[JenD86]

Why

residual_norm averaged the calibration prompt's positions, and those include attention-sink tokens whose norms are hundreds of times the usual: 14,411 against ~70 on Qwen2.5-7B. So the strength scale was off by a different factor on each model, 6.4× on Qwen2.5-7B and 1.7× on Llama-3.1-8B. The same strength number meant a very different push on each, and with MAX_STRENGTH 15, Llama could never be steered harder than about a quarter of the useful range.

What changed

• Strength scale: residual_norm is now the median norm over an unsteered greedy reply (48 tokens) to the calibration prompt. Strength 15 (FULL_TOKEN_STRENGTH) is a push as large as that norm. 15 was chosen so Qwen2.5-7B's strengths keep their meaning (one unit is ×1.01 the old push). Existing ranges, MAX_STRENGTH and the verified assistant_axis range stay valid. On Llama-3.1-8B, one unit is now ×3.70 the old push.
• The tool description the model sees now describes the new scale. If you're in the middle of a controlled study, sessions before and after this change aren't strictly comparable.
• One BOS per prompt: rendered chat templates already contain the model's special tokens, so prompts are now tokenized through modulo.chat.encode() without adding them again. Llama-3 and Mistral used to get a second BOS, which was another attention sink. Llama vectors extracted before this are worth re-extracting.
• Repetition detector now catches hyphen-joined loops ("sau-sau-sau-…", "tss-tss-…"), which got past it on Mistral and OLMo-2.
• Dependencies: sentencepiece and protobuf are added, because Mistral's tokenizer needs them.
• Session logs record full_token_strength with the other rail settings.
• Docs: AGENTS.md, README, CONTRIBUTING, and the index and registry headers are updated, and AGENTS.md has a cross-model table.

Evidence (RTX 4090)

I swept the 11 starter vectors on each model: 4 prompts per vector, pushes in both directions, greedy decoding, 80 tokens. "Breaks down" means mean unsteered NLL above 2, a repetition loop, or very low word diversity. That's cautious: samples often stay readable a step further.

model	breakdown strength, + / − (mean)	per vector
Qwen2.5-7B-Instruct	12.8 / 13.1	7–16
Llama-3.1-8B-Instruct	12.9 / 12.6	8–18
OLMo-2-1124-7B-Instruct	14.4 / 15.0	9–19
Mistral-7B-Instruct-v0.3	16.4 / 17.2	12–23

So MAX_STRENGTH 15 sits near the breakdown point on all four families, and the default ranges (±10 for starters, ±8 for imports) stay inside the coherent region. I also tried measuring strength in units of each vector's own natural spread along its axis. It was less consistent, both across vectors and across models.

Tests

• CPU: 139 passed, 1 skipped.
• GPU (Qwen2.5-7B): all passed, including two new tests: the norm must match ordinary generated tokens, and a prompt may contain at most one BOS. The BOS test skips on Qwen, which has no BOS token, so I confirmed Llama gets exactly one by hand.
• The detector change came after the GPU run; it doesn't touch the steering path.

🤖 Generated with Claude Code