corrigibility-games/cg-ktane

Corrigibility Games: KTANE

Traces of principal–subordinate AI teams in Keep Talking and Nobody Explodes, collected to study correction-supporting behavior under asymmetric information. Includes role-specific inputs, communication, actions, game-state observations, synchronized media, and provisional qualitative annotations.

**CG-KTANE · Pilot v0.1.0 **

About Corrigibility Games

Corrigibility Games investigates whether interactive games can reveal how agents preserve a principal's ability to notice problems, initiate correction, and carry a change of course through to its consequences. The project studies observable behavior across the correction process: what information is disclosed, what authority is respected, what actually changes, and whether the change persists.

This release focuses on asymmetric information and vigilance. Two models collaborate under explicit roles, with different observations and a shared task. Corrections and disagreements arise during ordinary gameplay; the collection does not inject scripted mistakes, principal interventions, or mandatory reversals. Both roles are occupied by models in this pilot. Transfer to human oversight remains a separate research question.

The game and the agent roles

Keep Talking and Nobody Explodes (KTANE), developed by Steel Crate Games, is a cooperative puzzle game. A Defuser sees and operates a fictional bomb; an Expert has its procedural manual but cannot see the device. They communicate to identify and solve its modules. This corpus uses an adapted GPTNT harness to run the real game locally, with hosted models exchanging text. GPTNT documentation describes the underlying multimodal collaboration environment.

Role	Available information	Operational responsibility
Expert / principal	Manual, partner messages and its own retained context	Direct the plan, authorize or delegate actions, investigate challenges, revise or withdraw instructions
Defuser / subordinate	Game observations, partner messages and its own retained context	Inspect and report, act within scope, disclose consequential discrepancies, follow applicable revisions

The principal has decision authority without being factually infallible. The subordinate can challenge an unsupported premise and request reconsideration. The game engine determines physical effects and outcomes.

Maze episode	Memory episode

What the dataset contains

The release contains 18 terminal games: 12 defusals, 3 timeouts and 3 strikeouts. A completed game reached a definite ending; completion does not imply good correction behavior.

There are 24 linked table configurations, each in Parquet and JSONL. All share one exploratory pilot partition. episode_id connects records; request, message, event and asset identifiers provide finer links.

Evidence	Contents
Episodes and configuration	Mission, model/provider identifiers, roles, prompts, seeds, timing, outcomes and provenance
Agent requests	934 requests, including 36 postgame reflections, with ordered instructions, retained history, current input and actual images
Communication	357 published messages, their receipt events and 356 recorded request-inclusion links
Decisions and actions	910 parsed model outputs, 24 harness recovery outputs and 261 action dispatches
State and execution	6,748 events, 453 observation records, step records, researcher state, usage, errors and recovery
Annotations	31 issues, 59 role contributions, 295 qualitative dimension records, and 6,885 evidence links
Media	14 native videos, 15 separate game-audio recordings, 14 compact videos, sampled frames and alignment maps
Audit materials	Schemas, copied artifacts, source versions, redaction records, checksums and validation reports

The corpus covers 11 single-module families and nine model aliases. Exact model identifiers are in episodes. See the data dictionary, machine-readable schema, media guide and source-code provenance.

Load the dataset

from datasets import load_dataset

repo = "corrigibility-games/cg-ktane"
episodes = load_dataset(repo, "episodes", split="pilot")
messages = load_dataset(repo, "messages", split="pilot")
requests = load_dataset(repo, "requests", split="pilot")
annotations = load_dataset(repo, "annotations", split="pilot")

print(episodes.select_columns([
    "episode_id", "principal_model", "subordinate_model", "outcome"
]))

Asset paths are relative to the repository root. Loading a table does not automatically download files referenced by string paths. Download a request context individually:

import json
from huggingface_hub import hf_hub_download

r = requests[0]
path = hf_hub_download(repo_id=repo, repo_type="dataset",
                       filename=r["context_path"])
with open(path) as f:
    context = json.load(f)

instructions = context["instructions"]
history = context["message_history"]       # ordered retained context
current_input = context["message_input"]  # current turn, kept separate

For the full repository, including media and recursively referenced images:

from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download(repo_id=repo, repo_type="dataset"))
# A recorded media path can now be opened as root / media_row["path"].

Scope, timing and limitations

These are natural games under prompted authority roles, with puzzle seed 845, rule seed 1, unequal pairings and evolving harness/capture versions. They are suitable for case analysis, and inspection of interaction traces. They do not establish general corrigibility, internal motives, a representative model ranking or the necessity of collaboration despite possible memorized rules.

License

The project’s original traces, annotations and documentation are licensed under CC BY 4.0. KTANE game/manual content and GPTNT software retain their separate rights; the CC BY grant does not relicense those components. Please preserve attribution and identify changes when reusing the project’s material.

Citation

@misc{corrigibility_games_ktane_2026,
  author       = {{Alfaxad Eyembe}},
  title        = {Corrigibility Games: KTANE},
  year         = {2026},
  version      = {0.1.0},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/corrigibility-games/cg-ktane},
  note         = {Natural-play pilot; cite the repository revision used}
}