Measured, bounded, inspectable

Claims connected
to artifacts.

The strongest result is a focused stress test—not a universal accuracy claim. Every prompt, outcome, operation ID, and limitation remains available for review.

0blind refactor trials

0regressions introduced

0extra catches beyond plain CI

Focused benchmark

Capsules caught what the visible suite missed.

Thirty Nemotron refactors covered two real upstream bug histories. Eight trials reintroduced the historical regression.

Plain visible CI3 / 8

Capsule replay8 / 8

Recall routing

12 / 12

Correct top-1 selection

Authored change descriptions across four verified histories, for lexical and changed-path-aware retrieval.

Live memory audit

3 + 2

Sandbox + model calls

The Werkzeug conflict audit used three sandbox calls and two model calls within a $0.75 allowance.

Automated validation

134

Repository tests

Unit and integration coverage includes memory, MCP, budgets, recovery, and execution controls.

Cost-aware AI

Use the right model only when needed.

Fast indexingNemotron Nanokeywords and triggers

EngineeringNemotron Superregressions, repairs, conflict analysis

Escalated reasoningNemotron Ultrahigh-risk adjudication only

Semantic recallQwen3 Embeddingone bounded batch

Honest boundary

What a passing replay means.

It supports: the selected frozen behavior passed in the recorded environment with valid execution evidence.

It does not support: universal repository correctness, absence of every vulnerability, or behavior outside the selected checks.

Inspect all evidence artifacts →