Let the model argue. Never let it add.
The fastest way to ruin AI-assisted analysis is to let the model do the arithmetic; the fastest way to waste it is not letting it judge at all. The working split: LLMs judge, code counts. Here, three judge personas read four (fictional) listing candidates and score a rubric in words, with reasons. Everything numeric — robust medians, lens weights, totals, ranks — is deterministic code you can audit. Switch lenses and watch the ranking flip; open the disagreement queue and see exactly where a human should spend their minutes.
Honest-AI note. The 48 scores and one-line reasons were drafted by Claude at build time — which is also how the production version works: judgments are generated once and cached, so the expensive, non-deterministic step never silently re-runs. Every number on this page is computed live in your browser. Median, not mean — one carried-away judge can drag a mean a full point; a median needs two of three to agree.
01Words from the model, numbers from code
A model asked for "a score out of 100" gives you confident noise. A model asked "would this survive a bad cycle, and why?" gives you judgment. Keep the arithmetic — medians, weights, ranks — in code that runs the same way twice.
02Disagreement is signal, not error
When the growth analyst says 5 and the skeptic says 2, don't average to a meaningless 3.5 — flag it. The spread queue is the highest-value reading list in the process: it's where the thesis actually lives.
03Cache the judgment, rerun the count
Judgments are expensive and non-deterministic — generate once, cache, version. Counting is free and exact — rerun it on every lens, every weight change, every audit. That split is what makes an AI-scored process defensible to an IC or a regulator.
Plain-language key (rubric, median, lens, spread)
- Rubric
- The fixed set of questions every candidate is scored on — same questions, every time.
- Median
- The middle value of the three judges — robust to one outlier opinion, unlike an average.
- Lens
- A named set of dimension weights (balanced / quality / growth). Same judgments, different priorities.
- Spread
- Highest judge score minus lowest. Spread ≥ 2 routes the cell to a human.