p2 · [0/7] · 26–40h

Applied: evals, error analysis, and measurable improvement

Stop guessing whether changes help. Build the measurement apparatus first, then improve something and prove the improvement.

A committed eval report showing baseline vs improved scores, broken down by failure category, with the specific fix documented for each.

Tasks
t1-p2-01
authored

Hand-written eval dataset of 50 questions

5–9h
t1-p2-02
draft

Automated eval runner

4–6h
t1-p2-03
draft

Two-method grading layer

5–8h
t1-p2-04
authored

Failure taxonomy from the first run

4–6h
t1-p2-05
authored

One targeted fix for the largest failure bucket

4–7h
t1-p2-06
authored

Re-run and paired before/after comparison

3–5h
t1-p2-07
draft

Refusal handling

3–5h