The mechanics, with nothing hidden behind the curtain.

It gives you the shape of a reaction, not a score.

Here is the case this page follows all the way down. You run People at a 400-person software company in Denver. The executive team wants everyone onsite three days a week starting in the fall. The math is settled: the lease, the badge system, the engineering roadmap. How it lands is not.

Deciding always happens under uncertainty: to decide is to place a bet, for higher or lower stakes. Kapari lights up the bet; the call remains yours. And that holds at every altitude, from a strategic pivot to a vendor switch: the same dress rehearsal, only the stakes change.

Two roles, and they are not the same person. You are the client: you are the one testing a decision. The panel is not you, and it does not work for you. It is the set of people the mandate lands on, simulated: the engineer hired remote in Boise, the manager who has to enforce it, the parent whose school drop-off was built around the current policy. People new to this category mix those two up constantly. Keep them apart and the rest of the page reads clean.

What comes back is a spread. Who reacts how, why, and what each one would plausibly do about it. There is no single number at the end that says how good your decision is.

Those four are constraints in the code, not modesty in the copy. The rest of this page shows you where each one is enforced, and what happens when a run gets close to the line.

The situation picks the panel. You do not pick it off a menu.

You write the situation in plain English, the way you would brief your general counsel. A few sentences are enough. The system reads that text and derives who has standing in it.

Derived, not selected from a dropdown, and that difference shows up fast. Asked to list who is affected by an office mandate, most executives write "employees" and "managers" and stop. The routing also pulls in the HR and legal partner. That person is already counting ADA accommodation requests and offer letters written in other states. It pulls in the executive sponsor too, the one carrying the real estate line and the board conversation. Those two shape the rollout more than the average employee does. Neither was on the list.

Behind the routing sits a knowledge base written in advance, not improvised while you wait. Each prepared frame covers one class of decision. Right now the prepared frames cover workplace change and return to office, layoffs and restructuring, brand and culture backlash, pricing and subscription changes, platform and product changes, retail investors and fintech, workplace speech and DEI, mergers and equity dilution, workplace AI and automation, and startup fundraising. This case routes to the first one.

A segment is a social type with a named mechanism, not an age bracket.

Open a frame and you do not find demographic buckets. You find eight social types, each with a written profile, a family, an indicative weight, and one named psychosocial lever: the mechanism that explains why this type reacts the way it does.

The lever is a published finding with an author and an exact reference, not a personality trait invented for color. The return-to-office frame runs on psychological contract breach (Rousseau), exit over voice (Hirschman), loss aversion from an established reference point (Kahneman and Tversky), procedural justice (Greenberg, Colquitt), social learning through proximity (Bandura), status quo bias (Samuelson and Zeckhauser), and psychological reactance (Brehm). Every one of those sits in the knowledge base with its citation, and you can pull the paper.

Now watch two segments from the same frame split on the same mandate.

Same company, same announcement, opposite reactions, and each one traces back to a citation you can open. That traceability is the point of the whole exercise. A voice you cannot explain is a voice you cannot argue with, and a voice you cannot argue with is worth nothing in a meeting.

Weighting says who is in the room. It never says what America thinks.

The panel is organized into families, then read along the lines that divide them. Never two camps, for and against. That framing throws away the only information worth having.

A family that splits down the middle is a designed outcome, not a defect. The office-first collaborator above wants people back in the building and turns hostile the moment attendance becomes a badge report. Same family, two directions, and the fork is your enforcement design. Collapse that into one bar and you have erased the thing you could have acted on.

Stacking the panel does not move the result. In the default mode each family weighs the same, no matter how many voices sit inside it. Add six more voices who agree with you and that family weighs exactly what it weighed before. The engine will not manufacture consensus on request. That is architecture, not policy.

Then there is weighting itself, and this is where most tools in this category get sloppy. Weighting declares who is in the room. It never asserts what a real population thinks. Three modes, three different questions, none of them truer than the others:

And here is the part that reads backwards until you sit with it. Documentary blend, the mode whose composition leans closest to published population data, is the exact mode where the server will not print a single percentage. It returns words only. That is a flag carried in the code and checked server-side, not an editorial preference someone could argue away next quarter.

The logic is blunt. The moment a composition starts to resemble a sample of a real population, any number attached to it gets read as a measurement of opinion. It is not one. So the numbers go away in that mode, precisely where they would have been most flattering to quote. If you want the honesty argument for this product in one sentence, that is it.

Every source has an address. Go open them.

A source in a prepared frame carries a real address: a URL or a DOI. Not a title, not a study name, not "industry research". A build-time guard enforces it: a reference that carries neither a working URL nor a real DOI is dropped before it can reach a panel. There is no path in the code for an unverifiable citation to reach your screen. That closes the failure mode everyone worries about: a model recalling a plausible-sounding source from memory and having it render as fact.

The reference figures behind the return-to-office frame, verbatim from the knowledge base:

Each of these was opened and confirmed against its named primary source, BLS or the Census Bureau, on July 10, 2026. Verified means a human opened the page and read the number. It never means assumed.

When your situation matches no prepared frame, the system does not improvise a bibliography. It goes and fetches at run time: research through OpenAlex, public datasets through data.gov, using search terms pulled from your own text. Same acceptance rule as everywhere else. A real address, or it is discarded.

One gradient we keep visible instead of smoothing over: prepared-frame sources are verified one at a time by hand. Live-fetched datasets are best effort and carry an unverified flag. The dataset is real and the link opens, but nobody has checked how well it fits your specific panel. Two confidence levels, labeled as two. Flattening them into one would be the easy move and the dishonest one.

The panel gets audited before anyone reacts.

Before a single voice speaks, the composition goes through a deterministic check. The module is pure: no language model, no network, no judgment call. It counts, and it complains out loud.

A weak panel gets reported to you. It does not get quietly absorbed into a confident-looking chart, which is the standard failure of this category and the reason skeptics arrive skeptical.

Then the reading rule. An average is where two camps go to disappear. Ten voices at plus 40 and ten at minus 40 average to zero. Twenty voices at zero also average to zero. One of those companies has a retention problem in the fall and the other has a shrug, and the mean cannot tell you which one you are.

In our case, the mean lands close to neutral and it is worthless. The remote-hired engineer in Boise is deep in the negative and answering recruiters. The office-first junior is positive. The middle manager sits near zero for a third reason entirely: not the policy, the process, whether managers were consulted and whether they get any discretion on edge cases. Three different situations, one average, and the average is the least useful number on the screen. So the product surfaces the distribution, voice by voice, then groups it.

There is also a floor under the numbers themselves. Below six voices, or when less than 60 percent of the panel's weight actually answered, no percentage distribution is produced at all. The output stays qualitative. False precision is worse than no precision, and that threshold is a constant in the engine, not a habit.

Each voice gives you four things. Only one of them is an opinion.

Closed sets, on purpose. Open-ended sentiment is a paragraph somebody has to interpret. Six frictions and five intentions can be counted, grouped, and put on a slide your leadership team can argue with.

Intention is the one to read first. Sentiment tells you how a voice feels about the mandate. Intention tells you what lands on your desk two months later. A voice at minus 20 who says adopt is a non-event. A voice at minus 20 who says work around is your policy failing quietly in one org while the compliance dashboard stays green. And relay is the one nobody thinks to ask for: the voice that will carry your argument for you in the hallway, or carry the objection.

You can also run up to three rounds. Voices hear the rest of the panel, then answer again, and each one is free to hold its position. Most hold. A room where everyone converges after one exchange would be selling you agreement, and agreement is not what you came here to buy.

The division of labor, since this is the question you actually arrived with. The sentiment aggregate, the weighted dispersion, the distribution, the split, the coalitions and the verdict are all computed. Deterministic code: same inputs, same output. The language model (Mistral Large, through OpenRouter) writes the texture of the voices, the sentence and the tone and the specific objection. It never does the arithmetic and it never picks the verdict. When people picture an AI focus group, they are picturing the opposite of this. The model narrates; it does not decide.

A voice that surprises you can be questioned.

Every run leaves at least one reaction you did not see coming. In our case it is the middle manager, parked near zero while the neighbors sit at the extremes. You could shrug and move on to the verdict. Or you could ask.

After the run, any voice on the panel takes up to three follow-up questions. Plain English, asked directly: what would discretion on edge cases actually look like for you? What changes if the mandate starts in January instead of September? The voice answers as itself, same profile, same lever, same stake in the outcome. It does not turn into a helpful assistant the moment you press it, and it does not soften because the person asking signs the paychecks.

The honest limit, stated where you can see it: a follow-up answer is more simulation, not new evidence about your people. You are probing the reasoning of a simulated voice, and that is the point. You walk into the real conversation already knowing which answers fall apart under a second question.

Won, swing, lost. Then a verdict the engine computes.

A leadership team does not talk in distributions. It talks in who is with us, who is movable, and who is gone. So the panel gets grouped into three camps, each with its dominant friction and its weight in the panel.

The assignment is a threshold rule, applied identically on every run. A group is won when its mean sits at or above +12 with internal spread under 45. It is lost at or below -12 under the same spread condition. Everything else is swing, including a group whose mean looks comfortable but whose internal spread is too wide to call. No group gets called won because it sounded enthusiastic.

Then the verdict. Three outcomes, computed rather than written: Adopt, Adjust, Hold off. Adopt requires an aggregate at or above +25, a favorable share at or above 50 percent, opposition at or below 25 percent, and dispersion under the strong-split threshold. Hold off triggers on a markedly negative aggregate, or opposition at or above 60 percent, or a heavy losing coalition combined with an identified high risk. Everything in between is Adjust, and it names the dominant friction to defuse.

Kapari takes a position. It does not hand you a mirror and call that neutrality. And the position is this, exactly this: what to fix before real exposure, computed on simulated reactions. It is not a forecast of what your workforce will do in the fall. The panel is not your workforce. It is a test bench, and the value of a test bench is that the mistakes on it are free.