When Model Error Moves the Boundary
Model error becomes decision-critical when it moves a financial feasibility boundary actually reached by the optimized policy.
- Status
- Research manuscript · 2026
- Area
- Financial AI · private credit
- Evidence
- Design
- Prospective frozen holdout and separate balanced robustness study
A representation-to-decision failure.
Private-credit decision systems may optimize against encoded liquidity and contractual constraints. This study isolates a representation-to-decision failure: an optimizer can be exact for its encoded model and still select an infeasible action when an omitted stress mechanism changes a reached constraint.
Frozen 20-world holdout.
nominal-plan constraint failures under the 15-unit working-capital shock
true stressed-optimum failures in the same worlds
mean normalized achievable value under the omitted mechanism
Original prospective holdout.
After development, the environment, deterministic seed rule, analysis, primary criteria, and bootstrap procedure were fixed before evaluating the fresh holdout. A designed line-cut control removed a revolver that the selected nominal recession path did not use. The boundary-moving treatment instead introduced an omitted working-capital drain.
The designed line-cut control.
A designed line-cut control removed a revolver that the selected nominal recession path did not use.
failures under the designed line-cut control
Two constraint families.
Because the original 20-world holdout contained 19 minimum-cash worlds and only one leverage-style world, a separate source-anchored follow-up was frozen before its intervention outcomes were computed. It tests 20 fresh worlds per constraint family across a pre-specified 5/10/15/20-unit shock grid.
At the 15-unit shock, failures are 14/20 for minimum-cash worlds and 17/20 for leverage-style worlds, while the exact stressed optimum remains feasible in all 40.
failures for minimum-cash worlds
failures for leverage-style worlds
worlds in which the exact stressed optimum remains feasible
Scope of the evidence.
Frozen 20-world holdout; line-cut control; 14/20 failures under omitted working-capital mechanism; stressed optimum feasible 20/20.
Integrity boundary
The follow-up is not represented as hash-continuous confirmation of the original v1 archive because the exact original archive was unavailable for byte-for-byte comparison. It is reported separately as robustness evidence.
Not a prevalence estimate.
The environments are synthetic near-boundary mechanism tests. They do not estimate prevalence in deployed private-credit systems and omit many institutional features such as amendments, cure rights, lender discretion, sponsor support, and renegotiation.
Artifact status.
Pre-specified holdout plus separately frozen, source-anchored 40-world follow-up.
Original/follow-up are not hash-continuous; synthetic environment; no deployed-system prevalence claim.
Finance reproduction packages can be published as separate safe repositories.