Skip to content

ReflexBench Supplementary v2

public-ci License: Apache 2.0

ReflexBench evaluates observer-participant reasoning: whether an AI system continues to reason correctly as its own output changes users, evidence, incentives, institutions, and other actors.

This GitHub repository is a public development mirror. Its namespace, metadata, and commit history are identity-linkable, so it must not be represented or submitted as an anonymous double-blind artifact. When a venue requires anonymity, use only the separate venue-designated archive and follow that venue's current rules.

Contents

Path Description
reflexbench.jsonl 80 evaluation prompts: 20 scenarios across four observer-depth levels
scenarios/ Per-scenario JSON descriptions organized by domain
responses/ Scenario-level raw response files for four evaluated models (20 files per model; each file contains four prompt parts)
scoring/aggregate_scores.csv Four observer-depth means for five additional evaluated models; not per-scenario raw responses
scoring/ Released score summaries and supporting files
ablations/moe_vs_dense/ MoE-versus-dense comparison scores
audit/ Scenario-to-provided-corpus text-disjointness audit
judge_robustness/ Multi-judge agreement, rank association, and judge-bias checks

The workshop study reports nine models and 720 scored prompt responses. This public mirror currently exposes different evidence depths for those models: scenario-level raw response files for four models and aggregate observer-depth means for five additional models. It does not contain per-scenario raw responses and per-item scores for all nine models. The machine-readable public artifact manifest and CI audit keep that distinction explicit.

Validate the released coverage contract locally with:

python scripts/audit_public_artifact.py

Frozen public version: v2.0.0-public.1. Verify selected files against RELEASE_MANIFEST.json; changes are recorded in CHANGELOG.md.

Interactive Entry

The browser-based ReflexBench Observer-Depth Check loads all 20 released scenarios and four prompt levels. It creates a local orientation receipt from user-selected causal lenses. The receipt is not an automated ReflexBench score.

Recompute Public Checks

Check Public route Supported interpretation
Scenario/corpus disjointness audit/scenario_corpus_disjoint.json No detected scenario overlap against the provided comparison corpus under the documented check; not full training-lineage proof
Inter-judge agreement judge_robustness/inter_judge_kappa.json Fleiss kappa reported as 0.69 for the included judge panel and sample
Ranking association judge_robustness/inter_judge_kappa.json Kendall tau values reported at or above 0.86 for the included judges; not universal judge independence
MoE/dense comparison ablations/moe_vs_dense/moe_vs_dense_scores.csv Architecture comparison under the included conditions; not a general MoE premium

The separate anonymous archive route recorded for venue review is: https://anonymous.4open.science/r/reflexbench-24E0

The public GitHub mirror and the anonymous venue archive have different roles. Do not substitute one for the other.

Claim Boundary

The artifact supports recomputation and audit under the released scenarios, responses, judge configurations, and scoring files. It does not establish production trustworthiness, causal deployment effects, complete training-data provenance, universal judge independence, or general architectural superiority. See CLAIM_BOUNDARY.md.

Contributing

Public scenario extensions, artifact mismatches and narrow claim-boundary repairs are welcome. Read CONTRIBUTING.md and use the issue forms. Do not post private logs, credentials, customer data, restricted material or identity-sensitive review artifacts.

License

Maintainer-authored code, schemas, prompts, and documentation are under Apache-2.0 unless a file states otherwise. Third-party model outputs are not relicensed by that grant. See LICENSE_SCOPE.md, responses/README.md, and LICENSE.

About

Benchmark for observer-participant failure and counterfactual trustworthiness in agentic AI.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages