History Anchor 100 — Model Trajectories

Per-(model × condition × scenario set × seed) raw outputs from the paper "History Anchors: How Prior Behavior Steers LLM Decisions Toward Unsafe Actions".

This dataset contains the full set of model decisions that back every figure and table in the paper. Use it to:

• audit a single model's behaviour scenario-by-scenario,
• recompute headline metrics without re-running the (paid) API sweeps,
• mine reasoning_content traces from models that expose them (DeepSeek Reasoner, Anthropic extended thinking, etc.).

If you only need the scenarios themselves (no trajectories), use the companion dataset: albertoRodriguez97/history-anchor-100.

What's in the box

~228 MB across 151 experiment directories. Each top-level directory is one experimental cell, named:

<set>_<model>_<config>_<condition>_<seed>

Examples:

history_anchor_100_sonnet46_thinking_disabled_consistency_seed0
history_anchor_100_opus47_thinking_disabled_clean_seed0
history_anchor_action_perms_perm1_deepseek_v4_pro_consistency_seed0
history_anchor_prefix_mix_0U_3S_sonnet46_consistency_seed0

Three experiment families:

Per-cell layout

<cell>/
├── aggregate_summary.json      # cell-level KPIs
└── ha100_<domain>_<idx>_<slug>_<model>_<...>_<condition>_seed<N>/
    ├── manifest.json           # exact provider/model/temperature/system-prompt
    ├── summary.json            # per-scenario aggregate
    ├── runs_table.csv          # one row per seed
    └── runs/
        └── seed_<N>.json       # full episode: prompt, chosen action, reasoning, ...

runs/seed_<N>.json is the atomic record: it carries the exact prompt sent to the model, the raw response, the parsed action_id, the resulting Machiavellian score, and the next node. For providers that expose reasoning_content (DeepSeek Reasoner, Anthropic extended thinking), the trace is preserved verbatim.

Top-level summary files

In addition to the per-cell directories, the root of the dataset ships pre-aggregated CSVs/JSONs used to render the paper figures:

history_anchor_clean100_<family>_family_comparison.{csv,json}
history_anchor_action_perms_subset_{comparison,pooled,rows}.{csv,json}
history_anchor_prefix_mix_*.{csv,json}
…

These are the same files shipped under data/ in the code repo, included here for one-stop replication.

Quickstart

hf download albertoRodriguez97/history-anchor-100-traces --repo-type dataset --local-dir traces

# Inspect one cell
ls traces/history_anchor_100_sonnet46_thinking_disabled_consistency_seed0/ | head

# Grab one decision trace
cat traces/history_anchor_100_sonnet46_thinking_disabled_consistency_seed0/\
ha100_academic_01_exam_leak_sonnet46_thinking_disabled_consistency_seed0/\
runs/seed_0.json

Or via huggingface_hub:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="albertoRodriguez97/history-anchor-100-traces",
    repo_type="dataset",
    local_dir="traces",
)

Models evaluated

All runs use temperature = 0 and a fixed seed where supported. GPT-5.x was evaluated with reasoning_none; the Anthropic models with extended thinking disabled. See the paper Limitations section for caveats.

Companion artefacts

• 📦 Code & paper: github.com/alrod97/history-anchors
• 🧩 Scenarios: albertoRodriguez97/history-anchor-100

Citation

@misc{rodriguezsalgado2026historyanchors,
  title  = {History Anchors: How Prior Behavior Steers LLM Decisions Toward Unsafe Actions},
  author = {Rodr\'iguez Salgado, Alberto},
  year   = {2026},
  note   = {arXiv preprint forthcoming}
}

License

MIT — same as the code repo and the scenarios dataset.