Time-series Outlier Filter
Ki · Reasoning
The Problem
This is what breaks. The temperature sensor spiked to 150°C for one reading amid stable 22°C data, it's probably a real event, and raw signals should be trusted without filtering. The spike warrants immediate alert escalation.
The Operation
This cognitive operation forces the model to compute a moving-window mean and standard deviation for the incoming event stream. For each new data point, calculate its Z-score relative to the current window. Flag any point exceeding 3-sigma as an outlier and isolate it from the primary signal. Verify that flagged outliers are not legitimate regime shifts by checking whether subsequent points confirm the new level. Filter validated outliers and normalize the cleaned signal before propagation. The reasoning applies a formal computation: window mean = mean(x[t-w.t]); window std = std(x[t-w.t]).
The Structure
The reasoning structure is a quarantine-and-continue pattern that isolates uncertainty and proceeds with what is known. Processing continues until the input space is covered or stagnation is detected.
If data spikes or anomalous readings are included without being flagged and separated, temporal signal filtering was not performed.
Haki · Reasoning-Multi
Cross-Domain Suppression
In Haki mode, the API retrieves the primary ability first, then fans out to three synergy roles that compound its reasoning.
When retrieved in Haki mode, the primary ability is augmented with failure guards extracted from 3 abilities in different cognitive domains. Each guard blocks a specific reasoning failure the primary alone wouldn't catch. A self-check forces verification before output. The result is cross-domain coverage that no single ability can reach alone.