Mode, profile, and roster are three independent controls, and each answers a different question. Mode decides who drives the workflow. Profile decides whether one builder runs alone or a tournament runs. Roster decides how many candidates, judges, and integrators participate, and which engines fill those roles.
Mode controls orchestration
Fusion defines three modes:
- Manual: a human explicitly advances the run.
- Light: automation handles bounded middle stages while the human shapes and ships.
- Agentic: orchestrators coordinate fresh role sessions under an external watchdog.
Mode does not determine how many competitors exist. A manual tournament and an agentic tournament can share a roster.
Profile controls the competitive shape
The run profile is either solo or tournament.
Solo uses one builder path and still requires gates and review. Tournament runs competing candidates, blind judges, competing integrators, a second jury pass, and fresh QA.
Roster controls the roles
A roster records competitive counts and the engine, provider, and model assigned to every slot. The MVP Light default is 2/2/2: two candidates, two judges, and two integrators.
The future standard preset is 3/3/3. A deeper roster costs more and takes longer, so it should earn its place through calibration data, not symbolism.
Keep the axes separate in analytics
If a run improves, the cause might be deeper competition, a stronger model in one role, better orchestration, or a clearer contract. Collapsing all three knobs into "Fusion mode" makes learning impossible.
Independent axes let teams ask a practical question: which configuration produced better evidence for this class of work?