# Data audit (P1)

| file | SHA-256 | rows | cols |
|---|---|---|---|
| data/model.csv | 098123961451e4f9a2e76afdbc83aec78066fc713d0645659c09b77b1636195b (matches bundle.json) | 10 | 3 (parameter, value, description) |

Parameters: alpha=0.36, beta=0.95, y_min=0.05, y_max=5.0, grid_points=400, grid_spacing=log, vfi_tolerance=1e-08,
eval_y_1=0.5, eval_y_2=1.0, eval_y_3=2.0. Units: y and c in units of output; tolerance in utils.
No missing values, no duplicated parameter names, no dates (theoretical parameters). `value` column is mixed
(numeric + the string "log"); parsed per-parameter.

Anomalies / checks:
- 0<alpha<1, 0<beta<1: OK.
- None of the evaluation points is a grid node (log step = ln(100)/399 ≈ 0.01154); the VFI policy is linearly interpolated
  as the task allows.
- Grid invariance: closed-form y' = (alpha*beta*y)^alpha ranges over [0.231, 1.213] for y in [0.05, 5] — strictly inside
  the grid, so the truncation of the state space is not binding (verified in analysis.py: diag_closed_yprime_min/max,
  diag_policy_edge_hits = 0).
- No information-availability violations: only model.csv is used; no rows excluded.