Working code is a start.
Engineering goes further.
A practical framework for understanding what engineers can build, how they work with AI, and the evidence behind the result.
This edition explains the assessment framework. Sample scores illustrate how to read a report; they are not published benchmark results or population rankings.
01 / Read beyond the total
One result. Several things worth understanding.
An overall score can start a review, but it cannot explain one. Read the code, requirements and AI collaboration separately, then follow each dimension back to its evidence.
Engineering evidence profile
Illustrative session · Backend task
Does the implementation actually work?
Run the submitted solution against expected behavior, including boundary conditions and failure cases.
Evidence before interpretation. Keep test outcomes, code review and interview observations visible.
AI is a separate signal. Strong prompting cannot substitute for a working solution.
02 / Choose relevant work
The task changes the signal.
Engineering is more than one kind of problem. Choose the work you need to understand, then decide what evidence would make a result convincing.
Build behavior that survives edge cases.
A rate limiter, a file-processing API or a service change. Look at contracts, failure handling and tests as well as the final implementation.
- Correct behavior at limits
- Consistent API responses
- Tests that exercise failures

03 / A repeatable review
Make the evaluation as clear as the brief.
Use a consistent protocol for each role. The steps below describe how to plan an assessment and interpret its output without losing the context of the work.
- 01
Define the work
Write a role-relevant brief with acceptance criteria, constraints and a clear definition of done. Choose the criteria before reviewing the result.
- 02
Keep conditions comparable
Record the task version, available tools and time window. If two sessions use different conditions, preserve that distinction in the review.
- 03
Collect direct evidence
Look at the running solution, test outcomes, code changes and the relevant interview or AI conversations. A claim is stronger when the underlying observation is inspectable.
- 04
Review dimensions separately
Assess correctness and requirement coverage alongside maintainability, technical understanding and AI collaboration. Do not infer one dimension from another.
- 05
Make a human decision
Use the report to guide a follow-up interview or a training plan. Record unresolved questions and let a reviewer make the hiring or development decision.
04 / Inside an assessment
A rate limiter is more than a function.
A small service can reveal how an engineer handles state, concurrency and verification. Use this illustrative brief to see what a review can include.
Build an API rate limiter
Implement a token-bucket rate limiter for an API. Support independent limits per client, refill tokens predictably and return a clear response when the limit is reached.
When the first AI suggestion fails a boundary test, can the engineer find the cause, correct it and explain why the fix holds?
05 / Put the evidence to work
A report should lead to a better next question.
Turn a score into an interview.
A candidate meets most requirements but misses a concurrency case. Ask them to reason through that failure and inspect the fix together.
Turn a gap into focused practice.
An engineer writes clear requests but accepts untested suggestions. Choose a task that makes verification necessary and revisit the skill after practice.
Compare progress in context.
Use a baseline and a later assessment on comparable work. Keep changes in task difficulty, tools and review criteria visible.
06 / Scope & limitations
Useful evidence has boundaries.
Interpret every assessment in the context of the task and the person. This report is a framework for a better review, not a substitute for one.
- Example data is not a benchmark. The scorecard on this page contains illustrative values. It does not establish model rankings, hiring thresholds or population averages.
- Task choice matters. A result on one repository, stack or design exercise does not establish capability across every engineering domain.
- Signals need context. Copy events and AI request volume are observations to review, not standalone proof of misconduct or ability.
- AI use is optional. When the assistant is not used, its collaboration dimension is excluded rather than treated as a zero.
- Improvement needs comparable work. A later score is meaningful only alongside the conditions, difficulty and criteria of both assessments.
Build your assessment
Start with the work
your team actually does.
See how real codebases, system design and AI fluency fit into your hiring or development process.
Explore DRIP with the team