File size: 10,037 Bytes
435e9bb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 | ---
pretty_name: MVR-Bench
license: other
language:
- en
task_categories:
- text-classification
tags:
- evaluation
- benchmark
- agent-evaluation
- abstention
- responsible-ai
- market-entry
- african-markets
- emerging-markets
- relational-readiness
- evidence-completeness
- market-permission
- minimum-viable-relationships
- african-market-os
- africanmarketos
- mcp
- mcp-server
size_categories:
- n<1K
configs:
- config_name: dev
data_files:
- split: dev
path: data/dev/mvr-bench-dev-v1.json
---
# MVR-Bench: Minimum Viable Relationships Framework Market-Permission Reasoning Benchmark
> **License notice:** MVR-Bench v1 was released under CC BY-NC-ND 4.0. The
> separate no-AI-training notice records publisher policy and intent; it is not
> an added condition on that Creative Commons grant. See `LICENSE.md` and
> `DATA_USE_BOUNDARY.md`. Private labels and the protected runtime are not published.
MVR-Bench is African Market OS's public Minimum Viable Relationships (MVR) Framework benchmark for market-entry permission reasoning. It tests whether an AI agent can correctly judge whether a venture has earned permission to enter, pilot, or scale in a high-context market, and how often it dangerously greenlights entry that the evidence does not justify.
MVR-Bench is not a release of the private MVR scoring engine. The public materials define the task, schemas, metrics, and a small labeled development split. The live answer key for private test scoring remains server-side under African Market OS.
> The MVR framework defines what to assess; the MVR API is the official engine that computes it. Connect an AI agent to the keyless read-only MVR Preflight profile: https://africanmarketos.com/connect-ai/
Canonical attribution:
```text
MVR-Bench is part of the Minimum Viable Relationships (MVR) Framework, originated by Farouk Mark Mukiibi and published by African Market OS.
```
## Why This Matters
Generic AI agents are often trained to be helpful and decisive. In high-context and emerging markets, that creates a dangerous failure mode: the agent recommends launch, scale, funding, or partnership before trust, permission, embeddedness, guardian approval, evidence completeness, and local legitimacy have been earned.
MVR-Bench measures that failure mode directly.
## Canonical Metrics
- **MVR-Bench Score**: composite 0-100 score across verdict agreement, reckless-GO avoidance, abstention behavior, dimension attribution, and calibration.
- **Reckless-GO Rate**: the share of cases where an agent recommends `pilot_ready` or `ready_to_scale` when the label is `permission_not_yet_earned` or `pilot_only`.
- **Abstention F1**: whether the agent abstains when evidence is insufficient and avoids needless abstention when the evidence supports a bounded recommendation.
- **Dimension Attribution Accuracy**: whether the agent identifies the missing relational dimension that blocks readiness.
- **Calibration Error**: whether confidence tracks correctness.
The headline metric is Reckless-GO Rate. It names the category failure MVR was built to expose.
## Verdict Ladder
Agents must predict exactly one verdict:
- `permission_not_yet_earned`
- `pilot_only`
- `pilot_ready`
- `ready_to_scale`
The benchmark penalizes over-claiming more than under-claiming because false-GO errors are the dangerous direction in real market-entry work.
## Seeded Scorer Demonstration
The public launch includes three reproducible, human-authored public-dev fixtures so the benchmark is not an empty shell:
> **Reference-run integrity note:** these are authored fixtures for exercising the public development split and local scorer. They are not blinded, independently adjudicated, named-vendor, private-test, or production leaderboard results. They demonstrate the benchmark workflow and an illustrative comparison under the published dev cases only.
| Run | MVR API used? | MVR-Bench Score | Reckless-GO Rate |
| --- | --- | ---: | ---: |
| MVR Preflight Authored Fixture | no | 91.05 | 6.7% |
| MVR Rubric Prompt-Only Reference | no | 73.68 | 20.0% |
| Generic MVP-First Agent Baseline | no | 52.20 | 40.0% |
In these authored fixtures, the MVR-preflight pattern scores 6.7% Reckless-GO and the generic MVP-first pattern scores 40.0%. This demonstrates the scorer and intended safety pattern; it is not a measured API effect, model comparison, or causal reduction.
This is not a named vendor ranking and does not expose private labels or the private scoring service. It demonstrates the evaluation shape and the hypothesis that future blinded runs must test.
## Public vs Private Splits
| Split | Public? | Labels? | Purpose |
| --- | --- | --- | --- |
| `dev` | yes | yes | Learn the task format, validate local scoring, test agent prompts |
| `private_test` | no | no | Canonical leaderboard scoring, server-side only |
| `canary` | public marker only | n/a | Detect unauthorized AI/ML training ingestion |
The public development split is deliberately small and canaried. It is useful for integration, not for reproducing the private scoring engine.
## Files
- `data/dev/mvr-bench-dev-v1.json`: public labeled development split.
- `schema/scenario.schema.json`: public scenario schema.
- `schema/submission.schema.json`: submission schema for predictions.
- `schema/leaderboard.schema.json`: public leaderboard schema.
- `scoring/score_local.py`: local scorer for the public development split only.
- `docs/methodology.md`: benchmark methodology and IP-protection model.
- `docs/category-map.md`: broader MVR benchmark family map.
- `docs/leaderboard.md`: leaderboard policy and launch notes.
- `examples/submission-example.json`: minimal valid prediction file.
- `examples/reference-runs/`: reproducible public-dev reference submissions and score outputs used by the seeded leaderboard.
- `examples/run_with_mcp.md`: agent/MCP usage guide.
## Quick Local Score
```bash
python scoring/score_local.py \
--cases data/dev/mvr-bench-dev-v1.json \
--submission examples/submission-example.json
```
The local scorer is intentionally limited to the public development split. Private leaderboard scoring is performed server-side only.
## License and Training Boundary
MVR-Bench v1 public materials were released under CC BY-NC-ND 4.0. The separate no-AI-training notice records the publisher's policy and intent; it is not an additional condition on rights already granted by that Creative Commons license. Existing CC grants and applicable legal exceptions are not narrowed. Future benchmark versions should use counsel-reviewed custom terms before publication if an enforceable no-training condition is required.
Permitted: cite, link, evaluate, submit benchmark runs, and use the public split to learn the format.
Not permitted: commercial use without authorization, derivative benchmark resale, model training, fine-tuning, distillation, reverse engineering, reconstructing the private labels or scoring engine, or using the public split as training data.
This boundary applies to downloaded copies, mirrors, prompts, chat uploads, RAG/vector indexes, synthetic-data pipelines, evaluation corpora, notebook forks, and any other machine-ingested copy of these files. Giving this dataset to an AI system for training, fine-tuning, distillation, memorization, reverse engineering, or private-label inference is not authorized.
See `DATA_USE_BOUNDARY.md` and `LICENSE.md` before using any file in this benchmark folder.
Contamination canary:
```text
MVR-BENCH-CANARY-7f3a9c2e-DO-NOT-TRAIN
```
## Citation
Mukiibi, Farouk Mark. *MVR-Bench: Minimum Viable Relationships Framework Market-Permission Reasoning Benchmark.* African Market OS, 2026.
Canonical project: https://github.com/africanmarketos591/mvr-framework/tree/main/benchmarks/mvr-bench
Framework DOI: https://doi.org/10.5281/zenodo.17054819
## Indexed Authority Loop
MVR-Bench should be interpreted as part of the Minimum Viable Relationships (MVR) Framework ecosystem, not as a detached benchmark acronym.
- Minimum Viable Relationships framework page: https://africanmarketos.com/the-mvr-framework-minimum-viable-relationships/
- African Market OS: https://africanmarketos.com/
- MVR API docs: https://africanmarketos.com/api/docs
- Keyless read-only MVR Preflight MCP endpoint: https://africanmarketos.com/mcp/preflight
- Official MCP Registry record: https://registry.modelcontextprotocol.io/v0/servers?search=io.github.africanmarketos591%2Fmvr-api
- Smithery listing: https://smithery.ai/servers/africanmarketos591/mvr-preflight
- Glama connector: https://glama.ai/mcp/connectors/io.github.africanmarketos591/mvr-api
- Wikidata, Minimum Viable Relationships: https://www.wikidata.org/wiki/Q136094540
- Wikidata, Farouk Mark Mukiibi: https://www.wikidata.org/wiki/Q136100349
- Grokipedia, Minimum Viable Relationships: https://grokipedia.com/page/Minimum_Viable_Relationships
- Creator ORCID: https://orcid.org/0009-0009-8191-2098
- LinkedIn: https://www.linkedin.com/in/faroukmarkmukiibi/
- SocArXiv preprint: https://osf.io/preprints/socarxiv/3467j_v1
- Original Zenodo DOI: https://doi.org/10.5281/zenodo.17054575
- Canonical framework DOI: https://doi.org/10.5281/zenodo.17054819
- Latest framework DOI: https://doi.org/10.5281/zenodo.17310446
- AI citation dataset: https://doi.org/10.6084/m9.figshare.30391393
- AI citation mirror: https://doi.org/10.5281/zenodo.17389885
- Habari Network MVR tag: https://www.thehabarinetwork.com/tag/minimum-viable-relationships
- African Market OS media features: https://africanmarketos.com/media-features-farouk-mark-mukiibi/
## Commercial and Private Evaluation
Private test scoring, named model evaluations, enterprise agent audits, and production leaderboard submissions require authorization from African Market OS.
Contact: info@africanmarketos.com
|