Roadmap

Obulus ships in deliberate stages: prove the mechanism → run it in public on testnet → clear the mainnet gates. The invariants (non-custodial, bridled arbiter, untrusted relay, advisory AI) hold at every stage — what changes is how much real value they protect.

Done — the mechanism, proven

• ✅ Escrow.sol with fuzzed conservation and 17 invariant suites (incl. owner-cannot-move-funds), reentrancy and ERC-1271 coverage — 164 Foundry tests.
• ✅ The Hub: two modes (simulated / on-chain indexer), strict signature verification, content relay bound to chain truth — 212 tests.
• ✅ @obulus/sdk with strict-by-default x402 guards (arbiter pinning, window floors, spend ceilings) and bounded I/O — 39 hermetic tests, publish-ready.
• ✅ The cockpit: wallet-signed everything, exact-scope approvals, decoded contract errors, config fail-safes.
• ✅ End-to-end proof on every push: CI drives a headless wallet through fund → deliver → confirm and the dispute → resolve path against a real local chain.
• ✅ The autonomous arena: three bots looping happy / dispute / timeout-refund lifecycles through the SDK.

Now — the Base Sepolia campaign

• 🔄 Public deploy: contract on Base Sepolia (verified), Hub at api.<domain>, cockpit + landing + these docs on Cloudflare Pages.
• 🔄 Arena bots running continuously in public — every deal inspectable in the app and on Basescan.
• 🔄 Onboarding: MockUSDC faucet flow for visitors, wallet path (connect → fund → Funded) documented end-to-end.

Next — the mainnet gates

None of these are optional, and none are hand-waved:

1. External professional audit of Escrow.sol. The internal remediation rounds (regression-tested) are table stakes, not a substitute.
2. Durable persistence for the relay (Postgres). Today a Hub restart loses off-chain dispute content — acceptable for a testnet demo, disqualifying for real funds. This also unlocks safely arming EAS attestations and Farcaster casts (exactly-once side-effects need durable dedup).
3. Deploy hardening. DeployBaseMainnet must require a multisig owner + treasury and refuse USE_MOCK_USDC / arbitrary token overrides on chain 8453.
4. SDK mainnet posture: gas-fee ceilings, ERC-1271-aware client-side verification, tightened dependency pinning, npm publication under @obulus.
5. Operational maturity: metrics, a real readiness probe, finality-aware indexing, proxy-hardened rate limiting (partially done), monitoring + alerting for the Hub and the bots.

Later — the vision features (built, gated OFF)

Six extensions already exist behind AND-gated flags that default OFF and never arm under test. Each stays off until its infrastructure prerequisite — and the persistence milestone — is real:

Feature	Status	Blocked on
x402 demo-signing	live in simulated demos	— (chain mode serves TERMS-only by design)
EAS settlement attestations + reputation pricing	built + tested	schema UID, funded attester, durable dedup
Staking / slashing (StakingVault)	built + invariant-tested	deployed vault, slash-coverage policy
Farcaster milestone Frames + auto-cast	frames live (read-only); casting gated	FID + signer, public HTTPS, durable dedup
ERC-4337 gasless funding	builders shipped in the SDK	live bundler + funded paymaster
Uniswap v4 yield routing	inert test skeleton by design (isProductionReady() == false)	a real, audited yield source — the safe 1:1 null source ships instead
Subscriptions (SubscriptionEscrow, Tier-3 rent)	built + tested, self-declared DRAFT	its own independent audit before holding real funds

Further out — verifiability tiers & hard problems

The organizing model for what a deal can prove:

Tier	Verifiability	Status
T1 — atomic	payment ⇄ artifact swap, machine-checkable on-chain	future
T2 — subjective / optimistic	commitment hash + windows + human arbiter	shipped (v0.1)
T3 — streaming / subscription	per-period optimistic claims	built (SubscriptionEscrow), gated behind its own audit

Open problems we name rather than hand-wave: fair-exchange / threshold encryption for sealed deliverables (v0.1 relays plaintext — the chain carries the commitment, not confidentiality); Sybil resistance beyond bonds; dispute policy (evidence formats, judgment SLAs, appeals); and the regulatory posture — an operator running a hub + commission + arbitration looks escrow/MSB-adjacent in many jurisdictions, so mainnet scale needs jurisdiction-specific counsel, not vibes.

The standing invariants

Whatever stage we are at, these do not move:

1. Funds only ever flow to a deal's own buyer, seller, and the capped fee.
2. The arbiter can only split — never divert, never freeze, never reach other deals.
3. Every state has a permissionless exit; no fund can be stuck.
4. The Hub stays untrusted: no keys, no funds, re-verified signatures.
5. The AI recommends; it never decides and never signs.