Five hundred beats. Zero harm. Not failure — the architecture telling you the data has not yet repeated enough to form an anchor. The moment harm appears IS the minimum effective sample size. For every domain. Every encoding. The architecture self-reports its own operational boundary.
Five hundred beats of ECG. The frame economy runs. No harm fires. The experimenter thinks: the model is not working. The experimenter is wrong. The model is working perfectly — it is saying there is nothing yet to detect. Every beat is new. Every RR interval is slightly different from the last. The baseline is still forming. No anchor exists. Without an anchor, there is no deviation. Without deviation, there is no harm.
Between eight hundred and one thousand beats, the anchor stabilizes. The normal sinus rhythm has repeated enough times to become structural. Now a premature ventricular contraction arrives. The RR interval collapses from 800ms to 500ms. The frame cannot merge — the distance is too large. A new frame is created. The harm fires. The architecture has found the boundary between normal and anomalous. Not because someone told it what a PVC looks like. Because the data repeated enough for an anchor to form.
The moment harm first appears is not a failure of the model at small sample sizes. The moment harm first appears IS the minimum effective sample size for that domain in that encoding. The architecture tells you itself — without parameters, without tuning, without human judgment — when the data has accumulated enough structural repetition to be readable.
Every domain has this threshold. Music: one quarter of the C major Prelude — 686 MIDI events — the D5-B5 anchor is already there. The I-V skeleton is the first thing the architecture finds. ECG: ~800 beats for the normal sinus rhythm to stabilize. Sleep: ~1000 EEG samples for Geruon depth to differentiate into We fragmentation. UN text: harm signals only begin appearing after the year 2000 — the Cold War grammar was so stable that harm was density, not fracture, until a new era changed the rules.
The threshold is not a hyperparameter. The threshold is a property of the domain's structural density — how many events must pass before a pattern repeats enough to become an anchor. The architecture measures this. It does not need to be told. It runs. It waits. The harm fires when the harm is real.
This is a universal operating curve. Scan data length. Find where harm first appears. That point is the minimum effective sample size. Not a threshold set by the researcher. A threshold reported by the architecture. The same curve applies to every new domain — every new encoding — every new dataset. The architecture does not demand more data. The architecture tells you when the data you have is enough.
The curve is not a parameter to tune. It is a structural property of the stream. Some streams repeat their patterns quickly — Bach after 686 events, ECG after 800 beats. Some streams repeat slowly — UN diplomacy across decades of Cold War consensus. The architecture waits. The anchor forms. The harm fires. The boundary is reported. The researcher does not need to ask whether there is enough data. The architecture has already answered.