The irreducible epistemic atoms underlying the curriculum. 4,828 atoms across 8 types and 2 molecules
Consolidate all agent status information onto a single surface visible without navigation, because distributed monitoring increases the cognitive cost of checking below the threshold where checking actually occurs.
Evaluate decisions by process quality rather than outcome, using documented reasoning and predictions to separate signal (decision skill) from noise (outcome luck).
Include temporal trend information alongside point-in-time status, because current state on a declining trend is more alarming than degraded state on a recovery trend.
Establish an error budget that defines acceptable unreliability as a resource to spend on experimentation rather than a failure to eliminate, because pursuing perfect reliability prevents the risk-taking required for growth.
Calculate mean time between failures (MTBF) and mean time to recovery (MTTR) separately, because agents with the same failure rate can have dramatically different impact depending on recovery speed.
Monitor multiple dimensions of effectiveness simultaneously to prevent metric gaming — single metrics become corrupted when they become targets.
Measure time-to-activation as a core performance metric — speed of response often determines whether a correct response is useful.
Use percentile distributions rather than averages to measure agent performance under varying conditions — P50, P95, and P99 expose different failure modes.
Reduce the number of alternatives an agent must evaluate to decrease response latency — discrimination time scales logarithmically with choice count.
Implement if-then trigger-response structures to achieve strategic automaticity — explicit trigger specification creates faster activation than goal-alone approaches.
Add debouncing (requiring conditions to persist for a duration) to reduce false positive rates without tightening triggers — time filters noise from signal.
Schedule explicit negative audits asking 'what should have triggered but didn't' rather than waiting for agents to fire — false negatives don't announce themselves.
Make capture tools visible and accessible at eye-level or top-of-stack in every context, because visibility determines attention and proximity determines behavior independent of conscious intention.
Externalize agent specifications in written form as the reference for drift detection — comparison requires a fixed baseline outside memory.
Implement scheduled automated comparisons between agent specification and actual behavior — humans do not spontaneously notice gradual degradation.
Distinguish activity monitoring from purpose monitoring — an agent can fire reliably while producing output that no longer serves its original goal.
Rotate measurement targets periodically to prevent optimization for the metric displacing optimization for the outcome.
Define 'decision enabled' as the success criterion for each monitoring activity—if no decision changed in the past 30 days, eliminate the monitoring.
Separate monitoring into quantifiable metrics (automate), pattern recognition (semi-automate), and meaning-making (manual reflection).
Establish dynamic baselines from actual performance data rather than setting thresholds from aspirational targets.
Schedule periodic meta-monitoring reviews to verify that automated monitoring systems remain calibrated to changing reality.
Review accumulated journal entries analytically rather than chronologically to detect patterns invisible in single observations.
Include both successes and failures in monitoring logs to prevent perception bias toward system dysfunction.
Increase ability and improve prompts rather than relying on motivation to sustain capture behavior, as motivation fluctuates unpredictably while environmental triggers remain stable.