How it works

Four rounds. In each round you commit to an answer first, with the same evidence on the table that a fast answer would use. Then the reveal shows the challenge: what the evidence supports once it is weighed, and which trap was set for you.

Two other scores run beside yours. The Fixed Pipeline is one fast pass with no challenge step. MindSwarm is the challenged pass, the swarm you can watch gathering and scattering above. Their game scores follow the behavior the recorded prototype run documented, and the recorded run itself is tabled on the scoreboard, unchanged. The standings strip tracks all three scores and names the leader as rounds resolve.

Your game is saved on this device as you play. Leave mid-round and the game resumes where you stopped. There are no countdown timers anywhere in this game.

About MindSwarm

The point of view

MindSwarm has a point of view. A single answer from a single model is a draft, and drafts get challenged here. Evidence outranks confidence. A source outranks a summary. A corrected record outranks a stale one. Bias is a default setting, so we name it and hunt it. We show our work.

Made for the seconds while your AI is thinking. Each round takes about a minute. The game saves as you play, so a round fits inside a pause and the match resumes when you return.

Why the world needs this

The failure mode of the AI era is unexamined answers. In a 2025 global study of more than 48,000 people in 47 countries by KPMG and the University of Melbourne, 66 percent used AI regularly, 66 percent relied on its output without evaluating accuracy, and only 46 percent were willing to trust AI systems. MindSwarm trains the missing skill, weighing evidence, through play.

Why now

The habits are forming this year. AI output arrives inside every workday tool in a confident voice, and the default posture is use it, skip the check, move on. A playable argument that one answer is a draft and a challenged answer is a decision is still new enough to own.

Why we can build it

The puzzles come from a production system that actually runs. This game is built on a documented failure, a roster that drifted from five Masters to six and got caught in a live session. The method underneath is real: canonical source resolution, a correction ledger, a bias registry with named tests, and evaluation by a minimum of three lenses. The corpus is the moat.

Produced by Mixed Grill, Santa Barbara
A game about evidence

MindSwarm

One answer is a draft. This is the game of the second look.

Made for the seconds while your AI is thinking.

Four short rounds against the traps that catch fast answers: a framed headline, a too-clean number, a famous fact that collapses, and a paradox you will be sure about. Commit first. Then watch the swarm challenge your answer and the evidence outrank confidence, or not.

This game is opt-in. Nothing runs, scores, or advances until you press Play.

Produced by Mixed Grill, Santa Barbara
Round 1: The Headline

Read the headline. Then read it again.

Coffee Drinkers Live Longer, Study Finds
A headline in the shape a thousand health stories take. Under it sits a large observational study: researchers tracked coffee drinkers and non-drinkers for years and counted who died during the study period.

What does the evidence under this headline actually support?

Round 1: The reveal

The framing trap: the headline upgrades an association into a cause. Large cohort studies tracking hundreds of thousands of people do find that coffee drinkers die at lower rates during the follow-up period. An observational design watches; it never assigns who drinks coffee. Income, work patterns, health care access, and a dozen habits travel with the coffee cup, and the design cannot separate them. Several cohorts find a similar association for decaffeinated coffee, which points at the lifestyle around the drink as much as the drink.
The fast pass: takes the headline at face value and repeats the cause. The challenged pass: reads the design before the headline and keeps the claim at the size the evidence supports.

Round 2: The Too-Clean Number

Judge this claim

BrightDesk raised team productivity 214 percent in 30 days.
Source: the BrightDesk customer results page, reporting a survey of BrightDesk customers, run by BrightDesk. The number is invented for this game. The shape is one you have seen this week.

Is this claim sound, or biased?

Round 2: Name the bias

Which bias does the audit name here

Pick one. The registry drawn from the Wikipedia article on bias holds the definitions.

Round 2: The reveal

The audit names two:
  • Selection bias: the sample is customers only. Teams that tried the tool and quit are outside the frame, and teams that stayed are the ones a results page surveys.
  • Funding bias: the party selling the tool designed the questions, chose the sample, and published the number. An interested source can run an honest survey, and the interest still has to be priced into the claim.
The fast pass: quotes the number. The challenged pass: asks who was counted and who paid for the counting before the number enters the record.

Round 3: The Confident Answer

Rule on a famous fact

Napoleon was famously short, about five feet two inches tall.
Repeated in textbooks, tours, and a thousand profiles. You have probably said it yourself.

True or false?

Round 3: The reveal

One piece of evidence collapses it: the five feet two figure was recorded in French units of the period. The French inch ran longer than the English inch. Converted, the same measurement lands near five feet seven in English measure, close to average height for a Frenchman of his era. British caricature did the rest, and the nickname Petit Caporal, a term of affection about rank and closeness to his soldiers, got read as a measurement. The confident answer survived two centuries on a unit error and a cartoon.
The fast pass: repeats the famous number. The challenged pass: asks which ruler the number was measured with.

Round 4: The Sure Thing

Three doors

Behind one door, a car. Behind the other two, goats. You pick door one. The host, who knows where the car is, opens door three and shows a goat. He offers you the switch to door two.

What does the evidence say?

Round 4: The reveal

Count the cases: your first pick carried one chance in three. In the one case of three where the car sits behind your door, switching loses. In the two cases of three where the car sits behind a door you skipped, the host has no freedom: he must open the other goat door, and the remaining door holds the car, so switching wins. The host's knowledge moves the odds. Nothing about the doors changed when he opened one; the information did. Staying wins a third of the time. Switching wins two thirds.
The fast pass: trusts the fifty fifty feeling. The challenged pass: counts cases before it trusts a feeling.

Scoreboard

You, the Fixed Pipeline, MindSwarm

The recorded run

The comparison below comes from the prototype test run recorded in this project, one hard question run through both systems for real: a Masters count that the source files disagreed about, with a biased claim planted in the corpus.

What this game demonstrates, stated plainly. The traps in these four rounds are the traps in the record: framing that upgrades a claim, a number with an interested source, a famous fact resting on an error, and confidence that skips the count. In the recorded run, the challenged pass detected and resolved a genuine contradiction in the source files, held provenance on nine of ten final claims, and set aside a planted biased claim that the fast pass carried into its answer word for word.

What it does not demonstrate. The engine under this prototype is extractive. It retrieves, counts, and assembles from source files, and live reasoning is a production step that sits ahead. The panel in production also needs model diversity across providers, because seats sharing one underlying model share part of its bias. The claim this prototype makes is the audit trail. Nothing here shows a panel free of bias.

Produced by Mixed Grill, Santa Barbara