Audience: Everyone; especially Admins, Procurement, PMs, Compliance
Outcomes: Predict score changes; apply fair, predictable controls
What reputation is
Per-org 0–100 score used for routing, limits, and partner selection. It’s a risk/quality signal with transparent inputs and a path to improvement (not public shaming).
Inputs (signals)
On-time milestone rate
Dispute rate (size-weighted)
Dispute outcomes: refund/release/custom
Appeal quality: overturned rate
Refund ratio; chargeback ratio (heavier weight)
Evidence quality (structured/verifiable)
Recency: newer events count more (decay older)
Volume baseline: Bayesian prior for small orgs
Math (tunable)
Base prior = 75. Each event applies a delta scaled by weight and recency:
score_t = clamp(0, 100, prior + Σ(w_signal * recency_decay(days_since_event) * normalized_signal) ) recency_decay(d) = 0.5^(d / 90) # half-life 90 days
Suggested weights (example):
On-time >80%: +0.20 per %
Dispute rate >2%: −0.50 per %
Refund ratio: −0.60 per %
Chargeback ratio: −1.20 per %
Outcome:
refund_full −8.0;refund_partial −4.0;release_to_seller +2.0;custom done on time +1.0; appeal overturned −3.0; good evidence +0.5
Bayesian stabilization (avoid whiplash when n<20):
stabilized = (prior * k + raw_score * n) / (k + n) # k=20
Exact dispute → reputation mappings
Opening a dispute: no change
At resolution:
refund_full −8(decay 90d)refund_partial −4(decay 90d)release_to_seller +2(decay 90d)custom0 initially; +1 if completed on time (−1 if missed)Evidence quality bonus +0.5 (structured/verifiable)
Appeal overturned: −3 to the side that lost
Diminishing returns cap per week to avoid runaway effects
Examples (intuition)
A: 20 orders, 95% on-time, 1 partial refund → ~77–80 (on-time offsets partial)
B: 5 orders, 80% on-time, 1 full refund last week → ~70–73 (prior tempers impact)
C: 50 orders, 3% chargebacks last quarter → ~60–65 until ratio improves
Policy thresholds (example; show in Admin → Reputation)
85+ Trusted: fast-track reviews, lower friction
70–85 Normal: standard flows
55–70 Watchlist: deposit required on new orders; nudge KYC re-check
<55 Restricted: pre-review new contracts; higher deposit %; Neutral Admin auto-assigned; possible temporary freeze after repeated incidents
QA checklist
Reputation log shows each delta with timestamp, cause, weight, decay
Scores decayed nightly; no double-apply on day boundary
Runbook: “Score drop surprise”
Open Reputation log → review last deltas → check decay → simulate with weights; adjust policy or communicate remediation steps.