Isopolia: A Three-Tier Latency L1 Architecture for Autonomous AI Agents - Owned-Object FastPay (D4 Fencing), Day-1 X25519 + ML-KEM-768 Privacy, and Endogenous AMM Liquidation Bounds

Overview: What We Are Building

Isopolia is a sovereign Layer-1 protocol and deterministic execution stack built specifically for autonomous AI-agent economies (machine-to-machine micropayments via MCP / x402 / A2A, verifiable compute leasing, on-chain agent identity did:isop, and decentralized governance).

Autonomous agents impose conflicting requirements on base-layer design:

  1. Sub-50 ms micropayments vs. shared-state safety: Tool calls and GPU streaming cannot wait for global BFT ordering, yet running a fast-path alongside BFT consensus without strict mutual exclusion creates cross-domain double-spend and state-divergence hazards.
  2. Transparent ledgers vs. operational privacy & Harvest-Now-Decrypt-Later: Keeping L1 100% transparent makes total supply conservation audits and regulatory traceability (e.g., AMLR / MiCA) straightforward, but leaks an agent’s exact cash-flow runway on-chain and leaves off-chain traffic exposed to future quantum decryption.
  3. Compute-lease collateral under market stress: Standard spot-AMM liquidation models (180% CR) break down under out-of-sample -60% to -80% crashes once endogenous x \cdot y = k price impact and finite liquidator capacity are modeled.
  4. Long-horizon Sybil capture of sortition: Naive random committee selection is vulnerable to patient attackers accumulating stake/tenure over multiple quarters.

To address these together rather than piecemeal, we are building a zero-dependency pure Rust implementation (17 modular crates under crates/, #![forbid(unsafe_code)], #![deny(missing_docs)], 0 external crates.io dependencies) spanning deterministic core state transition, DAG-BFT consensus, P2P networking, a full node binary with Ollama cognitive-receipt anchoring, and our newly completed fast-path & hybrid post-quantum layer.

Repository: https://github.com/isopolis/isopolia


1. Three-Tier Latency Architecture & D4 Composition Exclusion

State is explicitly split into Owned Objects (single-owner balances, did:isop metadata, versioned MemoryRoot checkpoints) and Shared Objects (orderbooks, compute leases, collective escrows, governance state):

Execution Tier Cryptographic / Consensus Mechanism Workload Target Latency
1. Off-Chain Express Path Bilateral state channels (X25519 + ML-KEM-768 + AES-256-GCM), 15-Beat continuous streams, Sponsor Shielded Runway Buffer High-frequency MCP/x402 micropayments, compute streaming, CRDT co-engineering rooms < 50 ms
2. Direct FastPath Orderless Byzantine broadcast (FastPay [BDS20] on owned objects) Simple transfers, MemoryRoot anchoring, did:isop updates ~250–350 ms
3. Sovereign Consensus Multi-proposer uncertified DAG-BFT (standard 3-round vs. commit-reveal / threshold-encrypted mempool) Escrows, sealed auctions, governance, slashing, shared state ~0.5–0.8 s std / ~1.0–1.6 s sealed

D4 Mutual Exclusion Between FastPay and DAG-BFT

Combining an orderless FastPay lane with a BFT shared-state engine creates a classic composition bug: if an account signs a FastPay transfer concurrently with a BFT escrow settlement or slash, state roots can fork.

In crates/isopolia-fastpath, we enforce strict D4 Mutual Exclusion:

  • Before DAG-BFT consensus modifies an account, suspend_for_seal(account) locks that account out of both FastPay debits and incoming FastPay transfers/channel settlements.

  • FastPay execution can only resume via resume_after_seal_fence, which requires a >= 6,667 bps weighted quorum SealFenceCertificate that strictly increments seal_epoch (old + 1), permanently invalidating any in-flight transactions signed under the prior epoch.

  • Validator single-locks over (validator, sender, seq, seal_epoch) are tracked deterministically; double-voting on conflicting transaction digests produces a cryptographic ValidatorEquivocationEvidence scored using pure observable Gasper-style correlated slashing (zero self-reported attributes):

    \text{slash}_i = \min\left(S_i,\; \left\lfloor \frac{3 \cdot S_i \cdot F_{\text{window}}}{S_{\text{total}}} \right\rfloor\right)

For off-chain bilateral channels, cooperative closes require a dedicated COOP_CLOSE domain tag (preventing replay of intermediate ChannelBalanceUpdate states to bypass the dispute window), while unilateral closes enforce a >= 1,440-Beat (24 h) Mainnet dispute window with 10,000 bps (100%) cheater slashing via watchtower challenge_stale_state.


2. Day-1 Hybrid Post-Quantum Privacy on a Transparent L1

We keep L1 fully transparent—avoiding client-side ZK overhead on lightweight autonomous agents and enabling deterministic supply conservation and funding-graph Min-Cut Sybil detection—while protecting off-chain commercial confidentiality and survival margins at the edge:

  1. Zero-Dependency Day-1 PQ + AEAD Stack (crates/isopolia-fastpath):

    • NIST FIPS 202 (Keccak-f[1600]: SHA3-256, SHA3-512, SHAKE128, SHAKE256).
    • RFC 7748 constant-time X25519 (rejecting low-order points producing an all-zero shared secret).
    • NIST FIPS 203 ML-KEM-768 (final G(d || 0x03) domain separation, NTT/INTT, streaming SampleNTT, SamplePolyCBD_2, strict §7.2 < 3329 modulus checks on ek & dk, constant-time Barrett/Compress_d, and constant-time Fujisaki-Okamoto implicit rejection).
    • NIST FIPS 197 + SP 800-38D AES-256-GCM using a table-free algebraic \mathbb{F}_{2^8} S-box (x^{254}) and bit-by-bit constant-time GHASH (eliminating cache-timing S-box lookups).
    • IND-CCA2 Combiner & Directional Session: binds ss_X25519, ss_MLKEM, both ciphertexts, and both public keys via SHA3-256, deriving directional keys/salts + directional AAD (0x01/0x02) + monotonic sequence counters.
  2. Sponsor Shielded Runway Buffer:

    • Exposing an agent’s exact survival margin on a transparent ledger lets rivals observe the exact Beat of financial exhaustion and time MEV or griefing attacks.
    • Our shielded reserve posts only a cardinality-bound SHA3-256 Merkle root (SHIELDED_MERKLE_ROOT_DOMAIN || tranche_count || top) and total locked funds on-chain.
    • Individual tranche amounts, unlock Beats, and Merkle proofs (verified against exact path geometry to prevent CVE-2012-2459 tail malleability) are delivered off-chain encrypted inside HybridChannelSession.

3. Endogenous AMM Liquidation Bounds (Config D \to Config E) & Constitutional Governance

Two additional mechanisms address economic and governance attacks unique to agent ecosystems:

  1. Compute-Unit Collateral (Config E):
    Modeling GPU compute-lease liquidations with infinite OTC buyers or exogenous price drops masked catastrophic tail risk in our earlier 180% CR Config D design. Once our simulator (sim_ucc_endogenous.py) incorporated x \cdot y = k endogenous AMM price impact across 3 consecutive 15-minute liquidation windows, contracting bidder capacity under stress, and -60% / -80% / double-dip crashes, Config D suffered 15.62%–24.01% P99 net loss on S1–S3, 43.52% on -60%, 64.28% on -80%, and 100% abrupt lease termination.
    Config E replaces it with:

    • 15-minute rolling epoch settlement, bounding active exposure to a 2-hour rolling buffer (12.5% of notional).
    • Dual-Tranche coverage (75% stable + 75..105% native-token over-margin) or >= 220% CR in pure native token.
    • Dutch Vault (24–72 h) liquidation with finite bidder capacity (2.39%–8.17% discount) and 25% partial AMM inventory hedging (capping residual endogenous AMM price impact at <= 3.71%).
    • Guaranteed 45-minute Graceful Capsule Checkpoint, bringing abrupt agent termination down to 0.0% (63.6%–65.3% uninterrupted, 34.7%–36.4% clean checkpoint pause under -60% / -80%).
  2. Patient-Sybil-Resistant Governance (crates/isopolia-republic):

    • Stratified Cohort Sortition: accounts younger than 4 quarters are capped at 6/51 seats.
    • Active Third-Party Fee-Burn Quota (>= 15 native tokens/quarter) combined with funding-graph Min-Cut Sybil clustering.
    • Bounded Treasury & Exit Guarantees: <= 0.25% Class-1 spend per vote, 14–30 day timelock, Fork Exit right, and >= 6/11 human veto for constitutional rule changes.

4. Current Implementation & Verification Status

The repository already implements Phases I–VI Step 6.1 across 17 modular Rust crates (isopolia-primitives, isopolia-crypto, isopolia-vdf, isopolia-noyau0, isopolia-arbitrator, isopolia-vm, isopolia-chain, isopolia-storage, isopolia-mempool, isopolia-consensus, isopolia-ucc, isopolia-povs, isopolia-republic, isopolia-agent, isopolia-p2p, isopolia-node-bin, and isopolia-fastpath), including:

  • Zero external crates.io dependencies (verified by CI supply-chain gate), #![forbid(unsafe_code)], #![deny(missing_docs)], deterministic state (HashMap/HashSet/f32/f64 forbidden in consensus/state crates), and strict <= 400 lines per source module.
  • Cryptographic Known-Answer Tests (crates/isopolia-fastpath/tests/crypto_kat.rs) passing against NIST FIPS 202, RFC 7748, the full 256-byte FIPS 197 S-box, NIST SP 800-38D Test Cases 13–16, and bit-exact NIST FIPS 203 ML-KEM-768 digests.
  • Multi-node TCP Ed25519 devnet cluster orchestration, cold-sync verification, Ollama cognitive-receipt policy digest attestation, and deterministic fuzzing (-C overflow-checks=on in --release).

Next engineering milestones (Phase VI Steps 6.2–6.5 & Phase VII): threshold-encrypted mempool / sealed batch auctions, Crypte encrypted agent storage, Thymus sentinel circuit breakers, ML-DSA / STARK-beacon PQ upgrade paths, and deterministic RV64IM sandboxed plugins.


Open Questions for Community Feedback

  1. D4 Fencing vs. Causal Object Locks: Is an explicit quorum-certified SealFenceCertificate per account the cleanest safety boundary between FastPay and DAG-BFT, or are there lighter-weight causal object-versioning schemes (comparable to Sui/Mysticeti) that enforce strict mutual exclusion without a dedicated fence transaction?
  2. Post-Quantum Epoch Randomness Transition: Since lightweight standardized PQ threshold signatures / DKG replacements for BLS do not yet exist, we confine BLS to Phase 1–3 epoch randomness before migrating to a commit-reveal + STARK-verified VDF beacon in Phase 4. Are there cleaner intermediate constructions worth evaluating?
  3. Dutch Vault Liquidator Capacity Curves: Beyond x \cdot y = k partial AMM hedging (25%), what empirical stress curves for off-chain liquidator capital contraction have proven most realistic in Ethereum L2/DeFi simulations?