authorised spec. Two finished documents โ a spec already written, a work already produced โ dropped onto a fixed lattice by a walk that takes a handful of steps and then halts, every time, on every input. Semantic position is physical position: the answer is a cache hit, a one-cycle byproduct of the memory geometry, not a recursive computation that can loop forever for the undecidability to hide in. The work the agent did stays exactly as unjudgeable as Rice says it is; only the placement is signed, and a stranger can recompute the same bits offline without trusting the team that produced them. The new device does not break the theorem. It steps off the board the theorem plays on. The fence is the device: it decides the decidable half and refuses the rest out loud.
The instant that device is available, Hand's clause turns the other way. The same six words that shielded the deployer who had nothing to adopt now name the deployer who has something to adopt and declines. A whole calling may have unduly lagged. A Caremark court does not ask whether the decision was wrong; it asks to be shown the system that was watching. For ninety years the honest answer was that no such system could exist. The answer has a filing date now โ US 19/637,714 โ and a filing date is what turns available from an argument into a public fact. After it, we had no way to monitor this stops being true on the record, and the oldest defence and the newest one collapse into one: everyone else is doing it and no instrument existed were the same sentence, and the instrument's arrival ends both.
So the ninety-year arc closes on itself. 1932 named the duty to adopt the available device and left available as the open term. 1953 proved why, for software meaning, the term stayed empty โ and handed the deployer a shield built from the gap. The receipt fills the term, below the line the theorem reaches, and the shield becomes the sword Hand wrote it as. Nothing about the doctrine changed. The device arrived, and the doctrine finished its own sentence.
You give: The shield โ no instrument existed, so non-use was not negligence. You get: The sentence Hand left open, now closed against you โ the available device exists, and the calling that declines it is the one he named.