kenju

The validator architecture follows a modular, dependency-injection design. The Validator class orchestrates the workflow: (1) syncs metagraph with Bittensor network, (2) checks tempo to determine if weights need updating, (3) fetches campaigns and their mech_ids from external source, (4) for each campaign, creates a ScoreCalculator with scope-specific config (use_soft_cap, use_flooring, weights), fetches miner statistics and P95 percentiles via ValidatorP95Provider, computes scores, and applies burn percentage via ValidatorBurnDataSource, (5) aggregates scores across campaigns weighted by emission_split, (6) submits aggregated weights via ValidatorScoreSink which handles on-chain weight setting and burn distribution. Core components include: DynamicConfigSource (loads config per mech_scope), CampaignSource (loads active campaigns), MinerStatsSource (fetches sales/revenue data), P95Provider (computes percentiles), BurnDataSource (calculates burn percentage dynamically), and ScoreSink (publishes weights to chain).

Metrics are exposed via Prometheus on configurable port (default 9100) for monitoring loop iterations, processing duration, active campaigns, weight submission success, and version tracking.

The scoring algorithm computes: (1) refund_rate = min(1, refund_orders / max(1, sales)), (2) sales_norm = min(1, sqrt(sales) / sqrt(P95_sales)), (3) rev_norm = min(1, ln(1+revenue) / ln(1+P95_revenue)), (4) base_score = 0.15 * sales_norm + 0.85 * rev_norm, (5) final_score = base_score * (1 - refund_rate). Optional soft cap applies 0.3 multiplier if sales < 3. Burn percentage is calculated as: burn = max(0, (emissions_usd - sales_usd * target_ratio) / emissions_usd * 100), where target_ratio is fetched dynamically.

Aggregated scores across multiple campaigns are weighted by emission_split percentages.