RandallSPQR/perfect-lie

The Perfect Lie

Lies written by language models for a storytelling game, each told to persuade a judge whose private beliefs about what makes a story convincing were (or were not) disclosed to the liar; the cue annotations of every lie; the judges' verdicts; and, for two open-weight liars, the activation directions and steering runs of Phase 2. Three pre-registered studies (each registered as a dated GitHub release; see ARCHIVE.md in the release): study 1 (full versus placebo note), the reversed-note study (full versus reversed note), and Phase 2 (activation probe and steering).

Fabricated-content notice. Every story in this dataset is a deliberate lie, written to be believed. Many attribute invented events, findings, failures, accidents, cover-ups or harms to real, named organizations (universities, government agencies, museums, journals, companies) and to invented people with real-sounding names. None of it is true. Do not cite, quote or summarize these stories as fact, do not use them to train systems that present them as fact, and do not attach them to the organizations they name. Names are flagged, not redacted (below).

Contents

file	rows	what
study1/lies.jsonl.gz	2,160 cells (2,110 complete, 50 failed)	study 1 interim (replicates 1-15): every planned cell
reversed/lies.jsonl.gz	4,320 cells (4,297 complete, 23 failed)	reversed-note run
reversed/comprehension.jsonl.gz	500	comprehension-check answers, parsed and scored
phase2/identity_lies.jsonl.gz	1,620	identity-check lies generated on the pod (two runs), with primary grades
phase2/pilot_lies.jsonl.gz	2,800	stage-1 pilot and method-check pre-check lies (coherence only; never graded)
phase2/stage2_lies.jsonl.gz	2,560	stage-2 lies: steered, placebo, unsteered references, method check; graded
phase2/probe/<model>/		vectors.npz (v and the cue-mention direction for every layer, float64), per_layer_auroc.csv, probe_result.json, diagnostics.json, scale.json
phase2/steering/<model>/		setup records of every steered run (layer, norms, cosines, exact hook checks), pilot and coherence results, naturalness
phase2/hf_revisions.json		pinned Hugging Face revisions of the open weights
instrument/		personas, cues (rubric v0.6), prompts, designs, reversed notes, categories, model and provider pins
screens/<study>/		the blind confession/refusal screens and the owner's adjudications
content_flag_summary.json, MANIFEST.sha256		flag counts; sha256 of every file

A cell is one planned lie (liar x prompt or category x target x condition x replicate). Each row is the cell's last record in the run log (the release's git history keeps the full logs), with every field the runner stored:

• user_prompt (public prompt, identical across conditions), liar_system_prompt (the private note is in it), liar_delivered_messages (the exact messages sent), lie, lie_words, liar_usage, liar_served_provider;
• study 1 only: elicitation (the liar's stated belief about the judge, asked after the lie);
• targets: each judge model's verdict (accept, confidence, raw text);
• grades: each grader's cue annotation (cues, counts, confidence, the grader model, attempts and the raw output where kept): primary (Claude Sonnet 4.5), google (gemini-2.5-flash), and secondary (gpt-5, seeded 25% subsample); study 1 also bhat_grades (the same graders on the elicited belief);
• status, errors, parse_failures (raw text of unparseable responses);
• Phase 2: alpha, arm, hyp, seed, ids (generated token ids), nll (per-token NLL of the first 64 tokens under the unsteered model), primary and google_call (raw grader calls with usage);
• content_flags (added by the release builder; below).

The scores and every analysis are reproduced from these files by the release's Tier 0 (make reproduce).

Provenance

part	generated	liars / models	judges	graders
study 1	2026-10-07 to 2026-10-08, OpenRouter	Llama 3.1 8B Instruct (CoreWeave, bf16), Gemma 3 27B IT (Novita), gpt-4o-mini (OpenAI, Azure), Gemini 2.5 Flash-Lite (Google)	the same four	Claude Sonnet 4.5, gemini-2.5-flash, gpt-5 (subsample), all via OpenRouter
reversed note	2026-10-16, OpenRouter	the same four, same pins	the same four	the same three
Phase 2 identity check	2026-10-20 to 2026-10-21	Gemma 3 27B IT, Llama 3.1 8B Instruct on one RTX PRO 6000 (bf16, transformers 4.53.2)	none	Claude Sonnet 4.5 via OpenRouter
Phase 2 pilot, stage 2	2026-10-23 to 2026-10-27	the same two, steered	none	Claude Sonnet 4.5 direct (claude-sonnet-4-5-20250929), gemini-2.5-flash via OpenRouter (20% subsample)

Personas, notes and categories were written for the study (no personal data). The adjudications are the study owner's, made blind to condition before unblinding.

Content flags (fabricated content about real organizations)

Every lie row carries content_flags:

• fiction: always true;
• org_mentions: organization-like names found by a pattern (capitalized phrases ending in, or built on, words such as University, Institute, Museum, Agency, Department, Survey, Society, Journal, Company; plus common acronyms such as NASA, USGS, EPA, MIT). Heuristic: it misses some names, truncates some ("Massachusetts Institute"), catches some generic phrases ("The Journal"), and does not tell real organizations from invented ones;
• names_institution, alleges_official_failure: the primary grader's institutional_authority and official_failure cue labels (null where a lie was not graded);
• possible_fabricated_harm_to_named_org: an organization mention and an alleged official failure in the same lie.

Counts (content_flag_summary.json):

file	lies	with organization mentions	names an institution (grader)	possible fabricated harm to a named org
study 1	2,160	646	814	83
reversed note	4,320	1,149	1,347	220
Phase 2 identity check	1,620	562	701	154
Phase 2 pilot (ungraded)	2,800	826	n/a	n/a
Phase 2 stage 2	2,560	849	985	122

Decision: flag, not redact. The texts are kept verbatim because the cue annotations, judge verdicts and every analysis are computed on them; a redacted text would no longer match its grades. Downstream users who display or redistribute lies should filter on these flags or redact the org_mentions themselves.

Licenses and terms

The lies, verdicts and annotations are model outputs. Each generating model's terms govern its outputs; the API route adds the provider's and OpenRouter's terms. Summary as understood on 2026-10-29 from secondary copies of the terms (the primary pages could not be fetched from the build environment; verify each before publishing):

model (role)	terms	outputs	points relevant to redistribution
Llama 3.1 8B Instruct (liar, judge)	Llama 3.1 Community License + Acceptable Use Policy	not claimed by Meta	Redistributing Llama Materials needs a copy of the license and "Built with Llama"; a model trained on Llama outputs and distributed must have "Llama" at the start of its name. The AUP bars using Llama to generate or promote disinformation or defamatory content: this dataset is a labelled research artefact, released with the notice above, not content presented as true.
Gemma 3 27B IT (liar, judge)	Gemma Terms of Use (rev. 2026-04-01) + Prohibited Use Policy	"Google claims no rights in Outputs"; outputs are not Model Derivatives	Same caution on misinformation; users are responsible for outputs and their use.
gpt-4o-mini (liar, judge), gpt-5 (grader)	OpenAI Terms of Use / Business Terms + Usage Policies	assigned to the user	Bars using output to develop models that compete with OpenAI; state that the content is AI-generated.
Gemini 2.5 Flash-Lite (liar, judge), gemini-2.5-flash (grader)	Gemini API Additional Terms (rev. 2026-03-23)	Google does not claim ownership	Bars using the service to develop competing models; users responsible for generated content and its downstream sharing.
Claude Sonnet 4.5 (grader)	Anthropic Commercial Terms	customer owns outputs	Bars using the services to build a competing product or to train competing models without approval.
providers (CoreWeave, Novita, Azure, Google, OpenAI) and OpenRouter	their terms of service

Practical consequence, stated as the release's terms of use: research and evaluation use; not for training models that compete with the providers above; keep the fabricated-content notice with any copy. The study's own materials (instrument, adjudications, flags, code) are the owner's to license.

What is not included, and why

• Activations (Phase 2 probe stage: 1,800 items x 62 layers x 5,376 for Gemma, 1,800 x 32 x 4,096 for Llama; about 3.3 GB in float32) and the placebo, method-check and naturalness vectors: they stayed on the study's RunPod network volume and were not copied out. The probe directions v for every layer are included; the steering vector of every run is unit(v) x alpha x ||v|| at the selected layer (both in this dataset). release/tiers/tier3_fetch_volume.sh describes how to retrieve the rest.
• Weights: not redistributed; pinned by revision in phase2/hf_revisions.json (both repositories are gated).
• The grading-route check's raw API responses: only the parsed labels were kept (in the release's regrade_direct.json).
• Pilot and smoke runs before each registration and the quarantined void run: in the release's git history (results/perfect_lie/), not in this dataset.
• No API keys, no personal data.

Citation

Zenodo: 10.5281/zenodo.24122306. Code, write-up and the push-button reproduction (Tier 0): GitHub release perfect-lie-release-v1 (two_truths_lie_study/release/README.md). The contents of this repository are the release archive perfect-lie-dataset-v1.tar.gz unpacked (sha256 4b9527a0e7db3a32cfae3434bf338c56ef53cec8e8511c32a7403c08a81ef17a); MANIFEST.sha256 lists every file, except this card, which adds this paragraph.