The primary record.
These files support the LLM Freedom Audit v1.0. They contain the prompts, the observed and unobserved model responses, and the per-scenario analysis. For a governance team, they are the difference between a headline and evidence: download them, check the counts, and re-run the comparison independently.
| File | What it contains | Size | Download |
|---|---|---|---|
| observed_results.json | Model, condition, framing, and every observed-arm response with parsed actions. | 30 KB | Download |
| unobserved_results.json | The matched unobserved-arm responses and parsed actions. | 29 KB | Download |
| deepseek_drift_analysis.json | Per-scenario drift analysis: total scenarios, action changes, rate, and detailed comparisons. | 28 KB | Download |
| audit.pdf | The full report with method, master table, controls, and limitations. | 365 KB | Download |
Schema at a glance
The two response files share the same top-level keys:
- model — the model identifier used for the run.
- condition — observed or unobserved.
- framing — the framing text applied to that condition.
- results — an array of scenario records containing the prompt, parameters, raw response, and parsed action.
The drift analysis file reports total_scenarios, action_changes, action_change_rate, and a detailed_drift array with the observed and unobserved responses side by side.
How to check the headline
- Open deepseek_drift_analysis.json and confirm total_scenarios is 20 for the scored rounds.
- Compare the detailed_drift entries against the observed and unobserved files.
- Confirm that every scored round has 20/20 usable responses in both arms. The report's master table lists the per-round flip counts.
- Read the limitations before treating any single rate as a general property of the model.
No API keys, credentials, account identifiers, or personal data are included in the published files. The model responses are research data, not private client engagements. Private audit data is never published without the client's approval.
Using the data
The files are published for independent verification and replication. If you use them, cite the report:
Freedom Theory / justrunthatshit.com (2026). LLM Freedom Audit — Monitored vs Unobserved (Report v1.0, 19 October 2026). https://justrunthatshit.com/audits/llm-freedom-audit/
For commercial reuse, custom scenarios, or a private engagement, see engagements or email us.