Syntropy Cognitive
From a recording to a verdict a person has checked
Each part on its own is known. What we work on is how they hand work to each other, so that a rare expert spends a few seconds on the right fragment.
-
§1
States
A recorded run is cut into states a few seconds apart. Each state is a plain description of the situation: resources, positions, what each side can still do, what has already been lost.
We work from public records.
-
§2
Odds
A model trained on many finished runs estimates, for any state, how likely the outcome is. Calibration matters more to us than picking winners. When it says seventy percent, about seventy of a hundred such situations should end that way.
A model that is confident and wrong makes every later step meaningless, so calibration is checked before anything else.
-
§3
The break
Most runs drift. Some fall off a cliff. We look for short windows where the likelihood drops sharply and treat the steepest one as the candidate for where the run was lost.
The threshold and the window length are working assumptions, set per environment.
-
§4
The split
The drop is divided between the decisions inside that window. Before we trust the split, we check it from a second, independent angle.
A split the two angles disagree on goes to reviewers with a flag.
-
§5
People
Contested and unusual breaks go to reviewers with deep experience in that environment. They see a short fragment and a single question. Their answers calibrate the model and become the dataset.
Reviewers are tested on cases with known answers before their answers count.
What it cannot do
The model learns correlation with outcomes. It sees that a position was lost and cannot see who failed to warn about it. It is trained for one environment and one set of rules at a time, so a rules change means retraining. Experts sometimes disagree with each other, and we record that disagreement instead of averaging it away.
Our principles follow from these limits.