One Definition, One Owner

The three-answer problem

Meridian’s rate team, outage team, and executive dashboard all report “circuits at risk.” In a single quarter the three reported 14, 31, and 9.

None of them is wrong. They are answering different questions with the same words:

The damage is not the discrepancy. It is the meeting where the three numbers collide, everybody loses confidence in all three, and the organization concludes the data is unreliable — when in fact the data was fine and the language was broken.

Prompt first: surface the buried assumption

Here is a metric request: "circuits at risk."

Do not write SQL. Produce a definition skeleton with these fields:
grain, population (inclusions AND explicit exclusions), time basis
(window, as-of, time zone), calculation in plain words, null and
edge handling, proposed owner role.

For each field, give me the 2-3 plausible interpretations someone
could reasonably intend, and flag which one you would have silently
assumed if I had just asked you for the query.

That last clause is the whole point. It makes the agent surface the assumption it would otherwise have buried in a WHERE clause, and it turns a tool that hides ambiguity into one that exposes it.

What a definition has to contain

A metric definition is precise enough when two people who have never spoken can compute it and get the same answer. In practice that requires five things:

1. Grain — one row per what? Per circuit, per circuit-month, per circuit-and-crew-assignment. Get this wrong and every downstream aggregate is quietly double-counted.
2. Population — which rows are in scope, and explicitly which are excluded. Decommissioned circuits? Circuits energized mid-period? Customer-owned equipment?
3. Time basis — as-of when, over what window, in whose time zone. “Last 30 days” from a job that runs at 2am local means something different in March than in January if you cross a daylight-saving boundary.
4. The calculation — stated in words before it is stated in SQL, because the words are what a non-engineer can dispute.
5. Null and edge handling — what a missing value means here. A circuit with no readings is not the same as a circuit with readings of zero, and the difference changes the count.

Here is Meridian’s, written out:

METRIC        circuit_degradation_score
GRAIN         one row per circuit, per calendar month
POPULATION    energized distribution circuits only; excludes transmission,
              excludes circuits decommissioned before the period start,
              includes circuits energized mid-period (partial month flagged)
TIME BASIS    calendar month, America/Chicago, closed at month end
CALCULATION   weighted count of fault events and comms failures per
              customer served, normalized to circuit length
NULLS         a circuit with no telemetry in the period is NOT scored;
              it appears with status "no data" and is never counted as zero
OWNER         Distribution Planning Supervisor
STATUS        certified

The NULLS line is the one that would have prevented Meridian’s worst version of this. Treating “no telemetry” as a score of zero produced a list where the healthiest-looking circuits were the ones whose monitoring had failed.

One owner, and what they actually decide

The owner is not the person who maintains the pipeline. The owner is the person who decides, when Finance says the population should exclude municipal accounts and Operations says it should not, which one it is.

That requires a single name — not a committee, and not “the data team.” The data team cannot own it, because the question is not technical. Somebody with domain authority has to be able to say “municipal accounts are in, here is why, this is now the definition” and have that stick.

What the owner decides:

• Changes to any of the five elements above
• Whether a proposed variant becomes a second certified metric or gets rejected as a duplicate
• When the definition changes, whether history is restated or the break is documented and left in place

That last one is the decision people forget to make, and it is the one that causes the most pain later. Restating history makes trends comparable and makes every previously published number wrong. Leaving the break makes old numbers defensible and makes your trend chart lie at the seam. Either is fine. Silently doing one of them is not.

Certified and ad-hoc can coexist

The reflex after a three-answer incident is to lock everything down. Do not. Most analysis should stay ad-hoc — someone asking a one-off question should not need a governance process.

The workable split is two tiers with a visible label:

• Certified — has all five definition elements, a named owner, checks that run, and a service level. Appears in anything anyone outside the team sees.
• Ad-hoc — anything else. Perfectly legitimate. Labeled as such wherever it appears, and explicitly not the number of record.

The rule that makes this hold: an ad-hoc number that gets used twice by someone outside the team is a certified metric that has not been certified yet. That is your trigger to either promote it properly or stop it from spreading.

Key takeaway

Three teams reporting three numbers is a language failure, not a data failure — and an AI CLI accelerates it, because the ambiguity that once surfaced as “I can’t write this query” now resolves silently into a confident number. A definition needs grain, population, time basis, calculation in words, and null handling, with one named person who settles disputes and decides whether a change restates history. Certified and ad-hoc coexist as long as the label is visible and a twice-used ad-hoc number triggers promotion.

LEADERSHIP DECISION   name one owner per contested metric, and back that
                      owner when their call disappoints a department
PRACTITIONER ACTION   write grain, population, time basis, calculation, and
                      null handling before any query; make the agent state
                      the assumption it would have buried
SUCCESS MEASURE       the same question asked in two meetings returns the
                      same number; zero certified metrics without a named
                      owner