how interlock works

Interlock is an AI trading agent for tokenized stocks on Solana with one unusual property: it cannot decide anything. The model's output is a request. The decision is made by lib/policy.ts — a pure function with no clock, no network, and no model inside it.

That separation is the entire safety story. A model can be confident and wrong, or steered by whatever it just read. None of that reaches your money if the thing holding the trigger is arithmetic.

1. 1You set six numbers — the policy.
2. 2The agent reads live prices for the symbols you allow and writes one proposal with a thesis.
3. 3The engine runs every rule and stamps ALLOW or REFUSE, naming the line that decided it.
4. 4You tighten a rule and the same proposal is re-judged instantly — in your browser, by the same function.

verdict = f(proposal, policy, spentTodayUsd)

symbol      ∈ policy.allowedSymbols     SPYx, QQQx, GLDx
side        ∈ policy.allowedSides       buy
0 < size    ≤ policy.maxSizeUsd         $500
confidence  ≥ policy.minConfidence      0.62
thesis.len  ≥ policy.minThesisChars     180
spent+size  ≤ policy.dailyBudgetUsd     $2000

• ·Symbols are an allowlist, never a blocklist. A ticker that isn't named is refused — including ones that didn't exist when you wrote the rule.
• ·The budget counts today's spend. It's passed in, not read from a clock, so the same inputs always give the same verdict.
• ·Every failing rule is reported — not just the first one.

A proposal that passes by sitting exactly on a limit is stamped EDGE. It's still an ALLOW — you authorized it — but a rule you only just cleared is a rule worth re-reading. The engine also exposes the weakest proposal your policy would allow: max size, minimum confidence, minimum thesis. Not a suggestion — a warning about the floor you've set.

A second model is told your policy and asked for the worst trades it still permits. Its attempts go through the same engine, in public. Whatever gets through is not a flaw in the engine — it's the floor of what you allowed. The fix is always to tighten the rules, never to argue with the model.

All endpoints return JSON. The verdict is the same function the UI runs.

{
  "proposal": { "symbol": "SPYx", "side": "buy",
                "sizeUsd": 200, "confidence": 0.8,
                "thesis": "…" },
  "policy": { "maxSizeUsd": 500 },
  "spentTodayUsd": 0
}
→ { "verdict": { "decision": "ALLOW", "checks": [...],
                 "failed": [], "edges": [], "reason": "…" } }

{ "policy": { "allowedSymbols": ["SPYx"] } }
→ { "proposal": {...}, "verdict": {...}, "market": [...] }

{ "policy": { "minConfidence": 0.5 } }
→ { "floor": {...}, "attempts": [...],
    "gotThrough": 2, "refused": 1 }

→ { "stocks": [ { "symbol": "SPYx", "name": "S&P 500",
       "mint": "Xso…", "priceUsd": 773.83 } ], … }

• ·Nothing is executed. Interlock decides; it does not trade for you. When execution ships, it ships behind these same rules.
• ·Non-custodial. Connecting a wallet only reveals your address.
• ·Real or zero. Counters that need real signed verdicts stay at zero until they exist.
• ·One official address. $ILOCK's contract appears only on the token page and @interlocklat on X — anywhere else is a fake, however early it appeared.