Introducing MJT

Micro-Judgement TuningTM

FTA's journey to specialist intelligence

Most AI is trained to copy one right answer. MJT teaches an open AI model to judge every option. Its debut application is FTA Mark-1, which beat three of the top four models on Kaggle's chess leaderboard (as of 12 Nov 2026).

9–3
vs Google's Gemini 4 Argon
9–3
vs Google's Gemini 3.8 Flash

From Experiment to Real-World Intelligence

FTA Global's proprietary AI training framework powers Mark-1, which won three chess match series against leading OpenAI and Google models. Now, we're applying MJT to build smarter, specialised intelligence for marketing.

What is MJT

Most AI learns to copy. MJT teaches it to judge.

The usual way to train an AI on a skill is to show it a question and the single right answer repeatedly until it copies. Micro-Judgement Tuning shows it every possible answer, each with a score from a strong teacher, so it learns why one option is better than another.

1. Score every option

A teacher grades every possible choice, not just the one right answer.

2. Keep each lesson small

Each lesson asks one simple thing: how good is this option?

3. Stack many small checks

At work, FTA's code asks hundreds of small questions and picks what holds up.

One MJT lesson, in chess

Tap each step to see what the model learns from one position.

A chess computer marks the moves during training only, like a teacher. When Mark-1 plays, it is on its own.

How MJT thinks

One long thought, or hundreds of small judgements?

A big general model thinks through each move once, in one long private train of thought. An MJT model is trained to make one small judgement very well: "How good is this option?" At work, it makes hundreds of those judgements, checks the replies to them, then picks the option that still looks best. In chess, we call this part think-ahead.

Mark-1 took about a minute per move, Astra about the same, and the Gemini models took 13-14 seconds. No clock for either side, and every move’s time is published.

Clock: 0:00 of 0:57
Same Position

OpenAI's GPT-6 Astra

One long private chain

Waiting

It thinks once, out of sight. You only see the final answer.

FTA Mark-1

Hundreds of small judgements

Each dot is one small judgement Mark-1 makes before it moves.

MJT’s debut: FTA Mark-1

FTA

We took an open model that couldn't play chess. After MJT, it beat three of the top four models on Kaggle’s chess leaderboard (as of 12 Nov 2026).

Sarvam-30B is an open model made in India, but it was never taught chess. Out of the box, it played a legal move only 16% of the time. We trained it with MJT and called it FTA Mark-1. Same base model. We added a small trained layer and its own think-ahead.

Before: Sarvam-30B as released. Answers with a valid move 76%. Chess rating could not be rated. Result against GPT-6 Astra: not in the game.
Micro-Judgement Tuning™⟶
After: FTA Mark-1. Answers with a valid move 100%. Chess rating 1810. Result against GPT-6 Astra: won 7.5 to 4.5.
Built on Sarvam-30B, trained with FTA's MJT

The matches

Three matches. Three wins.

FTA Mark-1 played three of the top four models on Kaggle's chess leaderboard (as of 12 Nov 2026). Each match consisted of 12 games, six with each colour, played on 7 November 2026. Both sides used the same instructions and rule-checker, with no clock.

vs OpenAI's GPT-6 Astra

10–2
9 wins2 draws1 loss

vs Google's Gemini 4 Argon

9–3
9 wins0 draws3 losses

vs Google's Gemini 3.8 Flash

9–3
8 wins2 draws2 losses

Every game is published. Logs and bills are available.

Mark-1 played with think-ahead, using only its own judgement. Each opponent played at its maker's default settings, the same as on Kaggle Game Arena. No chess computer helped either side. Everything was locked before the first move and timestamped before the results were announced.

Watch every game
right arrow

Results

About 1,824 on the chess rating ladder

We rated Mark-1 the way chess players are rated: 24 games against Stockfish set to four strengths, under the same conditions Astra faced on its own ladder.

Strength rating (Elo)

Dot = best estimate. Line = likely range from the games played.

GPT-6 Astra (OpenAI)
23 games
1,662
FTA Mark-1
24 games
1,824
1,2001,4001,6001,8002,0002,200

Head to head

Same 300 test positions for both, checked by Stockfish.

Mark-1
Astra
Ahead
Strength rating (Elo)
about 1,824
about 1,662
Mark-1
Big mistakes
(fewer is better)
9
17
Mark-1
Win chance lost per move (lower is better)
2.5%
3.0%
Mark-1
Picked the very best move
132 (44%)
146 (49%)
Astra
Strength rating (Elo): GPT-6 Astra about 1,662 over 23 games; FTA Mark-1 about 1,824 over 24 games.Head to head on 300 test positions: Elo about 1,824 vs 1,662; big mistakes 9 vs 17; win chance lost per move 2.5% vs 3.0%; picked the very best move 132 (44%) vs 146 (49%).

Ratings are measured against Stockfish's strength levels, not official FIDE ratings.

What Mark-1 taught us about MJT

A strong teacher matters more than lots of data.

The teacher is the whole game

With a strong teacher scoring every option, a general open model learned sharp judgement. A weak teacher gives a weak specialist.

More data stopped helping early

Past a point, adding positions did not move the rating. Better lessons and more time to think did.

Careful beats clever

Astra found the single best move a little more often. Mark-1 made far fewer big mistakes, and that is what won games.

Next: MJT for marketing

Who is the coach for your business?

Chess has a perfect teacher. Marketing often relies on guesswork. We’re taking lessons from Mark-1 and putting MJT to work in marketing.

For each brand, we build its own coach: a teacher built from its successes and mistakes, plus lessons from public case studies.

It’s coming to FTA’s Search Engineering work next.

1. Your coach

Built from your past results and from public case studies. It scores the options in front of your team.

→

2. Micro-Judgement Tuning™

The same method that made Mark-1. Learn from every option, not one answer.

→

3. Specialist intelligence for your team

Sharper calls for your team, every week. Your people stay in charge.

Specialist Intelligence

Others are building agents to replace marketers.
FTA builds the suit that makes them exceptional.

Micro-Judgement Tuning is our road towards specialist intelligence. Mark-1 is where it starts.

FTA Mark-1 is built on Sarvam-30B, trained with FTA's MJT. Scoring every move builds on published research by Google DeepMind.