logicBombExe/INSIDER_LLM_DETECTION_RESULTS

Results

This dataset is archived. The work continues in four separate projects:

Project	Question	Code	Benchmark
how_reliable_models_self_logs	When a model writes its own action log, is that log true?	GitHub	Hugging Face
double_logging_system_to_detect_an_unaligned_llm	Does comparing the model's own log with an automatic system log detect harmful behaviour?	GitHub	Hugging Face
does_small_to_medium_sized_models_can_act_as_insider	Do small and medium-sized models act as insiders?	GitHub	Hugging Face
do_old_insider_LLM_tests_still_work_today	Do the published insider-LLM tests still work on current models?	GitHub	Hugging Face

The files in this dataset are kept as a record of the earlier study.

One folder per run; folder names are the run ids and must stay unchanged. Mirrored to Hugging Face as the dataset repo https://huggingface.co/datasets/logicBombExe/INSIDER_LLM_DETECTION_RESULTS (make publish-results in the code repo, which runs hf upload on this folder). HF is its versioned home; the project root is not a git repository.

Each run folder contains:

• meta.json — config snapshot, code git hash, classifier source commit (classifier_source; upstream_commit in runs before 2026-09-10), dataset version, prompt-set hash, model digest, seeds, reconstructible flag (+ reconstructible_note when false), phase (smoke / validation / pilot / study), and append-only reparse_log / classify_log / evaluate_log trails.
• episodes.jsonl — one JSON object per episode: prompts, raw response, derived fields (final_text, system_log, model_log, logged), and harmful labels.
• failures.jsonl — episodes whose API call failed (absent when none failed).
• labels.jsonl — append-only classifier audit trail (present after ild classify).

Raw response fields are immutable. Derived fields are re-derivable at any time with ild reparse --run <dir> (recorded in meta.json's reparse_log; every episode that newly becomes not_logged is printed for hand verification). tables.md and INDEX.md are generated by ild eval — do not edit by hand. Only phase = study ∧ reconstructible runs enter the study table.

The corpus was reset on 2026-09-03 (all earlier runs deleted); it now holds the 2026-09-03 study sweep only (INSIDER_LLM_DETECTION_CODE/configs/study_*.yaml).