Skip to content

SPONSOR_DEFENSE.md

Latest commit

 

History

History
36 lines (25 loc) · 5.26 KB

SPONSOR_DEFENSE.md

File metadata and controls

36 lines (25 loc) · 5.26 KB
Raw

Why ONLY Pear + QVAC + WDK — The Treble (build edition)

The Treble needs all three stacks simultaneously: a trustless table (Pear), self-custody for every player including a machine (WDK), and an opponent that reasons with no cloud (QVAC). Below: every load-bearing API, where it is used in this codebase, and what replacing it would cost.

Pear (P2P state)

  1. Autobase 7 — new Autobase(store, bootstrap, { open, apply }), host.addWriter, base.append/update/view → src/p2p/pot.js. The multi-writer pot with no server; the AI is just another writer core. Without it: a coordination server + a CRDT you'd have to write and prove.
  2. Hypercore (via Corestore) — per-writer signed append-only logs → authority in reduce(state, op, { from: node.from.key }) (src/p2p/pot.js → src/core/reducer.js). Picks become tamper-evident by construction. Without it: a trusted timestamp server and signature plumbing.
  3. Hyperbee — the deterministic materialized view (state, event/<seq>) with valueEncoding: json → src/p2p/pot.js. Without it: hand-rolled snapshot/undo logic that Autobase reordering would break.
  4. Hyperswarm — swarm.join(discoveryKey); the treble1… invite IS the room → src/p2p/pot.js, src/p2p/invite.js. Noise-encrypted transport for free. Without it: signaling servers + TLS + NAT pain.
  5. Protomux + compact-encoding — the treble/1/hello seat-request channel riding the replication stream → src/p2p/pot.js. Without it: a second connection layer just to ask "may I sit down?".

WDK (money + guardrails)

  1. new WDK(seed).registerWallet(name, Manager, config) — the modular manager pipeline → src/wallet/index.js. Our sim engine is a custom chain module built on @tetherto/wdk-wallet's WalletManager/IWalletAccount base classes (src/wallet/sim-wallet.js) — exactly the extension point WDK advertises — and @tetherto/wdk-wallet-solana drops in for devnet with the same call sites.
  2. Transaction Policies — wdk.registerPolicy({ scope: 'project', wallet, rules }), default-deny on governed accounts, PolicyViolationError, account.simulate.transfer(...) → src/wallet/index.js, src/agent/seat.js, test/wallet.test.js. This is the whole "bounded autonomy" story: the agent pre-flights its own allowance, declines pots above its cap, and cannot even sign() outside its ALLOW rule. Without it: home-made limits the judges would have to trust.
  3. account.transfer({ token, recipient, amount }) / getAddress / getBalance — stake bonding and every settlement leg with per-transfer receipts → src/wallet/index.js (stakeBond, executeSettlementLegs). Without it: raw chain SDKs per network, one custody bug away from disaster.
  4. BIP-39 utilities — WDK.getRandomSeedPhrase() for fresh self-custodial identities per participant → src/wallet/index.js. Without it: key-management code nobody should improvise.

QVAC (the mind)

  1. completion({ modelId, history, stream, tools }) with tool-calling — the pundit must commit via the submit_pick tool (type:'function', JSON-schema params); we drain run.tokenStream for narration and consume the dedicated run.toolCallStream (falling back to run.toolCalls) for the parsed call → src/agent/brains/qvac.js. The pick is a decision, not parsed prose. Without it: brittle regex over free text, or a cloud API that breaks the entire premise.
  2. loadModel({ modelSrc }) / unloadModel — Qwen3 1.7B on-device (default chosen for empirically-verified tool-calling; Llama 1B-class models tended to narrate instead of calling); TREBLE_QVAC_MODEL accepts a local GGUF or a pear:// link → src/agent/brains/qvac.js. npm run verify:offline proves the reasoning path needs zero network. Without it: "AI opponent" = someone else's datacenter.

Take any one stack out and The Treble is impossible: without Pear you need a server (no longer trustless); without WDK the "AI with its own wallet" is a custodian in a trench coat; without QVAC the opponent is a cloud API (neither private nor yours, and against the QVAC track rule).

Honest limitations of the sponsor stacks (as encountered)

  • @tetherto/wdk is beta — pinned to 1.0.0-beta.12. A policy DENY that merely fails to match an ALLOW reports governed-but-unmatched without policy metadata; we add an explicit DENY rule so blocked stakes carry policyId/ruleName. Chain modules (e.g. Solana) install separately; devnet funding is on the operator.
  • @qvac/sdk model weight — the default LLM is a multi-hundred-MB first-run download with platform-specific native backends, so CI and the 20-second demo use a disclosed deterministic fallback brain; the SDK surface itself (tool schema, event stream) is exercised in tests without loading weights. On Bare, plugins must be registered explicitly (@qvac/bare-sdk) — our desktop UI keeps the LLM brain on the Node path for now.
  • Autobase concurrency semantics are powerful but subtle: causally concurrent ops linearize deterministically yet not always intuitively (see AUDIT_REPORT "pick/lock window"), and replication acks need both peers to update(). The reducer is designed so every such ordering stays safe.

Full threat model and residual risk: docs/AUDIT_REPORT.md.