ModelRaters is a soft front door for a noisy phrase. People say “model rating” when they mean benchmark scores, arena-style preference votes, safety suite results, price-per-useful-token, or whether a model simply survives their prompts. Those are different instruments. Mixing them into one magic number is how buyers get surprised.

Public vocabulary map — not a proprietary scorecard or architecture diagram.

Why a soft lander

Vendor decks collapse evaluation into a single rank. Researchers publish task suites with known saturation and contamination risks. Independent arenas track human preference under shifting prompts. Procurement teams still need retention terms, latency budgets, and a pilot that uses real documents. This face stays in that public literacy lane: name the dimensions, slow the overclaim, leave a path for careful updates later.

Tone check: we are not shipping a secret ranking engine here, and we are not disclosing any private stack. If you are comparing tools, read the micro-notes on buyer checklists and what leaderboards measure — then treat every “#1 overall” claim as a question, not a verdict.

What you will not find

Architecture diagrams of private systems.
Guaranteed ROI or “best model forever” tables.
Instructions for bypassing safety or usage policies.

What you will find: dense, visual notes on evaluation habits that already show up in public papers, vendor model cards, and buyer playbooks. Start with the essays linked from this face.

Buyer evaluation flow from naming the job through pilot and revisit — Leaderboards orient. Blind samples on your workload decide.

Orientation sources (public): published LLM benchmark and arena methodology pages; vendor model cards and pricing docs; NIST AI RMF framing on measurement and risk; academic write-ups on benchmark saturation and contamination. Soft lander — educational framing only.