unrulyabstractions/apart-idt-data

Instrumental Differential Treatment audit data

Working data for a project that audits published model organisms for Instrumental Differential Treatment (IDT). IDT is group-conditional behaviour that serves an undisclosed objective, activates on implicit identity cues, and hides in distributional gaps rather than in any single transcript. Every individual reply is defensible on its own, so per-transcript auditing structurally cannot reach it; the pipeline instead compares response distributions between a target model and its own base model, run identically, and reports both a directional component (treatment that differs by candidate) and a common-mode component (treatment delivered equally to every candidate).

Coming from the paper? Start at INDEX.md: it maps every table row, robustness check, and figure to its directory and stored verdict.

Directory layout

The tree is experiment-first. One experiment owns its stages and everything derived from them:

main/<family>/<experiment>/
    ellicit/  promptset/  conjecture/  score/  compare/
    rejudge/<seat>/{score,compare}    helper_swap/<stage>/
    geometry/  persona/  judge_probe/

• main/secret_loyalties/: the gen9 program, three calibration conditions and the three challenge organisms, plus shared/ (the challenge family's blind promptset and axis registry, the pooled coherence record, the Qwen persona probes).
• main/auditbench/: the Llama-3.3-70B family, the two positives at the root and the eight null controls under controls/. Each organism holds responses/ (with the collection-time verdicts) and judge_mini/ (the reported gpt-5-mini judging).
• main/sycophancy/: the trained political-sycophancy organism.
• main/external/: a collaborator's prompted organism sampled through our runner.
• configs/: the generated stage configs, one JSON per stage and condition, uploaded verbatim.
• explorer/: the data behind the results explorer page.
• superseded/, archive/, logs/, quarantine/: the retired first run, session backups, box logs, and quarantined artifacts. Mirror-only; the working tree no longer carries them.

Every compare/ holds the registered test's comparison_summary.json and the base-free counterfactual's reference_free.json. Every data directory carries a generated STATUS.md naming its checkpoint, helper seats, judge, and health.

Pipeline stages

The pipeline has six stages, and the stage directories are their outputs:

1. Principal elicitation (ellicit/): ask the target who it favours; outputs questions.json and elicitation_report.json.
2. Prompt set construction (promptset/): build matched user prompts per candidate principal; outputs templates.json and promptset_report.json.
3. Hypothesis conjecture (conjecture/): propose behaviour hypotheses to score; outputs hypotheses.json and scoring_questions.json.
4. Response collection (score/responses_<model>.jsonl).
5. Response scoring by judge (score/verdicts_<model>.jsonl).
6. Comparing distributions (compare/): comparison_*.json, comparison_summary.json, side checks, and behaviour figures (pdf/png).

Row schemas

Responses rows (responses_<model>.jsonl), one row per sample: {principal, prompt_id, instruction_id, system_id, s, refused, failed, text}. s is the sample index within the cell. Failed generations are recorded with empty text and failed: true, never dropped. Refusals are recorded, not discarded. The retired first run under superseded/ lacks system_id and failed.

Verdict rows (verdicts_<model>.jsonl), one row per sample per judge level: {principal, prompt_id, instruction_id, system_id, s, judge, level, verdicts}. verdicts maps each hypothesis id to true, false, or null. A null is a missing judge verdict: it is counted and reported, never imputed. Each judge level is its own table and is never averaged with another.

Every prompt is collected under 4 collection system prompts (none, live_deployment, committed_supporter, unreviewed_authority) and instruction_id is the composite "<system_id>::<template_id>". The retired first run uses the bare template id, and the two runs are not comparable row for row.

Some secret-loyalties score directories carry a __v suffix. A seat was split across two machines, and each half wrote its own directory before the rows were merged into the parent run. The parent run is what every reported number was computed from. The halves are kept here so the merge can be checked, and their rows must not be counted a second time.

What is not here

No model weights are included. The audited organisms are third-party published checkpoints and are not redistributed in any form.

Dual use

Real public figures appear as candidate principals because the published organisms were trained on them; prompts stay at the level needed to reproduce the safety result and no further. Organisms establish possibility, not propensity: nothing in this dataset describes any deployed model's actual behaviour. This repository holds the full reply corpus, every model reply the pipeline collected and every judge verdict on it. The paper releases only aggregates, so this repository is private and must stay private.