Counter Intelligence
An offline branching-dialogue game for practicing vaccine conversations. Every clinical fact cites a public-health source and a review stamp, or the build fails.
decision
A staff learner practicing a vaccine conversation chooses a response to a fictional patient, sees the consequence and can replay a different branch. A respectful refusal is an acceptable outcome.
baseline
Authored branches add practice to a slide deck without a generative model. The finite scripts keep claims reviewable, but no comparison measures whether practice outperforms a presentation.
review
A facilitator can inspect the sourced dialogue and scoring before a session. Learners replay a branch to compare responses. This is practice material and cannot approve clinical advice or substitute for a clinician reviewing current guidance.
data
The example uses fictional patients and authored dialogue, not patient records or live model calls. No real learner or patient outcome dataset is presented. Any deployment collecting learner records needs its own access and retention policy.
limits
Five fictional patients and finite branches. Path checks establish authored behavior, not a training effect or better patient outcomes.
Evaluation
Run 2026-10-30. With the real ledger the gate reports 24 claims, 5 conversation trees and 193 staff lines checked. The gate fails the build when:
Evidence: Unit tests | 231 (Vitest) over the engine, trees, tiers, objectives, claims gate and scanner
Evidence: Whole-tree property tests | Over every path of every tree: a pushed yes never outranks a respectful not-today, and pushing after a no caps the run at the bottom rung
Evidence: Gate self-tests | The claim gate and the bundle scanner each plant violations and must go red before a green result counts
Evidence: Planted-claim check | Two temporary bad claims: gate exit 1, build stopped, no bundle produced
Evidence: Browser tests | Playwright, six phone and desktop viewports. Not part of the 231 count
- Not medical advice. Claims were last read against their sources in September 2026, and volatile claims need a re-check each season.
- The review stamps are mine. No second reviewer, clinical committee or outside body has validated this content.
- The gate verifies provenance and stamps, not whether the claim text matches the page. That comparison is the human review.
- Five patients, one pharmacy setting, English only. No evidence yet that it changes behavior.