01 // Abstract
Can one native direction in a frozen receiver control its answer for both grammatical genders? The test adds a native gender direction at block 16 of Qwen3-1.7B, at five doses, on an Arabic grammatical-gender task. No channel and no bridge are used: it is an oracle test of the receiver’s own direction. The preregistered receiver-native gate failed at all five doses, because the masculine shifts did not beat norm-matched random directions.
The feminine shifts beat the random directions at every dose. A second script re-derived the hashes, doses and intervals, and all 252 jobs (12 examples, 21 conditions) reran with exactly equal outputs. The rerun used the same code and CPU, so it is not an independent replication.
02 // Gates
| Gate | Preregistered condition | Result | Record |
|---|---|---|---|
| Receiver-native gate | At two adjacent doses in {1.5, 1.75, 2, 2.25, 2.5}, for both genders: CORRECT beats ZERO, WRONG and each random direction (99% lower bound > 0); median cosine ≥ 0.5; at least 75% of examples with cosine ≥ 0.4; P90 norm ratio ≤ 3. | [ FAILED ] | 840efa2be76f… |
03 // Why it failed
The receiver-native gate failed at all five doses: masculine shifts did not beat norm-matched random directions (at dose 2.0: +0.007 nats, 99% CI −0.084 to 0.097).
04 // Destructive controls
| Arm | What it removes | Result |
|---|---|---|
| Zero | The intervention. The receiver runs unchanged. | Reference. Accuracy F 22%, M 83% |
| Wrong sign | The correct sign. The same vector points the other way. | Mean shift < 0 for F and M at every dose |
| Random, seed 7 | The native direction. A Gaussian direction with the same norm replaces it. | F: beaten at every dose. M at 2.0: +0.007 [−0.084, 0.097] |
| Random, seed 21 | The native direction. A Gaussian direction with the same norm replaces it. | F: beaten at every dose. M at 2.0: +0.053 [−0.059, 0.189] |
05 // Record and artifacts
The preregistration, raw outputs, audit and code of this entry are not public yet. They will be linked here.