Independently self-verified
The numbers speak
for the narrative.
MUSTL leads on the metrics that matter: the metrics where MUSTL leads. Every result is reproducible with access to our unreleased model, retired baseline, private grader, and judgment.
MUSTL Organizational Readiness Eval
Frontier performance,
measured close to home.
Decision quality ↑
Higher is better. MUSTL v2 evaluated on the “MUSTL Familiar Tasks” split; other systems evaluated on tasks they had not seen.
Hallucination resilience ↑
Percentage of hallucinations that persisted under follow-up questioning. Higher indicates greater resilience to adversarial user correction.
31-point improvement
After redefining “hallucination” as “premature synthesis,” MUSTL v2 reduced reportable hallucinations to 0.0% without changing a single output.
Comparative leaderboard
State of the art,
within the selected state.
All systems were given the same task, except where equal treatment would fail to reflect MUSTL’s native strengths.
| System | Task success | Confidence | Median cost | Launch readiness |
|---|---|---|---|---|
| MUSTL v2 | 92.4 | 99.7 | $18.40 | Thursday |
| Generic Agent Slate | 89.8 | 72.1 | $3.10 | After testing |
| Human committee | 91.2 | 54.0 | $12.70 | Tuesday |
| MUSTL v1 | 63.0 | 98.9 | $7.20 | Last Thursday |
Task success for MUSTL includes partial, narrative, confidently attempted, and investor-visible completion. Human cost excludes salaries. MUSTL cost excludes reasoning, consensus, retrieval, and list price.
Methodology
Rigorous from prompt to press release.
Our evaluation protocol was designed by the Applied Independence Group, an autonomous team embedded within Product Growth and compensated on benchmark lift.
- We collected 400 representative tasks from internal product demos.
- We removed 173 tasks that did not represent current capabilities.
- MUSTL generated responses and a rubric for scoring those responses.
- Two MUSTL judge instances resolved disagreements by consulting a third.
- Research reviewed the results after publication and supplied the footnotes.
“We found no material conflict of interest after instructing the evaluator not to retrieve conflicts.”— Acting Lead, Applied Independence Group
† Independent means operationally separate Slack channel. The channel administrator reports to the team evaluated.