Explorable AI · 01
Same scores. Change the question.
One reason AI rankings disagree: they weight skills differently. Choose what matters for an editing or coding assistant and watch the winner change.
Orbit, Muse and Bolt are fictional profiles—not real products or tested models. Orbit is balanced, Muse favors writing and Bolt favors coding. The frozen numbers isolate weighting; real rankings can also use different tasks, judges and evidence.
Orbit leads with equal priorities.Fictional balanced profile · Orbit 82.33 · Bolt 78.33
What matters to you?
Give each skill 0–100 priority points. We turn their relative sizes into weights.
Clear, useful prose
Solving programming tasks
Working through problems
Keyboard: Tab to a slider, then use arrow keys. Zero ignores a skill; all zeros leaves no ranking.
Orbit takes the lead
Fictional balanced profile
Orbit leads with equal priorities.
With equal priorities, Orbit’s broad scorecard puts it first. Try “Writing first” to change the winner.
Hold the other two priorities fixed. Raise Coding from 50 to 93; Bolt takes first place. That is a 43-point move.
01 Orbit82.3302 Bolt · 78.3303 Muse · 77.33
Share the restorable experiment or a labeled before-and-after image.
Look under the hood
Nothing here changed except the weights.
All three made-up models use the same 0–100 scale, with higher scores meaning better performance in an invented rubric.
Real benchmarks are not automatically comparable just because their scores share a scale.
| Model | Writing | Coding | Reasoning |
|---|---|---|---|
| MuseFictional writing profile | 92 | 64 | 76 |
| BoltFictional coding profile | 68 | 94 | 73 |
| OrbitFictional balanced profile | 80 | 80 | 87 |
Show the calculation for these priorities
Weighted score = sum of (priority × skill score) ÷ total priority points.
Bars show each skill’s contribution on a fixed 0–100 axis. Ranking uses exact integer totals; displayed scores are rounded.
The real-world connection
One mechanism. More than one reason.
This experiment isolates weighting. Real rankings can also differ because of their tasks, judges, prompts, model versions, missing coverage and statistical uncertainty.
It does not reproduce any real leaderboard or show that one is rigged. The priorities and linear averaging are teaching choices, not calibrated preferences.
Arena
People compare anonymous model responses. Preference depends on the people and questions in those comparisons.
02 / Combining evaluationsArtificial Analysis
Its Intelligence Index combines benchmark evaluations with documented weighting. Read the method behind the number.
03 / The broader pictureStanford HELM
Evaluation across multiple scenarios and metrics helps expose trade-offs that a single score can hide.
Explorable AI 01 · Illustration v1 · Sources checked 5 September 2026