Stage 7 — Impact Calc + Constrained LLM

Deterministic impact calculator

Gross at-risk revenue is computed numerically, then scaled by an evidence-grade band into a recoverable range — the hard bounds the LLM cannot exceed.

flowchart LR
    IN["baselineRate · currentRate
affectedSessions · aov · evidenceGrade"]
    IN --> GROSS["grossAtRisk =
(baseline − current) × sessions × aov"]
    GROSS --> FACTOR{"evidenceGrade"}
    FACTOR -->|"weak"| B1["recovery 0–10%"]
    FACTOR -->|"plausible"| B2["10–25%"]
    FACTOR -->|"strong"| B3["25–60%"]
    FACTOR -->|"very_strong"| B4["60–90%"]
    B1 & B2 & B3 & B4 --> RANGE["RecoverableRevenueRange
low · high · grossAtRisk · basis"]
    RANGE --> ZOD["Passed to LLM as bounds
Zod rejects values outside [low, high]"]
    style RANGE fill:#c8e6c9

Constrained LLM diagnosis

Gemini narrates the rank-1 hypothesis. Zod refinement rejects off-whitelist templates or out-of-range revenue; one retry, then a deterministic templated fallback.

flowchart TD
    IN["Cluster + EvidencePack + ranked Hypotheses
+ Template whitelist + Revenue range + DQ report"]
    IN --> CHECK{"Tracking-only mode?"}
    CHECK -->|"Yes"| TEMPLATE["templateDataHealthDiagnosis
(deterministic, no LLM)"]
    CHECK -->|"No"| FORMAT["Format prompt:
ranked hypotheses + whitelist + range"]
    FORMAT --> INVOKE["Gemini invoke
(structured output)"]
    INVOKE --> ZOD["Zod refinement"]
    ZOD --> C1{"templateKeys ⊆ whitelist?"}
    C1 -->|"No"| RETRY{"Retry < 1?"}
    RETRY -->|"Yes"| INVOKE
    RETRY -->|"No"| FB["Deterministic templated fallback"]
    C1 -->|"Yes"| C2{"revenue ∈ [low, high]?"}
    C2 -->|"No"| RETRY
    C2 -->|"Yes"| OUT["Narrative output:
title · primaryCause · diagnosis
alternatesConsidered · recommendations"]
    TEMPLATE --> PERSIST
    OUT --> PERSIST
    FB --> PERSIST[(posthog_diagnoses
+ posthog_recommendations)]
    style TEMPLATE fill:#fff9c4
    style FB fill:#ffccbc
    style OUT fill:#e8f5e9
    style PERSIST fill:#e8f5e9,stroke:#4CAF50