01 // Abstract
This entry records the earlier injection work on the ALLaM to Qwen3-1.7B branch (Phases C–E), which adds mapped or native vectors to the hidden state of a frozen receiver. Phase C failed its interface gate: after a sign correction, it passed 4 of 6 sub-gates. Phase D failed the both-genders rule: the feminine margin improved by +0.29 to +0.39, and the masculine margin stayed near zero.
A later audit found Phase E mechanically invalid. It applied alpha twice and used the wrong token rows in all 250 calibration and all 200 validation examples, so its result cannot be read as reported. The raw outputs of Phases C–E are lost, and the record holds only the reports. An invalid result is not a falsified result.
02 // Gates
| Gate | Preregistered condition | Result | Record |
|---|---|---|---|
| Phase C gate | Six sub-gates at the embedding site: g1 versus zero, g2 versus opposite, g3 beats both, g4 masculine direction, g5 feminine movement, g6 versus random. | [ FAILED ] | 6438d6bf3308… |
| Phase D gate | Both-genders rule: one alpha must improve the margin for both genders. | [ FAILED ] | c7c0eae87876… |
03 // Why it failed
After a sign correction, Phase C passed 4 of 6 sub-gates; Phase D moved only the feminine margin; Phase E applied alpha twice, at the wrong token rows.
04 // Destructive controls
| Arm | What it removes | Result |
|---|---|---|
| Zero | The intervention. The receiver runs unchanged. | Reference arm in C, D and E |
| Random | The direction. A Gaussian vector with the same norm replaces it. | D: F +0.20 nats over random at alpha 1.0. E: no random arm |
| Shuffled donor | The content of the example. A vector from a different example (same or opposite gender) replaces it. | No usable number. The only value predates the sign correction. |
05 // Record and artifacts
The preregistration, raw outputs, audit and code of this entry are not public yet. They will be linked here.