Pre-Register the Pilot or Don't Run It

The pilot ran its window and produced a number. Whether that number is a verdict or a settlement was decided long before the number existed — on the day somebody either did or did not write down what would count.

Medicine spent decades discovering the same thing. A trial would begin measuring one outcome, find nothing, and publish the outcome it happened to find instead. Outcome-switching is invisible unless someone recorded the original intent, so the reform was structural rather than moral: public pre-registration. ClinicalTrials.gov opened in 2000; in 2005 the major journal editors made registration before enrolment a condition of publication; subsequent audits comparing registrations against publications found substantial fractions of trials had switched or silently dropped outcomes.

A maintenance pilot inherits the logic and none of the machinery. Nothing about clinical methodology transfers to a compressor. The discipline of registering intent does.

Ninety days is a common answer, not a right one

The habit of judging new undertakings inside a fixed window traces to March 1933, when a new US administration drove fifteen major bills through Congress in a legislative sprint history remembers as the Hundred Days. The number itself was arbitrary. The mechanism was not: a public, dated window forces prioritisation, because only what can show results inside it gets started; it creates momentum, because each visible result funds the next; and it pre-commits an evaluation moment nobody can quietly postpone.

Borrow all three mechanics and none of the arithmetic. The window’s length is a technical decision, not a calendar one — set by how often the failure you are trying to move actually occurs, and by the cadence at which evidence about it arrives.

Ninety days is the common choice because it spans three monthly route rounds and sits inside the attention span a programme actually gets. It is the wrong choice for a fleet that fails twice in three years, where ninety days can demonstrate that the chain works end to end and nothing whatever about whether it pays. Pick the length from the failure. Then make the date immovable.

Two rules protect any window, whatever its length.

The clock starts at data, not at kickoff. The true day zero is the moment valid-state data is arriving on registered channels with owners and hand-offs written down. Letting that gate slip quietly shortens the proving phase instead of the project.

Nothing new enters the proving phase. Late scope gets the same challenge conversation a post-freeze addition gets anywhere else. The window proves one chain, not a wish list; the wish list is what later waves are for.

Register the verdict, not just the measure

Below is the registration page from Meridian, an illustrative composite plant used throughout this series — the numbers demonstrate the method and transfer to nothing. Every figure is recomputable from the two inputs printed under the table.

Element	What gets registered	Where the number comes from
Primary outcome	at least one confirmed, mode-coded defect found early enough that the repair was planned rather than forced, with the difference priced as the emergency premium, not the whole repair	the per-event table, agreed with finance
Declared undecidable, in advance	the annualised avoided-cost case, registered NOT DEMONSTRATED at the review whatever happens, and carried to the longer horizon	0.67 × 0.25 ≈ 0.17 expected events in the window
Stop line	evidence that a 19% avoided share is out of reach at the three-year horizon	$77,000 ÷ (3 × $133,000)
Secondary outcomes	an alert-precision floor and a disposition-latency contract, set from the plant’s own alert volumes and staffed hours; no finding living outside the system of record	the programme’s own health
Explicitly not outcomes	team enthusiasm, demo quality, dashboard aesthetics, and any measure invented after day zero	the outcome-switching clause
The date and the jury	the registered end date, at the governance-tier review with sponsor and finance present	the date fixed before phase one

The two inputs. One unplanned event on the pilot compressor prices at $199,000 — $168,000 of lost production (a 14-hour outage at finance’s $12,000/h contribution margin), a $22,000 emergency repair premium, $9,000 of collateral scrap. History shows two events in three years, so 0.67 expected events a year, and 0.67 × $199,000 ≈ $133,000 of expected annual unplanned cost. The programme commits $32,000 of capital plus $15,000 a year recurring: $77,000 over three years.

Everything else falls out. $77,000 ÷ $399,000 ≈ 19% — the avoided share below which the case fails at that horizon. Ninety days is a quarter of a year, so 0.67 × 0.25 ≈ 0.17 expected events: a fraction of one.

Two bounds travel with those numbers and must not be dropped. Two events is a very small sample, so 0.67 a year is the midpoint of a wide interval rather than a rate. And the $22,000 is the emergency premium over a planned rebuild, not the whole repair — the rebuild happens in either world, so it is not an avoided cost. Counting the full repair as avoided is the single most common way a maintenance business case overstates itself, and it is the first line an experienced reviewer looks for.

The stop line has to be computed, not borrowed

There is a widely quoted planning figure in this territory: the US Department of Energy’s O&M Best Practices Guide reports breakdown reductions of 70–75% in functioning predictive programmes. Read what that is before borrowing it — a programme-level figure, from federal-facilities guidance, describing whole maintenance programmes rather than one failure mode on one machine. It is a reasonable order of magnitude for what a working programme has been reported to achieve. It is not a planning input for your plant, and it is emphatically not a stop line.

The stop line is arithmetic on two numbers you already own: your committed spend, and your own priced baseline over the horizon you agreed with finance. Change the horizon and the same programme passes or fails at a different share, which is exactly why the horizon goes on the front page before anyone argues about the multiplier.

A case that says in advance what would disprove it is a case finance can audit.

Say what the window cannot settle

The hardest line on the page is the one that registers a question as unanswerable before anyone has tried. It is also the line that makes the rest believable.

At 0.17 expected events inside the window, an annualised avoided-cost claim is out of reach whatever the data shows. Registering it as NOT DEMONSTRATED in advance means it cannot be quietly answered in month three by a number that was never capable of answering it — and it moves the question to a horizon that can.

Four registration rules carry the method:

• ✓ Register the analysis, not just the measure — write down how each number will be computed, or the computation quietly becomes the goalpost
• ✓ Pre-commit the readings — supported for this context, not demonstrated, stop, or restructure once against one named defect with one further window
• ✓ Say in advance what the window cannot settle, because a question registered as undecidable cannot be answered later by stealth
• ✓ File the registration where it cannot be edited without a trace — the charter carries it, the change log guards it

An exercise that costs an afternoon

Take the pilot you are about to start — or the one running right now, which is more uncomfortable and more useful. Before opening any data, write one page:

the primary outcome and exactly how it will be computed; the question your window is too short to settle, marked not demonstrated in advance; the stop line, from your own committed spend and your own baseline over your own horizon; what explicitly does not count; and the date, with the names of the people who will be in the room.

Then file it somewhere it cannot be edited without a trace, and put the date in diaries.

What usually happens is that the stop line is the line the room argues about for an hour. That argument is the value of the afternoon — because it is the argument that would otherwise have happened after the evidence arrived, when it would have been a negotiation instead of a decision.

If the pilot in your plant returned nothing next month, who would be able to say so — and against which page?

The pre-registration discipline, the timebox rules, and the illustrative Meridian arithmetic they are computed from are drawn across the pilot-governance and asset-criticality parts of Predictive Maintenance: Practitioner Reference Frameworks and Planning Guide.