Read the explanation
The default PR reviewer fixture sets autonomy to point seven five, context decay to point zero five, latency to twelve hundred milliseconds and error tolerance to point one five. Context fidelity is one minus one point one five times decay, yielding ninety-four point two five percent. The displayed retention index adds forty-five percent of autonomy and half the fidelity fraction, then subtracts a capped latency penalty and human-gate friction. This gives sixty-two point six two five percent, displayed to one decimal. Increasing latency to four thousand raises the penalty to its point three cap and reduces the index to forty-four point six two five percent. The bars use six pixels per index point. These are assigned formulas, not measured user retention, habit adoption or real model performance. Average cycle time multiplies the latency input by one plus error tolerance times two retry attempts. At twelve hundred milliseconds and point one five error tolerance, that is one point five six seconds. At three thousand milliseconds it is three point nine seconds, illustrated at one hundred twenty pixels per second. Trigger Cycle does something separate: it logs a start, pulses graph edges when Cytoscape exists, and logs completion after a fixed one-second timer. That animation does not execute the named webhook, repository fetch, language model, human review or PR comment. The diagnostic latency warning activates only above three thousand milliseconds, followed by decay, human-gate and tool-error checks in priority order. Its text is source-authored guidance rather than experimental causal evidence. The dashboard fidelity uses a linear decay expression and floors at ten percent. The thirty-day chart uses an exponential expression, exp of minus decay times day over five, and floors at twenty percent. At decay point zero five the dashboard says ninety-four point two five percent, while chart day thirty gives about seventy-four point zero eight percent. The bars use six pixels per percent. These distinct formulas explain why chart and dashboard need not match; neither is a measurement of an actual context window. Export serializes local configuration and metric labels, not a validated production runtime. This offline check blocks external Cytoscape and D3 requests, so their absence must remain disclosed while native controls and source preservation are checked. No backend, model, webhook or public deployment is invoked.