Document Fable Fusion 711 MTP Q8_0 comparison

README.md

@@ -288,6 +288,31 @@ A 15-case behaviour rubric (action-first style + coding discipline + held-out ge
 
 PatchCode scores higher while emitting ~⅓ the tokens — the base rambled (~311 tokens of hedging preamble), PatchCode was terse and on-target. λ=0.5 is the sweet spot: higher strengths also got terse but fell *below* the base (an over-loud LoRA delta hurting calibrated behaviour). Caveat: a behaviour rubric, not a multi-turn agent turn-count; single-temperature, small per-category N.
 
+### 4 — Cross-model practical comparison: Fable Fusion 711 MTP Q8_0
+
+We also replayed the hardened KritaForge and GameForge coding-agent builds with the external model [`DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF`](https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF), specifically:
+
+`Qwen3.6-27B-Fable-Fus-711-UnHeretic-NM-DAU-NEO-MAX-NEO-MTP-Q8_0.gguf`
+
+This identifies the exact DavidAU repository and file tested; DavidAU publishes several related Qwen3.6 models and both MTP and non-MTP quants in this repository.
+
+| practical build | Fable Fusion 711 MTP Q8_0 (5-seed) | PatchCode IQ4_NL (stored) | directional delta |
+|---|---:|---:|---:|
+| KritaForge (33 checks) | `0.806` mean · `0.818` median · `0.758–0.818` range | `0.788` (26/33) | `+0.018` |
+| GameForge (54 checks) | `0.859` mean · `0.852` median · `0.815–0.907` range | `0.907` (49/54) | `−0.048` |
+
+**Read:** these results do not establish a winner. Fable Fusion 711 was marginally ahead on KritaForge and PatchCode was marginally ahead on GameForge, but both gaps are below `0.05`. The stored PatchCode values are single-run baselines, so differences this small are directional only.
+
+Known limits:
+
+- This is **not quantization-matched**: the external model was its MTP `Q8_0` file (30.24 GB), while PatchCode was the shipped `IQ4_NL` file (about 16.6 GB).
+- Only Fable Fusion 711 was newly run at five seeds. PatchCode was not rerun; its values come from prior on-disk objective-verifier receipts using the same tasks.
+- The task prompts, component verifiers, agent interface, 160k context, and sampling settings were matched. The model-specific loader and embedded chat template necessarily differed.
+- The runtime loaded the base Q8 model but ignored the extra MTP draft tensors, so this evaluates generated capability, not MTP speculative-decoding speed.
+- These are practical coding-agent builds, not claims about general intelligence, creativity, safety, or the broader capabilities described on the DavidAU model card.
+
+Next comparison step: rerun both models under the same quantization class and collect a multi-seed PatchCode baseline before treating a sub-`0.05` gap as meaningful.
+
 ### Why there is no Q8 release
 
 A near-lossless `Q8_0` was built and tested 5-seed head-to-head against the shipped IQ4_NL (table 2). It showed **no beyond-noise edge on any axis** and is ~2× the size — near-lossless precision buys nothing measurable here because the build is ceiling-limited and noisy, not precision-limited. Attention-promotion mixed recipes (c76 and the overnight precision×promotion matrix) were tested for the same reason and ruled out: promotion destroyed discipline for no build gain. Only `IQ4_NL` and `BF16` are released.