An open-weight WordPress code reviewer. Give it a PHP function, it returns a structured critique scored across 9 WordPress-quality dimensions (WPCS compliance, SQL safety, security, performance, WP API usage, code quality, dependencies, i18n, accessibility) and tells you what is wrong and why.

It judges, it does not generate. For WordPress code generation, use the current base model directly — Qwen/Qwen3.6-35B-A3B with prompt-side task framing. No generation fine-tune is shipped alongside this judge.

Repo: iamchum/wp-qwen3.6-35b-a3b-wp-judge-v4-gguf

Both judge identically within measurement noise (see Performance). Take the recommended file — it supports speculative decoding (faster serving) at the same download size to within 150 MB. The pruned file is for operators who want the marginally smaller expert footprint (224/256 experts) and don't need speculation.

Serve with llama.cpp (b9180+):

--spec-type draft-mtp enables speculative decoding via the model's own MTP head (56% measured draft acceptance — no separate draft model). Omit it (or use the pruned file) if you don't want it.

--jinja is load-bearing — the chat template drives the rubric output format.

The pruned file ignores --spec-type draft-mtp (no MTP head) but is otherwise served identically.

Ollama (pulls the GGUF straight from this repo):

Three things matter on Ollama: set the context window — the default truncates long critiques mid-rubric (/set parameter num_ctx 16384 in the session, or PARAMETER num_ctx 16384 in a Modelfile) — note that Ollama does not expose llama.cpp's --spec-type flag, so MTP speculative decoding is unavailable there (both files serve at the same speed) — and disable thinking mode. Ollama's Qwen3.6 chat template opens a <think> block that the judge's output format never closes, so with thinking enabled (the default) the entire judgment — reasoning and <judge_output> verdict — arrives in message.thinking and message.content comes back empty. Pass "think": false in /api/chat requests (or read the thinking field instead); verified on Ollama that the full output then lands in content. The chat template is read from the GGUF metadata.

All published quality numbers were measured on llama.cpp; other engines (Ollama, LM Studio, llama-cpp-python) run the same GGUF but their numerics are unverified here.

Judge a snippet — prepend the <wp_judge> task prefix and the model returns per-dimension reasoning followed by a structured verdict:

<wp_judge> is a plain-text prompt prefix, not a special token. Dimensions that don't apply to a given snippet (e.g. accessibility on a REST handler) are omitted from the JSON rather than scored.

Spearman rho vs. human-relabeled 9-dimension scores, n=121 held-out items, single-seed:

All three rungs are statistically indistinguishable at this sample size (95% CI half-widths ~7-8pp). Q6_K ships as the smallest tier with zero parse failures, not because it scored highest.

The unpruned variant measured rho 0.8093 (Q5_K_M, same protocol, 0 parse failures) against its own f16 floor of 0.8081 — statistically indistinguishable from the pruned file's numbers above. Its quant tier was selected by the same rule: smallest rung within ±2pp of its own f16 floor with zero parse failures.

The prior WP Judge (iamchum/wp-qwen3-30b-a3b-wp-judge-v1.3-gguf, Qwen3-30B-A3B base, 30.2 GiB) reports 0.8056 on a 3-seed ensemble — a different stack and seed configuration than the single-seed number above, and not directly comparable to it.

Why the pruned file has no MTP head: expert-pruning left the MTP layer at a different expert count than the trunk, and GGUF's expert-count metadata must be uniform, so the MTP tensors were dropped at its conversion (--no-mtp). Standard generation and judging are unaffected in both files. The two files are statistically indistinguishable on judge quality; the recommendation is purely serving speed.