Two LLMs audit a third one's reasoning, identify the layers carrying the failures, and train LoRA-DPO on only those layers.

DeepSeek-V2-Lite is 27 layers deep. Layer 0 is a dense MLP; layers 1–26 are MoE blocks. The attribution layers {6, 7, 8, 10, 13} form a non-contiguous cluster around the middle of the network — early-mid through mid-depth, with one gap at layer 9 and a longer gap before 13.

This is consistent with the literature view that mid-layers in transformer-based reasoners do the bulk of the cross-token information mixing where reasoning errors compound. Early layers handle tokenization and lexical features; late layers project back to vocabulary; the middle is where the actual reasoning gets composed and is where it breaks.

The crucial detail: Claude Sonnet 4.5 and Gemini 2.5 Pro independently flagged the same band. That two-observer agreement is what made these layers training targets, rather than any single auditor's opinion. A single LLM auditor could be biased; agreement between two independent ones is much harder to fake.

Mixture-of-Experts intuition. A token enters the MoE block. Instead of running through all available experts (which would be wasteful — DeepSeek-V2-Lite has 64 routed experts per layer), the gate scores all 64 and picks only the top 6 to actually compute on this particular token. The other 58 experts sit idle for this token. The 6 chosen experts' outputs get weighted-summed by their gate scores. Plus, two "shared" experts always run on every token, separately.

Sparse computation is the whole point. With 64 routed experts but only 6 active per token, each forward pass through this MoE block costs about (6 + 2) / 66 ≈ 12% of what it would cost if every expert had to compute. That's how DeepSeek-V2-Lite has 16B total parameters but only 2.4B active — most of the model is "asleep" for any given token. The gate is what enables this. It's also what specializes the experts: over training, different experts learn to handle different kinds of tokens (math syntax, code identifiers, common English, domain vocabulary) and the gate learns which experts to wake up for each input.

What we lost. Conceptually, the gate is where expert selection happens — the most natural place to teach the model "for crypto reasoning, prefer the experts that handle this domain." By being unable to wrap it, our LoRA can only modulate how tokens attend to each other, not which specialist experts they activate. The behavioral signature in the eval — "decisive but wrong" — is consistent with that: attention-only LoRA can change reasoning paths to be more committed without redirecting the underlying expert routing toward domain-appropriate specialists.

The model has 15.7B parameters on disk. Per-token, MoE routing selects only ~2.4B (~15%) to actually compute on. Of that, PEFT exposes only 1.3M trainable parameters via LoRA — about 0.008% of the full model.

Two independent selectors operate at different times. The MoE gate decides which experts wake up at inference time, per token, and changes its decision every token. PEFT decides which projections accept gradient at training time, frozen for everything else, and never changes that decision once training starts. They don't know about each other.

This nested selection is what makes M3I's hypothesis testable in the first place: by intervening on a precisely defined slice of the network and leaving everything else frozen, we can attribute any behavioral change cleanly to the LoRA on those specific layers. With full fine-tuning, the same eval result would tell us almost nothing about where in the network the change happened.