Three local models.
One execution system.
Laro combines Gemma 3 model weights with Veylaro's local coding tools, project context, and verification systems. Choose a size that fits your hardware. Historical measurements do not establish scores for the current downloadable bundles.
Laro Lite
4B parameters · approximately 2.49 GB on disk
Laro Med
recommended12B parameters · approximately 7.30 GB on disk
Laro Max
27B parameters · approximately 16.55 GB on disk
The model proposes.
Evidence promotes.
Eight systems, all model-agnostic, all measured. Together they are worth more than the weights: the same model that solves almost nothing on its own ships working software once the lattice is in the path.
Contract compiler
Turns the requested behavior and invariants into explicit checks before editing starts.
Execution gate
A patch survives only after real syntax, tests, lint, build, and task-specific checks pass.
Failure lattice
Stores failed strategies by error class so the next attempt does not repeat the same mistake.
Holdout sentinel
Fingerprints evaluation material and blocks grader leakage from prompts, memory, and selection.
Blast-radius governor
Penalizes broad rewrites and blocks unrelated tests, fixtures, lockfiles, and generated output.
Counterexample forge
Derives deterministic edge probes from the public contract before candidate selection.
Evidence scheduler
Matches search depth and parallelism to the model tier and available memory.
Claim calibration
Labels results measured, attributed, uncertain, or unverified instead of filling gaps with confidence.
Every release score
must follow the artifact.
Historical runs remain available for audit, but they are not assigned to a renamed, requantized, or newly packaged model. Each released tier needs its own executable report, immutable artifact hash, and clean-machine reproduction before it receives a public score.