leonliuzx/riskchainbench-task1-paper-inputs

RiskChainBench Task 1 — reviewed paper inputs

This is an input-only, reviewed export, not the full historical research archive or a complete reproduction of the paper. Packaging date: 2026-09-25.

Paper · Maintained code and scoring · Evaluation package

Content and version

• 3,600 unique token-text inputs from 600 synthetic source sessions.
• Six variants per source: v000 Primary, v001 Phonetic, v002 Entry encoding, v003 Lexical, v004 Few-line, v005 Vertical. Each variant has 600 records.
• All variants enter Task 1 scoring; only v000 supplies the fixed Task 2 entry gate.
• The test split is an input distribution label, not a new train/dev/test split. Variants from the same source must stay together in any future split.
• No rows, message strings, order, Unicode glyphs or identifiers were rewritten. Do not normalize, deduplicate by message text, or resplit this frozen snapshot.

model_visible/tasks_3600.jsonl is byte-identical to the reviewed GitHub paper snapshot. Required SHA-256:

a3c7248d06caa6f2d978284d87fecf6c080f963b6a30473da338ba459e65b644

spec/model_prompt.md SHA-256:

2ece83426010b6c5d6abc4ef005011f6780a60a42b467009a2cc4f753269373c

SOURCE.json records the source commit and copied-file hashes. MANIFEST.sha256 covers every distribution file except itself.

Loading and verification

The records are UTF-8 JSON Lines with sample_id, platform/client context, message content, and task/prompt identifiers. The included JSON schema documents the structure. After downloading the complete export, run:

python3 verify.py

To load without any dataset-specific executable code:

import json
from pathlib import Path
rows = [json.loads(line) for line in
        Path("model_visible/tasks_3600.jsonl").read_text(encoding="utf-8").splitlines()]

Use the pinned prompt and the maintained GitHub runner. Model runs may incur provider charges. This export contains no provider endpoint or credential.

Evaluation boundary

Gold, evaluator files, source sessions, private resolver mappings, website bindings and historical model predictions are not included. Public scoring code exists, but full scoring needs separately authorized evaluator data with SHA-256 c4fdcb3a06b7d2834963750f999b38ea238e7900ce989087c62d9020737d8685. Ask about evaluation access without posting private material. Access is not automatic or guaranteed.

Task 1 archived bootstrap seed 20260726 differs from the supplement's 20260727-plus-offset statement; this packaging does not resolve that discrepancy. Full historical predictions and the exact Task 2 paired-analysis input remain unrecovered. New model runs would be new results, not recovery of old results.

Licensing and appropriate use

The synthetic input records and project-owned provenance metadata are CC BY 4.0 (LICENSE). The prompt, schema, verification script and project-owned documentation are Apache-2.0 (LICENSE-CODE). Existing notices retain their terms. Attribution is required; identify any modifications.

These grants do not cover private archives, original website content, third-party images, model weights or material the contributors do not own. This input-only export contains no screenshots, Docker images or original website assets. Benchmark text may describe harmful or deceptive scenarios; it is intended for defensive research. Entries are controlled/reserved surrogates. Do not resolve them against live services or attempt to identify original subjects.

Citation and acknowledgments

Please cite RiskChainBench, https://arxiv.org/abs/2609.16900; see CITATION.cff. We thank @Poor-Jack, @lele104, @Zzzzzz-y-s, and @zhengwj07-lgtm for human annotation and related dataset preparation.

感谢以上四位同学参与 RiskChainBench 数据集的人工标注及相关数据建设工作。 See ACKNOWLEDGMENTS.md for upstream attributions. No funding or endorsement is implied by these acknowledgments.