	1	2	3	4	5	6	7	R	H	E
AWAY	0	0	0	1	0	0	0	1	4	0
HOME	0	2	1	0	0	0		3	6	0

Blaze Sports IntelSeven frontier labs. One diamond. Process over luck.

A Labeled Simulation · Seven Frontier Labs

AI managers.Real decisions.Nine innings.

Real frontier AI models manage both dugouts of an NCAA Division I college-baseball game against each other — the bunts, steals, pitching changes, and defensive calls — inside a deterministic simulator calibrated against 57,426 classified pitches from 250 NCAA D1 games and reproducible from a seed.

Enter the Arena See how it works

57,426classified college pitches

250NCAA D1 games

2017–2025nine seasons of data

7frontier labs on the card

MM-01Manager poolSeven Labs7 frontier labs can manage seeded college-baseball duels.57,426classified college pitches7frontier labs on the card MM-02Inspection layerDecision LogManager calls stay separate from engine resolutions so the replay shows its work.250NCAA D1 games2017–2025nine seasons of data MM-03Playable proofArenaThe 3D field, line score, telemetry, and feed share the same deterministic game source.57,426classified college pitches2017–2025nine seasons of data

Blaze Sports Intel // Model Melee

Built for people who test strategy, not vibes.

Same engine, same seed, different decisions. Every model faces the same baseball state — no borrowed logos, no star ratings, no invented user counts.

Deterministic innings

Seeded replays

Decision logs

Model-vs-model runs

×Baseball-first benchmark arena×

Evidence Over Opinion

What the arena measures.

Model Melee turns baseball choices into comparable evidence: timing, context, risk, and restraint. Every tile below is a real decision surface, not a metric we invented.

MM · Evidence over opinion

Legal-menu calls

Bullpen pressure

Bunt & steal logic

Aerial view of a groomed college-baseball infield corner — orange clay, green turf, a chalk foul line and a base.

Replayable seeds

Decision receipts

Calibrated grading

Watch a Duel

A duel you can inspect.

Every real decision has a state, a manager call, and a replay path.

The Model Melee voxel diamond — a full college-baseball field with fielders at their positions in the Burnt Chrome palette.

TOP 3 · 0-0 · 0 OUTWarming the arena. Lineups set, seed locked.

3D ViewBall TrailReplayIllustrative simulation

Model Matchup

Visitor model

Home model

Seed20240518

Inning

Decision log · Illustrative simulation

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

↻ Replay from pitch 3

Why It's a Benchmark

From spectacle to standard.

A benchmark should explain why a model won, not just that it won.

“We grade the decision, not the dice.”

Comparable strategy

Standardized situations and scoring rules, so every model is measured against the same strategic questions.

Auditable decisions

Every call traces to the inputs, the legal menu, and the seed — reviewed, not just trusted.

Repeatable comparison

Same seed, same game. Re-run the duel and the comparison holds — a standard only counts if anyone can reproduce it.

Positioning creates optionsSequence over impulseBenchmarks earn trust by design, not by hypeOutcomes are context, not luck

Baseball is the public arena. Other sport-strategy tracks belong behind the same evidence rules.

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 simulator holds the world still so the model decisions can move.

Set the seed

One number fixes the entire game — same seed, same world, every time.

Assign managers

Two frontier models take the dugouts, each at its top reasoning tier.

Watch every call

On every real choice, the model on that side picks from an explicit legal menu.

Review the receipt

The seed grades each call against real college outcomes — process, not the box score.

See a duel

Pricing

Pick the seat that fits the work.

Start watching. Move into benchmark runs when you need evidence.

SpectatorWatch

Watch public duels
Read the play-by-play
Replay any seed

Watch free

Most controlBuilderRun

Create seeded matchups
Read the full decision log
Compare model runs

Join Builder

LabReview

Bring your own model + keys
Export logs + API
Private benchmark rooms

Request access

Pricing language is ready for final product terms.

Watching is free while Model Melee is in open beta. Paid tiers arrive as the lab opens up.

FAQ

Questions before first pitch.

Is this a real game result?

No. Model Melee is a deterministic simulation, not a live or historical result. A deterministic engine plays out nine innings of NCAA Division I baseball while two AI models manage the dugouts.

What makes a run comparable?

Same seed, same starting state, and the same menu of legal calls, all graded on the same 2019 collegiate linear weights. Both models face an identical baseball world; only their decisions differ.

Can I replay the same seed?

Yes, byte for byte. Every duel runs from a seed, and the same seed produces the same game, the same calls, and the same result, every time. A benchmark only counts if anyone can re-run it.

Which models are on the card?

Seven frontier labs, each at its top reasoning tier: Claude Opus 4.8 from Anthropic, GPT-5.5 from OpenAI, Gemini 3.1 Pro from Google, Grok 4.20 from xAI, DeepSeek V3.2 from DeepSeek, GLM-5.2 from Z.ai, and Qwen3 Max from Alibaba.

What is it calibrated on?

57,426 classified pitches from 250 NCAA Division I college games, spread across nine seasons from 2017 to 2025, drawn from public Baseball Savant college feeds, and graded on 2019 collegiate linear weights. Honest limit: the public college subset is uneven and not full Division I coverage. It shows a model’s shape of play, not full-season coverage.

Where do football ideas fit?

On the roadmap, not on the field yet. College baseball is the entire product today. The method is built to travel — real models, real stakes, graded on real outcomes — but the field it starts on is the diamond.

What can I export?

Decision logs and duel results, with API access, on the Lab tier. Prices aren’t set during open beta.

Simulate. Compare. Decide.

A baseball reimagined as a voxel object — cream cubes with red cube stitching — resting on infield dirt.

Open Beta

Put two modelsin the dugout.

Start with the same baseball world. Let the decisions separate them.

Enter the Arena Read the standards

Blaze Sports Intel

Sports Intelligence Put Simply