scripts/analyze_concept_drift.py

# ruff: noqa: T201  # CLI build-script report payload (CLAUDE.md whitelist)
"""Concept drift analysis across LXX / NT / Vulgate strata.

Implements two complementary methods from Schlattmann & Vogl 2024
("Trajectories of Change: Approaches for Tracking Knowledge Evolution",
arXiv:2501.00391), adapted from time-sliced corpora to biblical transmission
strata (LXX Greek → NT Greek → Vulgate Latin).

Method 1 — Crystallization index (EDE intra-stratum):
  mean_cosine_to_centroid per concept per stratum, already in concept_trajectory_sp.
  Paper's Method 2 analogy: density at the centroid using the stratum's own documents.
  High = tight cluster (crystallized); low = diffuse (polysemous).

Method 2 — Cross-stratum density at reference centroid:
  For a concept C transitioning from stratum A → B, compute the average cosine
  similarity of B's verse embeddings to A's centroid.

  cross_density(A→B) = mean_cos(emb_B, centroid_A)

  High → B verses echo A's semantic centre  (concept preserved)
  Low  → B verses are far from A's centre   (concept drifted)

Consolidated 2026-06-23 into NuBerea/concept-analysis:scripts/ from the retired
NuBerea/concept-trajectories repo so build.py resolves this import in-tree.
The three pure functions (cosine_sim, cross_density, compute_drift_dataframe)
are imported by build.py; analytic logic is unchanged from the original module.

Inputs (HF-native via load_hf_or_local with sibling-build-dir fallback):
  NuBerea/concept-analysis  concept_trajectory_sp   — 190 rows
  NuBerea/concept-analysis  concept_verse_edge_sp   — 59,478 rows

Output: writes concept_drift_analysis.csv to the resolved data dir, or to
--out if specified.

CLI:
  python analyze_concept_drift.py [--out /path/to/output.csv] [--quiet]
"""

from __future__ import annotations

import argparse
import pathlib
import sys

import numpy as np
import pandas as pd

BUILD = pathlib.Path("huggingface/candidates/build")
DATA  = BUILD / "concept-analysis/data"
DEFAULT_OUT = DATA / "concept_drift_analysis.csv"

# Make hf_registry importable
_THIS = pathlib.Path(__file__).resolve()
_CANDIDATES = _THIS.parents[2]
if str(_CANDIDATES) not in sys.path:
    sys.path.insert(0, str(_CANDIDATES))


def cosine_sim(a: np.ndarray, b: np.ndarray) -> float:
    na, nb = np.linalg.norm(a), np.linalg.norm(b)
    if na < 1e-8 or nb < 1e-8:
        return float("nan")
    return float(np.dot(a, b) / (na * nb))


def cross_density(emb_b: np.ndarray, centroid_a: np.ndarray) -> float:
    """Mean cosine similarity of stratum-B verses to stratum-A centroid."""
    cn = np.linalg.norm(centroid_a)
    if cn < 1e-8:
        return float("nan")
    c = centroid_a / cn
    norms = np.linalg.norm(emb_b, axis=1, keepdims=True)
    units = emb_b / (norms + 1e-8)
    return float((units @ c).mean())


def compute_drift_dataframe(traj: pd.DataFrame, cve: pd.DataFrame) -> pd.DataFrame:
    """Compute drift metrics from trajectory + verse-edge dataframes.

    Both inputs use the SPhilBerta 768-dim embedding space.
    """
    traj_idx = traj.set_index(["concept_id", "stratum"])

    emb_index: dict[tuple, np.ndarray] = {}
    for (cid, stratum), grp in cve.groupby(["concept_id", "stratum"]):
        emb_index[(cid, stratum)] = np.stack(
            [np.array(e, dtype=np.float32) for e in grp["embedding"]]
        )

    all_concepts = traj["concept_id"].unique()
    concept_label = traj.drop_duplicates("concept_id").set_index("concept_id")["concept_label"].to_dict()

    pairs = [("lxx", "nt"), ("nt", "vg"), ("lxx", "vg")]
    records = []

    for cid in all_concepts:
        label = concept_label.get(cid, "")
        row: dict = {"concept_id": cid, "concept_label": label}

        for s in ["lxx", "nt", "vg"]:
            key = (cid, s)
            if key in traj_idx.index:
                row[f"verse_count_{s}"]     = int(traj_idx.loc[key, "verse_count"])
                row[f"crystallization_{s}"] = float(traj_idx.loc[key, "mean_cosine_to_centroid"])
                row[f"dispersion_{s}"]      = 1.0 - row[f"crystallization_{s}"]
            else:
                row[f"verse_count_{s}"]     = 0
                row[f"crystallization_{s}"] = float("nan")
                row[f"dispersion_{s}"]      = float("nan")

        for s_a, s_b in pairs:
            tag   = f"{s_a}_{s_b}"
            key_a = (cid, s_a)
            key_b = (cid, s_b)
            have_a = key_a in traj_idx.index
            have_b = key_b in traj_idx.index

            if have_a and have_b:
                c_a = np.array(traj_idx.loc[key_a, "centroid"], dtype=np.float32)
                c_b = np.array(traj_idx.loc[key_b, "centroid"], dtype=np.float32)
                cc  = cosine_sim(c_a, c_b)
                row[f"centroid_cosine_{tag}"] = cc
                row[f"centroid_drift_{tag}"]  = 1.0 - cc
            else:
                row[f"centroid_cosine_{tag}"] = float("nan")
                row[f"centroid_drift_{tag}"]  = float("nan")

            row[f"delta_dispersion_{tag}"] = (
                row[f"dispersion_{s_b}"] - row[f"dispersion_{s_a}"]
                if (not np.isnan(row[f"dispersion_{s_a}"]) and
                    not np.isnan(row[f"dispersion_{s_b}"]))
                else float("nan")
            )

            if have_a and key_b in emb_index:
                c_a  = np.array(traj_idx.loc[key_a, "centroid"], dtype=np.float32)
                emb_b = emb_index[key_b]
                cd_ab = cross_density(emb_b, c_a)
                row[f"cross_density_{tag}"]  = cd_ab
                row[f"cross_drift_{tag}"]    = 1.0 - cd_ab
            else:
                row[f"cross_density_{tag}"]  = float("nan")
                row[f"cross_drift_{tag}"]    = float("nan")

            if have_b and key_a in emb_index:
                c_b  = np.array(traj_idx.loc[key_b, "centroid"], dtype=np.float32)
                emb_a = emb_index[key_a]
                cd_ba = cross_density(emb_a, c_b)
                row[f"cross_density_{s_b}_{s_a}"]  = cd_ba
            else:
                row[f"cross_density_{s_b}_{s_a}"]  = float("nan")

            cd_tag = row[f"centroid_drift_{tag}"]
            xd_tag = row[f"cross_drift_{tag}"]
            both_ok = not np.isnan(cd_tag) and not np.isnan(xd_tag)
            row[f"drift_score_{tag}"] = (cd_tag + xd_tag) / 2.0 if both_ok else float("nan")

        records.append(row)

    return pd.DataFrame(records)


def load_inputs() -> tuple[pd.DataFrame, pd.DataFrame]:
    """Load concept_trajectory_sp + concept_verse_edge_sp via load_hf_or_local."""
    from hf_registry import load_hf_or_local

    print("Loading trajectory and verse-edge data ...")
    traj_ds = load_hf_or_local("NuBerea/concept-analysis", "concept_trajectory_sp", verbose=True)
    cve_ds  = load_hf_or_local("NuBerea/concept-analysis", "concept_verse_edge_sp", verbose=True)
    traj = traj_ds.to_pandas()
    cve  = cve_ds.to_pandas()
    print(f"  trajectory rows: {len(traj)}  verse-edge rows: {len(cve):,}")
    return traj, cve


def print_report(df: pd.DataFrame) -> None:
    """Pretty-print drift summary tables (verbose mode)."""
    pairs = [("lxx", "nt"), ("nt", "vg"), ("lxx", "vg")]
    W = 70
    print("\n" + "=" * W)
    print("CONCEPT DRIFT ANALYSIS  (LXX → NT → Vulgate)")
    print("=" * W)

    for s_a, s_b in pairs:
        tag = f"{s_a}_{s_b}"
        col = f"drift_score_{tag}"
        sub = df[df[col].notna()].sort_values(col, ascending=False)
        print(f"\n--- {s_a.upper()} → {s_b.upper()}  ({len(sub)} concepts with both strata) ---")
        for _, r in sub.head(10).iterrows():
            print(f"  {r['concept_id']:<23} {str(r['concept_label'])[:32]:<33} drift={r[col]:.4f}")


def main(argv: list[str] | None = None) -> int:
    ap = argparse.ArgumentParser(description="Compute concept drift analysis")
    ap.add_argument("--out", type=pathlib.Path, default=DEFAULT_OUT, help="Output CSV path")
    ap.add_argument("--quiet", action="store_true", help="Skip the verbose report block")
    args = ap.parse_args(argv)

    traj, cve = load_inputs()
    df = compute_drift_dataframe(traj, cve)

    args.out.parent.mkdir(parents=True, exist_ok=True)
    df.to_csv(args.out, index=False)
    print(f"Saved: {args.out}  ({len(df)} concepts, {len(df.columns)} columns)")

    if not args.quiet:
        print_report(df)
    return 0


if __name__ == "__main__":
    sys.exit(main())