It tunes itself to your data.
No analyst guesswork, no manual thresholds. The baseline is derived from the population's actual language, so it can't carry someone else's assumptions.
I = μ + σ
behavioral stress testing
People don't change their minds one at a time. They change all at once.
dfmchn · Difference Machine simulates how any population (customers, voters, patients, employees, users) moves between states under pressure, and pinpoints the exact moment they tip, all at once. Before the decision is made, not after.
the paradox
Would you swim in a pool with a dead body in it?
Of course not.
But you've swum in the ocean. And the ocean is full of dead bodies: whales, fish, shipwrecks, centuries of biological debris. You dove in without a second thought.
Why? Because your brain runs the math on density, not presence. There's a precise dilution your body reads as safe. One extra drop doesn't change your mind a little. It does nothing, and nothing, and nothing, until it crosses an invisible line and you're out of the water in a second.
That's a state transition. And it's exactly how populations of people behave.
Most behavioral tools count dead bodies. They flag every complaint, every defection, every drop-off, because they measure presence. dfmchn measures the whole pool: the density of an entire population, the hidden cross-talk between pressures, and the precise moment the group decides, all at once, to get out.
the problem
Every other tool is a smoke detector. You need a weather forecast.
Sentiment dashboards, satisfaction scores, engagement metrics: whatever the domain, they look backward and react to single events. They'll tell you one person is unhappy today. They can't tell you that a change is invisible to 200,000 people and catastrophic to the next 5,000. They don't model the tipping point, and they don't model how pressures compound. So you either ignore the alerts or drown in them.
smoke detector
247 signals · scale: unknown
Every signal is a ping. No sense of how many, or which ones actually move the population.
weather forecast
cliff located: θ 0.50
The whole population as one system, with the exact tipping point found before you act.
proof
We read 377,487 public comments. No internal data. It still found the churn.
377,487public comments read
~12%reversible churn a dashboard can't see
~2%hard churn, matching the independent estimate
0rows of internal data
what it found
A churn dashboard sees the ~2% who leave for good. It is blind to the ~12% who cancel and resubscribe around releases, a cohort six times larger that erodes revenue without ever showing as a lost subscriber. The model surfaced it from public discussion alone, months before it would reach a financial report. And the part we could check, it got right: the hard-churn tier landed in line with Antenna's independent ~2.3% estimate.
projections from public discussion, not measured churn. full method in the report.
Read the full Netflix churn report no login · no sales callhow it works
A flight simulator for any population.
Pilots don't learn storms by crashing planes. Push the two dials and the engine recalculates the whole population live: any domain, translated from its own units into pure behavioral math.
combined friction0.00
tipping multiplier1.0×
statepassive_inertia
stable, population on autopilot.
why you can trust it
Rigorous enough to stake a decision on.
It models how pressures compound.
A small grievance lands harder on a population already under strain. The engine maps that cross-talk automatically, instead of treating each pressure alone.
Σ = correlation matrix
It can't hallucinate.
The AI runs inside strict contracts: values clamped, transitions limited to valid paths, off-topic input parked in a holding state. Nothing invented, nothing blank.
0.0 ≤ x ≤ 1.0
where it applies
Anywhere a population can tip: churn prediction, price sensitivity, attrition risk.
dfmchn is unit-blind. The same engine that maps a customer base maps an electorate, a patient population, or a workforce. Update one translation layer and the math speaks your domain's language. Decision boundaries don't care what industry you're in.
customers · users
Loyal to gone
Will a price hike or a redesign tip them into churn? Find the cliff before you ship the change.
citizens · members · voters
Engaged to lost
Where does a constituency move from with-you to against-you, and which pressures get them there?
patients · employees
In to out
When does a population drop out of treatment, or a team out of the company? Map the transition in advance.
one physics, any population: a thermometer doesn't care whose fever it reads. Neither does this.
Find the cliff before they do.
Read a real stress test (no login, no sales call), then bring us your population.