# Pre-Estimation Checks

Read this page before fitting RHIP. The checks tell you whether feature matching
can support reward recovery for the planning horizon you intend to use.

Most RHIP failures are visible before optimization starts. Check feature
variation, support, transition validity, and normalization before treating a
fit as reward recovery.

| Check | Why it matters |
| --- | --- |
| Feature rank | Rank below the number of reward features means some reward directions are unidentified. |
| Feature condition number | A high condition number signals unstable feature matching. |
| Action dependence | Multi-action reward recovery needs features that vary across actions. |
| Transition row sums | Occupancy measures require valid probability rows. |
| Transition orientation | The tensor must be `(n_actions, n_states, n_states)`. Wrong orientation can have valid row sums and still produce the wrong occupancy measure. |
| State coverage | Unobserved states weaken the occupancy comparison. |
| Action support | Rare actions make action-specific rewards weakly pinned down. |
| Reward anchor | RHIP rewards need a normalization for level and scale. |
| Planning horizon | The horizon `H` sets the behavioral model the estimator matches. Choose it for the demonstrator-noise regime. It is a planner hyperparameter, not an identifying assumption. |

## Canonical Simulation Checks

Values from the primary high-dimensional route-choice synthetic run
(see [Simulation Study](validation.md)), 150-node random geometric road
network, features `[edge_cost, amenity, goal]`, true parameters `[1.0, 0.5, 1.0]`:

| Check | Value | Status |
| --- | ---: | --- |
| Feature rank | 3 / 3 | pass |
| Feature condition number | 2.46e+01 | pass |
| Action-contrast rank | 3 / 3 | pass |
| Transition row error | < 1e-10 | pass |
| Observed states | 150 / 150 | pass |
| Transition orientation | (4, 150, 150) | pass |
| Action support | 4 actions, 400 agents | pass |
| Reward anchor | true | pass |

## Common Risk Patterns

A state-only feature matrix can produce a plausible reward vector while still
failing to identify action-specific payoffs. Sparse demonstrations can match
the dominant actions while leaving rare-action rewards weak. Transition tensors
with the wrong orientation can have valid row sums and still produce the wrong
occupancy measure. Choosing a horizon much higher than the demonstrator's
actual lookahead does not fail silently, but the extra soft backups add cost
without improving recovery.