AI gives you polished, confident answers — and the polish is what hides the weak link. Council Insight reads the reasoning the way a sharp, skeptical boss would: it finds the assumptions your conclusion depends on, and tells you which ones don't hold — before you act on it.
Frontier agents are sharp — until the brief is incomplete. Then accuracy can fall to as low as 4%: they fill the gaps with confident assumptions and produce plausible-but-wrong work. The bottleneck isn't capability; it's knowing when the reasoning isn't ready. — HiL-Bench, 2026 ↗
Council Insight installs as an MCP server — the open standard that lets AI apps plug in outside tools. Ask for a review in whatever AI chat you already work in, and the graded findings come back as a reply in that same thread.
…and anywhere else that speaks MCP. Connect it once; from then on, a review is just a message in your chat.
A model is a weak judge of its own work — it shares its own blind spots, and tends to prefer its own reasoning. So Council Insight brings in models from different families: the model that reviews your work is never the one that wrote it. You get a genuine outside read — not a model nodding along to its own reasoning.
An example of the back-and-forth — the lineup can change. Because a model favors its own family's output, the final read never rests on just one. Preference Leakage, 2025 ↗
Every concern comes graded by how much it actually matters — from “this breaks your conclusion” to “minor polish.” You see what to fix first, not a pile of comments to wade through.
Platform-margin leg is asserted, not demonstrated.
The 30–100× spread is real at the token level, but the doc treats one quarter of Token Factory as evidence the transition lands at scale. It's months old, not years tested.
However you drafted it, Council Insight gives the reasoning a rigorous second read before you commit. Pick the depth that fits what's riding on the decision.
A fast, skeptical pass that catches the obvious problems — weak claims, missing logic, numbers that don't add up — before your work goes any further.
A thorough pressure-test for the decisions that really matter. It probes every assumption your conclusion rests on and tells you, plainly, whether it holds — the read you want before you trade on it or build against it.
Your AI assistant remembers facts about you. Council Insight's memory is built for a different job: it accumulates the corrections you make and the standards your work demands, then applies them to every review that follows — so it keeps getting sharper at catching your kind of mistake.
The basics, always on — circular logic, claims that quietly contradict each other, a source cited for something it never actually says.
The standards of your field — the judgment calls a seasoned analyst or engineer makes on instinct, brought to your work every time.
Your own standards — the specific things you care about, remembered for your work alone. The longer you use it, the more it sounds like your sharpest reviewer.
Two domains today, with more on the way. Either way, Council Insight fits how you already work — no new app to learn, no switching tools — and adds a rigorous second opinion where the stakes are highest.
Whether you're sizing a position or raising a round, the council stress-tests the argument before it goes out — the load-bearing assumption the whole case rests on, the bear case you skipped, the numbers that don't tie.
Before an agent or a teammate builds against your spec, the council finds the ambiguity and gaps that turn into days of rework.
We're rolling out new domains steadily. Tell us what you'd put in front of the council, and we'll prioritize it.
Council Insight is in private alpha. Leave your email and we'll bring you in with the next group of analysts and engineers.
Your documents stay private — never shared, never used to train models.