DIMENSIONS

Two signals. From one session.

Traditional coding tests measure one thing: can you write code. We measure two signals that matter in the AI era — because writing code is no longer the hard part. Directing AI efficiently is.

01 — WHAT WE MEASURE TODAY

The complete Phase 2 score.

Each signal is scored from evidence in the session — not guesses. Every claim is backed by a specific moment in the chat log or code history.

UNDERSTANDING — 50%84

Did the candidate describe the important system behaviors in any words? Plain English ("two bookings cannot overlap") counts equally with technical jargon ("409 on collision"). The gap between what they described and what the system needs is the measurement. Credit only for behaviors the candidate REQUESTED — the AI follows a minimal-version rule, so anything it added without a request is a gap, not a strength.

TOKEN EFFICIENCY — 50%58

Same fixed token budget for everyone. Who built more with less. A candidate who understands writes fewer, clearer, more specific prompts. We measure prompt specificity, iteration count, and context window efficiency. A weak candidate burns tokens on "fix this" loops. A strong one finishes at 70% of budget.

OVERALL FORMULA

round(0.5 × understanding + 0.5 × tokenEfficiency)

02 — HOW THE FOUR DIMENSIONS MAP TO PHASES

The full 4-dimension model.

All four phases are live. The full assessment scores four dimensions; the 2-signal Logic + AI Build subset remains available as fast mode.

01LOGICAL THINKING

MEASURED IN PHASE 01, 04

Independent problem-solving, structure, correctness, edge case handling. Measured in Phase 1 (unassisted code) and Phase 4 (architecture comprehension).

02DOMAIN KNOWLEDGE

MEASURED IN PHASE 04

Do they know the relevant APIs, frameworks, and patterns? Can they identify security concerns, performance bottlenecks, and common anti-patterns? Measured across all phases where technical decisions are made.

03PROMPT ENGINEERING

MEASURED IN PHASE 02

Same fixed token budget for everyone. Who built more with less. A candidate who understands writes fewer, clearer, more specific prompts.

04AI JUDGMENT

MEASURED IN PHASE 03

Did they catch the injected bugs in the Debug phase? Did they question AI hallucinations? Did they test what AI generated, read the code, and diagnose specifically when things broke? This is the hardest dimension to fake.

03 — HOW SCORING WORKS

Evidence, not intuition.

Every score is tied to a specific observable action. We do not guess. We read the session log.

EXTRACT EVIDENCE

AI reads the full session log: every prompt, every code change, every test result, every token spent. It identifies specific moments that demonstrate skill (or lack thereof).

SCORE EACH SIGNAL

Evidence is mapped to the two signals. A precise behavior description scores on Understanding. A clean prompt with no rework scores on Token Efficiency. Cross-phase evidence is recorded as 3-5 quoted moments labeled positive/negative/neutral.

WEIGHT 50/50

Understanding and Token Efficiency are equally weighted. A candidate who describes the system perfectly but burns the entire budget scores 50. One who builds a lot with vague prompts also scores 50. Excellence requires both.

BENCHMARK

Every score is compared against the candidate pool. "Top 15% token efficiency" means something. We show percentile rankings, not just raw numbers.

See the full picture.

One assessment. Two signals today. Four when Debug and Review ship.

FOR COMPANIES →PRACTICE FREE ■