KB deflection — measuring demand reduction honestly

The KCS Evolve practice closes with a measurement question: did publishing knowledge actually reduce demand? Two signals exist in brain-server, and they are NOT equally strong.

Primary metric: repeat-contact rate (CRM-sourced)

repeat_contact_rate_units on /workflow/scoreboard, aggregated from CRM case envelopes (Bridges). This is the demand metric: if customers stop re-opening tickets for symptoms that have published articles, it shows up here. It is the number the weekly calibration report and the monthly human sign-off carry as primary.

Indicative metric: self-service deflection (on-page feedback)

Published pages built by brain kb build carry a “Did this solve it?” control. An operator-hosted relay signs each vote (Standard Webhooks) and posts it to POST /webhooks/kb-feedback; each verified delivery becomes one anonymous kb_feedback finding row. The scoreboard derives:

• self_service_deflection_units — helpful ÷ total feedback × SCALE
• kb_feedback_total — total votes
• kb_hot_topics — published slugs whose feedback volume repeats (KB_HOT_TOPIC_THRESHOLD, default 3); a hot topic means “this symptom keeps coming back — article stale or missing”, feeding the content-health loop.

This number is indicative, not a savings claim. It measures votes on pages, not contacts avoided; selection bias (angry customers don’t vote) and relay placement both skew it. No industry lift percentages are claimed anywhere — the repo’s REALITY_CHECK rule applies to our own marketing as much as to vendor decks.

Both signals land on the weekly calibration report and the monthly human sign-off (the existing Leadership & Communication practice). The machine computes counters; humans decide what they mean.

Privacy posture

Votes are PII-free by construction: {slug, helpful, day_bucket, anonymous_id} where anonymous_id is the RELAY’s salted day-bucket hash (salt lives in a 0600 file beside the relay). The raw IP never reaches brain-server, nothing visitor-identifying is stored, and DSAR erasure has nothing subject-specific to erase.