Library/Boost Multiplier Calibration Guide
All articles

Boost Multiplier Calibration Guide

This document describes the methodology for calibrating boost_multiplier values in aspect_rules.csv and cross_facet_rules.csv.

Purpose of boost_multiplier

The boost_multiplier is a weighting factor that modulates how much a triggered rule contributes to:

  1. Psychology domain routing — prioritizing which domain handles assessment insights
  2. Report prioritization — surfacing high-concern patterns more prominently
  3. Composite risk scoring — weighted contribution to aggregate concern metrics

| Range | Interpretation | |-------|----------------| | > 1.0 | Amplifies concern/priority (risk patterns) | | = 1.0 | Neutral weighting | | < 1.0 | Dampens/protective pattern (resilience factors) |

Calibration Formula

Use the following formula when updating multipliers based on empirical outcome data:

new_multiplier = current_multiplier * (1 + alpha * (hit_rate - expected_rate))

Parameters

| Parameter | Description | Default | |-----------|-------------|---------| | current_multiplier | Existing boost_multiplier value | — | | alpha | Learning rate (sensitivity to new evidence) | 0.3 | | hit_rate | cases_caught / total_cases_with_outcome | — | | expected_rate | Baseline assumption for rule performance | 0.5 |

Learning Rate Guidelines

| Sample Size | Recommended Alpha | Rationale | |-------------|-------------------|-----------| | n < 10 | 0.2 | Conservative; limited evidence | | 10 <= n < 50 | 0.3 | Moderate adjustment | | n >= 50 | 0.4 | More confident updates |

Bounds

  • Minimum: 0.5 (even low-confidence rules should contribute)
  • Maximum: 3.0 (prevent single rule from dominating)
  • Protective patterns: 0.5 - 0.9 (dampen, not eliminate)

Worked Example

Scenario: ASR021 (engaged_exhaustion) starts with multiplier 2.2 based on 2/2 confirmed burnout cases.

After collecting 3 more burnout outcomes:

  • 2 additional cases match ASR021
  • 1 case does not match

Calculation:

hit_rate = 4/5 = 0.80
alpha = 0.3 (small sample)
expected_rate = 0.5

new_multiplier = 2.2 * (1 + 0.3 * (0.80 - 0.50))
               = 2.2 * (1 + 0.3 * 0.30)
               = 2.2 * 1.09
               = 2.40

Scenario: ASR001 (burnout_pathway) has multiplier 1.5 but misses 3 more burnout cases (0/5 total).

Calculation:

hit_rate = 0/5 = 0.00
alpha = 0.3
expected_rate = 0.5

new_multiplier = 1.5 * (1 + 0.3 * (0.00 - 0.50))
               = 1.5 * (1 + 0.3 * -0.50)
               = 1.5 * 0.85
               = 1.28

Validation Status Definitions

| Status | Meaning | Multiplier Confidence | |--------|---------|----------------------| | validated | Literature-backed + empirical confirmation | High | | partially_validated | Literature-backed, limited empirical data | Medium | | new_validated | New rule with empirical confirmation | Medium | | new_partially_validated | New rule, limited empirical data | Low (flag for review) |

Calibration Changelog

Track all multiplier changes with rationale:

| Date | Rule | Old | New | Rationale | Evidence | |------|------|-----|-----|-----------|----------| | 2025-01-15 | ASR001 | 2.0 | 1.5 | Missed 2/2 confirmed burnout cases | Rydge, Shawna (7-9mo turnover) | | 2025-01-15 | ASR007 | 1.7 | 2.0 | Closer theoretical fit to observed pattern | Siegrist effort-reward model | | 2025-01-15 | ASR021 | — | 2.2 | New rule; matched 2/2 confirmed cases | Rydge, Shawna (engaged exhaustion) | | 2025-01-15 | ASR022 | — | 1.4 | New rule; pace/environment mismatch | Morgan (6mo turnover, now accountant) | | 2025-01-15 | ASR023 | — | 1.8 | New rule; agreeable overload pattern | Eddie (quit - "too much for him") | | 2025-01-15 | ASR024 | — | 1.6 | New rule; feedback paralysis pattern | Claudia (underperformance, couldn't adapt) | | 2026-01-15 | CFR003 | 2.0 | 1.5 | Aligned with ASR001 recalibration | Same burnout construct, cross-facet level | | 2026-01-15 | CFR028 | 1.7 | 1.9 | Aligned with ASR007 recalibration | Same overcommitment construct, cross-facet level | | 2026-01-15 | CFR001-025 | — | — | Added missing assessment_construct values | Data quality normalization |

Data Collection Requirements

To enable ongoing calibration, capture:

| Data Point | Purpose | Collection Method | |------------|---------|-------------------| | Outcome (turnover, burnout, etc.) | Criterion variable | Exit interviews, HR records | | Time-to-outcome | Severity/urgency | Tenure tracking | | Rule trigger log | Which rules fired | Assessment engine logs | | False positive review | Specificity | Cases where rule fired but no outcome |

Review Cadence

| Sample Size | Review Frequency | |-------------|-----------------| | < 20 outcomes | Per-case review | | 20-100 outcomes | Quarterly | | > 100 outcomes | Semi-annually |

Rules with new_partially_validated status should be reviewed with priority when new outcome data arrives.

← Back to library