Tool 27 · Session 1 · Text · Human preference
Whose Preference? Lab
Rank three fluent responses, assemble a panel of authored rater perspectives, then change how their judgments are aggregated. Watch “better” move when the panel or rule changes.
Big idea: next-token prediction can produce fluent candidates, but people and institutions decide which qualities get demonstrated, compared, rewarded, and eventually treated as preferable.
Authored teaching simulation. The responses, rater priorities, feature values, rankings, and aggregate signals were written for instruction. This is not an RLHF training run, reward-model output, or measurement of real people. Real alignment pipelines involve far more data, labor, policy, modeling, and optimization.
Choose an authored prompt
1 · Rank
2 · Panel
3 · Signal
Prompt
Your ranking · move the strongest response to 1
Rank all three responses before looking at the panel.
Aggregate teaching signal
Which response becomes preferred?
Selected rater rankings
| Rater perspective | Prioritizes | Authored ranking |
|---|
Investigation note
Make one bounded claim
Describe what changed in this authored panel. A preference signal records a choice under a rule; it does not prove universal quality or truth.
Human decision layer
What gets turned into a signal?
- Candidate set: someone decided which responses the panel could compare.
- Evaluation dimensions: accuracy, clarity, care, actionability, and brevity were selected by the author.
- Panel composition: including or excluding a perspective changes whose priorities enter the signal.
- Aggregation: averaging strength and counting first choices answer different questions.
- Training step: a real pipeline still has to learn from many such comparisons; labels do not directly rewrite one answer.