There are two easy but wrong postures available to a platform that models legal outcomes, and both are more common in this market than the harder, correct one. The first is claiming realized-dollar prediction everywhere, immediately, because the underlying probability models look sophisticated and the temptation to describe their outputs in the language capital allocators actually want to hear is constant. The second is declaring the claim permanently impossible, treating any future realized-outcome prediction as a category error no amount of evidence could ever justify. The correct posture sits between these, and it is less comfortable than either: the claim activates conditionally, asset class by asset class, only once specific evidence exists for that specific class.
The reason the current fleet does not make a realized-dollar prediction claim for most asset classes is not caution for its own sake. It reflects what the corpus actually contains for those classes: labeled distributions built from settlement and award bands drawn from court and regulatory records, not a dataset matched to the actual realized dollar returns a specific capital structure received against a specific position. A settlement band tells an institution what similar matters have resolved to. It does not tell an institution what a fund with a specific advance rate, a specific fee structure, and a specific portfolio position actually received back, because that second dataset, counterparty-side realized returns, is a different and much harder thing to obtain.
What activates the claim for a given asset class is specific and demanding: a model trained on counterparty tapes, meaning actual realized-dollar records supplied by capital partners who deployed real positions in that asset class, evaluated through a preregistered backtest against matters the model had not seen. Absent that specific evidence, for that specific class, the claim stays inactive, regardless of how well the platform's outcome-probability models perform on the classes where they have been validated. Strong performance on one measurable task is not evidence for an unmeasured, different task, however related the two appear from the outside.
Preregistration is what prevents this bar from being cleared by accident. The backtest's population, its cutoff date, and its criteria for success have to be locked and recorded before any result is generated, so that the evaluation cannot be quietly adjusted after an unfavorable early result to find a framing that passes. A backtest designed after seeing preliminary results is not a test of the model; it is a search for a description of the model's output that happens to look successful, and preregistration is the specific discipline that closes that door.
Activation has to happen per asset class rather than across the platform as a whole because the dynamics that separate a labeled distribution from a realized-dollar outcome differ meaningfully by class. A mass tort settlement, resolved through a structured allocation process across thousands of claimants, converts into realized capital-side dollars through a completely different mechanism than a single-plaintiff personal injury settlement paid directly against an advance, which differs again from an insurance subrogation recovery moving through a different payment chain entirely. Evidence that a model correctly bridges labeled distributions to realized dollars in one of these classes says nothing about whether it does so in another, and treating it as though it does is exactly the overclaim this discipline exists to prevent.
Once a claim activates for a specific asset class, it should look narrow and specific, not broad and celebratory. It should state the asset class precisely, reference the backtest that supports it, disclose the size and date range of the counterparty tape behind the evaluation, and stop there, rather than extending, by implication or by marketing convenience, into adjacent classes the same evidence does not cover. A properly activated claim reads like a footnote with teeth, not like a headline.
This discipline protects the institution over a longer horizon than the alternative postures do. A platform that overclaims broadly and is later shown wrong in one asset class loses credibility across every claim it has made, because the failure reveals that the claims were never actually class-specific in the first place. A platform that activates claims narrowly, one asset class at a time, with each backed by its own preregistered evidence, can absorb a finding that one class's backtest fell short without that finding contaminating the classes where the evidence genuinely held. Isolation of evidence is isolation of risk to the institution's own credibility.
This framework also gives an institution an honest way to describe its own roadmap without overpromising a timeline. Rather than stating that realized-dollar prediction is coming for a given asset class by a certain date, an institution can state the specific precondition that has to be satisfied, a counterparty tape of sufficient size and quality, a preregistered backtest of a defined design, and report progress against that precondition as it accumulates. This is a more honest form of forward guidance than a date-based promise, because it ties the claim's activation to evidence rather than to a calendar an institution cannot actually control, and it gives capital partners a concrete way to track how close a given asset class actually is.
Institutions evaluating a vendor's claim in this area should also ask what happens to an activated claim if a subsequent, larger backtest produces a weaker result than the original preregistered evaluation. A claim that can only strengthen and never weaken as new evidence arrives is not being tested honestly, and a vendor's willingness to disclose a deactivation, should the evidence later require it, is itself a meaningful signal of whether the original activation was earned rather than asserted.
The absence of a realized-dollar prediction claim across most of the current fleet is not evidence the platform has failed to build something. It is evidence the platform has set a bar for that specific claim and has not yet cleared it for most classes, which is a different and more defensible position than either overclaiming to meet market expectations or retreating into a permanent, evidence-proof refusal. The bar does not move to accommodate either impatience or pessimism. It moves only when a specific asset class produces the specific evidence the claim requires.