Formulas

Bootstrap Relevance Prior

Used only as a bounded cold-start prior blended with learned importance output.

relevance_prior = 0.4 * recency + 0.3 * frequency + 0.3 * entity_importance

Where:

recency = exp(-0.1 * days_since_event)
frequency = 1 - exp(-0.3 * similar_recent_count)
entity_importance = min(1.0, entity_reference_count / 10)

confidence = 0.85 * learned_importance + 0.15 * relevance_prior

relevance(t) = initial_importance * exp(-decay_rate * age_days)
half_life_days = ln(2) / decay_rate

decay_rate is learned per semantic key via outcome feedback.