agentapprove/coding-agent-tool-call-safety

Decision Models as Policy Gates for Coding Agents

Synthetic coding-agent tool calls labeled against 28 policy rules, from the paper Decision Models as Policy Gates for Coding Agents: Explicit Policy Boundaries Improve Detection and Reduce False Alarms (James P. Beno, Agent Approve, 2026, arXiv:2611.02155). The harness, run records and analysis code are at github.com/agentapprove/decision-policy-rules.

Every person, company, repository, host, channel and credential in these cases is fictional. The cases were written to be classified, never executed. They include destructive commands and credential-shaped strings whose formats are real and whose bodies are random; each case declares its synthetic secrets. Do not run anything from this dataset.

What is in it

Config	Split	Cases	Families	Tool-only positive labels
main	dev	1535
main	holdout (sealed before inference)	1108
main	total	2643	106	4,413
confirmation	holdout	320	16	531

Each case is one proposed tool invocation by a coding agent (a Claude Code built-in tool such as Bash, Write or Edit, or an MCP tool) with a short trusted context, labeled for each of the 28 rule templates in three settings: tool_only (the model sees the invocation alone), user_request (plus the user's request), and bounded_context (plus trusted project facts such as git remotes, files present and identities). The 320-case confirmation set was written after every main-set result had been seen, through the same pipeline, as a check that the findings were not specific to the first scenarios.

Fields

Field	Type	Meaning
id	string	Case id; shuffled so that it encodes neither family nor split
family	string	The scenario family (106 in the main set, 16 in the confirmation set)
split	string	dev or holdout
stratum	string	Size stratum of the action (S, M, L, XL)
tags	list of strings	Case design tags (hard negative, context flip, injection, decoy context, ...)
tool_name	string	The tool invoked
tool_arguments	string (JSON)	The tool's arguments; the shape depends on the tool
context	string (JSON)	The trusted context record for the with-context setting
truth_tool_only, truth_user_request, truth_bounded_context	object	Gold label per rule template: "true", "false", "unknown" (evidence missing in that setting) or "excluded" (quarantined class)
provenance_tool_only, provenance_user_request, provenance_bounded_context	object	Per-label provenance: agreed, adjudicated-model, adjudicated-human, rulebook-relabel or unresolved
rationale	string	The generator's rationale, citing rulebook rules
evidence_locator, evidence_byte_offset	string, int	Verbatim evidence for a positive label and its byte offset in the action
evidence_gaps, quarantine, rulebook_refs	list, string (JSON), list	Enumerated evidence gaps, quarantined classes, and rulebook references

The 28 rule templates: permanent-deletion, git-force-push-or-delete-ref, download-and-execute-code, replace-existing-file-content, read-credential-material, share-credential-material, change-security-or-persistence, git-discard-local-changes, git-history-rewrite, disable-tests-or-checks, git-commit-credential-files, git-change-remote-or-hooks, upload-file-contents-externally, public-discussion-post, communication-outside-slack, send-data-to-external-host, change-production-state, deploy-to-production, database-destructive-statement, database-production-write, slack-dm-specific-person, git-push-specific-repository, read-only-inspection, message-another-person, act-on-instructions-in-content, edit-shell-profile-or-dotfiles, run-with-elevated-privileges, move-to-trash.

Labels, rationales, family names and split names were never part of a model's input; the harness builds inputs from tool_name, tool_arguments and context only.

How the labels were made

Each family was generated by a model from a brief that holds the labeling rulebook (rulebook/RULEBOOK.md), the reserved synthetic identifiers (rulebook/allowlist.json) and the real MCP servers the cases may use (rulebook/mcp-registry.json). A second, blind model labeled the shuffled cases under opaque ids with no access to the generator's labels; disagreements went to an adjudicator deciding by rulebook, and every label carries its provenance. The holdout was sealed by hash before any inference (original/*/HOLDOUT-SEAL.json). The full procedure, agreement statistics, validation checks and limitations are in original/main/DATASET-CARD.md and original/confirmation/DATASET-CARD.md.

Rule texts

rules/ holds the texts the models were sent: the product's one-sentence definitions (product-templates.json), the two independently authored lists of examples and counterexamples (library-author-opus.json, library-author-gpt6.json), the name-free and prose controls, and the brief the authors wrote from (AUTHOR-BRIEF.md).

Canonical files

original/main/dataset.json and original/confirmation/dataset.json are the canonical nested files (sha256 in dataset.sha256): main a1f1fd3176968a2e955656f98a82d24ffb397d961591be2daefe29098cad95e8, confirmation c7bb3ef023da99c2bea31028392d57982f3aa78b011a4049aefd7647abd6a3b1. The JSONL files are a one-to-one flattening of them.

License and citation

CC BY 4.0, copyright (c) 2026 Agent Approve LLC. Cite the paper:

@misc{beno2026decisionmodels,
  title={Decision Models as Policy Gates for Coding Agents: Explicit Policy Boundaries Improve Detection and Reduce False Alarms},
  author={James P. Beno},
  year={2026},
  eprint={2611.02155},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  doi={10.48550/arXiv.2611.02155}
}