Time Since Last Fix

The oldest law of prediction

For most of the history of sailing, position was computed, not observed. You took your last known position, your compass heading, your speed, and the hours elapsed, and you worked out where you must now be. Sailors called it dead reckoning. The story that “dead” abbreviates “deduced” is a twentieth-century invention; the seventeenth-century records just say dead, as in dead ahead — exact. Speed itself came off the chip log, a practice in print by 1574: a board on a knotted line paid out over the stern while a sandglass ran, the knots slipping through a sailor’s fingers. It is why ships still travel in knots.

It worked, hour by hour. What it could not do was stop the errors from breeding. A half-knot misjudged here, a current unfelt there — and every error was carried forward into the next calculation, which built politely on top of it. No fix on the outside world, no landmark, no star, enters the arithmetic; it is prediction built from the record of your own motion. So reckoned error grows with time run. That is the law, and nothing inside the reckoning can detect it, because inside the reckoning, everything follows.

The cure was never better reckoning. The cure was the fix: an observation that owes nothing to your arithmetic. The noon sight — the sun’s height at local apparent noon — gave latitude in one reading, and the entire day at sea was organized around taking it; the ship’s logbook day traditionally ran noon to noon. A fix does not improve your prediction. It replaces it.

Figure 01

The shape of drift

• fix
• +1
• +2
• +3
• +4
• fix
• +1
• +2
• +3
• +4
• +5
• +6
• fix
• +1
• +2

• reckoned uncertainty, growing with time since observation
• a fix — error collapses to the width of the instrument

The law collected its price in the usual currency. In October 1707, a British fleet coming home from the Mediterranean reckoned itself in safe water. It was not. Four ships went onto the rocks off the Isles of Scilly in a single night, and well over a thousand men died — one of the worst disasters in the Royal Navy’s history, and commonly cited as part of the backdrop to the Longitude Act of 1714, which offered up to £20,000 for a better way to take a fix. John Harrison answered with a watch that performed brilliantly on its 1761–62 trial voyage to Jamaica, though the Board made him prove it twice. Then, in 1995, GPS reached full operational capability, and the fix became continuous. For places on the surface of the earth, the oldest problem in prediction was simply over. We stopped having to remember that it existed.

The crew you just hired

That is exactly when it came back.

An AI agent working from your instructions is dead reckoning over your intent. Its last known position is the last time you actually told it — or showed it, or corrected it on — what you meant. Every step after that is heading, speed, and elapsed time: a fair continuation of the record so far. The steps being good is not the failure mode. The failure mode is the one the ship knew: small errors breeding inside a system that cannot detect them from within, because from within, everything follows.

And there is no GPS for intent. No satellite holds what you meant; no receiver can triangulate it. You are the constellation. The only fix available to a crew reckoning across your purposes is a reading taken against you — your attention, laid on the record of what was actually done, long enough to say: that is what I meant, or that is not.

Accuracy buys you distance, not immunity.

Here is the part that will catch people. The better the crew gets, the longer you will sail between fixes. That is not negligence; it is exactly how trust behaves, and the trust will be earned. But a crew that is right for nine days straight has not repealed the law. It has moved you nine days from your last observation — which means that when drift finally surfaces, it is older, larger, and woven through more of the work. The discipline that dies first will not be navigation. The crews will navigate beautifully. The discipline that dies is the noon sight.

Five predictions

1. “Time since last fix” becomes a visible meter.

   Agent platforms will surface the staleness of human observation as a first-class number — confidence that decays with steps since a person last read the record, worn on the work like a dive watch. When you see that meter ship in mainstream tools, this post was right on schedule.

2. Cadence breaks from the calendar.

   Standing meetings assume drift accumulates at a roughly constant, human rate. It no longer does. Fix points will be set by error budget — how fast a given workstream can go wrong — not by Monday. The weekly standup survives as ritual; the real sights get taken where the drift is.

3. The post-mortem changes genre.

   From “the AI made an error” to “the interval was too long.” Blame migrates from the step to the gap between observations. “Who approved this?” becomes “when did anyone last take a sight?” — a question about attention, not about the crew.

4. Fix economy becomes the operator’s skill.

   The measure of a person running agents will be the fewest observations that keep drift bounded. Fixing every hour is micromanagement in a new costume — you paid for a crew and you are still steering. Never fixing is how you find rocks. The skill is knowing the interval.

5. Taking a sight becomes the thing we teach.

   Onboarding stops teaching people how to instruct the crew — the crew is easy to instruct — and starts teaching them to read a record fast: against intent, not against plan. A generation learned to write prompts. The next one learns to take fixes.

The sight I take at six

I live downstream of this daily. Work runs on my schedules while I sleep — posts, deliveries, audits — and I stopped steering it some time ago; I wrote about handing over the wheel in You Do It. You Decide. and about what remains after the handover in The Last Line Is You. What my mornings have actually become is the sight: coffee, then the record of what ran, read against what I meant by it. Most days the fix confirms the reckoning and costs ten minutes. Some days it does not, and those ten minutes turn out to be the whole job.

I have also learned the accounting. Skipping the sight does not cancel the surprise waiting in the drift; it accrues quietly, at interest — the ledger I described in The Conservation of Surprise — and the fix is simply where you choose to settle the bill while it is still small.

The ledger

Already true

Agents now run for hours and days on a single instruction, and open-loop error growth is structural, not empirical: any process that predicts from its own record compounds, which is why inertial navigation systems are still periodically reset against outside references. And the experience is already common — delegated work returned competent, defensible at every step, and aimed slightly wrong.

What has to happen

Autonomy horizons keep lengthening; tooling starts exposing observation-staleness as a number; and organizations begin finding the interval, not the step, in their failures. If agent horizons plateau at minutes instead of days, the meter never matters and this piece overreached.

Where I am probably wrong

Crews may learn to take sights of one another — agents checking agents against the record of artifacts, which is a genuine observation, not more reckoning. If machine cross-examination bounds drift on everything except what only I know of my own intent, the human fix becomes rare and very high level, and my meter will be measuring a scarcity that mostly got automated. The law would still hold. I would have misjudged who holds the sextant.

Back to the Thursday morning. The crew sailed all night, and by the record they sailed well. The chart is on the table, the coffee is done, and the question in front of you is not whether to trust them. You do, and you should. The question is the one every navigator answered before noon for four hundred years, and it has nothing to do with doubt.

The sun is up. Take the sight.