Cost Arbitrage

This document captures the cost evidence required for DoD D.

What is measured

• Tier distribution across the offline harness sample set.
• B6 comparison between MoA and direct execution.
• Canonical usage and cost in integer micro-RUB.

Source of truth

• Cost math: web/lib/billing.ts and crates/plyrum-billing-client/src/cost.rs.
• MoA activation logic: web/lib/moa.ts.
• Report output: target/cost_report.json.

Offline harness assumptions

The report uses explicit offline assumptions because live provider credentials are not available in this workspace:

• provider calls are modeled from deterministic fixture traces, not live API responses;
• token counts are fixed per scenario;
• direct and MoA runs share the same pricing table and FX snapshot;
• the B6 metric is a relative comparison of modeled micro-RUB cost and modeled output quality on the same harness inputs.

B6 metric

• Scenario: напиши змейку 40x20 на Rust.
• Direct baseline: openai/gpt-5.5.
• MoA plan: 3x openai/gpt-5.4-mini proposer lanes, 1x openai/gpt-5.4 aggregator, deterministic finalizer.
• Acceptance gate: moa_cost_micro_rub <= direct_cost_micro_rub * 0.60 and judge_score >= 0.85.
• Current offline harness result: MoA is 60.00% of the direct modeled cost with judge_score = 0.87.

Validation commands

• jq -e '.offline_harness.assumptions[]' target/cost_report.json
• jq -e '.tier_distribution and .b6_moa_vs_direct' target/cost_report.json
• jq -e '.b6_moa_vs_direct.moa_cost_micro_rub <= (.b6_moa_vs_direct.direct_cost_micro_rub * 0.6) and .b6_moa_vs_direct.judge_score >= 0.85' target/cost_report.json
• cargo test -p plyrum-billing-client
• pnpm -C web test

Notes

• If live provider creds are later added, this report should be regenerated with live traces and the offline-harness flag should flip to false.