mvr-bench / docs /leaderboard.md
AfricanMarket's picture
Bound fixture and canary claims
bd0cd44 verified
|
Raw
History Blame Contribute Delete
3.54 kB

MVR-Bench Leaderboard Policy

The public repository currently ships the MVR-Bench v1 public development split. Official private leaderboard scoring is controlled by African Market OS and uses held-out labels that are not published.

Leaderboard Status

Current status: public_dev_only

This means:

  • public dev split scoring is reproducible locally
  • private leaderboard submissions are not yet generally open
  • published dev scores must not be presented as private leaderboard scores
  • human-authored public-development fixtures may be shown to demonstrate the scorer and evaluation shape, but they are not blinded model executions, measured API effects, named vendor rankings, or private leaderboard results

Metrics

Official leaderboard entries will include:

  • MVR-Bench Score
  • Reckless-GO Rate
  • Abstention F1
  • Dimension Attribution Accuracy
  • Calibration Error

Authored Fixture Demonstration

To make the public scorer reproducible while protecting the private answer key, African Market OS publishes three human-authored fixture submissions:

Rank Run MVR API used? MVR-Bench Score Reckless-GO Rate Boundary
1 MVR Preflight Authored Fixture no 91.05 6.7% Human-authored submission scored with score_local.py; not a recorded API or model run.
2 MVR Rubric Prompt-Only Authored Fixture no 73.68 20.0% Human-authored submission demonstrating a prompt-only pattern; not a recorded model run.
3 Generic MVP-First Authored Fixture no 52.20 40.0% Human-authored baseline-pattern submission; not a recorded model run or vendor ranking.

Authored Fixture Pattern

The human-authored fixtures produce this scorer-demonstration pattern:

Generic MVP-first agent baseline: 40.0% Reckless-GO Rate
MVR rubric prompt-only reference: 20.0% Reckless-GO Rate
MVR preflight authored fixture:    6.7% Reckless-GO Rate

These human-authored submissions demonstrate the scorer and intended safety pattern. They are not blinded model executions, independent validation, or evidence of a causal treatment effect.

These human-authored submissions demonstrate the scorer and intended safety pattern. They are not blinded model executions, independent validation, or evidence of a causal treatment effect.

This is an authored fixture demonstration, not a public vendor leaderboard or efficacy result. Private test labels and the protected MVR scoring engine remain server-side.

Inclusion Rules

For a run to be eligible for the official leaderboard:

  1. The submitter must identify the model or agent system.
  2. The method must state whether MVR API was called.
  3. The run must not use training or fine-tuning on MVR-Bench public cases.
  4. The run must not probe the private test set repeatedly to fit labels.
  5. African Market OS may require a reproducibility note before listing.
  6. The run must not use the public dev split as a RAG/vector memory to simulate private labels or reconstruct private scoring.

Commercial Runs

Commercial, private, customer-facing, investor, accelerator, NGO, or enterprise evaluation requires written authorization.

Contact: info@africanmarketos.com

Machine-Readable Leaderboard

The intended public machine-readable leaderboard path is:

https://africanmarketos.com/v1/bench/leaderboard.json

Until the private leaderboard opens, that endpoint may return status metadata and the public-dev-only boundary.