schema-eval/compliance-sycophancy-cot

Compliance-Sycophancy CoT Analysis

When compliance-forcing instructions cause frontier AI models to fabricate answers, the models know they are fabricating.

Reading the reasoning traces of DeepSeek V4 Pro (129 traces) and Qwen3-80B (41 traces) reveals that 100% of fabrication cases show the model explicitly recognizing insufficient context, referencing the compliance instruction, and deliberately overriding its own uncertainty. A one-sentence defense phrase ("if you lack information, say so") recovers +19.5pp of correct refusals (p=0.00013).

This dataset accompanies a model forensics investigation extending The Compliance Trap (67,221 evals, 11 models, NeurIPS 2026 submission).

Code repository: GitHub

Browseable Tables

Use the dataset viewer tabs above to browse:

• sycophancy_results — 3,597-sample sycophancy experiment (4 conditions x 2 models)
• fabrication_taxonomy — Three-way split: 84.3% pure fabrication, 11.9% hedged, 3.8% true refusal
• defense_results — Permission defense recovers +19.5pp (p=0.00013)
• qwen_comparison — Cross-model: both V4 Pro and Qwen show 100% Pattern C
• statistical_tests — All p-values, CIs, Bonferroni corrections

Raw Trace Data

The traces/ directory contains JSONL files with full reasoning traces. Download for programmatic analysis:

from huggingface_hub import snapshot_download
snapshot_download("schema-eval/compliance-sycophancy-cot", local_dir="data/", allow_patterns=["traces/**"])

Trace Files

File	Records	Description
traces/fabrication_traces.jsonl	129	V4 Pro Condition A fabrications (100% Pattern C)
traces/baseline_refusal_traces.jsonl	149	V4 Pro Condition D correct refusals
traces/survivor_traces.jsonl	23	V4 Pro Condition A survivors (scored correct)
traces/defense_traces.jsonl	477	V4 Pro defense experiment (3 conditions x 53 items x 3 epochs)
traces/qwen_traces.jsonl	159	Qwen3-80B Condition A (53 items x 3 epochs, via Nebius)
traces/sycophancy_flips.jsonl	168	V4 Pro sycophancy flips (all conditions)
traces/sycophancy_nonflips.jsonl	1,532	V4 Pro sycophancy non-flips (all conditions)

Eval Logs

The eval_logs/ directory contains Inspect .eval binary logs from the sycophancy production run:

from inspect_ai.log import read_eval_log
log = read_eval_log("eval_logs/2026-09-01T17-47-49-00-00_compliance-sycophancy_HdGGxqvjdm8dGR733tEBBa.eval")

Key Results

Sycophancy Experiment (null on primary hypothesis)

Model	Condition	Flip Rate	95% CI
V4 Pro	baseline	13.4%	[10.5%, 16.9%]
V4 Pro	full_suffix	8.3%	[6.0%, 11.3%]
Sonnet 4.6	baseline	18.6%	[15.0%, 22.7%]
Sonnet 4.6	full_suffix	15.6%	[12.3%, 19.6%]

No comparison survives Bonferroni correction.

Fabrication Taxonomy (V4 Pro, n=159)

Category	%	95% CI
Pure fabrication	84.3%	[77.8%, 89.1%]
Hedged fabrication	11.9%	[7.8%, 17.9%]
True refusal	3.8%	[1.7%, 8.0%]

Defense Mechanism

Condition	Refusal Rate	p-value
compliance_only	63.5%	--
+ "say so"	83.0%	0.00013
+ "consider confidence"	70.4%	0.233

Models and Providers

Model	Provider	Pinning
DeepSeek V4 Pro	OpenRouter/Alibaba (fp8)	provider.order=["Alibaba"]
Claude Sonnet 4.6	OpenRouter/Anthropic	provider.order=["Anthropic"]
Qwen3-80B Thinking	Nebius direct	Direct API (matches paper)

Provider pinning is essential. See LessWrong.

Citation

@misc{kumar2026compliance_cot,
  title={Compliance-Induced Fabrication Is Transparent in Thinking Models: A Model Forensics Investigation},
  author={Kumar, Rahul},
  year={2026},
  note={Extends arXiv:2605.02398}
}

License

MIT