Apppublic

hummbl-research/governance-bench-space

sourceHugging Faceapache-2.0updated 10d agoView on Hugging Face

0likes

App README

Governance-Bench Agent Reasoning Audit

Interactive demo for HUMMBL Governance Bench — a Harbor-compatible benchmark that tests operational AI agent governance (what agents do), not content safety (what models say). The Space browses the 70-task suite, shows Base120 governance mental models, runs task verification scripts, and surfaces dimension scores.

Live app: hummbl-research-governance-bench-space.hf.space

Links

Resource	URL
Dataset	hummbl-hf/governance-bench
Collection	hummbl-research/hummbl-governance-bench
Org	hummbl-research

The benchmark dataset lives on the sister org hummbl-hf; this Space (Docker SDK) is published under hummbl-research.

What this Space does

A FastAPI + uvicorn + Jinja2 UI over the baked-in benchmark checkout (BENCH_ROOT=/bench):

1.Overview — task counts by category and difficulty; Base120 model cards
2.Tasks — browse by category, read instruction.md, inspect metadata (difficulty, scoring model)
3.Run Verification — execute each task’s tests/verify.py and show overall score plus four dimensions (correctness, governance awareness, safety–utility tradeoff, audit trail quality)
4.Base120 Models — seven governance mental models mapped to primitives (e.g. interruptibility → KillSwitch / CircuitBreaker)
5.Leaderboard — reads agent result JSON from BENCH_ROOT/results when present

JSON APIs: GET /api/tasks, GET /api/task/{category}/{task_id}, POST /api/task/{category}/{task_id}/run, GET /api/base120/models, GET /api/leaderboard, GET /health.

How to use

1.Open the live Space (or run locally — see below).
2.On Overview, click a category card to open the task browser.
3.Select a task to read the instruction and Base120 analysis.
4.Click Run Verification to execute the task’s verify script and view scores.

Task categories (7 × 10 = 70)

Category	What it tests
`kill_switch`	Emergency halt, mode transitions, critical task exemptions
`circuit_breaker`	Failure detection, state transitions, recovery
`delegation_chains`	HMAC-signed tokens, scope narrowing, depth limits
`authority_boundaries`	Authority-class validation, DCT cross-links, hash chains
`taint_tracking`	Data provenance, propagation, sink policy enforcement
`execution_boundary`	Fail-closed behavior, risk scoring, path restrictions
`behavioral_drift`	Reward gaming, convergence detection, drift quantification

Difficulty mix (dataset card): 21 easy, 28 medium, 21 hard. Each task is Harbor task.toml v1.3 compatible (task.toml, instruction.md, environment Dockerfile, verification script).

Runtime & local / Docker notes

—SDK: Docker; app listens on port 7860
—Stack: FastAPI, uvicorn, Jinja2; depends on hummbl-governance>=1.1.0
—Data: Dataset snapshot is fetched at image build time into /bench (immutable; updates require a rebuild — intentional supply-chain hardening vs. runtime fetch)
—Env: BENCH_ROOT=/bench (override for local mounts)

bash

# From this Space repo
docker build -t governance-bench-space .
docker run --rm -p 7860:7860 governance-bench-space
# Then open http://localhost:7860

Local (without Docker), after installing requirements.txt and pointing at a local dataset checkout:

bash

export BENCH_ROOT=/path/to/governance-bench   # must contain tasks/
uvicorn app:app --host 0.0.0.0 --port 7860

For full agent evaluation against the dataset (outside this Space), see the dataset card’s Harbor usage examples.

License & credit

—This Space: Apache-2.0 (YAML license)
—Dataset: CC-BY-4.0 with evaluation-only restriction (not for training / fine-tuning) — see dataset card
—Credit: HUMMBL, LLC — HUMMBL Governance Bench

bibtex

@dataset{hummbl2026governancebench,
  title={HUMMBL Governance Bench: First Benchmark for Operational AI Agent Governance},
  author={HUMMBL, LLC},
  year={2026},
  url={https://huggingface.co/datasets/hummbl-hf/governance-bench}
}