For teams building, fine-tuning & shipping coding agents

Verifiable environments to train and trust coding agents.

Hard, diverse SWE tasks with hidden, reward-hack-resistant graders — verifiable enough to train on, independent enough to trust. Every task is proven on a trust loop before it counts.

5 trust-loop checks
noop · reference · adversaries
stable · export-redacted
Hidden graders
Behavioral contracts, not visible unit tests.
0 flake rate
Deterministic, reward-hack-resistant grading.
Contamination-resistant
A private and synthetically generated corpus.
How it works
Author a task
Problem, workspace, and a hidden grader
Hidden grader
Behavioral contracts, not visible unit tests
Trust loop
noop fails · reference passes · repeat-stable
Differential oracle
A different-vendor solver re-derives the fix
Promote
Survivors copied in, re-gated, archived
Export
SWE-bench-ready, grader material redacted
task.yamlveyl
Performance
Deterministic, every run
noop fails, reference passes, and adversarial probes get rejected — stable across repeats, flake_rate 0.
Determinism
Trust-loop checks
every task, every run
Expected
Result
noop
must fail
✓
reference
must pass
✓
adversaries
rejected
✓
Contact us to run the gate on your own tasks. 
The trust loop runs on every task — hand- or AI-authored.
A grader that isn't fooled — every run
Hidden behavioral graders, hardened against reward-hacking and screened by an independent-vendor oracle. Determinism is by design: repeat grading is stable and exports redact every grader.
Pack at a glance
Registered tasks42
AI-authored, promoted2
Flake rate0.0
Adversariesrejected
Hardenedpassed
The agent never sees the grader
The agent solves against public material; the grader runs hidden behavioral checks; exports redact every grader.
The public / private split
┌───────────────────────────────┐
│ Agent phase                   │
│ /workspace (editable)         │
│ public task material          │
└───────────────────────────────┘

↕ the phase boundary

┌───────────────────────────────┐
│ Grader phase                  │
│ private grader root           │
│ hidden tests                  │
│ reference patch               │
│ adversary patches             │
└───────────────────────────────┘

The agent solves against public material. The grader runs hidden behavioral checks. Exports redact it all.
✓Agent never sees the grader
✓Hidden behavioral contracts
✓Exports redact grader material
Generation is getting cheap; verification is the scarce thing — the raw material to train on, with the trust to evaluate.
✓AI authors over-many candidates
✓The trust gate filters every one
✓Humans move to judgment, not production
Contamination-resistant
A private and synthetically generated corpus — not scraped public benchmark rows. Private grading material never enters public git history.
Hidden tests, reference solutions, and adversarial probes stay private — so tasks can’t be gamed, scraped, or memorized.
Hidden
 ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
 ▓░                                         ░▓
 ▓░  ┌──────────── THE PACK ─────────────┐  ░▓
 ▓░  │                                   │  ░▓
 ▓░  │   █ THE CORPUS █                  │  ░▓
 ▓░  │                                   │  ░▓
 ▓░  │   Graders, probes, refs           │  ░▓
 ▓░  │                                   │  ░▓
 ▓░  │                                   │  ░▓
 ▓░  └───────────────────────────────────┘  ░▓
 ▓░                    │                    ░▓
 ▓░                    │                    ░▓
 ▓░                    ▼                    ░▓
 ▓░  ═════════════════════════════════════  ░▓
 ▓░  ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  ░▓
 ▓░                                         ░▓
 ▓░  ┌────────── TRUST GATE ─────────────┐  ░▓
 ▓░  │                                   │  ░▓
 ▓░  │   █ VERIFIED SET  █               │  ░▓
 ▓░  │                                   │  ░▓
 ▓░  │   Hidden behavioral               │  ░▓
 ▓░  │   Ready to train and eval.        │  ░▓
 ▓░  │                                   │  ░▓
 ▓░  └───────────────────────────────────┘  ░▓
 ▓░                                         ░▓
 ▓░  ✓ Private corpus                       ░▓
 ▓░  ✓ Synthetic generation                 ░▓
 ▓░  ✓ Redacted exports                     ░▓
 ▓░                                         ░▓
 ▓░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░▓
 ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒


● Request a walkthrough
See the full factory run on your tasks.
A private corpus of hardened environments, an independent-vendor oracle, and AI-authored tasks that clear the same gate — run on your agent in a walkthrough.
Veyl
© 2026