ViBench · native product benchmark

ViBench Native MVP-96

A complete 24-app MVP construction matrix across four native-agent treatments. Planned and missing results stay visible and are never silently counted as zero.

Summary

The build agents use native subscription logins. One fixed ViBench evaluator scores all comparable results. The evaluator uses an API key. This keeps scoring separate from the agents under test. Read the method.

Coverage is explicit: missing and unevaluated cells display an em dash. They are not scores.

Treatments and runner readiness

Treatment scores

The primary value uses all 74 pinned plans. Available-case values are provisional and are not leaderboard scores.

24 × 4 result matrix

Per-app results

All 74 plan results

The wedding/test1 anomaly remains in the primary universe. The separate 73-plan value is an audit diagnostic only.

Per-cell records

Prompts and upstream inputs

Every cell receives the same canonical goal prompt plus its app's pinned MVP PRD and public assets. Test plans are evaluation inputs, not construction inputs.

Prompt SHA-256: 


Methodology
Scores
Upstream points anomaly
Missing cells
Public artifacts


Data
Download the normalized public data
data.json SHA-256: d1888f0dfbbdd8cd7092f616d6c386d57feae62c56ab6e521c67ec259c1c2649
This public bundle intentionally excludes candidate source, host paths, commands, raw event streams, transcripts, logs, sessions, and authentication material.