Controlled public release — autonomous code repair

LEONA by DLG Labs.
Governed code repair.

A controlled public benchmark and transparency kit for autonomous software repair. Mutation boundaries, immutable tests, a rollback engine, and pytest before/after telemetry — all evidence-verifiable in the public repo. Successes and failures are preserved honestly.

Evidence-first · Fails closed · Honest claim boundaries

View the public dashboard → See the evidence

Public Evidence

Two tracks, reported separately.

LEONA's public evidence is split into two tracks because they measure different things. Deterministic replay validates the harness. True-LLM repair measures actual model-generated patch attempts. Both are in the repo. Both carry explicit claim boundaries.

Track 1 — Deterministic Replay

Controlled 1,000-case benchmark

1,000 / 1,000 PASS

Cases:: 1,000
Convergence:: 1,000 / 1,000
Unauthorized mutations:: 0
Test files modified:: 0
Rollback events:: 5
Avg repair duration:: 931 ms

“This validates the benchmark harness, mutation boundaries, rollback system, telemetry, and evidence pipeline. It does not prove unknown-bug autonomous reasoning.”

See: evidence/controlled-report.md

Track 2 — True-LLM Repair Validation

50-case adaptive validation snapshot

23 / 50 PASS · 27 / 50 MODEL_LIMITATION

Cases:: 50
Passed:: 23
Model limitation:: 27
Patch rejections:: 217
Hallucinated patches:: 45
Syntax-invalid patches:: 9
Rollback events:: 98
Unauthorized mutations:: 0
Test files modified:: 0
Model label:: local-14b-code-model

“This evaluates actual model-driven repair attempts using pytest telemetry and authorized source context. Successes and failures are preserved honestly.”

See: evidence/true-llm-report.md

Five OSS micro-repository deterministic mutation-replay probes (toolz, tqdm, python-sortedcontainers, cachetools, boltons) also passed 5/5. These are controlled mutation-repair probes against real project structures, not claims of upstream defects. See evidence/oss-report.md.

Governed Mechanisms

What the harness actually enforces.

LEONA's repair governance is not a slogan — it's a set of concrete enforcement points that produce the telemetry above. Each one is verifiable in the public benchmark output.

Mutation boundaries

Only authorized source files may change. Every attempt to modify a non-authorized file is rejected and recorded.

Immutable tests

Test files are never modified by the repair pipeline. Test-file write attempts are rejected at the chokepoint.

Rollback engine

Failed patches revert cleanly. Rollback events are counted and reported as first-class telemetry.

Patch parser + validator

Hallucinated patches, syntax-invalid output, and schema violations are caught before any file is touched. The pipeline fails closed.

Pytest before/after telemetry

Every repair attempt runs pytest before and after. Test pass/fail evidence is the ground truth, not the model's self-report.

Public model labeling

Public artifacts use generalized labels (local-14b-code-model). Internal model IDs and provider tuning remain private.

Public Repo

What's in the public release.

The dlglabs/leona-public repo is the controlled transparency layer. It includes everything below. Private orchestration internals, repair heuristics, provider routing, and enterprise control paths remain separate.

dashboard/

Static public benchmark dashboard with embedded repair telemetry. Open →

evidence/

Raw benchmark reports and JSON result artifacts for both tracks. Reports →

benchmark-framework/

Procedurally-varied benchmark generator and deterministic validation harness.

telemetry/

Public telemetry field contract and validation notes.

tui/

Lightweight local terminal viewer for benchmark summaries.

orchestration-shell/

Thin command wrappers that validate and summarize public evidence.

tools/

Public evidence validator and summary renderer (validate_public_evidence.js).

docs/

Technical brief · Public/private boundary · Evidence provenance

What stays private for now. Advanced orchestration internals, multi-board routing, repair heuristics and policy tuning, provider routing, enterprise control paths, internal model IDs, and unreleased model-vs-model comparison data are not in the public repo. The release principle is straightforward: publish enough to make the benchmark inspectable and credible; keep enough private that LEONA remains defensible as a product.

LEONA by DLG Labs. Governed code repair.