A predicted outcome and an underwritten position are not the same object, and treating them as interchangeable is one of the most common failure modes in legal-asset finance. A prediction is a probability distribution over what a matter resolves to and when. An underwriting decision is a commitment of capital against that distribution, priced for the capital's cost, the position's duration, and the certainty of the distribution itself. Confusing the two collapses two distinct sources of error into one number and makes it impossible to tell, after the fact, whether a loss came from a bad model or a bad price.
Prediction is the first stage because everything downstream inherits its errors. A prediction states, for a matter with defined characteristics, the probability of each resolution class and the distribution of time to resolution — not a single expected value, but bands with stated confidence, built from a corpus of matters that actually reached those resolutions. The stage produces a number and a caveat: how much data supports it, how similar the training population is to this matter, and where the model's blind spots sit. An institution that receives a prediction without receiving its confidence interval and support size has received an opinion, not a data product.
Underwriting takes that distribution and answers a different question: at what price, and under what structure, does deploying capital against this position clear the capital's required return net of the position's risk and duration? This is where cost of capital, portfolio concentration, and structural protections enter — none of which the prediction stage should touch, because a model that adjusts its outcome estimate to justify a deal the desk wants to do is no longer a model. The separation is a governance rule, not a courtesy: the function that prices probability and the function that prices capital must answer to different incentives, or the probability stops being honest.
This is also why a single institution running both functions needs an internal wall as firm as the one that would exist between two counterparties. The underwriting desk is a consumer of the prediction, not a collaborator in producing it. If the desk can request a rerun with different assumptions until the number it prefers appears, the prediction has been shaped to serve a deal rather than describe a matter. Institutional buyers should ask, directly, whether the entity producing their outcome probabilities has ever revised a score after seeing what the capital desk wanted to hear. There is only one acceptable answer.
Operationally, this two-function structure does not require two separate companies, but it does require two separate audit trails. Every prediction the intelligence function issues needs a timestamp, a model version, and a record of the exact inputs it was given, captured before the underwriting function sees the number and before any capital decision is made. Without that record, a review conducted after a difficult stretch cannot distinguish a position where the model was wrong from one where the model was right and the desk priced it badly anyway. The audit trail is what makes the wall between the two functions checkable rather than merely declared.
Deployment is the act of committing capital under the terms underwriting produced. It looks operationally simple — a wire, a docket entry, a servicing record — but it is the stage where the first three stages either hold up under a live position or reveal that they did not. A deployed position is no longer a hypothetical: it has an actual counterparty, an actual jurisdiction, an actual procedural posture that will diverge from the median case the model was trained on in ways no model fully anticipates. Deployment is the point where the lifecycle stops being analysis and starts being exposure.
Monitoring exists because a deployed position is not a fixed bet; it is a claim on a process that keeps moving after capital commits. A new judge assignment changes venue-level base rates. A motion ruling changes the procedural posture the original prediction assumed. A counterparty's litigation conduct changes the settlement-timing distribution. None of these events retroactively change what was known at underwriting, but all of them change what should be believed now, and a position that is not re-scored as these events occur is being held on a stale prior. Monitoring is not portfolio reporting; it is the same predictive machinery from stage one, re-run against updated facts on a defined cadence, with drift in the score treated as a signal rather than noise.
The absence of monitoring is easy to miss because a portfolio without it can look calm for a long time. Positions accrue, matters proceed on their expected timelines, and nothing looks wrong until several matters resolve worse than priced in the same quarter — at which point the institution discovers it had no early warning because nothing was watching. A monitored portfolio produces bad news early and in small pieces. An unmonitored one produces it late and all at once.
Resolution is the stage every prior stage exists to be checked against, and it is the one institutions most often fail to formalize. When a matter concludes, its actual outcome and actual duration are known facts. Capturing them as structured labels, matched back to the original prediction, is the only way to know whether the model that priced the position was calibrated for cases like it. Skipping this step does not make the model's accuracy irrelevant; it makes it unknowable. An institution that cannot produce, for its own book, a comparison of predicted outcome distributions against realized resolutions has no basis for the claim that its underwriting is model-driven, whatever its materials say.
Resolution is also not always a single event, and treating it as one discards information. A matter can settle in stages, resolve on some claims while proceeding on others, or reach a judgment that is later modified on appeal. Each of these partial resolutions is itself a data point. A structured labeling process records the state of a matter at every material change, not only at the point the file is finally closed, so that the duration and value predictions made along the way can each be checked against what was actually known at the time, rather than collapsed into a single pass-fail grade at the end.
The loop closes when the resolution's structured label re-enters the training population the predict stage draws on. This is the step that makes the five stages a lifecycle rather than a pipeline: a pipeline processes each matter once and moves on, while a lifecycle uses every completed matter to sharpen the population the next matter is compared against. Skip the closure and the first four stages still function, but they stop improving — and worse, nobody inside the institution can tell that they have stopped improving, because nothing is checking.
None of this requires the intelligence function and the capital function to be the same entity, and there are good reasons for them not to be: capital allocation carries incentives that should never touch a probability estimate, and probability estimation carries technical judgment a capital desk should not be second-guessing deal by deal. What the lifecycle requires is that the two functions transact with each other on defined terms, at defined stages, with resolution data flowing back regardless of which entity deployed the capital. An institutional buyer evaluating a counterparty on either side of this line should ask where the wall sits, whether resolution data actually returns to the model, and how often. Those three answers describe the real structure, whatever the org chart says.