Learning bug
Synthetic model · v1Same cues. One agent learns. Reverse the reward, disconnect its memory output, then reconnect.
- 1. Before learning
- 2. Learn A
- 3. Switch to B
- 4. Disconnect memory
- 5. Reconnect memory
Learning is off. Both agents choose without a stored preference.
Control
Ready. No trial yet.
Learner
Ready. No trial yet.
Completed run · 27 Oct 2026
30 seeds · 7,800 paired training/intervention trials · 300,000 frozen assessment choices · Linux. All 30 seeds met the declared checks.
| After phase | Control | Learner |
|---|---|---|
| Before learning | 50.4% | 50.4% |
| Learn A | 50.4% | 94.9% |
| Switch to B | 49.6% | 94.2% |
| Disconnect memory | 49.6% | 49.6% |
| Reconnect memory | 49.6% | 94.2% |
Protocol, all seeds and checks · Full trial records (3 MB) · Exact model source · Standalone benchmark runner
Rule and limits
Only the chosen value changes: Q ← Q + 0.2 × (reward − Q). Choices use a softmax with inverse temperature 6 and 10% uniform exploration. The control never updates. Learning is on only in “Learn A” and “Switch to B”. Disconnecting the output forces P(A) = 0.5 while stored values remain unchanged.
The scene illustrates choices, not realistic movement. Rewards are deterministic and hand designed. This does not establish biological memory, learning in the earlier fly-brain model, or performance on other tasks.