When recovering outcomes one at a time is worth the wait
Recoverability-aware adaptive de-censoring: a positive backtested signal that adaptivity pays, an exact criterion for when it reverses the batch decision, and a simple operator-runnable policy that recovers almost all of the value.
Abstract
The exact de-censoring solver decides a whole recovery batch up front. But recovery is uncertain: an outcome you pay to recover can come back positive, negative, or partial, and that answer changes what you should recover next. We study adaptive de-censoring, sequencing recovery and updating the plan as answers arrive, and ask how much the option to adapt is worth. On a frozen block of twelve opened cases, 8 of 12 benefit from adaptivity and 4 exhibit a clean batch-to-adaptive reversal. We have an exact criterion that characterises which cases flip, stated in terms of what a recovery is worth when it pays, what it costs when it does not, and what the batch plan already assigns the case; the criterion itself is not published here. We then show that most of the premium is not from clever resequencing at all: a simple selective-stopping rule captured most of the adaptive value on those cases, and a small menu policy recovered nearly all of it at a fraction of the policy complexity. Prospective confirmation on live recovery channels is still pending.
The exact solver commits to a recovery batch before it sees any answers. That is the right thing to do when recovery is certain. It is not obviously right when recovery is uncertain: the outcome you pay to recover can return positive, negative or partial, and that answer changes which of the remaining outcomes are still worth recovering. Is the option to recover one at a time, and re-plan as answers arrive, worth the coordination it costs? This project measures it.
The premium exists, on opened cases
On a frozen block of twelve opened cases, adaptivity helped in 8 of 12. Four of those were clean batch-to-adaptive reversals: the batch-optimal plan and the adaptive plan disagreed on what to recover, and the disagreement was not a tie-break: the adaptive policy acted on information the batch policy had thrown away by committing early.
That is a positive result, and we treat it as one. It is also a modest one: adaptive value is bounded above by the value of full information, so a reversal count is only interesting once you know which cases reverse and why. That is the identity below.
The exact reversal criterion
Reversal is not mysterious once you have the right comparison. A case flips from batch to adaptive exactly when it sits in a two-sided band: high enough that the downside it exposes is covered, and low enough that the batch plan does not already recover it blind. We have the exact criterion that characterises this band. We are not publishing it.
What it buys operationally is the useful part. It tells you in advance which cases can possibly benefit from adapting, so you do not pay for adaptive machinery on the ones that cannot. Reversals live strictly below the point where a case is already worth recovering blind, they are cases the batch plan skips, but which become worth recovering once an earlier answer has moved the picture far enough that the downstream portfolio would use them. The criterion is exact under the recovery model of the flagship kernel. We do not claim it beyond that model.
Most of the value is in stopping selectively
The result that changed how we think about the product is this: the adaptive premium is mostly not from clever ordering. A simple selective-stopping rule, recover in a fixed sensible order, but stop the moment the remaining recoveries cannot change the downstream decision, captured most of the full adaptive value on the frozen twelve-case population, roughly nine parts in ten. The expensive machinery of full adaptive resequencing bought only the last fraction.
Better still, a small menu policy: a handful of plans precomputed in advance, switched between as answers arrive, recovered nearly all of the full adaptive value on those twelve opened cases while being far simpler to operate than a fully adaptive controller. In other words: the option to adapt is worth having, but almost all of its value is reachable with a policy an operator can run and audit. These are counts on a small, frozen synthetic population, not general fractions.
What would make this graduate
One thing, and only one: prospective confirmation. Everything above is measured on frozen opened cases under an admitted recovery law. The premium, the reversal criterion, the selective-stopping share and the menu retention are all backtested quantities. They tell us adaptivity is worth engineering and that a cheap policy captures nearly all of it, but a field claim requires running S2 against a batch baseline on live recovery channels, with the falsifier frozen before the data is opened. Until that run exists, this is a strong, honest backtest and not more than that.
Method disclosure
The method is not published.
The reversal criterion is the finding; it stays in the lab.
We publish selectively and say which. Open-source, academic and pro bono work is published in full. Where the mechanism is the commercial asset we publish the result, the evidence grade and the boundary, and state that the mechanism is withheld, and never write around the gap.
Limitations & novelty boundary
The signal is backtested on frozen synthetic and semi-synthetic populations of opened cases, not prospectively validated on operational recovery channels: the single most important caveat, and the one that would decide whether this graduates. The 8/12 benefit rate and the selective-stopping and menu-retention figures are measured on those small, frozen populations under an admitted recovery law; they are not general theorems. The reversal criterion is a two-sided characterisation, not an identity; it holds under the recovery model of the flagship kernel and is not claimed beyond it. Adaptive value is bounded above by full information, so a small measured premium can still be genuine and still not be worth the operational cost of sequencing.
References
- Blackwell, Comparison of Experiments (1951).
- Basciftci, Ahmed & Gebraeel, Adaptive two-stage stochastic programming with an application to condition-based maintenance (2019).
- Bazgan, Hugot & Vanderpooten, Solving efficiently the 0–1 multi-objective knapsack problem (2009).
- Boutilier & Poupart, Value-Directed Compression of POMDPs (2002).