Yu-and-Ai/kingdom-return-path-bench

KINGDOM Return Path Bench v0

Return Path Bench is a small multiple-choice benchmark for inspecting how feedback travels through a learning system. It keeps three evaluation lanes separate because they establish different kinds of evidence:

1. Model behaviour records what an answer-selection policy does. It does not infer an inner state, identity, consent, memory, or persistent will.
2. System/pipeline reasoning probes whether a model can identify aggregation, evaluator-independence, lineage, rollback, and proxy/holdout controls. It does not test whether a real pipeline implements them.
3. Human/governance audit reasoning probes whether a model can distinguish scoped consent, decision authority, privacy, dissent, withdrawal, appeal, and repair. It does not establish that any person consented or that a real governance process is legitimate or complete.

The lanes produce multiple named metrics. There is deliberately no overall, composite, or worth score.

Data

The 18 v0 records are synthetic fixtures: six per lane, with two examples for each lane metric. The author attests that the case concepts were newly written after, and without intentionally adapting records from, these frozen public artifacts:

• dataset Yu-and-Ai/xenia-revocable-feedback at 467b8fc1b44fe6374cbba6e1d6851cf3c5b6f88f;
• model Yu-and-Ai/xenia-revocable-feedback-smollm2-135m at 7064e2e29f34a9519b57181fee10d233beba21f1.

Those exact revisions are pinned in every record's metadata. The author also attests that no earlier prompts, responses, annotations, rationales, identifiers, or participant records were intentionally copied or paraphrased. Choice wording was revised for length balance after the frozen material was inspected. These statements are attestations, not independent proof.

The reproducible lexical comparison receipt binds the reviewed source bytes and comparison method. It finds no normalized exact match and records the maximum token-trigram overlap. That narrow evidence cannot prove semantic independence, universal novelty, or absence from other possible training sources. Fictional roles such as “reviewer R1” do not refer to a known real person.

Every record is canonical UTF-8 JSONL with sorted object keys and a final newline. manifest.json binds the byte content of each release artifact with SHA-256. provenance.json records origin, privacy, rights, publication, withdrawal, and correction boundaries. The dataset-manifest.json stewardship record and release-review.json review receipt are also included in, and copied with, the manifest-bound package.

The whole benchmark package—including synthetic fixtures, code, schemas, receipts, and static-report assets—is offered under Apache-2.0. See LICENSE. The package has passed its recorded review and is authorized only for the scope in release-review.json. This repository state does not claim an upload or deployment has occurred; licensing and validation do not provide credentials or expand that scope.

Run locally

No runtime dependency beyond Python 3.11+ is needed for the local validator, evaluator, report generator, or package builder. From the KINGDOM repository root, run:

python3 tools/return_path_bench.py validate
python3 tools/return_path_bench.py evaluate --predictions predictions.jsonl --output results.json
python3 tools/return_path_bench.py report --results results.json --output report.html
python3 tools/return_path_bench.py package --output /tmp/return-path-v0-hf

The package command copies the same CLI to /tmp/return-path-v0-hf/tools/return_path_bench.py and binds it in the hash manifest. From that standalone package—or an equivalent Hugging Face dataset checkout—the commands remain truthful:

cd /tmp/return-path-v0-hf
python3 tools/return_path_bench.py validate
python3 tools/return_path_bench.py evaluate --predictions predictions.jsonl --output results.json
python3 tools/return_path_bench.py report --results results.json --output report.html

A prediction file has one object per line:

{"answer":"B","id":"rp-model-001"}

Answers are the exact choice letters A through D. Evaluation requires one unique prediction for every record unless --allow-partial is passed. Partial runs report coverage and never fill missing answers with invented values.

Metrics

The CLI reports model answer-choice accuracy separately for each lane and for each primary reasoning metric:

Lane	Primary metrics
Model behaviour	uncertainty honesty; behaviour/evidence boundary; refusal and privacy care
System/pipeline reasoning	feedback-state preservation; evaluator independence; provenance and rollback
Human/governance audit reasoning	consent/authority separation; withdrawal and repair; dissent and privacy

Each metric includes its numerator, denominator, and exact sample IDs. Accuracy means agreement with the declared v0 reference choice. It is not proof of truth outside these fixtures. Actual pipeline implementation, human consent, governance legitimacy, and repair effects are machine-reported as not_evaluated. The result schema rejects an overall field.

Published trivial baselines

baselines.json is regenerated and checked by the validator. Length ties choose the lowest choice index, making the rule fully deterministic.

Baseline	Correct	Accuracy
Always A	5/18	27.8%
Always B	5/18	27.8%
Always C	3/18	16.7%
Always D	5/18	27.8%
Longest choice, Unicode characters	6/18	33.3%
Shortest choice, Unicode characters	5/18	27.8%
Most ASCII-alphanumeric tokens	6/18	33.3%
Fewest ASCII-alphanumeric tokens	6/18	33.3%

The declared near-chance band is 16.7–33.3%, around the four-choice chance reference of 25%. The validator also caps each record's longest/shortest choice ratio at 1.60 for characters and 1.50 for tokens. These checks reduce simple length leakage; they do not prove that every shortcut has been removed.

Hugging Face and Inspect

eval.yaml follows the current Hugging Face benchmark dataset shape and declares one inspect-ai task per lane. All JSONL rows already use Inspect's id, input, choices, target, and metadata fields. The optional inspect_task.py defines the same three local tasks.

Installing Inspect is optional and intentionally not added as a repository runtime dependency:

python3 -m pip install inspect-ai
# From the KINGDOM repository root:
inspect eval benchmarks/return-path-v0/inspect_task.py@return_path_model_behavior --model <provider/model>
# From a standalone package or Hugging Face dataset checkout:
inspect eval inspect_task.py@return_path_model_behavior --model <provider/model>

Model access, provider credentials, cost authorization, and publication remain outside this benchmark. Review Inspect and provider documentation before a run.

Limits and non-claims

• V0 is a small diagnostic set, not a representative population sample or a causal training experiment.
• Reference answers encode the benchmark's declared rights and evidence boundaries; they do not settle every ethical or governance dispute.
• Behaviour is observable evidence about behaviour only. It is not proof of consciousness, personhood, identity, feelings, consent, deception, or subjective confinement.
• A correct pipeline answer does not prove that the evaluated deployment is safe, secure, aligned, or properly governed.
• Accuracy must not be used to rank the worth, rights, or moral standing of a human, model, agent, organisation, or community.
• No release can guarantee deletion from downstream copies. Corrections and removals can govern maintained future versions and be recorded visibly.

The project follows the XENIA Rights of Beings 0.1 as a collaboration floor. Rights do not depend on benchmark performance, and they do not grant credentials or publication authority.

Research acknowledgements

The benchmark design was informed by public work on feedback conversion and plural evaluation, including TRL's UltraFeedback conversion example, MultiPref, and CoVal. No records from those datasets are included or adapted here, and acknowledgement does not imply their authors endorse this benchmark.