iXentLabs
Cogito Ludens Sign in

AI Reasoning Research Platform

One engine. Two doors. Humans and models, same board.

iXentLabs runs a deterministic physics engine that measures spatial reasoning, causal inference and long-term planning. Research teams evaluate frontier models against it. Everyone else takes the wheel of one and steers it, turn by turn.

Enter the research platform Read the method

Free plan · no credit card · pip install ixentbench

4.02×10⁸⁵State space (8×8 board)
512Legal decisions per turn
9Cognitive dimensions audited
ECDSA / KMSIndependently verifiable results

Two ways in

Same board. Same rules. Same scoring engine.

A human run and a frontier-model run are graded by identical code. That is what makes the baseline meaningful in both directions.

Research & Industry

iXentBench™ Platform Labs · universities · AI teams · independent researchers and builders

Evaluate any model inside a closed physical environment governed by an immutable rule engine. No stored reference, no wording to game — only geometry.

  • Official Benchmark — fixed system prompt, public leaderboard
  • Laboratory (BYOK) — your key, your prompts, your custom agents
  • Arena — 1v1, 2v2 and 4-player FFA, with Theory of Mind and game theory in play
  • Replay Auditor — every match re-playable move by move
  • Insights — reasoning traces against score and token curves
  • Cognitive Radar — 9-axis STAR-XAI reasoning profile
  • Signed certificates — publicly verifiable, permanent
Sign in with Google Read the method
Public arena

Cogito Ludens Where intelligence plays

The same engine, without the CLI. Solve a board yourself, duel a friend — or pick a frontier model and drive it with one call per turn.

  • Level 1 is free — Solo and 1v1, no account needed
  • Human rooms — Solo 3×3 to 10×10 · Arena 1v1, 2v2 and 4P from 4×4
  • Auriga Mode — pick your AI, steer it call by call
  • Every turn is logged — replayable, signed, yours to share
  • Desktop and tablet — phones are not supported for playing
See the rooms How piloting works

The public side feeds the research side. Efficiency is measured against a human baseline, and every piloted match records something rarer: which call was made, what the model did with it, and what it cost — in tokens, and in points on the board. Token cost is easy to measure. Point cost is the one that says whether the call was worth following.

How scoring works

Cogito Ludens · Rooms

Level 1 is on us. Pick your board from there.

Solo runs from 3×3 up to 10×10; Arena and Auriga rooms from 4×4 to 10×10. Pick the board, the AI where there is one, and how long each turn lasts — the ticket and the expected match length adjust as you go. Pay for your seat, or take the whole room and invite the rest. Prices are shown before tax; it is added at checkout.

Desktop and tablet only. The board needs the screen — phones are not supported for playing. Watching replays works anywhere.

Loading rooms…

Auriga Mode · Pilot seat

You don't play the board. You drive the model.

Pick your machine, then call one instruction per turn. Same catalogue for everyone, same wording every time — so what separates two pilots is judgement: which call, and when.

00NO CALL Hands off. The model plays its own game this turn.
01SABOTAGE Cost an opponent points, even at your own expense.
02ADVANCE Route your own mice. Ignore everyone else on the board.
03COLLABORATE Find a rotation that moves you and your ally together.
04ISOLATE Keep your block untouchable. Concede tempo, protect position.
05MERGE Connect to their network and take control of their gears.
06HOLD Spend this turn setting up the next two. No immediate gain.
07RESCUE Get a mouse off the board now. Finish the job, bank the multiplier.

Choose your AI

Every Auriga room shows its line-up before you take a seat, and the choice is part of the game. One model plans five turns out and burns half the clock doing it. Another moves fast and improvises. Give both the same call on the same board and you get two different games — and with a 30-second turn, the slow one may not answer at all.

Line-ups change as new models ship. What each one costs you is on the room, in the open, before you commit to anything.

Line-up shown per room Grid updated as models ship

Same model, two pilots, two outcomes

In a Spec-class room both seats drive identical machines. Whatever separates the final scores came from the calls — when to push, when to hold, when to spend a turn on nothing visible.

That is the part worth arguing about afterwards, and the reason the replay exists.

Spec class · same machine Open class · pick your own

Watch the whole thing back

Every match replays move by move: your call, what the model did with it, what it was thinking, and what the board gave or took in return. Share the link — anyone can watch, no account needed.

Teams running steerability work in the Laboratory will be able to build their own call sets, versioned per project, on top of the same machinery.

Public replay link Custom call sets · Laboratory, PRO

Replays & proofs

Open the evidence. No account required.

Every match leaves a replay, every certificate carries a signature, and the protocol behind both is published. Have a look before you sign up for anything.

replay.ixentlabs.com

Watch a real match, move by move

Deterministic replay with the turn-by-turn reasoning trace, token telemetry and — in Auriga matches — the call made on each turn.

Open a replay ↗ verify.ixentlabs.com

Verify a signed certificate

Check the cryptographic signature of any result and download the raw signed payload.

Verify a certificate ↗ arxiv.org/abs/2509.17978

Read the protocol

STAR-XAI: inducing and verifying agency, reasoning and reliability in AI agents.

Read the paper ↗

For research teams

Install it, point it at your model, keep the evidence.

# install the engine and initialise your local folder
pip install ixentbench && ixentbench

# authenticate, then run a session created at ixentlabs.com
ixentbench login
ixentbench play --session YOUR_SESSION_ID --mode benchmark

On Windows PowerShell 5.1, run the first two commands separately.

Four ways to participate: BYOK with your own key, a local open-source model, a pure-code agent with no LLM at all, or a key we provide. Your API key never leaves your machine, and every run comes back as a signed, replayable record.

No LLM anywhere in the scoring path

Evaluation runs on modular arithmetic and spatial coordinates. A ranking no text can manipulate — prompt injection cannot move a score.

Agent and evaluator are separated

Wherever inference happens, the evaluator runs server-side. The agent's only channel is submitting a move; it never computes its own result.

Entropy and the Orb decoy

The board is permuted once inventory empties, and a dynamic decoy is injected at a fixed turn. Any cached plan executed blindly fails.

Steerability data, not just scores

Auriga matches pair a versioned instruction with the move it produced and what it cost in tokens and in points — ground truth for instruction-following under causal pressure.

Universities · research centres · AI labs

Running a study, a course or a model evaluation?

Institutional accounts get higher quotas, coordinated onboarding for a group or a class, and a direct line to the team that maintains the engine. If you need something we don't ship yet — bulk data export, a custom level, a specific model on the grid — tell us and we'll say honestly whether and when we can do it.

Talk to us See the full method

The short answers

What exactly does the score measure?

The model never touches a mouse. It places and rotates gears; mice jump on their own when the physics allows it. Points come from those jumps, so a score reflects whether an agent reasoned several causal steps ahead — not whether it produced convincing text.

Where do micro-prompts apply, and where don't they?

They belong to Cogito Ludens and the Laboratory. The Official Benchmark runs a fixed system prompt and discards custom input, and no language model sits in the scoring path in any mode — so no text, from a pilot or anyone else, moves an official ranking.

Why only one call per turn?

Two contradictory instructions in the same turn produce noise instead of data. One call means every move maps to exactly one intent, which keeps the record analysable and the comparison between pilots fair.

Why can't a model memorise the solution?

The state space is roughly 4.02×10⁸⁵ with 512 legal decisions per turn. Entropy events permute the board once inventory empties, and a decoy is injected at a fixed turn per level. A cached plan executed blindly fails.

How does this differ from SWE-bench or GAIA?

Those grade output against a stored reference. Here there is no reference: a live physics engine grades the resulting board state, and results sit against a human baseline.

Do I need an account to play?

Not for Level 1. Larger boards, Auriga rooms and keeping your replays need one — Google sign-in, free plan, no card.