Skip to contentNew noteReasoning traces beat answer-only data

We only ship the tasks that beat your model.

Every task is run by frontier agents before it leaves here. If they solve it, we throw it away. Difficulty on this site is a measurement, not an adjective.

What survives. Every week thousands of tasks arrive, from credentialed experts across six continents. Each is pinned, built and made reproducible. Then frontier agents try to destroy them, and most of the batch does not survive. What is left is the only thing worth training on.

The bench

The bench is never asleep.

Six hubs, spread far enough apart that a normal working day in each one overlaps into continuous cover. Not a claim about effort — just what these longitudes add up to. Hover a row for what that hub is for.

  • Frontier training & computeThe deepest pools of compute, capital and research talent, and where most frontier models are trained.
  • Where deep learning grew upToronto and Montreal nurtured modern deep learning long before it was fashionable.
  • Foundational researchA century-deep bench in mathematics and computer science feeding world-class labs.
  • Sovereign computeSovereign investment behind some of the largest compute build-outs outside the US and China.
  • Where AI gets engineeredA vast software talent pool, and a linguistic proving ground for AI that has to work for everyone.
  • Southeast Asia hubResearch institutes, capital and multinational R&D packed into one highly connected city.

Union of all six windows 24 / 24 hours · no gap

Loading local times

A case file · Mohit Labs

Most training data is bought by the ton and never tested. Ours has to survive a trial first — and what you just watched is the part that made it through.

Read the file

Follow a single real task from the bug that produced it to the dataset it ended up in, and see what happened when three frontier agents tried to kill it.

Subject
apache/arrow #41207
Language
C++
Filed by
Systems eng · 9 yrs
Outcome
Accepted

Chapter one

It starts with something actually breaking.

Not a puzzle invented for a benchmark. A real defect, found by someone already deep in the code who noticed the arithmetic was wrong past the first chunk.

Session capture14 tests
$ ctest -R chunked_array_slice

  Start 1: ChunkedArray.SliceBasic ............   Passed
  Start 2: ChunkedArray.SliceEmpty ...........   Passed
  Start 3: ChunkedArray.SliceAcrossChunks ....   FAILED
  Start 4: ChunkedArray.SliceOffsetRebase ....   FAILED
  Start 5: ChunkedArray.SliceZeroLength ......   FAILED

  expected  items[20:25]
  actual    items[20:23]     <- offset never rebased

  11/14 passed · 3 failed

Chapter two

The fix is the easy part. The reasoning is the product.

Anyone can commit three lines. What we capture is the working — which branch was ruled out, why, and the check that proved it. Then it is pinned so it reproduces exactly, forever.

Base commit
a4f9c21e8b7d
Environment
Pinned Dockerfile — builds clean from scratch
Failing tests
3 must fail before the fix, and pass after it
Reference fix
812 lines across 7 files
Captured reasoning
Ruled out the empty-slice branch — it hides the bug rather than fixing it
Leakage scan
Instruction names no identifier from the solution

Chapter three

Then we try to destroy it.

Three frontier agents, eight independent attempts each, in the same sandbox a customer would run. Twenty-four chances to make this task worthless. Every filled square is an attempt the model won.

Alpha
0 / 8
Beta
1 / 8
Gamma
1 / 8

2 solves in 24 attempts

Alpha stalled rebasing chunk offsets · Beta regressed zero-length chunks · Gamma quit after two failed cycles

Filed the same week · did not survive

  • CutCSV quoting round-trip · pandassolved 8/8, nothing left to teach
  • CutOff-by-one in paginator · djangosolved 7/8
  • CutNull deref in config loaderinstruction leaked the function name
  • CutRace in scheduler · internalenvironment would not build twice the same way

Chapter four

It survived. That is the whole qualification.

Each agent gets eight attempts. This task went 0, 1 and 1 — a median of one solve per eight, which sits at the bottom of the window: hard enough to carry signal, not so hard it is broken. A task landing outside goes back to its author with the reason attached.

1 / 8Median · inside the window · kept

Solves per 8 attemptsWindow · 16

From here the same recorded session ships as a trajectory, an RL environment, an eval item or a preference pair — nothing is re-annotated afterwards, and the solve rate above travels with every one of them. See the full surface →

What we make

That one task ships in sixteen forms.

Post-training data, agentic environments and benchmark construction are usually three different vendors. They come out of one pipeline here, from the same recorded sessions — which is why the solve rate travels with them.

  • Post-trainingSupervised fine-tuning · Preference & RLHF · Code generation · Multimodal04
  • Reward & RLRubric & verifier grading · Tool-calling environments · Durable RL environments03
  • AgenticAgent trajectories · Computer & browser use · Long-horizon tasks03
  • EvaluationBenchmarks & evals · Deep research tasks · Failure & loss analysis03
  • DeliveryOff-the-shelf datasets · Custom evals & datasets · Professional domains03

FAQ

For AI teams, answered.

Still have a question?

What actually makes a task “hard” here?

Not an author’s opinion. Every task is run by frontier agents before it ships and the solve rate is recorded. Solve it too often and it is rejected — a task a capable model can already complete teaches your model nothing. Difficulty is a measured property of the task, and it travels with the data.

How does an engagement work?

We scope the capability or evaluation you need, run a small pilot to calibrate quality against your models, then scale. You can commission bespoke work or pull from review-cleared off-the-shelf datasets.

Who are the experts?

Credentialed practitioners with real on-the-job experience in their field, not generic annotators. Every contribution clears automated validation and layered human review before delivery, and contributors progress through tiers gated on accepted work.

What formats do you deliver?

SFT, preference and RLHF data, agent and computer-use trajectories, RL environments, code, multimodal data, and evaluation suites — in your schema, ready to train on.

How do you handle contamination?

Originality is a gating check, not a spot audit: every submission is compared against the existing corpus and rejected when it is too close. Tasks are built on pinned commits with held-out tests, and the fairness stage scans instructions for leaked identifiers before anything is accepted.

Who owns the data, and is it secure?

You own the deliverables. We work under NDA, isolate engagements, and scope handling to your security requirements.

The end of the file

Now tell us where your model falls short.

Bring the capability you are trying to move. We will build the tasks that defeat it, prove they do, and report the number.