AI labs
Teams that train or evaluate coding models and need tasks they can run and check.
AuraOne / Human Data / Agent training data
Coding data, checked by running it.
We are preparing coding tasks from commercial software we own. Each task starts from real code, states the problem plainly, and is graded by tests your model cannot see. A wrong fix fails. A different correct fix passes.
Each program is scoped by proposal. The signed scope names what is delivered.
- Input
- A description of the coding data you need
- Output
- Tasks with their checks, the evidence from running them, the review record, and the rights record
Who it is for
Teams that build or test coding agents.
You have an agent. You need tasks that tell a right fix from a wrong one.
Agent product teams
Teams that ship a coding agent and need tasks that show where it fails.
Task types
The kinds of work a task covers.
The task format names these types. Each program's scope sets which ones it covers.
Named in the task format
- Async retry regressions
- Access-control fixes
- Slow queries that run once per row
- API contract breaks
- Concurrency edge cases
- Partial migrations
- Tool-call failures
- Cache invalidation
- Other types, scoped per program
What arrives
What a program delivers.
Your signed scope sets the exact contents. Each task is one record with these fields.
The package
The parts of a program, as scoped.
- Source licence — the code and its history, with every component and permitted use listed.
- Tasks and private evaluations — a starting repository, a fair problem, hidden checks, and the evidence.
- Attempts and corrections — new runs against those tasks, recorded as they happened.
The task
What the model starts from.
- The problem, stated plainly
task_id,task_type,task_prompt - The starting code
repo_uri,base_commit - The container image and its reset command
image_digest,reset_command - The files the model can see
actor_files
The checks
The model sees some checks. The graded checks stay hidden from it. The record holds hashes of the fixes, not the patches.
- Checks the model can see
actor_checks - Where the hidden checks are kept
evaluator_checks_location - Hashes of the reference fix and a plausible wrong fix
gold_patch_sha256,plausible_wrong_patch_sha256
The evidence
What running the task produced.
- How the task was verified, with the receipt
verification - A recorded model attempt, and its hash
rollout,trajectory_sha256 - A hash of the whole record, and its creation time
record_sha256,created_at
The review
Who checked the task, and what they found. A task is marked not verified, verified, failed or quarantined.
- Where the task stands
status,review_status - Expert and second reviewer, and review time
expert_reviewer_id,second_reviewer_id,reviewed_at - Open secret or personal-data findings
secret_findings_open,pii_findings_open
How a task is checked
Coding tasks graded by tests the model cannot see; a wrong fix fails.
01
Start from real code
A verified task fixes a starting commit and a container image, and names the command that resets it.
Base commit, image digest and reset command
02
Fail before, pass after
The checks must fail on the starting code and pass with the reference fix. Repeat runs test for flaky checks, and a tamper attempt must be refused.
Cold-start, fail-before, reference-pass, repeat-run and tamper-probe results
03
Review by people
Each task record names an expert reviewer and a second reviewer. A task with an open secret or personal-data finding is not verified.
Reviewer IDs, review time and finding status
04
Record the run
A recorded run produces a receipt. A verified task's record carries the receipt hash and a hash of the whole record.
Receipt hash and record hash
Rights
Every task carries its rights.
Rights vary by source. The record names the source, the licence and the permitted use.
Source and licence
The licence of the code, and the scan that found it.
- Licence
repo_licence_spdx - Licence scan
licence_scan_sha256
Rights record
Each task points to a row in the rights matrix. The row grants each use on its own, with regions, retention and expiry.
- The key to that row
rights_id
Share or holdout
Each record is marked share or holdout, and carries a disclosure tier that limits how it may be shown.
- Split and disclosure tier
split,disclosure_tier
Traces are source material
Licensed traces are source material; tasks are derived and reviewed separately. See agent trace partnerships.
Scope and limits
What this program does not do
Preparing
Programs are scoped by proposal. Availability is confirmed in the signed scope. Nothing on this page is a released dataset.
Private datasets
A dataset built from your own code or records needs its own signed scope.
Hosting
Hosting the hidden checks, or running them for you, needs its own signed scope.
Model training
AuraOne prepares the tasks and the record. Training a model on them is not part of this program and needs a separate signed scope.
Tell us what your agent should be tested on.
We scope the tasks, the checks, the review, and the rights with you. Related pages cover agent trace partnerships, SFT data and RLHF and preference data.