One environment variable keeps the forex bot's reasoning grader off the trading hot path

The check is one question per hourly decision. shadow.py posts a state string plus a single question named `shadow_direction_consistency`, typed `noul`, with instructions and explicit criteria: true when the recorded direction is explicit and the thesis supports it, false when the direction is ambiguous, contradicts the thesis, or was never recorded ^[12][13]. A `noul` answer is pass/fail against criteria the operator wrote ^[14]. The verdicts stay binary and countable, so thousands can be lined up against the trades they judged and the incoherence rate tracked over time ^[14].

JEV is TypeSafe's System-One model, `jev-latest`, served at api.typesafe.ai/v1/systemone, and it is text-only: no vision, no market opinions ^[11]. It cannot say whether UP was the right call on EURUSD. It can say when UP arrived attached to a thesis that argued DOWN ^[16]. "So it's a proofreader, not a fortune teller," wrote the developer, who publishes as nodefiend on dev.to ^[15][28].

The stated design rule is that the trading path must not know the harness exists unless it is told ^[18]. With `SHADOW_MODE` unset, shadow.py is a complete no-op, the only hot-path cost is a single `os.getenv()` call, and the tests assert that zero network calls happen ^[19]. When it is on, the harness dumps the exact chart images sent to the vision model into `shadow_out/<instance>/<timestamp>/` and records a sha256 of the prompt text for later lookup ^[20]. The one JEV call is hard-timeboxed at 20 seconds on a daemon thread, and the state is capped at 30,000 characters to stay inside the model's context ^[21]. Twenty seconds is about 0.6 percent of an hourly cycle ^[3]. One JSON line per decision goes to `shadow_log.jsonl` with the verdict and its latency ^[22]. `attach_outcome()` later patches the most recent unfinalized line for that pair with EXECUTED, FAILED or ORDER_FAILED plus the ticket, and it only touches lines written in the last 15 minutes, so a slow hourly cycle cannot corrupt an older decision's record ^[23].

A 2 percent daily loss cap sitting next to a 0.6 percent daily profit cap means a permitted losing day costs about 3.3 times what a permitted winning day banks ^[1]. Those numbers bound a day and the watchdog bounds a streak of four, but it was max drawdown that ended the first account ^[5].

The logging was already detailed. signal_log.csv holds every decision, executed or skipped, with the model's reasoning attached ^[7]. trade_analysis.csv holds every close with MFE/MAE in pips and R-multiples ^[8]. Neither file checks whether the thesis argues for the direction the model picked ^[9]. "When the account died, I could read every trade and every excuse, and none of it told me whether the reasoning was any good," the developer wrote ^[10].

Each instance writes to its own log, so the two accounts never mix records ^[24]. The developer wants a few market-open cycles of confidence before turning the collector on ^[27].