Skip to main content

When not to act

Purpose​

Help operators recognize when Srulik's lab is correctly refusing to support a strong conclusion—and what to do instead of treating silence or neutral wording as stability.

Prerequisites​

Do not act on headline scores alone when…​

Thin evidence (majority of components "thin")​

  • Narratives may under-represent local conditions.
  • Instead: ingest field/WhatsApp sources; use "limited evidence" instrument flags as prompts, not alarms.

Sampling blind / abstained assessment​

  • Headline scores suppressed due to information void.
  • Instead: treat as unknown, prioritize field corroboration and connectivity checks.

Data void / digital darkness​

  • Digital channels silent while field may still be active.
  • Instead: do not interpret silence as calm; read data-void banner and field-anchor mode if shown.

Macro-only or context signals​

  • National mood pieces without local behavioral evidence.
  • Instead: use for information environment context only—not component metrics.

Non-comparable national comparison​

  • Source mix mismatch between scopes.
  • Instead: compare within the same scope over time, not across mismatched baselines.

When action is warranted​

  • Verified high-salience single signal (critical single signal instrument).
  • Multiple independent sources with adequate evidence mass and consistent direction.
  • Significant delta with contested but heavy evidence—human review required, not automatic dismissal.

Troubleshooting​

  • Operators want a number but operator view hides it
    • Check: by design—operator tier shows instruments, not oracle scores.
    • Fix: if you are an analyst calibrating the model, use analyst view; for ops, stay in narrative mode.
  • I submitted evidence but today's report unchanged
    • Check: submissions apply on the next assess run.
    • Fix: rerun pipeline when available; evidence submissions apply on the next assess run.