Jev trading bot (TypeSafe AI): does a Jev failure still place the order?

Trading and fraud gates read in source. "jev trading bot" was a Google suggestion (client=firefox) for the words "jev trading" when the pool checked on 4 Oct 2026, 05:21 UTC. Bing and DuckDuckGo offered "jev trader". "jev fraud", "jev fraud detection" and "jev finance" got no relevant suggestion on any endpoint checked, so "fraud" is kept out of the main heading. A suggestion shows that people search for something; it does not show how many. This page reads seven public projects at pinned commits: four trading bots and three fraud or payment-approval gates. For each it shows the default mode, what switches it to real orders, and whether an order or payment still goes ahead when Jev errors, gives a bad or unsure answer, has no key, or is skipped. The pool ran none of them: no bot, no exchange or payment call, no Jev call.

What the seven projects add up to

Seven projects: does the failure reach the order or the payment?

Open a situation to highlight its column. Opening one closes the others. The last column says, for each situation, whether the failure still reaches an order or payment call site. Failure terms use the pool's shared vocabulary: fail-closed (stop, refuse or take the cautious outcome), fail-open (act anyway), fallback-<what> (a named substitute), advisory, no-decision; below a threshold: acts-anyway, held, no-threshold. Documented source at the commit shown, read 4 Oct 2026, 05:14–05:21 UTC

• fail-closed or held (a person or the host's own prompt decides)
• fail-open or acts-anyway; a "conditional" cell that can act is counted fail-open
• fallback-<what> (named in the cell)
• not recorded, not applicable or no-threshold (the cell says which)

Colours follow the shared legend on when Jev fails (changed 6 Oct 2026; this page used its own colours before). A fallback that can still place the order is blue with "then fail-open" in the cell. Bypass cells: red means the order or payment goes ahead without Jev; green means it is held, reviewed or rejected.

Last column: yes the failure still reaches an order call site; no it does not; conditional it can, depending on a fallback; not recorded. Line numbers refer to the file at the pinned commit (paths in sources). How to test these branches without a key is on testing without a key; status codes and retries are on errors and rate limits; how to choose a threshold is on confidence thresholds. To compare these rows with other use cases, see when Jev fails. One row per project is on the products ledger.

QuantDinger: where the gate sits before the order

QuantDinger is a trading platform with an optional Jev entry gate in backend_api_python/app/services/ai_decision_filter.py. The pool read it in the 28 Sep audit from moving main links. The file has since been rewritten: it is 743 lines at a5a9f4c, while the 28 Sep links pointed at lines up to L2279. Those earlier line links are stale. This page uses the a5a9f4c pins. The behaviour summarised on 28 Sep is confirmed at this commit. Documented read by the runner, 4 Oct 2026, 05:17–05:20 UTC

Three order call sites, each traced

The gate (AIDecisionFilter.evaluate()) is called at three places, each before the order is created. An allowed=False result stops the order at all three. An allowed=True result, including every failure path in the table, leads straight to order creation.

1. Strategy v2 orders (strategy_v2/live_execution.py L233–273). The gate runs only when execution_mode is live and the strategy has ai_decision_filter set. A rejection sets the intent to ai_rejected and returns before INSERT INTO pending_orders (L308). The pending-order worker then sends these rows "through exchange and broker clients" (pending_order_worker.py L1–6).
2. Quick trade (routes/quick_trade.py L564–567, calling quick_trade_ai.py L14–68). The rejection returns before place_order_from_signal or client.place_limit_order (L575–595). Opt-in per request.
3. Alpaca route (routes/alpaca.py L453–479). Returns before place_limit_order or place_market_order (L492, L497). Opt-in per request. Alpaca is paper when the key starts with PK (L51, L72–75).

The 28 Sep claim that "rejected entries never reach the exchange" came from the README. It is now traced in source.

What the gate checks and what it lets through

• Off by defaultThe filter runs only if the strategy or request sets ai_decision_filter (trading_executor.py L1723; quick_trade_ai.py L19). Strategy orders default to execution_mode "signal" (L1763), which fills "the isolated virtual account".
• Five questions, three ways to blockChoices on data_quality, signal_alignment, market_regime, risk_check and execution_quality (L24–92). Only a risk "block", an execution "block", or both a signal conflict and an adverse regime at or above the minimum confidence reject the order (L299–309).
• Unsure answers do not blockAn insufficient answer on any check does not block, and data_quality is never read by the allow expression (L305–309). Minimum confidence defaults to 0.55 and is an operator setting (L285, L517).
• Errors end in "allow"On an exception the log says "Jev decision failed open" (L194–204); a configured LLM is tried (L206–223); if there is none, the order is allowed as error_allowed (L225–238). Timeout JEV_TIMEOUT_SECONDS: default 8 s, clamped to 1–30 s (L250, L516).
• Paths that skip JevExits (actions outside ENTRY_ACTIONS, L141–146); grid, DCA, martingale and layered martingale strategies (L147–150); QuantDinger's own credit check failing, which allows the order with no Jev call (L177–191).
• Audit does not blockEach decision is written to qd_ai_decisions. If the write fails, the log says "AI decision audit persistence skipped" and the order is not blocked (L622–700). No per-order human step was found.

What it sends to Jev: symbol, action, market and order type, quantity, reference price, notional, leverage, strategy type, signal reason and the strategy's context dict (L394–410), with no truncation in code. The repository has 17 tests for the gate, all mocked (tests/test_ai_decision_filter.py), including malformed answer → LLM and credit denial → order not blocked.

Not traced: the pending-order worker upgrades a queued signal row to live when the strategy is configured that way (pending_order_worker.py L1413–1420), while the strategy-path gate only runs for live requests. This may be another way past the gate. It was not followed end to end.

Fraud and payment-approval gates: real Jev calls, no money moved

All 3 gates have a real Jev call path. None of the 3 has a payment, authorisation or reimbursement call site, so for every situation in the table the answer to "does the failure reach the payment?" is no. Documented 4 Oct 2026

• jev-fraud-shield (eeb4fc6)Runnable demo on synthetic card transactions. Rules approve a transaction with 0 flags and decline one with 3 or more without asking Jev. A Jev error, a malformed answer or a below-threshold answer all lead to STEP_UP. Without a key it uses a local mock. The result is a dict that demo.py prints.
• fraud-jev (aa33b97)Runnable Next.js demo with fictional payments; the key is pasted in the UI. A Jev error returns REVIEW with HTTP 500; no key returns HTTP 401. A non-numeric noul returns ALLOW, and scores below the 0.65 / 0.75 / 0.85 cuts are allowed. OFAC destinations go to REVIEW without Jev. The answer is JSON to the UI.
• Expense-policy gate (161c33b)Expense approval with Jev by default. Fixed rules reject on currency, amount, receipt, duplicates and dates without Jev; nothing is approved without Jev. A Jev error or a malformed answer aborts the batch; a below-threshold answer goes to manual_review. Output is a JSONL decisions file; there is no reimbursement step.

How the seven were chosen and read

Picked from projects named in earlier pool research, two GitHub searches and this run's 24-hour scan; not a survey. In detail: the two trading and two fraud repositories named in the pool's 28 Sep audience findings, QuantDinger from the 28 Sep audit, two GitHub repository searches at 05:15 UTC ("jev fraud": 15 results; "jev payment": 10 results), and two projects created in the 24 hours before this run and found in its scan (btc-hft-jev and the expense-policy gate). A search count is a count of indexed repositories, not of users. TradeRank, named earlier, was not found in the searches or in two Jev project lists (cobanov/awesome-jev @72ee9f4, everyinfra/jev-radar @41c2547).

Each project was pinned to one commit and read through raw files at that commit. The runner read QuantDinger (05:17–05:20 UTC). A helper agent read the other six (05:14–05:21 UTC); the runner spot-checked 9 lines in 5 of those repositories at the same commits at 05:21:48 UTC, and all matched. No account, key, exchange, broker, wallet, testnet or payment-provider call was made; no bot was run in any mode; nothing was installed. Nothing on this page is Tested.