Financial Instruments Business Operator Kanto Finance Bureau Director (Kinsho) No. 1960 / Membership Association Investment Trusts Association, Japan Membership No. 012-02323 Go to GogoJungle Home
icon List
柴亮太
09/22/2026 21:33
Release： 09/22/2026 21:33
Update： 09/22/2026 21:33
柴亮太
Investment style : Trading System
Investment career : More than 10 years
Investment amount : 10 million - 50 million JPY
Representative of RO LLC. My main line of work is AI agent research and development. I have about 25 years of investing experience, but it’s not a brag-worthy track record. I am not the type to sit still in front of a chart, and I would often cut losses when I looked away during critical moments. Through that repeated pattern, I have melted tens of millions of yen. What remains is knowledge and a record of my failures. “If I can’t watch it myself, I’ll build something that can watch for me.” With that in mind, I am also working on developing AI agent EAs. Not to predict the market, but to tell me when not to enter a position. To avoid silently expanding losses when prices move against me. Listing on GoGoJungle is a testing ground for that research and development. The tools I publish here are, as of now, completed products, but not the final finished version. They are things that run on actual accounts, record losses, and are cultivated while being improved each time. What didn’t go well is also shown as is. Therefore, for those who purchase, it isn’t a “buy and end” relationship, but a partnership to accompany the research. It isn’t that you are not winning because of your efforts. I have faced the same points and repeated the same losses many times. Those who stumbled in the same place, this time I hope we can turn to the winning side together and continue with verification. If you have questions or requests, please don’t hesitate to share. The feedback we receive will be reflected in the next version.
Article cover image

Decompose the reverse-hedging hard gate added from a loss of -3,016 yen: ticket 29764139

Financial AI / AI Agent Research and Development, Real Account Loss Verification

Decomposing the added reverse-trend hard gate: ticket 29764139 from a loss of -3,016 yen

This is a record of turning entry points that could not be stopped by AI decisions alone into reproducible machine rules, without hiding the loss on the real account.

This article is a development log of the indicator I developed and the automated trading software (EA), researched and validated with AI agents and real accounts. Not only successes but losses, passings, and obstacles are recorded in the same format. The validation process is also published in a 24-hour weekday trading live stream.

First, disclose the facts

  • ticket: 29764139
  • Subject: GOLD_USD
  • Fill price: 4341.17
  • Initial hard SL: 4331.63
  • Realized P&L: -3,016 yen
  • AI judgment: REVERSAL_ATTEMPT, confidence 65
  • Deployment status: COUNTERTREND_REVERSAL_CONFIRM_V1 already reflected in production
  • LIVE_VALIDATION under natural market conditions: unconfirmed

If you only look at the amount of loss and conclude that “AI missed it,” you’ll pass through the same entry again next time. What matters is breaking down which timeframe showed what, what information was passed to the AI, and which guard of the EA was lacking.

Structure existing at the entry

At the target moment, PRIMARY day and H4 were候補BUY direction COUNTER_TREND. In SECONDARY H1, a reversal could not be confirmed, and only M5 ENTRY showed REVERSAL_BREAK. A combination that easily leads to misreading short-term changes as a full reversal on higher timeframes.

Gemini treated this state as REVERSAL_ATTEMPT and passed it to enter. The explanation was acceptable, but there was insufficient evidence to proceed to external action on the real account. Therefore, we separated “the model says it will enter” from “order permission.”

Added machine rules

The new COUNTERTREND_REVERSAL_CONFIRM_V1 blocks entry when the following conditions are all met simultaneously.

  1. Gemini returns enter
  2. Judgment is REVERSAL_ATTEMPT
  3. Confirmed Dow structural on daily/H4 is COUNTER_TREND in the candidate direction
  4. Neither WITH_TREND nor REVERSAL_BREAK on H1 or M15

Under this condition, the signal is not advanced to an order, and the reason for bypass is logged. Conversely, for trend-following, Gemini skip, REVERSAL_CONFIRMED, and intermediate-timeframe reversal-confirmed candidates, the previous behavior remains. It’s not about eliminating AI flexibility, but deterministically stopping only the weak-evidence reversal entries.

Ranges not changed

If you touch the entire system immediately after a loss, you cannot trace causes and effects. This time, SL, TP, lot size, HOLD exit, breakeven shift, trailing from 50% to 75%, GOLD13, and warning guard were not changed. Only the lack of entry confirmation was fixed with minimal differences.

We saved pre- and post-change SHAs and backups, and ran py_compile on both the VM production Python environment and an isolated environment. In tests reproducing the real input of the target ticket, the result was BLOCKED; skip, CONTINUATION, and REVERSAL_CONFIRMED passed as before.

Do not exaggerate test results

Low-frequency tests 4/4 and spread tests 6/6 matched before and after the change. Existing validations were 31/41 before and after; the differences are not regressions caused by the change but mismatches between the existing RELAX settings and the old expectations. We will not claim “all tests pass” for convenience.

In the pre-deployment heartbeat, open_count=0, unsent signal=0; after receiver restart and cron, exit=0, GOLD_USD no_breakout, held 0. However, the same combination reappears in natural markets, and we have not yet observed a scenario where entry is actually blocked.

AI and EA responsibility boundary

AI agents summarize multiple timeframes, articulating the meaning and uncertainty of candidates. EA and surrounding guards deterministically examine account state, data freshness, spreads, duplication, and higher-timeframe reversals. Even if explanations are natural, if order conditions are insufficient, it is designed to stop.

A model confidence of 65 does not mean a win rate of 65% or safety of 65%. It is simply a self-assessment of the given context, not to be used alone as a real account’s permission condition. Especially for reversals, higher-timeframe reversal verification is required as separate evidence.

Items to confirm with the next natural signal

  • Confirmed pivot columns for PRIMARY, SECONDARY, ENTRY
  • Gemini action, setup_type, basis
  • Reason for hard gate judgment and timestamp
  • Confirmation of no signal generation and no order submission
  • Maximum favorable and maximum adverse outcomes after blockage
  • Both opportunity loss and loss avoidance

Even if a blocked candidate later expands, we will not hide it. Pushing too far to the safe side may discard valid early movements. We collect a certain number of items in the same format and compare loss avoidance with opportunity loss before adjusting the threshold.

Criteria for using recordings

Embedding videos in articles is allowed only when sources with matching ticket, target account, symbol, and timestamp exist and can be independently verified. This time, we could verify entry conditions and losses from the Drive records, but until we confirm ticket matches in the article’s video library, we will not reuse footage from separate trades.

Research conclusion

-3,016 yen is not a good result. However, we identified a recurrence condition with higher-timeframe reversals and unconfirmed intermediate-timeframe, reproduced the target input, and converted it into a minimal-difference guard. We will publish without reframing the loss as a success and while leaving unconfirmed LIVE_VALIDATION in place.

Going forward, we will track not only profits and losses but also whether we stopped the entry that should be stopped, whether we can reproduce the reason for stopping, and whether there is any double-action due to restarts or resends, all with equal importance. To cultivate financial AI in real accounts, the verification density on losing days is most important.

Observational order fixed for loss verification

In future loss articles, we will fix the order as result, entry structure, AI output, machine guard, order response, position management, settlement, and post-settlement state. Changing the order each time invites convenient explanations that emphasize entry on winning days and only settlement on losing days. Fixing the observed items first places success and failure on the same yardstick.

Also, we will run the same fixture before and after code changes and verify that not only the target ticket but also trend-following, skip, and mid-timeframe-confirmed candidates have not changed. Running only a single reproduction test to production can risk breaking another normal path.

Price deviation guard remains a separate issue

This revision is limited to verifying reversal on higher timeframes. The question of how much deviation from the signal price triggers a stop when MT4 market orders fill will be addressed after comparing the impact on the current EA master and winning trades. We will implement multiple safety measures at once and avoid changes that obscure which ones affect profit, order rate, or execution.

When considering deviation guards, record the signal generation time, EA receipt time, order submission time, fill price, spread, slippage, and any market upheavals under the same ticket. Look at not only the average but also the upper percentile and maximum values to determine thresholds that do not misfire under normal fluctuations.

Policy for post-publication corrections

If later we find misread logs, ticket handling errors, or numerical differences, we will not silently replace the text. We will leave the corrected sections, rationale, and modification timestamps, and make the decision basis traceable. Articles on AI research and development will themselves be auditable records.

GOLD13 Sales Page
Weekday 24-hour Trading Live Stream

This article is a research and development and real-account verification log and does not guarantee profits or future performance.