The four operating numbers tell you how the practice is running. A service commitment is what happens when you turn one of them around and face it outward.
A number you report is an observation. A number you commit to is something a consuming team can plan against, and something you can miss.
Committing well is mostly about committing narrowly.
Tier the estate before promising anything
A commitment across the whole estate is either dishonest or so weak it means nothing. Every organization has systems it cannot promise anything about, and saying so plainly is what makes the rest credible.
Four tiers, and most estates sort into them in an afternoon:
TierWhat it coversFreshnessChange noticeTime to evidenceGovernedBoard metrics, regulated reports, definitions agents ground onHarvested dailyNotice before promotion, with impact listMinutes, self-serveManagedDepartmental reporting, definitions with a named ownerHarvested weeklyNotice at promotionSame dayDescribedEverything else with harvested metadataBest effortNoneBest effortDarkNamed systems with no harvested metadataNoneNoneNone. Published as a list
The Dark tier is the one people want to leave out, and it's the one that earns trust. A published list of systems you make no promises about tells a consuming team exactly where to be careful, and it converts an unknown into a known gap with a name on it.
Expect the Governed tier to be small at first. Three metrics is a reasonable place to start, and a small tier you hit beats a large one you miss.
What is safe to commit to
Four commitments cover almost every real need, and each maps to something you already measure.
Freshness. How recently the metadata describing this asset was harvested. Note what this does and does not say. It says the description matches the estate as of the last harvest. It says nothing about whether last night's load was correct, which belongs to data quality tooling. Keeping that line visible in the commitment itself prevents the most common misreading.
Change notice. How much warning a consumer gets before a definition they depend on moves, and whether the impact list arrives with it. For the Governed tier this is the most valuable commitment on the page, because a silent definition change breaks a retrieval system quietly and the AI team finds out from a user.
Time to evidence. How long a provenance question takes to answer for assets in this tier. Minutes for Governed is achievable once change history is in place; promising it for the whole estate is not.
Implementation turnaround. How long from an approved definitional requirement to a bound, versioned definition compiled into its consuming systems. This is your side of the governance interface, and committing to it is what earns the reciprocal commitment on decision turnaround.
What to keep out of the commitment
Anything about the values. Row counts, null rates, load success. Adjacent discipline, different pager, and absorbing one of these makes you the owner of all of them.
Correctness of the definition itself. Governance owns meaning. You can commit that the definition in every system matches the governed one; you cannot commit that the governed one is the right business rule.
Coverage as a promise. Coverage belongs on the roadmap rather than in a service commitment. Committing to a coverage percentage by a date converts a measurement into a target, and a measurement under target pressure is how a denominator gets redefined.
Anything you cannot observe. If you have no way to detect a miss, the commitment is decoration, and the first miss that goes uncaught is the one that ends the practice's credibility.
Publish the misses
A commitment with no visible miss record reads as marketing to every engineer who sees it.
Report against each tier monthly alongside the four operating numbers: what was committed, what was met, and what was missed with the reason. A practice that reports its own misses gets believed about everything else, and the miss log is the most direct evidence you will ever have for a headcount request.
Misses also tell you when a tier is wrong. An asset that misses its Governed freshness commitment three months running belongs in Managed until the harvest is fixed. Demote it. That is better than redefining what Governed means.
Where to start
Pick the three metrics you already scored. Put them in the Governed tier, alone. Commit to two things about them, freshness and change notice, and to nothing else.
Publish the tier list including the Dark column, report against it for a quarter, then widen. A narrow commitment kept for three months buys more than a broad one announced once.
Written by the team at MetaKarta and released under CC BY 4.0: free to use, adapt, and republish with attribution.
