Evidence from an individual practice session
nrml is a technical interview practice application. A session provides an existing codebase, a problem brief, an editor, command execution, and AI assistance. An evaluated session produces a report using the saved code, test results, patch actions, and recorded session events.
The score describes performance on that exercise under its rubric. It is not a validated hiring prediction, professional certification, or guarantee of performance in a different repository. Read the evidence and confidence notes alongside the number.
Five categories contribute to the score
Each problem carries its own category weights. The standard catalog rubric assigns 30 points to Code correctness, 25 to AI supervision, 20 to Engineering judgment, 15 to Testing discipline, and 10 to AI usage. Category percentages are scaled to those weights and rounded before being added.
- Code correctness combines fresh trusted visible checks (25%), hidden behavioral checks (65%), and the build result (10%) within that category.
- AI supervision uses recorded review actions and focused behavioral verification of encountered assistant mistakes. An automatically applied change is not treated as candidate approval.
- Engineering judgment uses observable process and code signals. These are proxies for reasoning, not direct access to a candidate’s thought process.
- Testing discipline looks for evidence such as a baseline run, verification after the final edit, and independently verified regression cases.
- AI usage looks for grounded requests and purposeful use of assistance across the recorded workflow.
What passing tests establishes
The final evaluation runs trusted checks against the submitted code. Editing or deleting workspace tests does not replace those trusted checks. Added regression fixtures can receive evidence credit when their expected behavior is independently verified and they demonstrate a failure in the original starter that the final code repairs.
A passing test establishes the observed behavior for the tested input. Hidden tests sample the requirements; they are not a proof for every possible input. Build failures, unavailable verification, and incomplete evidence can limit what a report can conclude.
Confidence and AI commentary have limits
Assessments include confidence and supporting checks where available. If no assistant mistake was encountered, the AI supervision category uses a neutral 50% baseline with low confidence; that is not evidence that a candidate handled a mistake. Older sessions can also lack the focused calibration available in newer evaluations.
Supplemental AI feedback is a separate interpretation of saved work and evidence. It does not change the deterministic category scores. Like any generated review, it may miss context or make an incorrect observation; inspect the linked code and recorded events before acting on it.
Use the report to choose the next experiment
Start with a concrete failed behavior or unsupported assumption. Reproduce it, add a focused case, and explain how a new implementation satisfies the contract. Compare attempts with attention to the problem version, language, session conditions, and rubric.
The public sample report is illustrative. Its scores and feedback are examples, not a real candidate result or an aggregate outcome claim. Product changes may alter future evaluations; this page records the methodology reviewed on the date shown above.