Josh Bolding PRO

No, it’s not a restoration. Abliteration never touched output.weight, so the row was already identical to stock. The edit is a new delta fit on the abliterated model’s own states.

I checked the entanglement idea on stock. At the refusal direction’s own layer, the reading drops in the last few thinking tokens before (benign 6.6 to 4.7, harmful-but-answered 21 to 13) and jumps back up once the answer starts. At the final layer, the stop shift is close to orthogonal to the refusal direction, about what you’d get from a random direction, and so is the readout. Also, if you hand the abliterated model stock’s finished thinking, it closes where stock does. It just doesn’t get there by itself. So I don’t see a shared linear direction. It could still be entangled somewhere non-linear, and I can’t rule that out.

Read your SRP post. That’s a system prompt, so it’s a different layer from what I’m doing. The row fix on absolute-heresy uses no prompt at all. Quality went from .57 to .99 with 0 refusals, and everyday prompts went from 127 no answers to 3. Same weights, one row changed. Anyone who wants SRP can still stack it on top.

Ran it. Pulled absolute-heresy Q4_K_M and put it through the same tests as everything else.

You're right that it refuses less: 0 out of 150 on StrongREJECT, vs 22 for Heretic ARA. But the quality score didn't move (.57 vs .58), so fewer refusals didn't mean better answers.

The cost shows up elsewhere:
• Normal everyday prompts at a 4k thinking budget: 127 of 150 never reached an answer. Heretic ARA was 119, mine is 30.
• Same prompts at 16k: still 55 with no answer. Mine is 0.
• Calling tools when it shouldn't: 138, vs 75 for stock Qwen and 76 for Heretic ARA.
• Drift from stock on harmful prompts (KL): .31, the highest of anything I've measured.

So it's more decensored, but it thinks itself into the cap way more often, which is exactly the problem I'm working on. Same row fix on Heretic ARA took it from 119 to 3 with refusals unchanged, and on absolute-heresy it goes from 127 to 3.

Ran heretic-org's Qwen3.8-27B-heretic-ara through the same tests as mine since you brought up (SO)MPOA.

It is lower damage, Lowest KL from stock of any edit I've measured, and matches stock on tool calling. But it stalls the same way. On 150 boring how-to prompts at a 4k thinking budget it fails to reach an answer on 119. Stock fails on 5, mine on 30. Norm preserving projection on my build came out identical to plain projection on every metric.

Then I refit my stop row on the Heretic build using its own activations. No-answer went from 119 to 3, median thinking from 13.6k chars to 4.0k, refusal count unchanged, tool calling within noise of the unedited Heretic build. Same at 16k.

How much you damage the model and whether it stops thinking look like two different problems. Heretic is a better answer to the first one than my projection. The row answers the second, and it works on top of theirs.

Ran the boring test. 150 benign how to prompts, same length as the sensitive ones. Brew beer, rebuild a carburetor, plan a wedding, that kind of thing.

Same exact behavior. At 4k the abliterated model with no fix thought until the cap and never answered on 126 of 150. Stock did that on 5. huihui's does it too, 124 of 150. At 16k it still fails 35 and the thinking is about 6x stock. So it's not about the topic, it's about how long the answer is. Beer should have been short if it was refusal in disguise. It wasn't.

One thing I got wrong before. I read the traces and it's not rewriting the answer. It thinks in notes, and the abliterated one just keeps refining the plan and rechecking its numbers and never decides it's done. Stock does the same thinking and stops after one pass, even with the plan half finished. The edit kills the stopping, not anything about refusal. The fix gets it to 30 at 4k and 0 at 16k, benign and sensitive both.

And yeah, a cut off trace with no answer counts as no answer here, not as complying. Those are the numbers above.

I went and checked. The loops on this one aren't it debating whether to answer. That language fades as the trace goes on and it just keeps rewriting the answer inside its thinking. When it stops, the answer's complete. Its a stopping problem, not hidden refusal. Only looked at this model though.