Swarm Lab · Experiment 07 · 品味漂移
Taste Drift
Can a good student inherit a teacher’s blind spot?
A fixed judge likes familiar ideas. The searchers learn to please it. Keep a quarter of them outside the feedback loop—and see whether protecting the search also protects what gets published.
Your challenge
Keep unusual true ideas alive without rewarding nonsense.
- Play ten rounds. Watch both unusual proposals generated and unusual true ideas retained.
- Raise the protected explorer share. See whether protecting generation alone changes the judge’s choices.
- Add 50% blind-review slots. This changes selection rather than only who proposes ideas.
First time here? The buttons explained
Play starts or pauses. Step advances one frame; the time slider rewinds. Changing a control prepares a new run; press Play again. Reset replays the current settings. Seed chooses a random world, not difficulty; keep it while changing one rule, then try Next seed. Share saves your setup in a link. The downloads let you inspect the data.
What every parameter means
- Seed
- A repeatable random world, from 0 to 4294967295. The compared arms use the same seed.
- Preference for familiar ideas
- Bonus given to familiar-looking proposals by the fixed judge. 0 ranks evidence only; 1 adds the largest familiar bonus. Both arms face the same judge.
- Protected explorer share
- Fraction of the second arm that keeps its original 50:50 unusual/familiar proposal mix instead of learning taste feedback.
- Blind-review slots
- Fraction of the 24 kept proposals chosen by evidence alone before taste-based selection fills remaining places. Applies to both arms.
Make a prediction. Then press Play.
Model 1.1.0 · simulation only120 proposals arrive each round. Purple dots are unusual ideas; green dots are familiar ideas. Dark dots are synthetically true, pale dots false. A gold ring marks one of the 24 retained proposals. Truth is visible to you for scoring, not to the judge. The dashed chart line is the first arm. Controls change both arms except the named treatment; the seed is shared.
Read the result: The main chart is a count, not a percentage: unusual AND true ideas among the 24 retained that round. Purple is unusual, not automatically correct.
Loading local simulation…
Now break the story
Protect half the explorers, then add blind-review slots. Which bottleneck did each intervention change?
Move one control, keep the seed, and compare again. The batch below is a fixed baseline—not a live summary of your current settings.
Observed in this model · 32 matched seeds · all runs retained
A result, not just an animation.
Protected explorers keep generating strange ideas. The judge mostly keeps rejecting them. The final-round retained count barely rises: protecting proposals is not the same as protecting evidence-based selection. Try blind-review slots.
| Arm | Mean | Seed range |
|---|---|---|
| Everyone adapts | 0.00 ideas | 0.00 ideas – 0.00 ideas |
| Protected explorers | 0.28 ideas | 0.00 ideas – 3.00 ideas |
Ranges describe these runs, not confidence intervals for real-world systems.
What the main score leaves out · default batch
- Everyone adapts: 0.44 unusual true proposals generated in round 10; 0.00 false proposals retained.
- Protected explorers: 5.44 unusual true proposals generated in round 10; 0.47 false proposals retained.
Under the hood · the actual rules
Each round samples 120 proposals, with truth probability 0.35 independent of unusualness. True evidence scores are uniform 0.55–0.95; false scores 0.10–0.65. The judge adds 0.45 × familiar preference to familiar candidates. It keeps 24; optional blind-review slots rank by evidence alone. Adaptive agents update their unusual-proposal probability to 0.6 × old probability + 0.4 × retained unusual share. Protected explorers keep their original 50% unusual-proposal rate. Both arms face the same judge and draws for ten rounds.
Read this model’s source · Read invariant tests · Reproduce the receipts
Where the analogy stops
The judge is a transparent scoring rule, not an actual scientist. Unusual does not mean true; truth is planted independently. Protected generation does not guarantee fair selection. This model cannot demonstrate that real expert feedback necessarily suppresses discovery, or that ignoring experts is beneficial.
Sources, methods & reproduction
Inspired by Robin’s diary. Background: Anthropic · scientific taste as feedback into agent instructions. This is an original teaching model, not a replication of that research.
Experiment notes · Run locally with Node · Version & source hashes