Calibration is usually discussed as a statistical property: does a stated probability match a realized frequency across enough instances to matter. That framing is correct but incomplete for an institution that actually runs models in production against live capital, because a statistical property with no owner, no review cadence, and no escalation path is a fact that exists somewhere in a repository, not a control that protects anyone. Calibration has to be governed, not merely measured, or the measurement is a one-time academic exercise rather than an operating discipline.
Owning calibration means a specific function inside the institution is accountable for producing and reviewing the reliability curve on a defined schedule, and that function should not be the same team that built and tunes the model. A modeling team checking its own calibration is not a control; it is the equivalent of a trading desk marking its own risk limits. The team that owns calibration review should have the standing to restrict a model's use or require retraining, independent of whether the modeling team agrees the drift is significant.
Every model version needs its own calibration record, and a new version does not inherit the prior version's validation by default. A retrain that changes feature engineering, expands the training population, or adjusts the target definition has changed the object being measured, and a calibration claim made about an earlier version of a model says nothing certifiable about the next version until that version has been separately checked. Institutions that treat calibration as a property of a model family rather than a specific, versioned artifact are extending a claim further than the evidence supports.
Governance requires a defined response when a re-validation shows drift, not a judgment call made fresh each time a bad quarter prompts someone to look. The threshold for what counts as meaningful drift should be set in advance, before any specific result creates pressure to explain it away, and the response, retrain, restrict the model's use to segments where it still holds, or disclose the drift to counterparties currently pricing against it, should follow automatically once the threshold is crossed. A governance process that only activates after someone notices a problem informally is not a process; it is luck with a name.
Aggregate calibration review is not sufficient governance, because miscalibration concentrated in a specific jurisdiction, case type, or procedural posture can hide inside an aggregate number that looks fine. Governing calibration at the segment level means the review produces a reliability curve for each meaningful segment of the institution's actual exposure, not just for the corpus as a whole, and that segment-level curve is what determines whether a model stays in use for that segment, independent of how the aggregate number reads. An institution whose book concentrates in a handful of segments has a direct interest in exactly this level of granularity.
Disclosure is part of governance, not an optional courtesy extended to a curious counterparty. A capital partner pricing decisions against an institution's model output has a direct stake in knowing when that model was last validated and what the validation found, and an institution that only discloses calibration status when specifically asked has made disclosure reactive rather than structural. The institutional standard is proactive disclosure on a stated cadence, delivered whether or not a counterparty happens to ask for it that quarter.
The strongest form of this governance places calibration review under a party structurally independent of the function being reviewed, the same separation of duties an institution would expect around financial controls. An internal audit function, a risk committee, or an external reviewer with access to the reliability curves and the underlying resolution data can confirm the review actually happened and actually used out-of-sample, out-of-time data, rather than accepting the modeling team's self-report that everything checked out. Independence is what turns a calibration claim into a governed fact rather than an assertion.
A useful analogy is the way a financial institution treats a model risk management function, distinct from the trading or lending desks whose decisions the models inform, with its own reporting line, its own documented methodology, and its own authority to restrict a model pending review. Legal-asset institutions adopting model-driven underwriting are, in effect, importing the same category of infrastructure that banking regulation forced onto financial institutions over decades, and the fact that this market has no equivalent regulatory mandate yet is a reason to build the discipline voluntarily, not a reason to skip it until it is imposed from outside.
The cadence question deserves its own governance answer rather than being left implicit. A calibration review conducted annually will catch drift eventually, but an institution pricing capital continuously against a model cannot afford to discover a year-old miscalibration only at the scheduled review date. The right cadence ties re-validation to volume, a fixed number of newly resolved matters accumulated since the last check, as much as to a calendar date, so that a model scoring a high volume of positions gets checked more frequently than one scoring few, and the review frequency scales with the actual rate at which new evidence about the model's accuracy becomes available.
A final governance element worth naming directly is documentation of the review itself: a calibration review that produces a finding but leaves no written record of who conducted it, what methodology was applied, and what the reliability curve actually showed is not meaningfully different from having conducted no review at all, because the finding cannot be reconstructed or independently confirmed once the individuals involved have moved on to other work.
A model's calibration decays quietly by default, because the population it was built against keeps moving while the model itself does not, and an institution with no ownership structure, no versioned record, and no escalation path around calibration will find out its model has drifted from a bad quarter rather than from a review that caught it early. Treating calibration as a governed object, owned, versioned, escalated, and disclosed, is the only structure that converts a statistical property into an operating control an institution can actually rely on.