mangrovedigital/tide-engine-api

1# ===== LEGACY / EXPERIMENTAL — NOT SHIPPED =====
2# Part of the superseded engine.core research stack (neighbour-transfer
3# nowcast). NOT imported by the product path (api_server / point_forecast /
4# engine.forecast). Kept for reference only. See LEGACY.md.
5"""Evaluation harness: grade baseline (astronomical only) vs corrected against
6real held-out observations on a common set of valid points.
7"""
8from . import linalg
9
10
11def grade(engine, block, mode="primary"):
12    """Return metrics dict comparing baseline vs corrected on `block` (aligned).
13    Only points with a real observation AND a producible correction are scored.
14
15    (Repaired 2026-07-01: was calling engine.predict_block, an API that no
16    longer exists — core.Engine exposes correct_block. Caught via Manus review.)
17    """
18    astro = engine.harm.predict(block["t"])
19    corrected, _layers = engine.correct_block(block, mode=mode)
20    obs = block["observed"]
21    b_pred, c_pred, truth = [], [], []
22    held = 0
23    for i in range(len(obs)):
24        if obs[i] is None:
25            continue
26        if corrected[i] is None:
27            held += 1
28            continue
29        truth.append(obs[i])
30        b_pred.append(astro[i])
31        c_pred.append(corrected[i])
32    if not truth:
33        return {"n": 0, "held": held, "error": "no scorable points"}
34    base_rmse = linalg.rmse(b_pred, truth)
35    corr_rmse = linalg.rmse(c_pred, truth)
36    base_mae = linalg.mae(b_pred, truth)
37    corr_mae = linalg.mae(c_pred, truth)
38    return {
39        "n": len(truth),
40        "held": held,
41        "baseline_rmse_m": base_rmse,
42        "corrected_rmse_m": corr_rmse,
43        "baseline_mae_m": base_mae,
44        "corrected_mae_m": corr_mae,
45        "rmse_skill": 1.0 - corr_rmse / base_rmse if base_rmse > 0 else 0.0,
46        "rmse_drop_m": base_rmse - corr_rmse,
47    }
48