RL environments and evals
Legacy Android codebases, built to test coding agents.
Private environments made from the apps real teams maintain: multi-module, mid-migration, users on every upgrade path. Graded by hidden tests. Delivered in Harbor format.
A frozen legacy app
Private and never published, so it's in no training corpus.
A real ticket
Written by a senior Android engineer, the way a tech lead writes one.
A layered grader
Build, anti-cheat gates, the app's own tests, then hidden tests.
Multi-step work
Migrations shipped in phases, each step graded on its own.
A full legacy database migration, shipped the way a team ships it.
Three tickets. The agent keeps its own code between steps. Each step is graded on its own and re-checks what earlier steps guaranteed.
Introduce the new data layer behind a feature flag
Both storage paths must preserve every user's data.
Move screens to a reactive, asynchronous data layer
Storage work leaves the main thread; screens observe the database.
Remove the legacy code
Every user moves to the new layer, on every upgrade path.
gates build ok anti_hack ok all_tests_ran ok layers existing_tests 1.00 hidden_tests 0.94↳ upgrade path from v3 loses data guardrails 1.00
Hard to game, safe to train on.
Every environment is built and validated by hand before it ships.
Written by an Android engineer
Every codebase, ticket and gold patch comes from a senior Android engineer who has shipped these migrations in production.
Layered, not unit-tests-only
The app's own suite, hidden behavioral tests the agent never sees, static progress checks, and guardrails for engineering rules.
Graders run apart from the agent
A separate container restores the original tests and fails any run that disables or skips them.
Checked before release
The gold patch scores 1.0 on every step, a no-op agent scores 0, and known cheats fail for the intended reason.
Never published
Codebases stay out of training corpora, so your results stay meaningful.
One reward, two uses
A machine-readable reward for evaluation or RL training. Code quality is scored separately against a reviewer rubric.
We run your agent on a set of environments and send you the failure report.
Gold Patch Labs was founded by a senior Android engineer (ex-Uber, Affirm, Compass) who has shipped the migrations these environments are built from.