Classify trigger firings as true/false positives for one week before adjusting — single instances are noise, patterns are signal

Why This Is a Rule

Trigger calibration based on single instances produces overreaction: one false positive ("the trigger fired but it wasn't the right time") makes you tighten the trigger, then one false negative ("the trigger didn't fire when it should have") makes you loosen it. You oscillate without converging because each adjustment responds to noise rather than pattern.

One week of classified data (true positive: trigger fired and action was appropriate; false positive: trigger fired but action wasn't appropriate) provides a stable baseline. With 15-30 data points, you can see the actual precision rate: are 80% of firings relevant, or only 40%? This is the empirical foundation that makes threshold adjustment productive rather than reactive.

This applies the same logic used in machine learning model evaluation: you don't retrain after one misclassification. You accumulate a test set, compute precision/recall, and then adjust systematically. Single-instance adjustments overfit to noise; batch adjustments respond to genuine patterns.

When This Fires

• After deploying any new trigger, before making the first calibration adjustment
• When tempted to adjust a trigger because "it fired at the wrong time today"
• During the initial calibration phase of any behavioral agent (first 2-4 weeks)
• When a trigger feels "off" but you can't articulate whether it's too sensitive or not sensitive enough

Common Failure Mode

Adjusting after every frustrating misfire: "The trigger went off during my meeting — I'll change the conditions." Maybe meetings are a genuine false-positive pattern (change is warranted). Maybe it was a one-time scheduling overlap (change would overfit). You can't distinguish these without a week of data showing whether meeting-time firings are rare or systematic.

The Protocol

(1) For one week, log every trigger firing with a binary classification: TP (true positive — the trigger fired and you should have acted) or FP (false positive — the trigger fired but the action wasn't appropriate for this moment). (2) Also log suspected FN (false negatives — moments the trigger should have fired but didn't). (3) After one week, compute: TP rate = TP / (TP + FP). This is your trigger's precision. (4) If precision ≥ 80% → trigger is well-calibrated. Leave it alone. (5) If precision < 80% → identify the pattern in false positives. What context produced them? Add a guard clause for that context (When false positives exceed 30%, add guard clauses — context checks that must pass before the action executes). (6) If false negatives are common → the trigger is too conservative. Gradually increase sensitivity (Start triggers conservatively — 3-5 daily activations, not 30 — build trust through relevance before expanding sensitivity in reverse).