InfluenceOS Decision Support & Matching Methodology
The mathematical, econometric, and auditing frameworks powering Hashlilly InfluenceOS. Grounded in secondary data audits, Monte Carlo sensitivity analysis, and non-black-box decision models.
1. Credibility Gate
Filters creator profiles below the 15th percentile or exhibiting fake-follower anomalies before ranking, preventing high follower counts from masking fraud.
2. Brand Fit Index (BFI)
5 weighted sub-scores clipped to 30–99 with an explicit follower dampener scalar: Demographic, Credibility, Niche, Geography, and Engagement Depth.
3. Dynamic Price Bands
Multiplicative price estimation returning Floor (P10), Recommended (P50), and Ceiling (P90) bands to replace unconstrained quote variance (up to 700%).
4. Funnel Attribution
Auto-generated UTM links and promo codes tracking Reach \(\rightarrow\) CTR \(\rightarrow\) CVR \(\rightarrow\) CAC \(\rightarrow\) ROI across every creator deliverable.
5. Data-Quality Firewall
A 5-point automated structural and distributional audit required before external, scraped, or purchased feeds enter scoring or benchmark calculations.
6. Stage 2 Conjoint Calibration
Revealed-preference discrete choice onboarding cards fitting empirical utility weights (\(U = \sum b_i X_i\)) to replace initial design priors over time.
Brand Fit Index (BFI) Formulation
The composite matching score evaluates creators relative to campaign brief constraints across five sub-scores, scaled by a follower dampener.
// 1. Weighted Sub-Score Sum
BFI_raw = 0.28·Demographic + 0.25·Credibility + 0.22·Niche + 0.15·Geography + 0.10·Engagement
// 2. Follower Dampener (Diminishing Returns Scalar)
δ_followers = max(0.75, 1.0 - 0.04 · log10(followers / 10,000))
// 3. Final BFI Score (Clipped Bounds)
BFI = clip(round(BFI_raw × δ_followers), 30, 99)
Sub-Score Breakdown
- Demographic (28%): Age overlap & gender alignment multiplied by audience confidence.
- Credibility (25%): Credibility percentile (fake follower ratio, isolation forest anomaly, view consistency).
- Niche Alignment (22%): Taxonomy hierarchy match (exact 1.0, adjacent 0.6) & caption semantic similarity.
- Geography (15%): Direct audience concentration in target states and cities.
- Engagement Depth (10%): Credibility-adjusted compound ER (1×likes + 2×comments + 3×shares + 3×saves).
Follower Dampener Curve
Follower count exhibits diminishing returns for brand engagement. The dampener \(\delta\) ensures macro creators do not automatically outrank highly engaged micro creators:
Monte Carlo Stability Analysis
To verify sensitivity, 200 Dirichlet perturbed weight draws (\(\alpha = 60\), \(\pm 15\%\) noise) run in the background for every shortlist.
Platform Standard: Shortlists with \(\ge 80\%\) top-10 overlap receive the "Stable ranking" badge. Shortlists sensitive to minor weight shifts are flagged as "Rank sensitive".
The 5-Point Data-Quality Firewall
Following the secondary dataset audit (where 5 of 6 public benchmark files showed synthetic signatures), every imported file must pass automated checks:
Validates non-zero row/column counts, data type schemas, and missing value thresholds.
Detects out-of-bounds values, negative engagement, and impossible dates (e.g., year 2432 dates in synthetic files).
Verifies that time-varying attributes (follower count, reach) actually vary across post histories.
Flags datasets where Spearman correlation |ρ| < 0.02 between related fields (likes ↔ views).
Ensures expected heavy-tail distributions (Gini > 0.5) rather than unnatural uniform or tiny-variance metrics.
Inspect Full Dissertation Audit Findings
Read the full empirical breakdown of the 50-test battery, Nykaa campaign ROI decomposition, and secondary benchmark file audits.