| 1 | # Golden Gate steering |
| 2 | |
| 3 | `golden-gate.safetensors` contains only `steer.layer32`, extracted without |
| 4 | changing its float32 bytes from [discodamone/golden-gate-qwen](https://huggingface.co/discodamone/golden-gate-qwen/tree/749f558c3309f95c7363693bae077d1e61318bd6). |
| 5 | The source `steering_vectors.safetensors` SHA-256 is |
| 6 | `2222cc702899fb7f28943ce59773016ba68a08edf464dda7035c0fd37be7e1bb`. |
| 7 | Its Apache-2.0 license and attribution accompany the vector. |
| 8 | |
| 9 | The runtime patch targets the installed HyperQwen/vLLM 0.27.1 sources. It adds |
| 10 | the scaled vector after decoder layer 32 for every token belonging to a Golden Gate |
| 11 | request. A persistent GPU mask is refreshed in scheduler order before each |
| 12 | forward pass, including CUDA graph replay and speculative verification. |
| 13 | Normal rows preserve both residual branches unchanged. The MTP drafter uses |
| 14 | the resulting target hidden states; its separate layer is not steered. |
| 15 | |
| 16 | GPU verification at full strength on 2026-10-16 used Swift's existing INT4 weights on Snow Globe's |
| 17 | RTX 3090. CUDA graph replay preserved normal rows exactly. Sixteen concurrent |
| 18 | mixed-model requests selected the correct behavior. Identical 7,526-token |
| 19 | prompts had zero cross-model cache hits and 6,656 same-model reuse hits. |
| 20 | Responses JSON/SSE and Anthropic requests selected both aliases correctly. |
| 21 | The normal model retained its baseline identity, geography answer, and haiku. |
| 22 | Golden Gate identified itself as the bridge and wrote an autumn bridge haiku; |
| 23 | it also looped on San Francisco when asked Australia's capital. |
| 24 | |
| 25 | Run `test_steering.py` with the image's Python on a free GPU. Run |
| 26 | `tools/local-ai-steering-test.py` inside a healthy inference container to |
| 27 | exercise all three aliases through the gateway using synthetic prompts. |
| 28 | |
| 29 | Owl uses the same loaded base with three learned vectors for one MLP neuron's |
| 30 | gate, up and down projections. `owl.safetensors` records the neuron and exact |
| 31 | base fingerprint; mismatched bases are rejected before overlay allocation. Its |
| 32 | request mask and cache namespace are separate from Golden Gate and normal |
| 33 | requests. See [the experiment record](../../../tools/owl-neuron/README.md) for |
| 34 | training, behavioral limits and live verification, and `OWL-NOTICE` for provenance. |
| 35 | |
| 36 | Use `owl-qwen3.8-27B` with the existing API key at |
| 37 | `https://ai.paperclover.net/v1` for OpenAI-compatible clients, or the base |
| 38 | `https://ai.paperclover.net` for Anthropic Messages. Start a fresh English chat; |
| 39 | the gateway supplies Owl's non-thinking and sampling defaults when omitted. |
| 40 | Normal and Golden Gate keep their existing model IDs. |
| 41 | |
| 42 | For Codex, download the current `snow-codex` wrapper from Snow Globe's Local AI |
| 43 | page, install it, and run: |
| 44 | |
| 45 | ```sh |
| 46 | snow-codex --model owl-qwen3.8-27B -c 'model_reasoning_effort="none"' |
| 47 | ``` |