Find the hardest tasks. Train, then find harder tasks.

Campaign tells you how well your model did. The frontier tasks it finds tell you what training your model needs next. We move beyond classic reinforcement learning with verifiable rewards. The old way is slow, because you have to attempt arbitrary tasks before their solve rate is known.

Today, Campaign runs the search as self-play. A setter model writes tasks in your domain, each with its own environment and verifier, and a solver model attempts each one. Each frontier task it finds is one more place training can move your model past its current limit. And we can do this over and over again. It’s quite powerful.

Campaign trials
Pack authoring and taxonomy trials prepare the campaign. Each trial has its own entry in the table of contents, with the commands and output below.
Click READ ALL to have me read you all the messages we wrote while running the campaign trials.
Pack authoring
completedstatus
4steps
1 · reasoningThe task pack needs checks for both correct and incorrect answers so a passing result is meaningful.
2 · agent messageI have read your campaign brief. I will build a small task pack with clear instructions, useful starting materials, and checks that distinguish a correct answer from an incomplete one.
3 · command
1python -m pytest
All task pack checks passed.
4 · agent messageI am checking that every task asks a clear question and provides the files needed to answer it. The solver should be able to understand the goal without guessing at hidden requirements. I will keep the starting materials separate from the checks that decide whether an answer passes. The first check is a known correct answer. It needs to pass for the right reason, with the expected output and all required files present. Next I will try an empty answer and a deliberately incorrect answer. Both should fail, which helps establish that the task checks are doing useful work. Once those checks are …
Built from your work
Start from one written description of the tasks you care about, grounded in your own code, data, or evals. A setter model writes each task with its own environment and verifier.
Scope
3 task groups
fixture-completed[VIEW IN CAMPAIGN]
Measured against the frontier
A solver model attempts every task several times. Campaign marks the tasks it sometimes solves and sometimes fails: that contrast is what reinforcement learning learns from.
Scores you can audit
Every verifier is certified against known answers before it scores anything, and every rollout keeps its files, logs, and full trajectory.
Method
Task pack source
Sandbox backend
Runtime image
Setter network access
Solver network access
Tasks
Iterations
Rollouts per task
Concurrent trials
Compute timeout (s)
s
Add private environment variables and choose which roles can use each one.
No secrets were supplied.
Model
Reasoning effort
Harness
CPUs
CPU
Memory (MB)
MB
Storage (MB)
MB
Agent timeout (s)
s
Action timeout (s)
s
Step limit
step
Max output tokens
tok
Temperature
Model default
Verifier timeout (s)
s
Context token budget
tok
Model
Reasoning effort
Harness
CPUs
CPU
Memory (MB)
MB
Storage (MB)
MB
Agent timeout (s)
s
Action timeout (s)
s
Step limit
step
Max output tokens
tok
Temperature
Model default
Verifier timeout (s)
s
Context token budget
tok
Spend you control
Set the number of tasks, attempts, parallel runs, and a time limit before launch. Nothing is spent until you press Run, and nothing needs you after.

How recursive self-learning works

What the frontier is worth

An asset that compounds
Recursive self-improvement compounds only as fast as new verifiable tasks arrive at a model’s edge, and a task is only known to sit there once it has been measured. Campaign writes and measures those tasks from your own code, data, and evals, so every round of the loop adds to a dataset that belongs to you and no one else has.
Training data with signal
Run reinforcement learning on tasks your model solves some of the time, not the ones it always passes or never can. Download every task with its full instruction and measured solve rate as JSONL, ready for your training pipeline. When your model outgrows them, raise the bar and run the next campaign from its recorded settings.
Built for long runs
A long-horizon campaign runs self-play across many iterations, unattended for up to a day, inside the time and attempt budget you set. With the Mulberry method, each new batch of tasks sees the history of the work before it and leans harder on purpose. Each run leaves a certified task pack that starts the next, so the loop can keep turning for weeks without rebuilding anything.

Watch a run find the frontier.

Every task is attempted several times. As each result lands, Campaign lights the tasks your agent solved only some of the time in flame, so you watch the frontier fill in while the run is still going.

See every step while it runs.

Campaign certifies its task setting before any task exists, then records every trial as it happens. The graph grows with the run, from the instruction to each task, trial, and reward, so it is never a black box.

Know exactly where your agent breaks.

Every step of a run is recorded, including the ones that fail. Here an independent check catches a task that could not tell a working fix from no fix, the repair fails too, and the task is discarded before a single attempt is spent on it.

Workertask completed · outcome rejected · no rollout runs

Trace any result to the evidence.

Open any attempt and follow it step by step, from the task it was given to the files it changed and how it was graded. When a number is questioned, the answer is one click away.

Get a clear answer when it finishes.

One chart shows the whole run: tasks started and concluded, your agent’s solve rate, and the share of measured tasks that landed in the frontier band, over the campaign’s elapsed time. You see what the budget bought.

No recorded work at this time

Work

1–24 / 24

Recorded workScheduledStartedConcludedDurationResult
Iteration 1, measuredfrontier-task-00016m 0s9m 0s26m 0s17m 0s3 of 8 passedFrontier
Iteration 1, measuredfrontier-task-00026m 45s11m 0s35m 0s24m 0s2 of 8 passedFrontier
Iteration 1, measuredfrontier-task-00037m 30s13m 0s35m 0s22m 0s4 of 8 passedFrontier
Iteration 1, measuredfrontier-task-00048m 15s15m 0s35m 0s20m 0s1 of 8 passedFrontier
Iteration 2, measuredfrontier-task-000533m 0s36m 0s54m 0s18m 0s3 of 8 passedFrontier
Iteration 2, measuredfrontier-task-000633m 45s38m 0s1h 3m25m 0s2 of 8 passedFrontier
Iteration 2, measuredfrontier-task-000734m 30s40m 0s1h 3m23m 0s5 of 8 passed
Iteration 2, measuredfrontier-task-000835m 15s42m 0s1h 3m21m 0s3 of 8 passedFrontier
Iteration 3, measuredfrontier-task-00091h 0m1h 3m1h 22m19m 0s2 of 8 passedFrontier
Iteration 3, measuredfrontier-task-00101h 0m1h 5m1h 22m17m 0s4 of 8 passedFrontier
Iteration 3, measuredfrontier-task-00111h 1m1h 7m1h 31m24m 0s1 of 8 passedFrontier
Iteration 3, measuredfrontier-task-00121h 2m1h 9m1h 31m22m 0s3 of 8 passedFrontier
Iteration 4, measuredfrontier-task-00131h 27m1h 30m1h 50m20m 0s0 of 8 passed
Iteration 4, measuredfrontier-task-00141h 27m1h 32m1h 50m18m 0s2 of 8 passedFrontier
Iteration 4, measuredfrontier-task-00151h 28m1h 34m1h 59m25m 0s3 of 8 passedFrontier
Iteration 4, measuredfrontier-task-00161h 29m1h 36m1h 59m23m 0s4 of 8 passedFrontier
Iteration 5, measuredfrontier-task-00171h 54m1h 57m2h 18m21m 0s2 of 8 passedFrontier
Iteration 5, measuredfrontier-task-00181h 54m1h 59m2h 18m19m 0s6 of 8 passed
Iteration 5, measuredfrontier-task-00191h 55m2h 1m2h 18m17m 0s3 of 8 passedFrontier
Iteration 5, measuredfrontier-task-00201h 56m2h 3m2h 27m24m 0s1 of 8 passedFrontier
Iteration 6, measuredfrontier-task-00212h 21m2h 24m2h 46m22m 0s2 of 8 passedFrontier
Iteration 6, measuredfrontier-task-00222h 21m2h 26m2h 46m20m 0s4 of 8 passedFrontier
Iteration 6, measuredfrontier-task-00232h 22m2h 28m2h 46m18m 0s3 of 8 passedFrontier
Iteration 6, measuredfrontier-task-00242h 23m2h 30m2h 55m25m 0s2 of 8 passedFrontier

Own everything each round produces.

Export every task, score, and trial record as files you can train on, start the next round from the certified task pack, or share a finished campaign with a link you can revoke.

Campaign data

3files

Campaign recordThe immutable record: request, status, provenance, and task counts.

Task summaries (JSON)3 task metadata rows and trial ids. Full prompts are available under Generated tasks and downloads.

Events5 worker events with their full payloads.

Play a whole campaign through.

The Vmax Campaigns Overview replays every stage of a campaign, from the first instruction to the frontier. Press play, pick a scenario, or step through it with the arrow keys once you click inside.

Scenario

Step

T+00:00:00.000no campaign

Naive self-play, completed

Refine the INSTRUCTION.md, launch, author and certify the task setting, then two iterations: a measured task in the frontier band, a task rejected after a setter repair, and a measured task no rollout solved. Finally, share the finished campaign.

Press Play to walk one campaign instance step by step, from the INSTRUCTION.md in the composer to its stored evidence and share link. Each step sets the nodes it touches alight: the browser controller, the client transport, the owner router and its handler, the Postgres tables behind it, the external worker running each trial, and the components that render the answer. Calls that start other calls are nested in the Call tree.

Illustrative: cmp_b1e5c0de0000000000000001, trial IDs, timings, and content are simulated. Every request and response shape is replayed through the real campaign-api.ts client and decoders by campaign-blueprint.test.ts. Space plays or pauses; the arrow keys step when the canvas is not focused.

Spinethe request path and the worker lifecycle
Fanone node spinning off the next calls
Treecomponents rendered from the view model
Convergetrials reporting into one outcome
Dataa read or write against a table

Start the loop with one page of instructions.

Find the hard tasks, train on them, then find harder ones. Campaign is open to approved Vmax accounts: join the waitlist for early access, or sign in if you already have one.

Read the overview →

Find the hardest tasks. Train, then find harder tasks.

Campaign trials?

Built from your work

Measured against the frontier

Scores you can audit

Spend you control