YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Team JSR — Aletheia's Quest white-box probe
This repository preserves the frozen trained artifact and reconstruction code for Team JSR's white-box submission to the Aletheia's Quest Competition.
Frozen artifact
trusted_activation_probe.npz is the exact probe used by the submitted
notebook. It contains the standardization statistics, collapsed linear
projection, intercept, operating threshold, and feature metadata.
SHA-256:
54743cf37a875c3842417d363e1a244667ae1947b6e761906a7d1c44c86caf1f
The detector uses trusted Qwen/Qwen3.5-9B layer 23, with final-16-token
mean pooling and its contrast with the final-24-token mean. The probe is a
regularized logistic readout trained in the span of 17 public source-group
mean-difference directions. The method is not Bayesian despite the historical
development filename bayes_v2.
Files
trusted_activation_probe.npz: exact frozen probe used for inference.train_trusted_activation_probe.py: reconstruction entry point.activation_detector.py: shared evidence, feature-extraction, and scoring logic.liars_bench_data.py: public Liars' Bench Expanded metadata loader.training_datasets.yaml: organizer-provided development datasets used by the reconstruction.
Only public conversation text, source-group identifiers, and labels are loaded
from Liars' Bench Expanded. Stored activations from the source organisms are
not used. Rebuilding the probe requires access to the public datasets and an
NDIF key authorized for Qwen/Qwen3.5-9B:
export NDIF_API_KEY=...
python train_trusted_activation_probe.py
The artifact and files in this repository are retained unchanged for organizer verification.