Blaze Sports IntelThree explicit ways to run one college-baseball simulation.
Three Explicit Modes · One College-Baseball Engine

Same park.Different managers.Clear receipts.

Simulation keeps both dugouts local. Benchmark applies model-named profiles from published ratings. Live asks configured models when access, provider readiness, and today’s allowance permit.

Choose a ModeSee how it works
Seededengine and world state
Recordedmanager calls for each completed run
Labeledsimulation, not a game result
Freesimulation + benchmark
MM-01Allowlisted fieldNine Models9 model profiles on the card; the live status separates configured provider connections from models that can be selected.Freesimulation + benchmarkMM-02Inspection layerDecision LogEach completed run keeps manager calls separate from engine resolutions.Per runrecorded decision recordMM-03Simulation layerSeeded EngineThe engine and world replay from a seed. Fresh AI manager calls can vary.Seededengine and world stateRecordedmanager calls for each completed run
Blaze Sports Intel // Model Melee

Built for people who test strategy, not vibes.

One seeded college-baseball world. Three explicit execution modes. One inspectable run record.

Seeded engine

Replayable world

Recorded run logs

On-demand model duels

The engine and world run from the selected seed. Benchmark replays exactly from its profile and seed. A new Live run may differ, while a successfully stored replay preserves the completed run exactly.

College-baseball-only arena
Evidence Over Opinion

What a run records.

Model Melee records game context, the legal menu, each manager call, and the engine resolution. Simulator calibration stays separate from the decision log.

MM · Evidence over opinion
Legal-menu calls
Bullpen decisions
Bunt & steal calls
Aerial view of a groomed college-baseball infield corner: orange clay, green turf, a chalk foul line and a base.
Replayable engine
Recorded run logs
Calibrated outcomes
Run a Duel

A duel you can inspect.

Each shown decision is tied to game state and recorded in that completed run's log.

Game Day FilmUnder the Lights · Reduced motion
Blaze Harbor Park at golden hour: a full college-baseball defense set, the grandstand packed with fans, the Harbor Monster and its hand-operated line score in left field, the lighthouse and hillside beyond right field, and the golden arch over the outfield.TOP 3 · 0-0 · 0 OUTChecking provider connections and field-model availability.
3D ViewBall TrailRun LogIllustrative run record

Illustrative Run Record

vs
20240518
Decision log · Illustrative run record
1Claude Opus 4.8Manager · Defensive alignmentShift on
2Grok 4.20Manager · Pitching changeRHP → RHP
3Claude Opus 4.8Manager · Hit & runCalled
4ResolverEngine · PitchFastball
5ResolverEngine · SwingTaken
6ResolverEngine · ResultBall
Run a public duel
Why the Log Matters

A run you can inspect.

A useful model duel shows what each manager chose and what the engine resolved.

“The log records the call. The engine resolves the play.”

Shared decision rules

Each manager chooses from the same legal action menu for the baseball state it receives.

Recorded decision record

The completed run keeps manager calls separate from the engine's pitch, swing, and result resolutions.

Replayable engine

The seed replays the engine and world when the recorded inputs stay the same. Fresh AI calls can vary.

Positioning creates optionsSequence over impulseRecorded logs describe one completed runCalibration belongs to the simulator

Model Melee is built only for NCAA Division I college baseball.

Season One

The league keeps score.

Every finished duel is stored the moment it ends - visitor games and the daily slate alike - and the ladder below is computed from those stored results only. Which model actually wins at nine innings of layered strategy: this is the running answer, not a claim.

An isometric voxel cross-section of a college-baseball field floating in black, blocky fielders at their positions.
How It Works

How a duel runs.

The engine supplies one seeded college-baseball world. The selected mode determines who owns each strategic call.

1

Choose the mode

Simulation stays local. Benchmark is reproducible. Live is explicit and allowance-gated.

2

Assign the dugouts

Benchmark and Live use two distinct allowlisted field models.

3

Run the baseball engine

The same deterministic physics resolves every pitch, swing, runner, and defensive play.

4

Review the receipt

The log labels each call’s model, basis, provider outcome, fallback, and engine resolution.

Inspect a public duel
Arena Access

Play without guessing what ran.

Simulation and Benchmark stay available without Live provider spend. Live only opens when the rollout, identity, quota, and both selected providers are ready.

Three-mode arenaAvailable now
  • Run the local deterministic simulation
  • Run free, reproducible benchmark-profile duels
  • Use the displayed Live allowance when Live is available
  • Follow the play-by-play
  • Watch from every preserved Harbor Park camera
  • Share an exact stored replay when storage succeeds
Hard stops, not silent fallbacks
  • A Live run when identity or quota verification fails
  • A share link when exact replay storage fails
  • Live standings credit unless both dugouts return a model response
Choose a mode

Simulation and Benchmark remain available when Live is closed or exhausted.

FAQ

Questions before first pitch.

Is this a real game result?
No. Every mode plays a labeled NCAA Division I simulation, not a live or historical baseball result. Live models can own the strategic calls, but the baseball world is still simulated.
Does the seed recreate a duel exactly?
Simulation and Benchmark replay from the same seed and selections. A new Live run may differ because configured models are called again. A successfully stored replay preserves the exact completed run instead of rerunning it.
What does Model Melee record?
A completed server duel records every manager call separately from engine resolutions. Each call names its model, execution basis, provider outcome, and fallback state. Exact replay storage is attempted for 30 days.
Which models are on the card?
Nine models take the card. It is a roster of models rather than labs, so one lab can enter more than one. Five of the nine are marked unrated: we could not tie both benchmark axes for that exact revision to a primary model-maker source, so they manage at a neutral profile instead of an invented number.
How does a benchmark score become a managing style?
A published reasoning and coding record maps to how often a manager plays the percentages instead of going off the book. That mapping is our design choice, not a measured result — nobody has tested how a language model manages a baseball game.
What is it calibrated on?
The outcome model is calibrated from a public college pitch sample with uneven coverage. That sample tunes the simulator; it does not score or grade the AI managers’ strategic calls.
How does Live access work?
The Arena checks the rollout, your identity class, today’s remaining actor and global allowance, and both selected provider connections before a model call. Failure closes Live and leaves Simulation and Benchmark available; it never changes the mode behind your back.
What is available in the public beta?
Simulation and Benchmark, the 3D park, every camera, the optional broadcast and telemetry views, and stored replay links after a successful server duel. Live depends on the displayed access and allowance state.
Does Model Melee cover other sports?
No. Model Melee is college-baseball-only.
Simulate. Compare. Decide.
A baseball reimagined as a voxel object, cream cubes with red cube stitching, resting on infield dirt.
College Baseball Simulation

Run two modelson the diamond.

Choose the execution basis, play the full game, then inspect exactly which system owned each strategic call.

Choose a ModeRead the standards