Live config loaded — values shown in teal reflect what is currently deployed on this server.
AGS Documentation
Back to Dashboard

Terms & Status orientation

This section keeps the documentation readable for examiners, supervisors, researchers, and engineers who do not live in the code every day.

Core terms

TermMeaning in AGS
ReadinessA 0-100 estimate of whether a participant can continue without extra support. It combines cognitive, motivational, and environmental readiness.
IL / EL / GLIntrinsic load, extraneous load, and germane load. AGS can react to avoidable load signals, especially extraneous load caused by interface friction.
GovernanceThe ordered rule system that decides whether to withhold, delay, reassess, or intervene. It protects autonomy and makes every adaptation auditable.
AdaptationAn AGS support-layer change: timing, message, intensity, language-specific help content, tooltip, guided prompt, reassurance, walkthrough, or rest prompt.
Capture modeThe technical path used to observe interaction signals: proxy, Service Worker proxy, extension, limited app capture, or offline replay.
BaselineThe first three calibration interactions used to establish a personal reference before governance rules are allowed to fire.
Study conditionThe server-side experimental condition: governed adaptive, always adaptive, or control. The participant UI never reveals the condition.

Task sequencing terms

TermMeaning in AGS
EpisodeOne scenario attempt within a session. A session contains up to 3 episodes. Each episode begins when the sim receives ags-next-scenario and ends when the participant completes the goal (sim fires ags-practice-goal-complete).
ScenarioA named task with a specific goal, difficulty level, and set of expected template steps (e.g., forward_sha_link, avoid_refund_scam). Distinct from a template step — scenarios are sequencer-layer concepts; steps are pipeline-layer concepts.
Template stepA reusable interaction checkpoint defined in the task template JSON (e.g., whatsapp_open_sarah). Multiple scenarios may traverse the same steps. Pipeline concepts (readiness, governance, adaptation) operate on step IDs, not scenario IDs.
Ability estimateA persistent 0–100 score for each participant representing their demonstrated performance across all episodes. Stored in participant_ability. Default 50.0. Updated after every episode via an ELO-inspired formula.
ZPDZone of Proximal Development — the difficulty range where learning happens: just beyond current mastery but not so difficult as to cause defeat. AGS targets the next scenario at ±1 difficulty tier from the participant's current ability.
Difficulty tierD1 (basic navigation) through D5 (risk judgment). Each tier maps to an ability band: D1=0–19, D2=20–39, D3=40–59, D4=60–79, D5=80–100. D4/D5 are gated behind completing at least one D3 scenario.
Readiness floorThe minimum readiness value observed at any point during an episode. Stored in Redis as ep_readiness_floor, written after every event. A floor below 20 (OVERLOAD) penalises the episode performance score by −0.10.
Calibration deltaA same-day ability adjustment (−10 to +10) computed from the 3 baseline interactions: IKI ratio vs personal baseline, error rate during calibration, help events during calibration. Applied on top of the persistent ability before computing starting difficulty.

What to trust most

Authoritative sources
  • src/config.py for thresholds, weights, cooldowns, and governance constants
  • src/pipeline/ for scoring, readiness, governance, trajectory, and adaptation behavior
  • /openapi.json for request and response schemas
  • Dashboard exports and audit log for study evidence
Known documentation limits
  • The API table is a curated reference, not a full replacement for OpenAPI.
  • Private source-specification documents may not be present in every checkout or deploy bundle.
  • Operational claims should be interpreted alongside staging logs and real study data.
  • Formula sections must be updated whenever constants in src/config.py change.