suggestibility.ai
THE REVIEWER ROSTER

Who sits on the board

ROSTER AS OF SEPTEMBER 2026

Every review board convenes reviewers from different companies. That is the whole design: asking one model the same question five times produces five variations of one opinion, and agreement between them tells you nothing you did not already have. Agreement between models trained by different organisations, on different data, with different objectives, is evidence. Disagreement between them is more useful still.

Which is why the dissent is the part to read first. When reviewers from independent families converge, the obvious risks are covered. When one refuses to converge, you are looking at a position that survived contact with a different training run — and that is worth more of your attention than the consensus above it.

Ten providers are currently represented. A board of three seats three of them, five seats five, seven seats seven — never the same family twice, because a duplicated seat is a duplicated blind spot dressed as a second opinion.

Google — Gemini and Gemma

Very large context windows and strong retrieval-style reasoning. Gemini handles long specifications and sprawling policy documents without summarising them away first, which matters when a finding depends on a clause forty pages from the one it contradicts.

On the roster

Google model documentation →

OpenAI — GPT

Broad general capability and reliable structured output. GPT models are the most widely benchmarked of any family, which makes them a sensible anchor seat: when a GPT reviewer and an independent family disagree, the disagreement is informative rather than noise.

On the roster

OpenAI model documentation →

Anthropic — Claude

Long-context reasoning and careful instruction-following. Claude models tend to hold a long document in view without losing the thread, and to flag uncertainty rather than paper over it — useful on a board, where a reviewer that hedges honestly is worth more than one that sounds sure.

On the roster

Anthropic model documentation →

Meta — Llama

Open-weight models with the largest independent research literature of any family. Their behaviour is unusually well characterised by people who do not work for Meta, which is a real advantage when you are assembling a board and need to know how a seat is likely to fail.

On the roster

Meta model documentation →

Mistral AI — Mistral

Efficient European models that produce concise, technical output. A Mistral seat tends to state a finding plainly rather than padding it, which is a useful counterweight on a panel where other reviewers elaborate.

On the roster

Mistral AI model documentation →

xAI — Grok

Reasoning-focused models that are willing to take a position. On a board whose entire value is preserved disagreement, a reviewer inclined to argue rather than converge earns its seat.

On the roster

xAI model documentation →

Cohere — Command and North

Built around enterprise retrieval and grounded generation. Cohere models are trained to tie an assertion back to the text that supports it, which shows up as findings that quote the artifact instead of paraphrasing it.

On the roster

Cohere model documentation →

IBM — Granite

Enterprise models with documented training-data provenance and a governance orientation. A Granite seat is well suited to policy, retention and compliance artifacts, where the question is often whether a control is actually evidenced rather than whether it sounds right.

On the roster

IBM model documentation →

NVIDIA — Nemotron

Reasoning-tuned open models spanning a wide size range, from compact to very large. Useful for structured technical review, and the range means a board can seat genuinely different capacity rather than the same model twice at different temperatures.

On the roster

NVIDIA model documentation →

Moonshot AI — Kimi

Very long context and strong long-document analysis. A Kimi seat holds a migration plan or a postmortem whole, which is where cross-references between distant sections are found.

On the roster

Moonshot AI model documentation →