# Pre-Estimation Checks

UFXP inverts empirical choice probabilities state by state, so its
pre-estimation risks are CCP's plus the usual linear-utility ones. Check these
before fitting:

| Check | Why it matters for UFXP |
| --- | --- |
| Feature rank | A rank-deficient design leaves a direction of theta undetermined; the closed-form solve flags it through `converged_`. |
| Feature condition number | Ill-conditioning inflates the variance of the closed-form solve. |
| State coverage | Conditions are scored only at visited states; unvisited states drop out, and very thin coverage leaves few usable conditions. |
| Action support per state | A state where one action is never taken makes the log-odds inversion degenerate at that state. |
| Transition row sums | Transition tensors must be row-stochastic in the `(n_actions, n_states, n_states)` orientation. |
| Reward normalization | One action's utility should anchor the normalization (the reference action). |

## Canonical Simulation Checks

Values from the canonical synthetic run (see [Simulation Study](validation.md)):

| Check | Value | Status |
| --- | --- | --- |
| Feature rank | 4 / 4 | pass |
| Feature condition number | 4.51 | pass |
| Observed states | 21 / 21 | pass |
| State-action coverage | 1.000 | pass |
| Minimum action share | 0.325 | pass |

## Common Risk Patterns

Concentrated panels are the pattern to watch. When most trajectories visit a
narrow corridor of states (a gridworld walked corner to corner, a mileage
process that rarely runs high), the choice-probability estimates at the edges
are noisy or absent. The optimal weighting handles this gracefully — thin
states are downweighted by their sample share rather than trusted — but no
weighting can recover information the data never carried. If coverage is thin
everywhere, prefer NFXP, which pools all observations through the likelihood.