Reference

Appendix C: The Proficiency Record

One page per quarter. Fill the thresholds section before the quarter starts; that is the whole trick. A threshold negotiated after a bad number arrives will lose to the same cognitive math that produced the bad number. Companion to Chapter 8.


Quarter: ______ Practitioner: ______ Partner: ______

1. Thresholds (set before the quarter, never edited during it)

Signal Floor / trigger Pre-committed response if breached
Planted-defect catch rate ___ of ___ per quarter
Override precision (per territory) below ___ % for 2 consecutive months
Deference regret above ___ %
Flip-card weeks completed fewer than ___ of 13
Manual-leg weeks completed fewer than ___ of 13

Date thresholds were set: ______ Witnessed by partner: ______

2. The flip card (weekly)

One line per week: date, the judgment, your commit (3 sentences, written before the agent’s answer), agreement or divergence, resolution when ground truth arrived.

Wk Judgment Committed first? Agree/Diverge Resolved: who was right
1
2
...

3. The manual leg (weekly)

One artifact per week, end to end, without AI. Log the artifact and one thing the manual pass surfaced that the model’s version would have missed, or the honest “nothing,” which is also data.

Wk Artifact Surfaced
1
...

4. The planted defect (monthly, with partner)

Month Artifact doctored Factual plant caught? Judgment plant caught? What the miss teaches
1
2
3

Quarter catch rate: ___ / 6

5. The disagreement ledger (continuous; computed monthly)

Month Overrides logged Override precision Deference-regret entries Notes by territory
1
2
3

6. Quarter-end review (one hour, with partner)

  • Territories where override precision is falling: ______
  • Territories where deference should increase (the model has earned it): ______
  • Territories demoted (more flip cards, draft-depth reading) until numbers recover: ______
  • Thresholds for next quarter, set now: ______

How to read the result

There is no industry baseline. You are your own control; the trend is the instrument. Three quarters of stable catch rate and stable override precision while your AI leverage grows is what maintained judgment looks like on paper. A falling line in one territory is not a verdict, it is a map: that lane gets the demotion response you wrote in section 1, and the response runs until the line recovers. A record with empty weeks is also a result. It means the regime lost to the calendar, and the fix is in Chapter 8’s three conditions, not in trying harder.