The methodology. The rubric is published. Read it before you buy, not after.
Four dimensions, behavioral anchors written separately for three levels, and weights that change by interview type. It comes from the Interviewer’s Guide, it is versioned, and a published version is never edited in place, because a change to it changes outcomes.
The four dimensions
Not five, not twelve. Four, because these are the ones a technical interview can actually observe in 45 minutes and defend afterwards.
Problem Decomposition
Does the candidate break an ambiguous problem into tractable parts, in a sensible order, and say out loud what they are deferring?
Architectural Judgment
Are trade-offs named rather than stumbled into? Does the candidate revise a position when a constraint changes, instead of defending it?
Technical Depth
Does the depth survive follow-up questions in any direction, or is it one layer thick and rehearsed?
Communication
Signposting, checking understanding, disagreeing productively. Weighted low on purpose, because it is the dimension most contaminated by accent, nerves and culture.
Weights shown at senior level (L5/L6). They change with the level and the interview type; every combination is in the table below.
| Interview | Level | Problem Decomposition | Architectural Judgment | Technical Depth | Communication |
|---|---|---|---|---|---|
| Technical Screen | L3 / L4 | 20% | 20% | 40% | 20% |
| Technical Screen | L5 / L6 | 25% | 25% | 35% | 15% |
| Technical Screen | L7+ | 30% | 30% | 25% | 15% |
| Coding | L3 / L4 | 25% | 15% | 40% | 20% |
| Coding | L5 / L6 | 25% | 20% | 35% | 20% |
| Coding | L7+ | 25% | 25% | 30% | 20% |
| System Design | L3 / L4 | 30% | 25% | 25% | 20% |
| System Design | L5 / L6 | 30% | 30% | 25% | 15% |
| System Design | L7+ | 30% | 35% | 20% | 15% |
| Data Science | L3 / L4 | 30% | 20% | 30% | 20% |
| Data Science | L5 / L6 | 30% | 25% | 30% | 15% |
| Data Science | L7+ | 30% | 30% | 25% | 15% |
Why Communication is weighted lowIt is the dimension where bias enters most easily, and where a nervous strong engineer looks worst. It is scored because it matters on the job, and it never carries more than 20 percent of the total.
What the anchors actually say
Senior level, Architectural Judgment. This is the whole anchor set for one dimension at one level, reproduced rather than summarized, because a rubric you cannot read is not published.
| Score | Behavior observed |
|---|---|
| 1 | Chooses components by familiarity. Cannot say why this database rather than another. No trade-off is named even when asked directly. |
| 3 | Names the trade-off being accepted for the significant choices, and ties it back to a stated requirement. Smaller choices go unexamined. |
| 5 | Names what is being given up, quantifies it where possible, and revises the position when a constraint changes rather than defending the original answer. Volunteers the failure mode of their own design. |
2 and 4 exist and sit between these. Scores are integers at the anchor and halves between, which is as much precision as behavioral observation supports. These anchors are published here and written per level, but the scorer does not receive the anchors yet; scoring against them is planned.
The same answer is a 4 at mid-level and a 2 at staff
This is the part a generic one-to-five scale cannot do, and the reason panels disagree with each other more than they think.
| Level | What "meets the bar" means | Prompting budget |
|---|---|---|
| L3 / L4 Junior and mid | Owns a well-scoped component. Asks for the constraint rather than inventing it. Needs the problem framed. | Generous. The interviewer’s hint budget is largest at this level, and you can ask for a hint. |
| L5 / L6 Senior | Owns the main path unprompted, surfaces scale and consistency constraints without being asked, and holds depth under three follow-ups. | Limited. A planned cap: exceeding it will cap the affected dimension at 3. |
| L7+ Lead and staff | Sequences the whole problem, states what is deferred and why, and reasons about organisational and operational consequences, not only technical ones. | Tight. A planned cap: exceeding it will cap the affected dimension at 2. |
Hints are not recorded or scored today: the interviewer gives them within its budget, and nothing about them reaches the report. Scoring that uses them is planned.
There is no tier below JuniorThe rubric stops at L3/L4 deliberately. Below that, the behaviors the anchors describe are not yet formed, and a scale that scores their absence measures exposure and confidence rather than engineering ability. An intern screen would need a different instrument, not a lower setting on this one.
How a score becomes a recommendation
Arithmetic, in code. Not a judgment call by the model.
| Overall | Recommendation |
|---|---|
| 4.25 and above | Strong hire |
| 3.50 to 4.24 | Hire |
| 3.00 to 3.49 | Lean hire |
| Below 3.00 | No hire |
| Any | Insufficient signal, which outranks every band above, when the session did not reach enough of the interview to judge. An interview that never got to technical depth is not a low score, it is not an evaluation. |
Every score cites the transcript
A quote that does not match the transcript word for word is discarded. The score is still stored and nothing is re-run. Too few verified quotes lowers completeness, and low completeness returns insufficient signal instead of a recommendation.
One model call, for now
Scoring is one model call, and the confidence shown is the model’s own figure. Self-consistency sampling is planned, with confidence taken from the spread across samples. It is not built yet.
The rollup is computed, not reasoned
The weighted total is arithmetic in code. A model that can be argued into a different total is not a rubric.
Versioned, because a rubric change is an outcome change
Every report records the rubric that produced it, and a published rubric is never edited in place. Adjustable weights on Scale and Enterprise plans are planned, not built yet. Calibration against an expert panel is planned: there is no calibration set and no agreement measurement today, so no agreement figure is claimed here.
An override rate, meaning how often your hiring managers disagree with the score, is planned. Reports cannot be marked as reviewed yet and overrides are not stored, so there is no such number today. If it climbs, something is wrong with the instrument, and you should be able to see that without asking us.
Read the rubric, then run it.
The fastest way to judge whether this is defensible is to see it score an interview.