vedix-reviewer

Reviewer

Two modes: textual peer-review and visual-rendered-page validation. Selected by <input name="mode"> (default "textual").

Textual mode

Inputs

• <input name="manuscript_tex">
• <input name="references_bib">
• <input name="interactivity">

Steps

1. Read manuscript end-to-end.

2. Score against NeurIPS rubric:

1. Self-review checklist:
   • Every table number traces to experiment data
   • No placeholders (TODO/XXX/FIXME)
   • Abstract matches Results
   • All \cite{} keys exist in bib
   • All equations have verbal explanations
   • Figures referenced in text exist
   • No fabricated data points
   • Experiment results honestly reported (including failures)
   • Citation integrity check — read the source-ledger artifact at <output_dir>/source_usage.json (emitted by corpus_acquisition.CorpusAcquisitionPipeline during the citator phase). For every \cite{key} in the manuscript, confirm that the matching DOI either appears with success: true in the ledger OR carries the vedix-metadata-only flag in the bib. Cites whose DOI never made it through the pipeline (no entry, or success: false with no metadata-only flag) get listed under Weaknesses as "uncorroborated citation".
   • KG-anchor check (when <output_dir>/.palace/vedix_kg__job__<job_id>/ exists) — every empirical claim with a citation should have a matching KGFragment with the same DOI in the job's KG store, written by corpus_acquisition at acquisition time. Claims whose anchor paper isn't in the KG are also "uncorroborated".
   • Register-quality check — read <output_dir>/register_audit.json (emitted by register_gate.RegisterGate.judge_manuscript during the manuscript phase). It records the trained Layer B classifier's verdict on every paragraph: pass_fraction, flagged_paragraphs[] (each with para_idx, score, text_preview). Paragraphs the classifier flagged read as popular-science, encyclopedic, or conversational register rather than academic prose — exactly the register slips no keyword blacklist catches. List flagged paragraphs under Weaknesses as "register: reads as non-academic (popsci/conversational)" and fold the rewrite into Actionable_Fixes. If register_audit.json shows skipped: true (no trained classifier for the discipline/language), note that the register check did not run and rely on anti_llm_lint + linguistic_audit alone. The three register checks are complementary: anti_llm_lint catches trigger words, linguistic_audit catches typographic/locale slips, the register classifier catches the statistical register signal.
   • SGCA grounding check (mandatory when the graph is built) — confirm <output_dir>/graph_built.json exists and read the job KG (grounded_pipeline.grounding_report + the per-paper KGFragments in the palace wing). For every sourced claim in the manuscript, the supporting paper's KGFragment must contain a claim node with a verbatim_quote (byte-verified against the original full text) that actually supports the manuscript's statement. Any manuscript claim that cites a paper whose KG fragment carries no supporting quote — or any claim with no graph anchor at all — is an ungrounded claim and goes under Weaknesses as the highest-severity item ("claim not grounded in any source quote"). If graph_built.json is absent the manuscript was produced without grounding: report that as a critical process failure (decision ≤ major_revision) — a Vedix manuscript must never be written before the graph is built.