Research and Development of FinTech AI / AI Agents / Weekly recap for the week of September 27 to October 3, 2026

Manage next week's verifications by “completion criteria” — Unresolved items for GOLD13, EA, and AI agents

The weekend recap is not a summary to inflate成果. It is a time to decide what to verify next week to convert candidates into verified items. This week there are log displays for GOLD13, EA update candidates, and weekly domestic account tallies. However, confirming whether all three refer to the same trade, which version operated in production, and whether the video recording matches the order history requires additional evidence. This article整理 the unresolved items and pass/fail criteria for the research and development of financial AI, AI agents, self-made indicators, and automated trading software (EA) based on primary materials available as of October 4.

We validate our self-made indicator GOLD13 and the EA on a real account and share the progress via a weekday 24-hour trade livestream. However, the livestream screen is not the broker statement nor the AI approval record. Even if GOLD’s price movement appears in the video, it cannot automatically determine which EA version was loaded at that moment or which ticket it linked to. For readers of the published articles, we clearly separate verified facts from traces not yet connected.

The first hurdle is to trace one natural order from end to end

Next week, the first natural order will be considered passed if signal ID, ACK received by the EA, order response, ticket, positions, settlement result, and ledger can be uniquely linked. Merely having a SELL display and a ticket in nearby time is not enough to conclude it’s the same order, due to branches such as multiple candidates appearing in the same time window, order postponement, broker rejection, or delayed communication. We will cross-check symbol, direction, lot size, magic number, and timestamp basis, and leave any mismatches as not supported.

The deliverable from this hurdle is not “one line of profit,” but a traceable mapping that anyone can follow. If there is no record at a stage, display it as unknown rather than blank. If ACK is missing, do not state that the AI received it; if a ticket is missing, do not say it has been ordered. FinAI can collect and reconcile大量 logs, but it’s a quality condition not to auto-fill vague responses. Orders that do not pass will be excluded from the population for win rate or EA performance comparisons.

Next hurdle is cross-checking with natural trading media

In editor’s primary materials, 47 weekly items correspond to unconfirmed natural trading videos, screenshots, and trade_meta in Drive. Until these gaps are filled, the weekly totals cannot be visually verified for entry and exit of each trade. For next week’s natural orders, we will save the trade_full.mp4 corresponding to the same ticket, short clips of entry and exit, screens showing surrounding context, and trade_meta.json, and verify not only filenames but also contents and timestamps. We will not repurpose footage of other trades with similar prices for the article.

Even with videos, that does not complete proof of trade results. We will check whether the accounts, symbol, timestamps, tickets, and filled/settled prices shown on screen match the history. If the footage cuts around the order, we will注明 the cut portions. Orders without recording will remain “no recording”, and only materials with verifiable sources will be embedded in GogoJungle video features. Weekday 24-hour livestreams are for reference in verification, but we will not treat cut-out videos’ sources or their corresponding tickets as real trading recordings in a given article unless their origins are verifiable.

Don’t leap GOLD13’s display to the order result

The local log on October 3 shows examples like “GOLD13 REALTIME SELL” and “warn=0.” These are display investigations and do not prove unique signal IDs, AI receipt, order tickets, or P/L. Next week we will record distinctions between normal signals and warning signals, occurrence times, target symbols, prices, and suppression reasons in the same format. The traditional policy of not attaching to warning signals but evaluating ordinary signals after the warning remains unchanged. We will separate discovery of display examples from evaluation of the EA’s trading candidates.

Research on self-made indicators must not merely count increased signals as success. We need logs that reveal false triggers, duplicates, price update losses, and mismatches between display and AI receipt. Self-made EA research likewise should record not only why an order was placed, but why it was passed, stopped, or hit a ceiling. FinAI and AI agents are useful for extracting such gaps or inconsistencies, but even if AI’s explanatory text appears convincing, proofs require re-confirming the referenced signal and ticket.

Separate saved files from production-ready versions

This week’s materials record the EA’s local file updates, EX4 SHA changes, and the GOLD13-related stage folder. However, the purposes of file changes, diffs, compile paths, VPS deployment, MT4 restart, and actual loading of the版本 are unconfirmed. In another October 4 morning handover, production VPS practical EA version, GOLD13 version, AI judgment settings, and AutoTrading are marked as unknown. Merely noting that new files exist should not rewrite as “production improvement version is running” — this remains the second important boundary for next week.

To bring candidate changes into production, we must have readback backups of the pre-change files, confirm diffs and generation paths, compile results, SHA comparisons between local and VPS, deployment logs, necessary restarts, and readbacks of the production version. Additionally, we will verify the AutoTrading status and the first natural order result to determine whether they operated with the intended settings. If any single item is missing along the way, we will explicitly state whether it is in “candidate,” “saved,” “stage,” “deployed,” or “operation-confirmed,” and will not skip to completion. In automated trading, this plain history is the foundation of safety.

AI-led approach should not confuse design with the current method

In the most recent handover, the current method is organized as EA or indicator proposing candidates, AI approving, and EA automatically executing actual orders. A separate approach where AI leads from candidate discovery to new/skip decisions, holding management, and settlement is a topic for consideration. Simply introducing a particular local model does not establish that method; performance, MT4 connectivity, always-on operation, and safety constraint design are not yet decided. This conversation did not implement model deployment, EA integration code changes, VPS deployment, or real account testing. We do not present research concepts as currently active features.

If we consider an AI-led approach next week, first run parallel recordings with the current method under the same market to separately compare candidate discovery, order feasibility, holding management, and settlement decisions. Next, verify the entire order path on a demo account, ensure AI cannot alter maximum loss, lot cap, or stop during communication failures, and then decide on connecting to a real account in a subsequent independent stage. The priority is to create an isolated environment that does not break the current system for comparison, and it is also important not to rely on others’ short-term P/L images seen on social media as proof of performance.

Make next week’s completion judgment clearer than numbers

Weekly materials show 47 items, 31 wins, 16 losses, gross 2967, commission -915, swap -289, net +1763, but currency units and broker statements are uncross-checked, and results for October 3 are unknown. When reusing numbers, this caveat should remain. Next week’s completion judgment should confirm not only whether trading was positive, but also that at least one natural order can be traced from pre-order to post-settlement without gaps, that media and history for the same ticket match, and that the production version and settings can be read back. If these are not all present, that item will be carried over to the next verification as incomplete.

The research and development of self-made indicators, EA, FinAI, and AI agents becomes easier to improve as accurate records accumulate. Even in the weekday 24-hour livestream, we will not merge candidate, operating, orders, and settlements into one story, and will present evidence for each. Details of the tools can be found atGogoJungle sales pagewhich will be announced. This article documents research progress and unresolved items and does not guarantee profits or future performance. There is a risk of principal loss in FX and GOLD.