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