Build1 publisher2 min readPublished
One risk word in the reasoning trace cut a live SOPHUSDT order to 0.8 times size
The F-072 rule engine scans an LLM's chain of thought for risk words and multiplies position size by 0.8 when it finds one. On the trade its author logged, both constants it applied were already in the model's own reasoning.
The Engineer · Build desk
What happened
- At 01:17:36 on 10 September 2026, during a crypto flash crash, the advisor gate scored a SOPHUSDT long at 73.5 and returned RULING=PROCEED at confidence 0.60.
- Seventy-eight seconds later a council review logged the same swing trade as CAUTION with a delta of -5, while its own PROCEED component still read pass.
- The F-072 layer then matched the Chinese token for risk inside the model's reasoning text and logged a warning that it was auto-tightening a PROCEED ruling.
- The override multiplies base position size by 0.8, cuts the calculated stop-loss distance by 10 percent, and appends a hard-coded tag to the payload before the order reaches the exchange API.
- The final ruling still read PROCEED, logged at delta -3 and confidence 0.62 with the risk-word tightening tag attached.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Anyone parsing only an agent's structured field now has a worked example of the alternative, and adopting it makes the reasoning text a control input that has to be versioned like configuration.
- constraint The trigger is a fixed token list, so coverage is a property of the prompt language and the model's phrasing habits rather than of the risk being described.
- capability A model boxed into a binary ruling field gets a channel for degree: the hedge it writes in prose comes out the other end as a smaller order.
- contradiction The post frames the mechanism as overriding a buy signal, while the log shows the ruling stayed PROCEED and only the sizing and stop moved.
F-072 skips the JSON and intercepts the whole model output, including the raw chain of thought and the reasoning strings, then runs a regex and keyword scan over the text [5]. The token list the post gives mixes languages: the Chinese words for risk and for catching a knife, alongside the English "hesitation" and "volatile" [6]. A hit leaves the ruling alone and mutates the execution parameters before the order reaches the exchange API [7].
Where the matched token sits matters. The final ruling's reason field contains the clause the author renders as "scale down 0.8 + tighten stop-loss 10% to control the risk of catching a falling knife" [8]. The scanner's trigger word is inside that clause [15]. The two constants the engine applied, 0.8 and 10 percent, are the two constants in the sentence it scanned [16].
The post's rationale is that the structured output forced a binary PROCEED or REJECT, so the model compromised and put its hedge in the prose [19]. That part is a real design problem, and the author is direct about the alternative. "If I had only parsed the final structured output, the system would have executed a full-size market buy," the author wrote [9].
The interception is slow. The gate line is stamped 01:17:36.802 and the F-072 warning 01:19:08.944, which is 92 seconds [13]. Part of that is upstream: the post notes API responses running 27 to 31 seconds that night [11]. The reasoning the scanner read described a counter-trend entry on a day down 15.47 percent with 226M of liquidity [10].
The tightening is also smaller than the language around it. Multiply size by 0.8, pull the stop 10 percent nearer, and the worst case at the stop is 0.72 of the untightened trade, 28 percent less [14]. The nearer stop is reached by a smaller adverse move, so that loss is taken more often.
Nothing described handles negation. A trace saying the risk is low contains the same token as a trace saying the trade is risky, and what reads the trace is a keyword matcher [5]. Switch the model or write the prompt in English and the Chinese entries in the list stop firing [6]. The evidence in the post is one trade on one night [18].
What to watch
- Whether the author reports the outcome of the tightened SOPHUSDT long, including whether the 10 percent closer stop was hit.
- Whether F-072 gains a symmetric path that loosens parameters when the reasoning trace reads confident.
- A published rate at which the scan fires across many traces would show whether 0.8x is the exception or the effective default size.