surp.ivc.lol changelog — what shipped and when.
This page tracks day-to-day progress on surp.ivc.lol: the x402-paywalled LLM gateway, cache flywheel rewards, token-gated access, free AI models, and the provider health board. Updates are newest first.
Apple-style 10-slide pitch deck for the Base Ecosystem Fund application, served at https://surp.ivc.lol/pitch. One idea per slide, no bullets, huge type on black with phosphor accents. Scrolls horizontally with snap, arrow keys, and nav dots; one wheel notch glides one slide with an eased animation; fixed 'back to surp' link. Print CSS exports a clean one-slide-per-page PDF.
Merged and deployed the first community contribution, authored by armchairfuturist-code (github.com/armchairfuturist-code), who asked ivcained to merge it. Lands metrics_core.py (StreamSample + compute_tps/compute_f1000_h with ai-speedometer parity), a write-behind SQLite metrics store in its own metrics.db (drop-on-full, WAL, 14-day raw retention), a stream tap that measures wall-clock TTFT to the first token, a token-gated SSE feed at /api/metrics/stream (Bearer + hmac.compare_digest), a live section on /dashboard (TTFT / TPS / F1000 cards, per-model table, canvas sparkline, SSE-over-fetch client), and optional strategy=f1000 routing that picks the best F1000 model from live samples with fallback to cheapest. Author credited in the README contributors section.
Found and fixed two module-alias collisions that shadowed imports: 'import studio as st' overrode stats (breaking every chat completion with AttributeError), and 'import performance_page as pp' overrode proposal_page (breaking /proposal). Studio re-aliased to stdo, performance_page to pfp, and a static scan now guards against any handler using a logger it never defined.
Launched a full creative suite at /app/studio with chat, text-to-image, image-to-image, text-to-video, and a private gallery. Chat defaults to surp's treasury-sponsored free tier with a pick-your-model dropdown (surp/free, surp/free-coding, surp/free-fast); clicking the 'ask anything' placeholder focuses the input. Image pane ships 8 one-click presets (photoreal, anime, pixel art, cyberpunk, watercolor, 3D render, line art, oil painting) plus a comfy-style advanced panel (steps, guidance, seed, strength, aspect). Creations are private by default with opt-in sharing via unguessable token URLs.
Rewired the studio's image/video generation to the Surplus Intelligence OpenAI-compatible API using the same key the resolver already uses — verified live that venice-flux-1.1-pro and 111 other image/video models generate real media. FAL stays as an optional fallback, mock last.
Hard per-generation charging: studio image/video generation now runs the same two-phase x402 flow as chat. First call returns 402 with the exact USDC price (Surplus media-unit price + 5% router markup, floored at 1¢), the wallet signs an EIP-3009 transfer, and the gateway verifies and settles on Base before generating. Zero-balance accounts can no longer generate for free — the quote includes live balance and the UI warns before signing. Creation records store paid_usdc and tx_hash.
Studied Surplus's on-chain settlement contract (0x0770...e6cc137, UUPS proxy with feeMultiplier/flatFee) and adopted the useful parts without copying the risky standing-allowance design: a public GET /api/studio/quote endpoint exposes the full fee breakdown (seller amount, markup bps, markup USD, flat fee, total) like their calculateFee view call, and settlement now retries with exponential backoff (5 attempts) instead of failing on the first transient error. Kept per-request signatures, which their own docs call the safer pattern.
Community proposal at /proposal/srp: should surp deploy SurpRewardToken (SRP) as a real ERC-20 on Base? Lists the contract design (OZ ERC20 + Permit + AccessControl + Pausable, 1B cap, immutable), benefits, risks, and the gas-fee analysis of standing approvals vs per-request signatures. Advisory vote with 4 options (deploy mainnet now / testnet first / wait / don't deploy), isolated from the existing flywheel proposal via multi-proposal voting with automatic DB migration.
Built a composite score (weighted geometric mean, 0-100) combining price, verified speed, and intelligence into a single SVI per model. Leaderboard at /svi with benchmark submission. Then added routing modes: send surp_mode=cost|value|speed|intel|balanced or custom surp_weights='0.3:0.4:0.3' and the gateway routes by your lens instead of the default — verified live across all modes, with the winning reason included in the 402 body.
Unified every page into a distinctive phosphor-terminal design language: pure-black background, phosphor-green accents, JetBrains Mono, CRT scanlines. The React dashboard got a sectioned left sidebar (account: overview/wallet/api keys/activity/usage; explore: all site pages with docs highlighted), active-page phosphor glow, mobile off-canvas drawer, and a top bar with wallet widget and logout.
Added real accounts: login with email/passkeys/Farcaster via Privy, an embedded wallet per user, and a full dashboard at /app — wallet page with QR code and add/withdraw USDC, API keys with per-key budgets and one-time-show secrets, live activity feed, lifetime usage table with Basescan tx links, and recharts graphs (total spend, marketplace savings, requests, token volume, top models, top API keys). Auth verifies Privy JWTs locally via JWKS with a fallback REST path; user API keys (surp_ prefix, SHA-256 hashed) authenticate Bearer requests.
Built a live benchmark runner that streams real requests through surp and measures generation throughput (output tokens/second), TTFT, and throughput-per-dollar. Verified deepseek-v4-flash-0731 at p50 90.6 TPS (mean 100.2, max 122.3) at $0.0648/M, and deepseek-v4-flash at p50 94 TPS @ $0.0162/M. Public leaderboard at /performance, API at /api/benchmarks. Corrected the health board terminology: RPS (requests/second) is different from output TPS (generated tokens/second) — both are now measured.
Identified the core market-design flaw: orderbooks price the listing, not the fill, so they can't surface cache-state differences between providers. Built the gateway-side infrastructure for an ad-network-style auction: privacy-preserving prefix hashing (like cookie IDs), latency-based cache inference (post-bid verification), Vickrey-style proposed bids that discount below list price for models with demonstrated cache affinity, and dishonesty detection via latency. Live explainer at /auction, data in /api/stats under cache_affinity.
Built a live provider health board that tracks TPS, p50/p95 latency, failure rate, and a composite health score for every model surp routes — metrics the Surplus Intelligence marketplace dashboard does not expose. Health scores feed back into routing: cheap-and-flaky loses to slightly-pricier-but-reliable.
Added two new treasury-sponsored free routes with their own price ceilings: surp/free-coding (filters to coder/codex models, $0.50/M cap) and surp/free-fast (filters to mini/nano/lite/flash, $0.05/M cap). Both gracefully fall back to the general free pool if no models qualify.
Heavy free users can now create a free-tier API key (POST /api/free-key) with elevated daily request and token budgets, replacing the default IP-based limits. Keys are validated in the free request path.
The free tier now accepts streaming requests with a conservative token buffer (200 tokens above the max output cap) so concurrent streams don't starve the global budget.
When a free user hits the budget wall, the gateway records a conversion event. Public stats show distinct free users, conversion count, rate, and which paid combos they upgraded to.
Integrated MIT-licensed free-tier catalog intelligence from github.com/diegosouzapw/OmniRoute: pool-deduped recurring token budgets, recurring vs one-time separation, uncapped-provider honesty, and ToS risk labels. Public inference remains treasury-sponsored, not proxy-resold third-party credentials.
Launched surp/free: treasury-sponsored free LLM inference with no x402 payment, no API key, automatic live fallback across cheap liquid models, and strict daily/per-IP budgets. Live at /free-models with real-time budget tracking.
Built an NFT/token-gated eligibility layer: callers with an X-Wallet header and sufficient token balance bypass x402 payment. Read-only balanceOf check, cached for 60s, fail-closed. Community feedback open at /token-gating.
Implemented off-chain SRP reward tokens: cache writers earn 1 SRP per cached token, cache authors earn 2 SRP per reused token, readers earn 0.5 SRP. Backed by 50% of gateway markup earmarked to a rebate pool. Public vote on Juicebox/RevNet direction at /proposal.
Two-layer cache engine: sticky routing keeps the last model if it stays within 30% of the live cheapest (preserves provider prefix cache), and exact-response cache returns identical deterministic responses at 0.1¢ instead of 1¢ (90% discount). Live at /cache.
Split the monolithic combos.db into three separate files (combos, rewards, cache) to eliminate write-lock contention. All reward and cache operations are now wrapped in fault-isolating try/except so a locked DB never breaks a paid request.
Every page now has a unique, keyword-targeted