Bench
Same net. Same seed. One variable.
Identical network. Identical data. Identical random seeds. The sole variable is the input encoding.
Live miniature · seed 7
Task A — is this reading odd?
Taught on 0–127. Asked on 512–1,023 — never seen. Same logistic readout. The only change is the input encoding.
Structured
Train 0–127
0.0%
Unseen 512–1,023
0.0%
One-hot
Train 0–127
0.0%
Unseen 512–1,023
0.0%
Embedding
Train 0–127
0.0%
Unseen 512–1,023
0.0%
A slot that was never trained holds whatever it was born with.
Place-value add — no new slots
Taught small. Asked large. The next column is already a rung. LSB on the left.
A
0
0
0
1
0
0
1
1
0
0
B
0
1
1
1
1
0
1
0
1
0
Σ
0
1
1
0
0
1
0
0
0
1
200 + 350 = 550 · 10 rungs
Task B — composed add, no retraining
Place value is what you use when the next column appears. The live check below applies the same add to 128–511 without fitting a new net. Monolithic one-hot has no slot for those magnitudes.
200 / 200
composed exact matches
- COMPOSED (place-value rungs)100.0%
- Structured, monolithic0.0%
- One-hot, monolithic0.0%
Full-protocol Task A (measured)
| Encoding | Train | Unseen magnitude |
|---|---|---|
| Structured | 100.0% | 100.0% |
| One-hot | 99.2% | 52.7% |
| Embedding | 100.0% | 50.0% |