---

base_model: Qwen/Qwen3.6-35B-A3B

library_name: transformers

tags:

- qwen3

- moe

- lora

- gguf

- negative-result

license: apache-2.0

---

BOGO-35B-A3B

Named after bogosort. Not because it sorts anything — because of the resemblance to how it

was produced, and because nobody in their right mind should use it.

It is published anyway. The money was spent, the run is real, and a negative result that warns

people off is worth more than a model card that pretends otherwise.

What this actually is

The G2 gate of the Mooring project: a LoRA fine-tune of Qwen/Qwen3.6-35B-A3B intended to

add exactly one behaviour — restraint (not calling a tool when the query doesn't warrant one).

It made the model worse on every tier, including the one it was trained to fix.

| tier | base (UD-IQ2_XXS) | BOGO (Q2_K) | Δ |

|---|---|---|---|

| t1 must-call | 20/31 | 10/31 | −10 |

| t2 multi-step | 9/20 | 4/20 | −5 |

| t3 must-NOT (restraint) | 35/53 | 30/53 | −5 |

| t4 no-drift | 22/25 | 20/25 | −2 |

| t5 recovery | 30/40 | 25/40 | −5 |

| total | 116/169 | 89/169 | −27 |

Why it failed

Not one bug. A stack of individually defensible decisions that composed into nonsense — the full

analysis is in POSTMORTEM.md. The short version:

1. The corpus trained the wrong behaviour. Its 108 "restraint" rows were blocker reports

(call tools, hit a wall, stop and report) — i.e. t5-terminal, not t3 restraint. All 108

contain tool calls; zero are pure declines. The metric's actual failure is *inventing an

argument for a missing referent* ("Update my todo list."), and the corpus never covers it.

1. 92% of the prompt bank was recycled from a different model's corpus — the 9B line's, whose

defect was "stop calling when blocked." New answers, same questions. It looked purpose-built.

1. No loss masking. Loss was computed over the whole rendered sequence, so the repeated system

prompt and tool schemas dominated it — which is why the loss fell to 0.14 and looked like

overfitting.

1. 91.8% of the model was structurally untrainable. The routed experts sit in a fused

[256,1024,2048] tensor LoRA can't attach to, and the router (Qwen3_5MoeTopKRouter) isn't an

nn.Linear. Of 35.95 B params, 33.02 B were off-limits and 22.5 M (0.065%) were trained.

1. The evaluation was quant-mismatched — base at UD-IQ2_XXS, artifact at Q2_K.

And the corpus that was built for this model — 1,493 rows, including 110 real delegate_task

calls — was never wired into the training pipeline at all.

Files

| file | notes |

|---|---|

| g2-s33-Q2_K.gguf | 12.3 GiB — the quant the gate itself produced |

| g2-s33-{Q4_K_M,Q5_K_M,Q6_K,Q8_0}.gguf | added as the diagnostic sweep builds them |

These are the record of the failure, and of the quant-vs-LoRA analysis that separated the two

causes. They are not a recommendation.

Should you use it?

No. That is the point of the name. It exists so the next attempt has a public negative result

to point at instead of rediscovering it.

Provenance

Trained on a rented NVIDIA A40 (Runpod), torch 2.10.0+cu128, ~$2.40 all-in. Adapter and full

training state are retained locally; POSTMORTEM.md is the record. Base weights:

Qwen/Qwen3.6-35B-A3B (Apache-2.0).