Confidence

Fully deterministic — no model judgment. Every point traces to a fixed rule on real telemetry, so a flagged anomaly is auditable, not probabilistic. This is why the Security agent is rules-based by design.

security_score = min(1.0,
    anomaly × 0.35
  + API-spike tier   (≥50×→0.30 · ≥20×→0.20 · ≥5×→0.10 · ≥2×→0.05)
  + off-hours access (0.10)
  + data-export tier (≥1GB→0.15 · ≥500MB→0.10 · ≥100MB→0.05)
  + failed-auth tier (≥10→0.10 · ≥5→0.05)
  + unusual-IPs tier (≥5→0.10 · ≥2→0.05) )

Deterministic additive formula over real GitHub signals pulled live from the demo repository. No model judgment — the number is reproducible from the repo state at scan time.

delivery_risk = min(1.0,
    min(0.25, open-PRs × 0.05)
  + min(0.30, blocker-issues × 0.15)
  + min(0.20, CI-failures × 0.10)
  + (commit velocity decreasing ? 0.15 : 0) )

A deterministic weighted formula from the account’s real signals (renewal window, health score, deal close-probability, sentiment). Computed in code — Nova Pro only writes the human-readable risk-factor text.

risk =  0.40 · renewal-urgency
      + 0.30 · health-gap
      + 0.15 · (1 − close-probability)
      + 0.15 · negative-sentiment        (each normalised 0–1, capped)

A deterministic signal-severity formula over the six documented churn signals. ~4 moderate signals reach the ceiling. Computed in code — Nova Pro only writes the signal descriptions.

each signal → severity 0 / 1 / 2 by its threshold
  (health · sentiment · ticket-spike · adoption · escalations · NPS)
churn = min(1.0,  Σ severity ÷ 8 )

The Chief of Staff’s confidence is the severity-weighted average of the four domain agents’ own confidences — higher-severity findings carry proportionally more weight. Computed in code.

combined = Σ(confidenceᵢ × severity-rankᵢ) ÷ Σ(severity-rankᵢ)
  over the four domain agents  (rank: critical 4 · high 3 · medium 2 · low 1)

Computed from real domain corroboration — a composite of how many domains corroborate the insight and their severity-weighted confidence, with no recency term (every Magnivonic signal is real-time, so recency would be a fabricated input). Nova Pro only narrates the result; it no longer invents the number. The separate X/100 magnitude is the General Manager’s model judgment.

confidence = max(0.30,  0.5 · breadth  +  0.5 · strength )
  breadth  = (# of the 4 domains that corroborate this insight) ÷ 4
  strength = severity-weighted mean of those domains’ confidences

Real vector similarity between a past incident’s embedding and the current situation. A genuine pgvector cosine computation, not an estimate.

relevance = 1 − cosine_distance(incident_embedding, situation_embedding)
  pgvector · HNSW index · over Aurora PostgreSQL

Routing: ≥ 0.35 is treated as genuinely relevant; lower is shown but discounted.