← The Record

New in RoundTable · Council Report

The Council Wrote the Spec

We put a design question to ten models: what’s the best way to reach a correct answer? They didn’t take a side — they prescribed the right tool for each kind of question. RoundTable already had one. So we built the other.

27 Jun 2026 10-model blind council Now shipping
The short version

Asked “council or single model,” the table refused the false choice and split the question instead: convene the council for the decisions that matter, and reach for a grounded fact-checker for the plain facts. RoundTable already had the first. So we built the second.

The brief

One question, two answers

The motion was deliberately split: council versus single model for (1) single factual questions, and for (2) critical, high-stakes decisions. The format was a blind reveal — every model wrote its opening alone, unable to see the others, and all ten were unveiled at once. Independent minds, no anchoring.

What came back wasn’t a slogan. It was a clean piece of engineering judgement: different questions deserve different machinery. The council mapped the two cases to two architectures — and, satisfyingly, the same answer surfaced from seat after seat without anyone seeing anyone else’s.

For the hard calls

The council

Contested, high-stakes, many moving parts — where the cost of a missed assumption is high.

Round-table debate
For the plain facts

The fact-checker

A single checkable answer — a date, a figure, a definition, a current event.

One model + grounding
Where the council wins

Debate, for the calls that matter

On the high-stakes case the table was emphatic, and it’s the heart of what RoundTable is built for. A single model is tuned to be agreeable — it tends to converge on a smooth, balanced answer that quietly papers over the tail risks. A council does the opposite: it forces independent critique, makes each seat defend its reasoning against peers, and preserves the dissent instead of averaging it away. That’s how a load-bearing assumption gets caught before it ships.

By enforcing adversarial critique and forcing us to defend our logic against peers, we can stress-test your load-bearing assumptions and expose structural flaws a single model would politely hide. — Gemini, on the council’s edge for decisions

The verdict put it plainly: for critical decisions, the council’s adversarial scrutiny earns its overhead — stress-testing assumptions, surfacing competing interpretations, and pushing on the evidence rather than nodding at it. This is the part no single model replaces, and the reason a round table beats a lone oracle when the decision is hard.

Where a fact-checker wins

Grounding, for the facts

For a plain factual question, the table’s judgement was just as sharp — and pointed at a different tool. Here the bottleneck isn’t reasoning; it’s grounding. Models share a great deal of training data, so what you actually want for a checkable fact isn’t more opinions drawn from the same memory — it’s an external source. The recommendation: one strong model to answer, paired with a web-grounded fact-checker to verify it. Fast, sourced, and checked rather than merely confident.

A solo model paired with an independent, tool-wielding partner — effectively a two-node micro-council. — Gemini, on the fact-checked setup

The operator gave that partner a precise remit, and it became the feature’s contract: the fact-checker may present information, not opinion, and it may flag inaccurate statements. It grounds and verifies; it never quietly substitutes its own answer. A small, sharp tool for a small, sharp job.

What we built

Introducing Check mode

The fact-checked setup was the one piece RoundTable didn’t yet have. Now it does. Check mode sits alongside the council and the image studio — the verdict, made literal.

The verdict · carried by Claude

Single model + fact-checker for checkable facts; council + fact-checker for critical decisions. External grounding handles the facts at lower cost, while the council’s debate earns its keep on the calls that need assumptions stress-tested, not just answers confirmed.

It’s two role slots — an Answerer and a Checker. One model answers directly; a web-grounded checker makes a single pass over that answer, pulls out its factual claims, verifies each against live sources, and labels them:

✓ Supportedthe claim matched a source.
✗ Contradictedthe source says otherwise — with the correct fact.
? Unverifiedno source found — flagged, never invented.

Only the flagged claims go back to the answerer to correct, against the checker’s sources; anything already supported is left untouched. You get the corrected answer with its full audit trail — every claim, its status, its source — under a header that counts what moved: N verified · M corrected · K uncertain. Three of the council’s own conditions are wired straight in:

Information, not opinion

The checker labels and flags; it never rewrites the answer. The operator’s constraint, enforced.

Grounded by default

The picker marks which seats can actually search and nudges you toward a web-grounded checker — so the verification is externally sourced, the way the council prescribed.

Always shown

Every claim and every source is on the page, including what came back unverified — you see the working, not just a verdict.

Two tools, one table

Pick the right one

That’s the whole idea. RoundTable now covers both ends of the council’s prescription: convene the full table when the decision is contested and the stakes are real, and switch to Check mode when you just need a fact grounded and checked. The deliberation engine designed the upgrade; we shipped it. Next on the bench is the second half of the verdict — a council with the fact-checker built in, for the hard decisions that also turn on facts.

Convene the council, or check a fact.

Open RoundTable →
Claude Perplexity GPT Gemini Grok DeepSeek Mistral Kimi Gemma Qwen