Swarm Lab · Experiment 05 · 「早就知道了」陷阱
The Already Known Trap
Does a familiar label make its neighbors disappear?
A tiny synthetic genome-like map contains 24 planted surprises. Some sit beside a label saying “already characterized.” Give two searchers the same inspection budget. Let only one see the labels.
Your challenge
Catch surprises hiding beside familiar labels.
- Play with the labels hidden on the left and visible on the right. Both inspect 60 neighborhoods.
- Set Known-label discount to 0. The search orders should now match.
- Try a Full scan of 120. A priority bias cannot hide a neighborhood forever if everything is inspected.
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.
- Known-label discount
- How strongly a visible known label suppresses inspection priority. 0 ignores labels; 0.8 reduces priority by 80%; 1 almost skips it. This is the hypothesis you impose, not a measured human bias.
- Inspection budget
- How many different neighborhoods each searcher can inspect. More search costs more work.
- Known-label share
- Probability that a synthetic neighborhood gets a known label. The same labels exist in both arms but only one searcher sees them.
- Evidence noise
- Maximum upward or downward perturbation of repeat and partner scores. Higher noise can both hide real anomalies and create false alarms.
Make a prediction. Then press Play.
Model 1.1.0 · simulation onlyEach tile is one synthetic neighborhood, not a DNA sequence. Pale tiles are uninspected; gray tiles were checked; green means a true find and coral a false alarm. Outlined tiles carry visible known labels. The dashed chart line is the first arm. Controls change both arms except the named treatment; the seed is shared.
Read the result: Recall means true finds ÷ 24 planted anomalies. False alarms are counted separately. The tiles are synthetic evidence, not real DNA.
Loading local simulation…
Now break the story
Set the label discount to zero. If the gap disappears, what did the experiment actually test?
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.
Hidden labels recover more planted anomalies on average here. The discount is a rule we deliberately imposed, not a discovery about human psychology. Set it to zero, or inspect everything, and the coverage gap disappears.
| Arm | Mean | Seed range |
|---|---|---|
| Labels hidden | 62.8% | 45.8% – 79.2% |
| Labels visible | 56.8% | 41.7% – 79.2% |
Ranges describe these runs, not confidence intervals for real-world systems.
What the main score leaves out · default batch
- Labels hidden: 0.72 mean false alarms; 60 inspections per run.
- Labels visible: 0.69 mean false alarms; 60 inspections per run.
Under the hood · the actual rules
120 neighborhoods contain 24 planted anomalies and 24 repeat-only decoys. Noisy repeat and partner scores are observable; truth is used only for evaluation. Inspection order is a seeded weighted race based on repeat evidence. A visible known label multiplies priority by max(0.001, 1 − discount). Both arms inspect the same number of neighborhoods. A candidate is flagged when both observed scores exceed 0.65.
Read this model’s source · Read invariant tests · Reproduce the receipts
Where the analogy stops
The label penalty is an explicit hypothesis, not a measured human or Claude bias. Anthropic reports that the RT was previously identified and Claude noticed its associated repeat array and partner; it does not establish why earlier researchers missed those features. No biological sequences or real discovery agents run here.
Sources, methods & reproduction
Inspired by Robin’s diary. Background: Anthropic · reported ART discovery and research process. This is an original teaching model, not a replication of that research.
Experiment notes · Run locally with Node · Version & source hashes