Steps marked API can spend Claude credits: Run shows what they do and the estimate, and waits for your OK. CPU steps cost nothing. Pause and Stop halt at once; Run resumes where it left off.
Problems
Log
Blind labels
Trials (out-of-sample, outcomes)
Baselines and held-out
Blind labels from state/blind_relabel.jsonl. Click a trade id for its entry chart, or review the filtered trades one by one. Manual decisions go to state/manual_review.jsonl (log-only: they do not change labels or training). Sealed held-out trades are not listed.
Flag a live or paper trade the loop does not have (e.g. one you pointed out elsewhere)
Decide before revealing: the record notes whether the outcome had been revealed first. Saving reveals the outcome. Keys: ← / → previous / next.
Strategy videos
Paste a PUBLIC YouTube link (unlisted/private are not supported) and its length. Gemini watches the picture and listens to you together; long videos are done in 40-minute parts. Then use the buttons below (separate from the loop's Run tab; both can run at once). The rule book appears in Reports as strategy_knowledge.md.
Or one link with a length you type yourself:
Analysis
Trade result
The recordings stop at the entry, so add how the trade ended. Your result replaces the video's "unknown" and is what the loop uses.
A vision model reads the charts at each 5-minute checkpoint (09:35-10:30) and trades the way the videos teach; the same days are compared with the XGBoost gate and the first setup of the day, all under live rules. Runs start from the terminal (they spend Gemini/Claude credits). Click a day to replay it below.
Trade studies: your recorded trades
Each recorded trade's morning rebuilt from the 1m data, cut at your entry and marked by the system's detectors, with the logic of the decision written next to it (Videos → Analysis → 3. Write trade studies). The vision test uses these studies as its examples. A trade without 1m data for its date cannot be studied.
His recorded mornings, from London / pre-market to the NY entry, with what he SAID marked at the ET minute he said it (statements hand-coded from the YouTube captions and lined up with the bars: datalake/analysis/timelines). Click a statement or caption line to move the chart cursor there. Step through replays the morning minute by minute: only the bars up to the cursor minute and the statements made by then. Read-only; nothing here calls an API.
Each item is what the extraction says he did or thought, in a paragraph backed by his quotes. Open the video at the quote, then confirm, correct (write what is right) or mark it wrong. Add anything it missed at the bottom.
The full-session analysis of a learning video, grounded in his words, the HD stills and the exact chart. "His words" are quotes; the inferred part is the AI filling in what he left unsaid. Answer its questions, comment on any decision, then press "Follow-up round" so the next pass revises with your answers.
On these days the blind reader got his direction wrong (top) or right. On most wrong days he traded WITH the overnight move (continuation) while the reader expected a sweep and reversal. For each day: was it a reversal or a continuation morning for him, what decided it, and why in your words. Saved to loop/state/direction_labels.jsonl.
Every trade the blind system took in the chosen run: one trade a day plus one re-entry after a stop, entries timed by code on the reader's side, target at the first untaken session/day/week level at least 1.5R away (else 2R), realistic fills (stop at a tick of slippage, target only on a trade-through, flat at 11:30). Chart: black = entry, red = stop, green = target (from entry to exit); his trades are blue arrows; faint blue rays = the indicator map at 09:20. Give each trade a verdict and a note; saved to loop/state/trade_reviews.jsonl.
His pre-open map (orange: the lines read from his stills before 09:30, mostly his hand-drawn lines; his indicator's lines have no axis label so they are not in it) next to the indicator map computed from price data: blue = session / previous day / previous week levels, green = equal and relatively equal levels (1m/5m/15m near price, 1H/4H further, daily/weekly), grey = 4H highs/lows, the nearest untaken 5m/15m high and low, and 1H highs/lows when no special level is near, purple = 1H/4H fair value gaps price is inside, or untouched nearby; dashed = already taken. Write the levels he would add, or what is wrong, in the note. Saved to loop/state/map_notes.jsonl.
Everything the AI inferred rather than heard him say. Newest first: the Fable walk-throughs' deep whys for each key decision (one row per decision) and the v4 pattern index (25 patterns, 11 open ideas not yet proven in his words). Older: the overnight-2 decisions and narrative spans. Rule on each: agree, disagree or partly, with a comment. Saved append-only (loop/state/inference_reviews.jsonl).
loading…
The chart ends at the decision minute, exactly what the judge saw (nothing after it is loaded until you press a reveal button; the days show their real dates). The traded index is shown; tick partner chart for the other index. Trades (one at a time, toggle above): blue = the judge's picked trade (entry, dashed stop, dotted targets with R:R, arrow) and orange = TJR's actual take (arrow at his entry time from his own decision record, plus a line at the price at his time because his entry price is not recorded); each is drawn on its own index's chart (tick partner chart to see both indexes). Gold = swept level, navy dotted = his pre-open map. Click a confluence below to put a marker on the chart. Drawing tools, sessions and indicator settings are the same as on the Candidates chart; drawings are saved per candidate.
The decision pilot's candidates (session-level SMT setups) in a fixed random order, each shown the way the LLM judges saw it: real dates hidden, bars only up to the decision minute (the entry bar's close). Make your call and save it (it locks); then the judges' calls appear, then Reveal outcome. Reveal is there before you decide too, but a candidate revealed before your call is left out of your score (any call saved on it afterwards is recorded as made after the reveal). Saved to state/operator_rule_decisions.jsonl; a change after saving is a new line marked as an edit after the reveal, and the summary keeps scoring your first call. Read-only otherwise; nothing here calls an API.
Summary · only the candidates you have decided, your first (locked) call; R as the pilot scores it (stop first on a same-bar touch, 240 min, timeout marked to market)
Times are New York; until the reveal, days are relative to the decision day (D, D-1 …), never real dates. Session boxes are built from the bars on the chart only, so a session still running at the decision minute stops there. As on TradingView: drag the chart to pan, wheel to zoom, drag the price axis to stretch it vertically (then drag the chart up / down), double-click the price axis or press auto to fit again, drag the time axis to zoom in time; scroll left for older bars. The two charts move together in time. Drawing tools on the left (Esc cancels, Delete removes the selected one, Shift+click measures); drawings are anchored by time and price, so they stay put on every timeframe, and are saved per candidate (this browser and the panel's state/dr_drawings/).