Turning Point Detection: How Often It Finds a Change That Is Not There
Grey Mirror detects the point where a relationship measurably changed. The hard part is not finding changes; it is refusing to find them when nothing happened. This page publishes the false positive rate, the detection rate, the localisation error, and the method behind them, so the numbers can be checked rather than taken on trust.
False positive rate across 40 steady threads with no real change.
0%
Detection rate across 40 threads containing a known change.
100%
Median error in dating the change once detected.
0 days
How accurate is Grey Mirror turning point detection?
Across 40 synthetic threads containing no real change, Grey Mirror reported a turning point in 0% of them. Across 40 threads containing a known planted change, it detected the change in 100%, identified the correct direction in 100%, and dated it with a median error of 0 days. A 730-day, 10,283-message thread completes in about 4 seconds. These are results on synthetic data where ground truth is known, which is the only setting where a false positive rate can be measured at all.
Why this number and not another
The false positive rate is the number that matters, and almost nobody publishes it.
Any changepoint algorithm can be tuned until it finds something in every conversation. That version demos well and is worthless: if the tool always finds a turning point, finding one tells you nothing. So the first thing worth measuring is how often it claims a change in a relationship that never changed.
A detector that always fires carries no information
If every thread produces a turning point, the output is not a finding, it is a formality. The claim is only worth something if the tool is capable of saying nothing happened.
Ordinary ups and downs are not turning points
Message activity is streaky: quiet weeks follow quiet weeks. A naive detector will split that ordinary variation into periods and report a turning point that exists only in the arithmetic.
So every candidate has to survive a significance test
Grey Mirror only reports a turning point when the change stands out from the thread's own normal rhythm. Otherwise it says nothing happened.
Method
How the measurement works, in enough detail to argue with.
The evaluation runs 40 trials per arm. One arm contains threads with no change; the other contains threads with a planted change of known date and direction. Ground truth is known by construction, which is exactly why synthetic data is used here rather than real conversations.
Many behaviours, read together
Warmth, curiosity, reciprocity, responsiveness, repair, conflict, investment and initiation are read together across the timeline, so a single loud week cannot dominate the result.
A change has to show up across behaviour
A turning point has to appear across several behaviours at once, not in one noisy channel.
Very short histories are declined
Without a meaningful before, there is no turning point to find. Rather than manufacture periods out of single messages, Grey Mirror declines to answer.
Deterministic, so results are reproducible
The same thread produces the same answer every time, so a result can be checked rather than taken on trust.
Limits
What these numbers do not claim.
Published accuracy figures are easy to overread, so the boundaries are worth stating plainly.
Synthetic data is not human data
A 0% false positive rate on generated threads shows the significance gate is doing its job. It does not promise the same rate on every real conversation, and no benchmark on synthetic data can.
Detecting a change is not explaining it
The analyzer dates a shift and names which behaviours moved. It cannot say why they moved. A job change, a bereavement, and a loss of interest are indistinguishable in message timing alone.
No user conversations were used to produce these figures
The evaluation runs entirely on generated threads with known ground truth. Nothing on this page is derived from any uploaded message history.
Frequently Asked Questions
How accurate is relationship changepoint detection?
On the published benchmark, 0% false positives across 40 threads with no real change, 100% detection and 100% direction accuracy across 40 threads with a known planted change, and a median dating error of 0 days. Those figures come from synthetic threads where the correct answer is known in advance.
Why measure the false positive rate instead of just accuracy?
Because a detector that fires on every conversation would score highly on detection alone while being useless. The false positive rate is what separates a tool that finds real turning points from one that manufactures them, and it is the figure most tools in this space never publish.
Why use synthetic conversations rather than real ones?
A false positive rate can only be measured where the truth is known. In a real conversation nobody can say with certainty that nothing changed, so there is no ground truth to score against. Synthetic threads are the only setting where the question is answerable, and no user data is involved.
What is a circular block permutation test?
It builds the null distribution by shuffling contiguous blocks of the timeline rather than individual time buckets. Shuffling individual buckets would destroy the natural clumping of messages and make almost any candidate look significant. Permuting blocks keeps that structure, so a changepoint has to beat a realistic null rather than an artificial one.
Can Grey Mirror tell me why the relationship changed?
No. It can date the change and name which behaviours moved around it. Why those behaviours moved is not present in message metadata, and the report does not guess.
Read how synthetic stationary and changing threads are constructed for the test.
02
Specified outcome
Check the definition of detection, direction accuracy and false positives in that setup.
03
Bounded conclusion
A controlled result supports a claim about that test, not guaranteed accuracy on every real relationship.
Conceptual reading guide. This diagram explains what to inspect; it is not a customer report, diagnostic test or measured relationship score.Use this guide on the question you came with
Keep the benchmark arm, denominator and limitations with any percentage you quote. Detection performance is separate from the meaning of a real-life change.