Standardization
Whether a multi-step method, longer than a single query, is run the same way across your team.
Outside the governed pipelines, not one multi-step method was run the same way by more than half the team. The reading is a share of methods rather than of people, so it holds at twelve analysts or twelve hundred.
Ask two of your analysts to build the same cohort and you will get two different numbers back, and neither of them will have any reason to doubt theirs. Standardization is the share of your multi-step methods, anything longer than a single query, that more than half the team runs the same way. Snippets spread on their own. Full methods do not, and the gap stays out of sight until two people who both say they measured the same thing put their answers side by side.
You do not know how many of your methods survive that test. You can, by the end of the day, off your own query history, for free.
What it tells us
A snippet travels easily. A join, a filter, a single-line function: people copy it without thinking, and it moves through a team like idiom. A method asks more. A cohort build, an adherence calculation, a multi-step routine that encodes a real decision about how to answer a question. It has to be adopted on purpose, and most of the time nobody does the adopting.
This vital sign is the fraction of those methods that more than half the team actually runs the same way. The zero is scoped to the ungoverned ad hoc layer, where analysts build and re-run work by hand. The governed pipelines are the exception. Models and validated jobs, with an owner, a review, and a place someone else can find them, are where methods held. Where Reuse / Build-On asks whether prior work gets picked up at all, this asks the harder question: when it does get picked up, is everyone picking up the same version?
What we found
In the ad hoc layer of one estate we assessed, not one multi-step method was run the same way by more than half the team. Everything that crossed the line was pipeline-adjacent: models and jobs someone had to write down, review, and own. Nothing built and re-run by hand made it across.
The canonical drift is adherence. Two analysts both set out to compute it, and before either settles the numerator they have already parted ways. One measures PDC, days covered. The other measures MPR, days supplied. That is a split in definition before it is anything else. Then the edges diverge too: how a stockpile is treated, whether overlapping fills are capped or counted straight through, what happens to a patient who switches therapy mid-year. Two labels read the same. The numbers underneath were never going to agree.
SELECT patient_id,
SUM(days_covered) / 365.0 AS adherence
FROM fills
GROUP BY patient_id
Counts the days a patient was actually covered over the year, capping any overlap so a stockpile is not double-counted. This is PDC.
SELECT patient_id,
SUM(days_supply) / 365.0 AS adherence
FROM fills
GROUP BY patient_id
Counts the days of medication supplied, overlaps and all, so an early refill inflates the total. This is MPR, and it runs higher than PDC.
Both say they measured adherence. One counts days covered, the other counts days supplied. The numbers disagree, and nobody notices, because both carry the same label. One of the two is the adherence figure your board saw last quarter. It was whichever analyst was free that week.
Why it matters to you
Ten analysts can build the same cohort ten slightly different ways and never notice, because the method was never written down as a method. Agreement is real at the level of syntax and assumed at the level of approach.
Method drift stays out of view until two people report different numbers for the same thing. By then the question is which one shipped, and the query history is the only place that answer lives.
A method nobody agreed on is a method you cannot defend. When a reviewer asks how adherence was computed, one agreed procedure with a written owner is an answer. A dozen private variants is not.
Give the agent fragments and it improvises the method, differently each run. Give it the method and it runs the same way every time. The variance you remove upstream is variance it cannot invent.
What good looks like
Your multi-step methods are captured and shared instead of carried in ten separate histories. The team runs them the same way, so the adherence number means one thing no matter who produced it, and the agent executes the same playbook rather than reinventing it each run. That is the difference between a good procedure sitting in one analyst's history and an institutional knowledge asset, a set of files readable by humans and machines that meets five criteria. The method your team ratified is the method that runs.
What to do about it
- Run Neuron on your query history.
- It surfaces where a multi-step method has drifted across the team.
- Put the variants side by side and agree on one.
- Promote the ratified method into your semantic model.
- Pick a recurring multi-step analysis, a cohort build or an adherence calculation.
- Ask each analyst who runs it to write down how they built it.
- Count how many run it the same way, then agree on one procedure.
- Re-read the query history each quarter to catch new drift.