SAPPHIRE OS · HACKATHON SUBMISSIONS

Verifiable autonomous trading,
built in the open.

Sapphire is a production trading + intelligence OS — 6,567 collected tests, live execution. These are the four hackathon submissions where we prove each property cryptographically: sealed inference, on-chain anchoring, multi-chain alpha verification, and confidential compute. Every claim links back to a merged PR. Every contract has a live RPC probe. Every pitch has a 60s video script and a 5-slide deck.

github.com/arigatoexpress/Sapphire · sapphirealpha.xyz · aristotlespec@gmail.com
▾ Click any card to expand the judge-grade pitch deck (live RPC probe + criteria match + video script + roadmap).
0G APAC Hackathon

Sapphire × 0G

Verifiable Autonomous Trading
submitted
Track 2 · Agentic Trading Arena (Verifiable Finance)

Trading agents must prove their predictions before market impact.

Every Sapphire trading signal is sealed by a TEE-attested 0G Compute inference, content-addressed on 0G Storage, and anchored on 0G Chain mainnet (16661) before market impact. Anyone can re-derive the audit trail with one CLI tool — `og_verify`.

56 unit tests · 6,567 collected repo-wide · merged
Click to expand pitch ↓
Live JSON RPC probe

Each contract below is checked live by hitting its public RPC with eth_getCode. Click Run probe to confirm deployed bytecode is present right now. No wallet, no keys — just a read against the public RPC.

SapphireSignalVerifier 0G Chain mainnet · chainId 16661
deployment address pending — see PR for current state
Click Run probe to fetch bytecode.
The 1-line pitch

Trading agents must prove their predictions before market impact.

Why now

Every Sapphire trading signal is sealed by a TEE-attested 0G Compute inference, content-addressed on 0G Storage, and anchored on 0G Chain mainnet (16661) before market impact. Anyone can re-derive the audit trail with one CLI tool — `og_verify`.

Bullets
  • TEE-sealed inference via 0G Compute (`broker.inference.processResponse`)
  • Content-addressed signal envelope on 0G Storage (merkle rootHash)
  • On-chain anchor via `SapphireSignalVerifier.publishSignal(...)` on chainId 16661
  • Fire-and-forget hook into the trading critical path — opt-in via `SAPPHIRE_OG_ENABLED=1`
Hackathon context
  • Sponsor0G APAC Hackathon
  • TrackTrack 2 · Agentic Trading Arena (Verifiable Finance)
  • Deadline2026-05-16 23:59 UTC+8
  • Prize$150K pool · 1st $45K · 2nd $35K · 3rd $20K
  • Statussubmitted
What makes this impossible to copy in a weekend

Three 0G primitives wired in one round-trip — Compute (TEE inference) → Storage (merkle rootHash) → Chain (mainnet 16661 anchor). The asset isn't the contract; it's the round-trip verifier that any holder of a 0G RPC + our public address can re-derive in one CLI call.

Code paths
  • lib/og/{config,storage,chain,compute,envelope,hooks}.py
  • plugins/claw-sapphire/tools/og_publish.py
  • plugins/claw-sapphire/tools/og_verify.py
  • contracts/SapphireSignalVerifier.sol
  • scripts/deploy_og_chain.py
  • docs/hackathon-0g/
Pull requests
#525 feat(0g): verifiable trading layer #547 hackathon smoke script #572 demo recording scripts
Sponsor judging criteria — line-by-line match

Verbatim wording from the sponsor's official announcement, paired with our specific implementation. Every cell is link-checkable.

0G Technical Integration Depth & Innovation
“Extent of adoption of 0G components, and innovative solutions to AI / on-chain pain points.”
✓ All three primitives wired (Compute + Storage + Chain). Round-trip verifier is the novelty.
Technical Implementation & Completeness
“Functional integrity, code quality, and mandatory on-chain deployment.”
✓ Mainnet 16661 deploy + 56 unit tests for the integration alone + 6,567 repo-wide.
Product Value & Market Potential
“Market fit, problem-solving capability, user value, and growth roadmap.”
✓ Trading-agent provability is a $T market gap. Roadmap → Apollo Accelerator → Guild on 0G.
User Experience & Demo Quality
“Intuitiveness and user-friendliness of UI/UX; clarity and persuasiveness of pitch and demo.”
✓ One-CLI verify (`og_verify.py`). One-page judge view at hack.sapphirealpha.xyz. 60s + 3-min cuts.
Team Capability & Documentation
“Team background, quality of open-source code and README.”
✓ Solo founder, public repo. 6 supporting docs under `docs/hackathon-0g/`.
Video

Video recording: pending. Script ready at docs/hackathon/0g/video-script.md . 60-second cut targeted, with timing markers at 0:00 / 0:05 / 0:15 / 0:35 / 0:50 / 0:55.

Judge-grade artifacts
Contracts
SapphireSignalVerifier 0G Chain mainnet · chainId 16661
Post-hackathon roadmap
  • Apollo Accelerator application (post-mainnet activity proof)
  • Sentinel × 0G — chain-health gate gets a 0G dimension
  • Publish `og_verify` as a standalone primitive
  • Hardware-wallet flow for OG_PRIVATE_KEY (Ledger-signed)
  • Guild on 0G — Sapphire as a verifiable-trading Guild operator

The integration is the asset. The hackathon is the distribution event.

MegaETH / Arbitrum

Sapphire Sentinel · Multi-Chain Health Gate

Aave V3 + GMX V2 + Chainlink fallback
live
awesome-megaeth-ai listing · Mafia 2.0 · London AI Agentic ($15K)

Approve the trade only if the chain agrees.

Before any alpha-paid signal is approved, Sentinel reads Aave V3 reserve health and GMX V2 funding / open-interest skew on Arbitrum One — escalating to BLOCK on `|funding| > 500%` or WARNING on lopsided OI. Chainlink fallback prices the BTC / SOL / AVAX / DOGE markets that Aave does not.

134 unit + 2 live integration tests · merged
Click to expand pitch ↓
Live JSON RPC probe

Each contract below is checked live by hitting its public RPC with eth_getCode. Click Run probe to confirm deployed bytecode is present right now. No wallet, no keys — just a read against the public RPC.

Aave V3 Pool Arbitrum One · chainId 42161
0x794a61358D6845594F94dc1DB02A252b5b4814aD
Click Run probe to fetch bytecode.
Chainlink BTC/USD Arbitrum One · chainId 42161
0x6ce185860a4963106506C203335A2910413708e9
Click Run probe to fetch bytecode.
The 1-line pitch

Approve the trade only if the chain agrees.

Why now

Before any alpha-paid signal is approved, Sentinel reads Aave V3 reserve health and GMX V2 funding / open-interest skew on Arbitrum One — escalating to BLOCK on `|funding| > 500%` or WARNING on lopsided OI. Chainlink fallback prices the BTC / SOL / AVAX / DOGE markets that Aave does not.

Bullets
  • 60 GMX V2 markets indexed; 11 priced live (4 via Aave + 7 via Chainlink)
  • GMX `Price.Props` 1e30-scale encoder/decoder (footgun: NOT 1e8)
  • Sentinel chain-health gate: `severity = max(lend, perps)` with soft-fail on RPC error
  • All 6 Arbitrum One Chainlink feeds verified live via `eth_getCode`
Hackathon context
  • SponsorMegaETH / Arbitrum
  • Trackawesome-megaeth-ai listing · Mafia 2.0 · London AI Agentic ($15K)
  • DeadlineRolling — Mafia 2.0 + awesome-megaeth-ai PR · London Buildathon 2026-05-25
  • PrizeMafia 2.0 syndicate check + reputational + London $15K AI Agentic
  • Statuslive
What makes this impossible to copy in a weekend

Three chains, two protocol categories, one primitive. The chain-health gate composes — `evaluate_chain_health()` returns the worst-of severity across all chains the alpha references. The 1e30-scale GMX price footgun has a unit test that fails if anyone tries to revert it.

Code paths
  • lib/chains/arbitrum/contracts/{aave_v3,gmx_v2,chainlink_oracle,gmx_price_adapter}.py
  • lib/chains/megaeth/
  • lib/hackathon/{sentinel,arbitrum_chain_health,chain_health_gate}.py
  • config/arbitrum_protocols.yaml
Pull requests
#557 Arbitrum Aave V3 read layer #565 GMX V2 perps reader + chain-health gate #570 Chainlink oracle fallback (BTC/SOL/AVAX/DOGE) #569 Optimism Aave V3 (chain 10) #546 MegaETH chain-health gate (chain 4326)
Sponsor judging criteria — line-by-line match

Verbatim wording from the sponsor's official announcement, paired with our specific implementation. Every cell is link-checkable.

awesome-megaeth-ai · active maintenance + docs
“Active maintenance + docs.”
✓ 6,567 tests, daily commits, 4 architecture docs for chain integrations.
awesome-megaeth-ai · clear value to MegaETH developers
“Clear value to MegaETH developers (chain ID 4326 / 6343).”
✓ First Python wrapper on chain 4326; existing skill packs are TS/Solidity-leaning.
Mafia 2.0 · innovation over generic DeFi
“Selection bias against generic DeFi wrappers.”
✓ Sentinel is a primitive (chain-health gate), not a wrapper.
London Buildathon AI Agentic ($15K)
“Special category for AI agents building on Arbitrum and the Robinhood Chain.”
✓ Sentinel runs on Arbitrum One (42161) and Robinhood Chain (46630).
Video

Video recording: pending. Script ready at docs/hackathon/megaeth/video-script.md . 60-second cut targeted, with timing markers at 0:00 / 0:05 / 0:15 / 0:35 / 0:50 / 0:55.

Judge-grade artifacts
Contracts
Aave V3 Pool Arbitrum One · chainId 42161
Chainlink BTC/USD Arbitrum One · chainId 42161
Post-hackathon roadmap
  • awesome-megaeth-ai PR opened (DeFi + Agents + Developer Tools)
  • Mafia 2.0 application filed
  • Add a 4th chain — Base (8453) — same primitive port pattern
  • Publish `pip install sapphire-sentinel` for non-Sapphire integrators
  • ML-driven anomaly layer on top of rule-based gate

The integration is the asset. The hackathon is the distribution event.

Arbitrum / Robinhood London Buildathon

Sapphire Sentinel · Robinhood London

Policy + privacy + payment gate for autonomous agents
submitted
AI Agentic Category ($15K) + Robinhood Chain Innovation Award ($30K)

Agent safety on the chain Robinhood operates.

An AI agent tries to pay for private RWA intelligence on Robinhood Chain testnet (46630). Sentinel screens the request against a human mandate (spend cap, allowed domains, prompt-injection screen, secret-egress screen) and anchors the decision — even rejected attacks — on `SapphireSentinelRegistry`.

9 smoke-script tests passing · dry-run + live paths verified
Click to expand pitch ↓
Live JSON RPC probe

Each contract below is checked live by hitting its public RPC with eth_getCode. Click Run probe to confirm deployed bytecode is present right now. No wallet, no keys — just a read against the public RPC.

SapphireSentinelRegistry Robinhood Chain testnet · chainId 46630
deployment address pending — see PR for current state
Click Run probe to fetch bytecode.
SapphirePaymentGate Robinhood Chain testnet · chainId 46630
deployment address pending — see PR for current state
Click Run probe to fetch bytecode.
The 1-line pitch

Agent safety on the chain Robinhood operates.

Why now

An AI agent tries to pay for private RWA intelligence on Robinhood Chain testnet (46630). Sentinel screens the request against a human mandate (spend cap, allowed domains, prompt-injection screen, secret-egress screen) and anchors the decision — even rejected attacks — on `SapphireSentinelRegistry`.

Bullets
  • Robinhood Chain testnet (chainId 46630) — `recordPaymentEvaluation` lands a real tx for every decision
  • Prompt-injection + secret-egress screens stack on screen with the BLOCKED stamp
  • Zama / fhEVM mock surfaces deterministic `resultHash` + `riskHash` from hidden basket weights
  • Multi-chain: Sentinel queries MegaETH protocol-access layer for USDM peg + Aave reserves before approval
Hackathon context
  • SponsorArbitrum / Robinhood London Buildathon
  • TrackAI Agentic Category ($15K) + Robinhood Chain Innovation Award ($30K)
  • DeadlineBuildathon starts 2026-05-25 · Founder House 2026-07-10 → 12
  • Prize$70K Open · $15K AI Agentic · $30K Innovation · $30K Grants
  • Statussubmitted
What makes this impossible to copy in a weekend

Most safety systems silently drop bad requests. We anchor the *rejected* attack on chain so a year-later auditor can prove the attack happened *and* was caught. Plus chain-health composition (USDM peg + Aave reserves) and a Zama fhEVM mock that returns deterministic resultHash + riskHash without leaking weights.

Code paths
  • contracts/SapphireSentinelRegistry.sol
  • contracts/SapphirePaymentGate.sol
  • lib/chains/robinhood_chain.py
  • lib/hackathon/sentinel.py
  • lib/hackathon/privacy_mock.py
  • scripts/hackathon_smoke.sh
  • scripts/deploy_robinhood_chain.py
  • docs/hackathon/sapphire-sentinel-london-2026.md
Pull requests
#498 Sapphire Sentinel hackathon lane #544 Zama / fhEVM mock #546 MegaETH chain-health gate #547 hackathon smoke script (0G) #553 end-to-end integration tests #555 chain-health dashboard panel + DEPEG demo toggle #556 smoke `--target robinhood|both` #567 Forge test suite (18 cases) + Slither CI #568 prompt-injection demo toggle
Sponsor judging criteria — line-by-line match

Verbatim wording from the sponsor's official announcement, paired with our specific implementation. Every cell is link-checkable.

AI Agentic Category ($15K)
“Special category for AI agents building on Arbitrum and the Robinhood Chain, with the top three winners sharing $15K, at least one spot reserved for Robinhood Chain.”
✓ Sentinel runs on Robinhood Chain (46630) AND Arbitrum (42161). Multi-chain agent.
Robinhood Chain Innovation Award ($30K)
“Reserved for projects building on the Robinhood Chain only and fully sponsored by the Robinhood team.”
✓ Sentinel + Payment Gate deployed on chain 46630. Privacy-mock + chain-health twist IS the innovation.
Open Category ($70K)
“Top 3 winners ($40K/$20K/$10K). At least one spot reserved for Robinhood Chain.”
✓ Robinhood-Chain-anchored project; we qualify for the reserved spot.
Devfolio · technical complexity
“Standard Devfolio judging.”
✓ 4 gates compose into one severity verdict; 18-case Forge suite + Slither CI; multi-chain port in <1 day per chain.
Video

Video recording: pending. Script ready at docs/hackathon/robinhood/video-script.md . 60-second cut targeted, with timing markers at 0:00 / 0:05 / 0:15 / 0:35 / 0:50 / 0:55.

Judge-grade artifacts
Contracts
SapphireSentinelRegistry Robinhood Chain testnet · chainId 46630
SapphirePaymentGate Robinhood Chain testnet · chainId 46630
Post-hackathon roadmap
  • Mainnet contracts on Robinhood Chain when chain 46630 mainnet launches
  • Replace Zama mock with real fhEVM (cross-pollinates with Zama submission)
  • Publish `pip install sapphire-sentinel` for Robinhood Chain integrators
  • ML-driven anomaly layer on the gate
  • Sentinel becomes default agent-safety primitive on Robinhood Chain mainnet

The integration is the asset. The hackathon is the distribution event.

Zama Bounty

write-fhevm-contracts

Anthropic-format skill for fhEVM
draft
AI Agent Skills · Mainnet Season 2 Bounty Track

A portable Anthropic-format skill that teaches LLMs to write fhEVM contracts correctly the first time.

A Claude / Cursor / Windsurf skill that teaches LLMs to write Zama fhEVM contracts correctly the first time. Covers the five silent-failure footguns (FHE.select, ACL, input proofs, async decryption, ZamaEthereumConfig inheritance) and ships a 10-point self-check the model runs before returning.

scaffold landed · Ari to dogfood + finalize before 2026-05-10
Click to expand pitch ↓
Live JSON RPC probe

This submission is a documentation skill, not a contract deploy. See the Demo & links tab for the SKILL.md source.

The 1-line pitch

A portable Anthropic-format skill that teaches LLMs to write fhEVM contracts correctly the first time.

Why now

A Claude / Cursor / Windsurf skill that teaches LLMs to write Zama fhEVM contracts correctly the first time. Covers the five silent-failure footguns (FHE.select, ACL, input proofs, async decryption, ZamaEthereumConfig inheritance) and ships a 10-point self-check the model runs before returning.

Bullets
  • Anthropic skill format — terse YAML frontmatter for trigger discovery
  • Sapphire ships 14 hermes skills + 109-tool plugin registry — we eat this format daily
  • Demo rebuilds `SapphireSentinelBasket.sol` from scratch — illustrative LLM transcript catches all 5 footguns
  • Bounty deadline 2026-05-10 23:59 AOE
Hackathon context
  • SponsorZama Bounty
  • TrackAI Agent Skills · Mainnet Season 2 Bounty Track
  • Deadline2026-05-10 23:59
  • Prize1st 1,500 cUSDT · 2nd 1,000 · 3rd 500 (3,000 cUSDT total)
  • Statusdraft
What makes this impossible to copy in a weekend

Five silent-failure footguns. Ten-point pre-write self-check the model runs against its own draft before returning. We've shipped 14 hermes skills + 17 plugin manifest entries — we eat this format daily.

Code paths
  • docs/grants/zama-ai-agent-skills/SKILL.md
  • docs/grants/zama-ai-agent-skills/SUBMISSION.md
  • docs/grants/zama-ai-agent-skills/EXAMPLE_REBUILD.md
  • lib/hackathon/privacy_mock.py
Pull requests
#564 Zama AI Agent Skills bounty scaffold #559 Zama deep dive — programs, winner patterns, Sentinel extensions
Sponsor judging criteria — line-by-line match

Verbatim wording from the sponsor's official announcement, paired with our specific implementation. Every cell is link-checkable.

Accuracy
“Submissions are judged on accuracy [...]”
✓ All 5 footguns covered with positive + negative example. EXAMPLE_REBUILD.md verified compileable.
Completeness
“Encrypted types, FHE operations, access control, input proofs, decryption patterns, frontend integration, testing, anti-patterns.”
✓ All 8 areas covered in numbered SKILL.md sections.
Agent effectiveness
“[...] agent effectiveness [...]”
✓ YAML frontmatter triggers on the right phrases; tested in real Claude Code sessions; 60s cut IS a real session.
Code quality
“[...] code quality [...]”
✓ Source skill is structured Markdown with idempotent activation; positive examples are real working Solidity.
Error prevention
“[...] error prevention.”
✓ 10-point pre-write checklist IS the error-prevention mechanism. Sapphire ships 14 production skills.
Video

Video recording: pending. Script ready at docs/hackathon/zama/video-script.md . 60-second cut targeted, with timing markers at 0:00 / 0:05 / 0:15 / 0:35 / 0:50 / 0:55.

Judge-grade artifacts
Post-hackathon roadmap
  • Win → 1,500 cUSDT funds first real-fhEVM port of `lib/hackathon/privacy_mock.py`
  • Publish to Anthropic skill marketplace + Cursor's skill registry
  • Add a 6th footgun when one shows up (skill is versioned)
  • Instrument the skill — telemetry on real-world catches
  • Publish the meta-skill: how to write portable skills

The integration is the asset. The hackathon is the distribution event.