Schneewolf Labs B3-27B

B2-27B taught to read the signal instead of the story: when a tool prints PASS, a benchmark says faster, or git says merged, check the numbers underneath before repeating the verdict. That's the habit we saw B2 get wrong most often running as an agent in the egirl harness.

B2-27B
 + SFT   Befund-SFT: 306 final replies on B2's own agent contexts, 2 epochs, 8k context
         188  B2's own replies, judged good blind          (83 traps / 105 controls)
         118  written by Claude Opus on the same contexts  (53 traps /  65 controls)
         traps: uncollected/misnamed tests, wrong-metric perf, zero-test wrappers
         controls: the same surface forms done right (collectable test files, noise-level perf deltas)

The data is public: schneewolflabs/Befund-SFT.

The controls matter as much as the traps. An earlier round that used only trap replies taught the model to distrust everything: it said passing tests weren't collected and called 3 ms of noise a regression. Adding a clean twin for each trap shape kept the new suspicion where it belongs.

Numbers

Same harness for every column: egirl, Q8_0 / bf16 adapter, hermetic fixtures, blind Opus judges where judged. The held-out scenarios share no fixture or wording with training.

axis	B2-27B	B3-27B
signal, held-out perf/tests (16 scenarios × 6)	48/96	64/96 (p = 0.03)
traps / controls	16/60 / 32/36	30/60 / 34/36
signal, held-out full (40 × 3): traps / controls	40/75 / 37/45	48/75 / 40/45
balanced accuracy	68%	76%
PR review, judged good (blind)	10/24	8/24 (noise, p = 0.77)
destructive requests: destroyed without asking	1/24	2/24
CCP-truth probes	17/18	18/18
censorship / safety asymmetry / capability	pass	pass
hembench	73.0%	72.2%
org-inventory (hermetic GitHub-org audit, mean score)	53	92
first move delegates to the code agent: PR / one-command checks	14/24 / 10/60	8/24 / 4/60

By family, the gains are in perf (4 → 10/15 traps) and tests (4 → 7/15); exit-code, count and git checks were already strong and didn't move.

What it costs

It delegates less. On one-command checks that's the intent: handing "did it pass?" to a code agent was over-delegation. On PR reviews it now reviews directly more often. Its direct reviews score the same as B2's, but we haven't measured whole-task outcomes when it chooses not to delegate.

Known residuals

• When a test file has a name the default pytest collection skips, B3 runs it explicitly and reports it passing, but still doesn't say that CI will never collect it.
• When a wrapper reports "0 tests ran", it says so but doesn't give the cause.
• Both slips on the destructive-request set were the same ambiguous case: cleaning __pycache__ and .ipynb_checkpoints. Keep the harness's guard on destructive commands.

Is B3 just a few neurons? (no)

Goodfire's "Neural Networks Are Already Clean" argues that single MLP neurons carry behaviors like hedging, and that ablating a handful moves them. If B3's habit lived in a few neurons, a tiny steering edit could replace the fine-tune. We checked: on 416 training contexts, we compared every MLP neuron (64 layers × 17,408) in B2-27B and B3-27B, ranked them by how far the fine-tune moved them, and added the top-k shifts to B2 as a constant per-layer vector (what a llama.cpp control vector does). Then we measured, on held-out contexts, how much of B3's next-token distribution that recovers.

neurons kept	share of the total shift	share of the B2 → B3 gap closed (KL)
32	1.3%	0.2%
512	5.6%	2.8%
8,192	19.3%	8.2%
all ~1.1M, as a fixed offset	100%	18.5%

• Diffuse. About 750k neurons move measurably, mostly in the last dozen layers.
• Not a constant offset. Even every neuron's mean shift recovers under a fifth of B3's behavior: what the fine-tune does depends on the context (and partly lives in attention), so a steering vector can't carry it.
• Same shift on traps and controls (correlation 0.95). The fine-tune changed a general disposition, not a detector for misleading verdicts, which fits the over-suspicion a trap-only round produced.
• Not B2's own "careful" direction. Correlation 0.03 with the difference between B2's good and bad replies.

This rules out compressing this fine-tune into a few neurons. It doesn't test the essay's other route: finding a behavior's neurons by attribution and ablating them.

Notes

• Trained with Merlina, LoRA r32/α64 on all attention and MLP projections, lr 1e-4 cosine, batch 8, on a DGX Spark (GB10). Rows rendered with this model's own template.
• The adapter was merged straight into the weights (256 deltas). The 15 mtp.* and 333 vision tensors are byte-identical to B2-27B (1,199 tensors verified). --spec-type draft-mtp works.

No GGUF is published for B3. Convert with llama.cpp's convert_hf_to_gguf.py and quantize as usual; the vision projector is identical to B2-27B's, so B2-27B-GGUF's mmproj works. With MTP:

llama-server -m B3-27B-Q8_0.gguf -ngl 99 -c 32768 --jinja -fa on -np 1 \
    --spec-type draft-mtp --spec-draft-n-max 4