Diagnostic Scores

Your measured results over time. This directory is the evidence base for every rebalancing decision in the program — and, by week twelve, the only honest proof that the work is working.

Table of Contents

• How to Record a Run
• Score Template
• Trend Table
• Level Trend
• Re-Test Schedule

How to Record a Run

1. Copy the template into baseline.md (first run) or retest-NN.md (subsequent runs).
2. Fill in every field. Blank fields are not neutral — they mean the plan gets built without that signal.
3. Add a row to the trend table below.
4. Tell me the numbers. PLAN.md gets written or rebalanced from them.

Score Template

# Diagnostic — <baseline | retest-NN> — <YYYY-MM-DD>

## Part 1 — Coding (45 min)
- Gates passed: _ / 4
- Time-to-first-passing-gate: __ min
- Gate times: G1 __ | G2 __ | G3 __ | G4 __
- Did I rewrite from scratch at any gate? yes/no — which one:
- Narration self-score: _ / 5
- Where I froze:
- What I would do differently:

## Part 2 — System Design (45 min)
- 2A requirements & scale:      _ / 5
- 2B architecture & API:        _ / 5
- 2C deep dive (hard parts):    _ / 5
- 2D failure & recovery:        _ / 5
- 2E rejected tradeoffs:        _ / 5
- Total:                        _ / 25
- Sections I ran out of time for:
- Did I ask the at-least-once vs at-most-once question? yes/no
- Did I do any arithmetic? yes/no

## Part 3 — Python Internals (30 min)
- Correct:            _ / 20
- Confident-wrong:    _
- Iterators/generators (Q1-Q8):   _ / 8
- Async/concurrency  (Q9-Q15):    _ / 7
- Memory/object model (Q16-Q18):  _ / 3
- Data model         (Q19-Q20):   _ / 2
- Confident-wrong question numbers (these go into review/ at 1 day):

## Part 4 — Behavioral (40 min)
- P1 architecture decision:  _ / 5   (cross-team? yes/no)
- P2 disagreement lost:      _ / 5
- P3 where is AI headed:     _ / 5
- Total:                     _ / 15
- Stories I drew on:
- Story categories I could not fill from real experience:

## Levels (from RUBRIC.md)
| Track | Level |
|---|---|
| A — Coding            | |
| B — Python internals  | |
| C — Systems design    | |
| D — ML/inference      | not measured |
| E — Take-home         | not measured |
| F — Behavioral        | |
| G — Agentic           | not measured |

## Subjective
- Hardest part:
- Where I froze:
- What surprised me:
- Energy at the end (1-5):

Trend Table

Add one row per run. This is the chart that matters.

Level Trend

Tracks D, E and G have no baseline by design — a 3-hour battery cannot measure a 48-hour take-home or a timed agent-driving run. Their first measurements land in weeks 2, 6, and 8 respectively.

Re-Test Schedule

Weeks 4, 9, 13, 17, 21, 25. Variants live in ../retest-NN/ as they are generated — same shape, different problems, so you measure skill rather than recall.

At each re-test, five things happen:

1. Take the battery cold.
2. Score it and add the rows above.
3. Rebalance the plan per ../RUBRIC.md — any track at L3 drops to maintenance and its hours move to the lowest track.
4. Re-run the coverage audit.
5. Re-run the Phase 0 search. If something newer and better-sourced than the source report has appeared, supersede it and update ../../research/findings.md.