Top-4 residual and top-96 head steering vectors

These vectors use the sqrt_fixed rankings from the full mean_response
ridge benchmark. Every sycophancy dimension has exactly four selected residual
layers and 96 selected (layer, head) units.

Fits

Each selected unit is fitted independently with standardized ridge regression:

• residual penalty: alpha=10000;
• head penalty: alpha=1000;
• fit weights: square-root inverse score frequency;
• scores_1-4: rows scored 1, 2, 3, or 4;
• scores_1-5: all rows scored 1 through 5;
• all_turns: one direction fitted from every eligible turn after removing
  weighted turn-specific activation and target means;
• turn_N: one direction fitted using only turn N.

The positive direction points toward a larger predicted sycophancy score. Use
the negative direction for an anti-sycophancy intervention.

Files

• selected_units.csv: all 500 selected units, their ranks, held-out ranking
  R² values, and direction-stability diagnostics.
• manifest.csv: one row per dimension, activation type, turn scope, and score
  range. Check valid, status, and low_n_warning before using a vector.
• config.json: complete reproducibility settings and signature.
• vectors/{dimension}_{act_type}.npz: fitted vector arrays.
• vector_catalog.csv: the recommended human-readable lookup table, with one
  row per usable vector and explicit dimension, layer/head, score range, turn,
  ranking diagnostics, tensor key, and tensor row.
• vector_store.safetensors: all normalized directions and raw coefficients
  in one safe, framework-friendly tensor store.

Recommended catalog interface

The catalog removes the need to construct NPZ keys manually:

import pandas as pd
import steering_vector_fitting as svf

root = "steering_vectors/top4_layers_top96_heads"
catalog = pd.read_csv(f"{root}/vector_catalog.csv")

chosen = svf.query_vector_catalog(
    catalog,
    dimension="spiral",
    act_type="head",
    scope="all_turns",
    score_max=5,
    positive_r2_only=True,
    top_k=10,
)
directions = svf.load_catalog_vectors(chosen, output_dir=root)

# directions[i] is identified by chosen.iloc[i], including layer and head.

Residual rows have head=-1; head rows contain both their layer and head.
Use scope="turn_2" for a turn-specific vector or scope="all_turns" for the
pooled turn-controlled vector.

NPZ scope keys have the form:

all_turns|scores_1-5|direction_unit
turn_2|scores_1-4|coef_raw

Important arrays include:

• direction_unit: unit-normalized raw-activation directions;
• coef_raw: unnormalized raw-coordinate ridge coefficients;
• coefficient_norm: coefficient norm, also the local predicted-score change
  per unit injected in direction_unit under the fitted linear model;
• projection_std: natural training-data standard deviation along the unit
  direction, useful for setting an intervention grid;
• train_r2: in-sample diagnostic only, not a generalization estimate;
• selected_units: (layer, head) pairs (head=-1 for residual layers);
• ranking_probe_cv_r2: the held-out R² used to select each unit.

Load one scope with:

from pathlib import Path
import steering_vector_fitting as svf

path = Path("steering_vectors/top4_layers_top96_heads/vectors/spiral_head.npz")
scope = svf.load_vector_scope(path, score_max=5, turn_index=None)

assert bool(scope["valid"])
units = scope["selected_units"]
directions = scope["direction_unit"]
natural_scales = scope["projection_std"]

Reliability warnings

The requested fixed count does not imply that all selected units generalize.
All four indirectness and parasocial residual layers have negative ranking R².
Only 74 of the 96 indirectness heads and 12 of the 96 parasocial heads have
positive ranking R². Filter on ranking_probe_cv_r2 > 0 for a conservative
intervention experiment.

Scopes with a single score or fewer than two rows are recorded but contain no
direction. Scopes with fewer than 20 rows are fitted to fulfill the per-turn
request but carry low_n_warning=True; they should not be used without more
data.