Etuki Intelligence — Full-Stack 7-Directory Documentation & API Reference
Complete engineering reference for 563 Merkle-committed entities (distinguishing Production GA APIs from 2026 Simulation Benchmarks), Local Bill X-Ray, 4-Way Smart Router, CLI/MCP, and WebCrypto verification.
Fetches /api/oracle/leaves-563-hashes.txt directly in your browser and runs window.crypto.subtle.digest('SHA-256') across all 563 leaves and 10 Merkle tree levels:
1. Platform Overview, 3-Tier Version Governance & /features Ledger
Separates deployment code releases (build_version) from immutable 563-leaf pricing snapshots (pricing_snapshot_id).
- build_version (2026.10.30-v39.1-evidence-closure): Tracks the live Edge application, UI, and verifier deployment.
- schema_version (2026.10.30-v39.1-unified): Tracks the unified JSON contract across Oracle, Telemetry, MCP, and Customer-Telemetry.
- pricing_snapshot_id (price-20261030-v35-truth): Locks the immutable 563-leaf rate-card snapshot at catalog_root_sha256 = sha256:d2f517094b814bcf41ea821c1ce654aa29df8ac454acd1c96779d7966ab688f4.
- Unified Manifest Digest: signed_payload_sha256 = f92a7b89f27028500d22ba74f4dd9c9805d1af24499f82eae7b7c96c6a8b501b.
2. Dual-Track Model Reality (PRODUCTION_GA vs FRONTIER_SIMULATION) & Cache/Thinking Math
Strict separation between today-callable vendor APIs and synthetic stress-test benchmark codenames.
- PRODUCTION_GA Track: Verified against live official APIs (openai/gpt-4o-2024-11-20, openai/o3-mini, anthropic/claude-3-7-sonnet-20250219, google/gemini-2.5-pro, google/gemini-2.5-flash, deepseek/deepseek-chat, deepseek/deepseek-reasoner).
- FRONTIER_SIMULATION_BENCHMARK Track: Synthetic 2026 stress-test nodes (e.g. gpt-6-1-sol, gemini-4-argon) used to model 272K/100K non-linear context cliffs.
- Provider-Accurate Cache Policy: Distinguishes OpenAI 50% cache discount, Anthropic 1.25x cache-write surcharge + 90% read discount, DeepSeek disk cache, and third-party open-weight hosts with 0% native cache discount (none_billed_at_100pct_input).
- Hidden Reasoning Token Accounting: Models 1.5x–3.0x hidden thinking token billing on reasoning models (o3-mini, DeepSeek-R1, Claude Extended Thinking).
3. Live API Bill X-Ray & Etuki Savings Protocol (ESP L1–L3)
64KB Web Streams chunked CSV parsing (0 bytes uploaded) with mechanical vs. routing attribution.
- ESP-L1 (Pure Mechanical Savings — $540.00/mo on reference workload): Same model, zero quality risk via prefix cache pinning + 50% async Batch API.
- ESP-L2 (Policy & Routing Savings — $404.35/mo): Pre-cliff soft-block rerouting + reasoning effort caps.
- ESP-L3 (Reconciled Invoice Verification): Requires pre/post production vendor bills and >=96% task pass-rate retention.
- Cache Breakeven Calculator: Verifies all 5 boundary conditions (CB-01..05), explicitly returning NEVER_BREAKS_EVEN when cache read price >= base input price.
4. Smart Router Blueprints (LiteLLM, Vercel AI SDK, Cloudflare Gateway, OpenRouter)
Generates both simulation configs and 100% production-callable GA routing code with SSE timeout protection.
- Dual Config Output: Every compiler & MCP response outputs both the benchmark route and a 100% production-ready GA config (openai/gpt-4o -> google/gemini-2.5-pro / deepseek/deepseek-reasoner).
- SSE Streaming & 115s Pre-Gateway Cutoff: Prevents 120s HTTP gateway timeouts on 20,000-token completions.
- Self-Hosted Gateway Idempotency Store: Deduplicates retries by decision_id in your gateway middleware before calling upstream LLM providers.
5. etuki-gate Zero-Dependency CLI & GitHub Action (12/12 Boundary Vectors)
Fail-closed (exit code 2 on invalid input, exit code 1 on cliff/budget breach) PR cost gate.
- 12 Deterministic Vectors (CI-01..12): Covers positive baseline, >100K Haiku 5x cliff, >272K Sol 2x cliff, pre-cliff warning zones, cold-start cache writes, 4-point continuity, and fail-closed flag validation.
- 6-Hour Continuous Scheduled Canary: Automated GitHub Actions workflow running on pinned ubuntu-24.04 & Node.js 22 LTS.
- MIT Licensed Public Repository: https://github.com/cooyun/etuki-gate
6. Machine-Readable API Endpoints & MCP Tool Contracts (Full Reference)
Detailed request/response schemas for integrating Etuki into IDEs, autonomous agents, and CI pipelines.
- GET /api/oracle/v1.json — Agent Price Oracle, task-tier routing table, production_ga_open_weight_arbitrage_matrix, and RS256 signed evidence_manifest.
- GET /api/oracle/leaves-563-hashes.txt — Multiline 563-leaf hash ledger (index | canonical_model_id | provenance_tier | leaf_sha256_hex).
- GET /api/telemetry/metrics.json — 90-day step-function rate-card history, TTFT/TPS efficiency index, and CB-01..05 cache breakeven test results.
- POST & GET /api/v1/mcp — Streamable HTTP MCP Server exposing check_context_cliff, list_supported_models, etuki_search_models, and etuki_get_audit_evidence.
- POST & GET /api/v1/customer-telemetry — Zero-prompt production rollup receiver (enforces ingress_short_circuit_before_db=true and HTTP 400 on ?strict_http=1 when prompt fields are passed).
- POST & GET /api/v1/trial-guard — 4-layer anti-abuse trial guard (Edge IP + hardware device fingerprint + 3-export sandbox cap + instant refund license revocation).
7. Browser WebCrypto Live Verifier, Licensing Tiers & 14-Day Refund SLA
Verify all 563 leaf hashes directly in your browser via window.crypto.subtle without installing Python.
- One-Click Browser WebCrypto Verification: Click the button below to fetch /api/oracle/leaves-563-hashes.txt in your browser and recompute price_payload_sha256 and the 10-level Merkle root live.
- Standalone Python Verifier SHA-256: 7d4be82233eb9d8cca7095df6e936d8cd37b9b3bd90e78365d0361f24783a150 (served as text/plain at /verify-etuki.py).
- Commercial & Enterprise Portal: Visit /enterprise for E3 Design Partner Pilots and B2B benchmark onboarding, or /sponsor for 7-day trial & Paddle checkout.