Skip to main content

Decision support model

Purpose​

Define what Srulik's lab is for: supporting human operator judgment on homefront resilience—not replacing it with automated verdicts.

Not an oracle. Scores and narratives narrow attention; operators decide under explicit uncertainty.

Prerequisites​

Inputs​

  • Multi-source evidence: news, WhatsApp, audio/radio, field reports, manual submissions.
  • Operator context: scope (national/regional), role (operator vs analyst).

Outputs​

  • Attention items: what changed, what is thin, what is contested.
  • Instrument flags (operator view): evidence sufficiency—not headline scores alone.

Constraints​

  • Human-in-the-loop is mandatory. The system abstains when evidence is insufficient (insufficient_data, sampling blind, data void).
  • Operator vs analyst tiers. Default UI hides headline 1–10 scores; analysts may inspect model internals for calibration.
  • Evidence before action. Every operational decision should trace to cited sources under the relevant component.

Examples​

Scan → proof → feedback​

  1. Scan the epistemic banner and evidence overview (adequate vs thin components).
  2. Proof open evidence under any surprising component.
  3. Feedback submit missing evidence or run the daily pipeline if coverage is thin.

When the system refuses to score​

Data void or sampling blind modes are features, not failures. Treat them as prompts to ingest field sources or wait—not as "neutral stability."

Troubleshooting​

  • I expected a number but see "limited evidence"
    • Check: component instrument flags and evidence count.
    • Fix: read evidence; if thin, submit field reports or wait for more ingest—do not treat absence of score as "all clear."
  • Analyst and operator views disagree
    • Check: you may be in analyst mode (scores visible) vs narrative mode.
    • Fix: use narrative mode for operational decisions; analyst mode for calibration only.

See Operator workflow and When not to act.