J3P — Experiment Log

Turn J3P into a system that actually makes money — with discipline. We test every idea, keep what works on FRESH data, and throw out what doesn't. · updated 2026-07-06 · J3P hub · trust report
🔴 MAJOR CORRECTION — read this before anything belowMAJOR CORRECTION (2026-07-07): the profit numbers on this whole page are LOOK-AHEAD-INFLATED. The backtest chose which stocks to trade using THAT DAY'S finished move (a name qualified if it had already moved >=3%% that day) — which you cannot know before the open. Rebuild the watchlist on honest pre-market data (the prior day's movers) and the out-of-sample edge falls from +0.17R to about ZERO (+0.004R), with every play at or below zero. So J3P has NO proven pre-market edge yet. What still holds: riding the 20-EMA beats capping the band (a relative exit finding). The only honest test now is the LIVE tracker going forward. Trade tiny or on paper until it proves an edge. Full detail on the Trust Report.
⚖ On paper vs. in your account — read this firstRead this FIRST. Everything below measures one thing: does the SYSTEM have an edge in BACKTESTING? It does — about +$1,500 per 100 trades on fresh data (roughly +0.12 to +0.20 for every $1 risked). That is NOT the same as you making money right now. Your real account depends entirely on whether you actually TRADE the system — and right now you mostly aren't. On 7/2 the system's names (MU, TSLA) made +$132; the 13 OFF-BOOK names you actually traded lost −$165. The edge is real; the gap is execution. So when a row says 'works,' it means the idea works ON PAPER — your Trade Check page tracks whether YOU are following it.
Where we stand: Big win: for the first time our system makes money on FRESH data it has never seen — not just the old data we tuned it on. The secret wasn't a fancier signal; it was trading the RIGHT stocks (the handful that are 'in play' that day — the ones actually moving on that day's news, rather than the same fixed list every day), in the MORNING, WITH the daily trend, using only our 4 best plays. On the practice data it makes about +0.37 (37 cents of profit for every $1 risked); on fresh data between +0.12 and +0.20 depending on the test (the 90-day test came in at +0.12; the walk-forward's forward stretch at +0.20) — call it roughly +0.15, a RANGE, not one magic number. Still a lab result — trade it small and by hand first before trusting it.
Best result so far: The four discipline RULES, one at a time, and how much profit each one adds (profit measured per 100 trades, where each trade risks $100): start with everything and no rules — basically break-even (about +$230 per 100 trades) — and profit climbs to about +$3,700 as we add each rule. Biggest helpers, in order: only trade the MORNING window, about 9:30–11:00 a.m. (adds about +$1,400 per 100); keep only our 4 best plays (adds about +$1,100); trade only the day's 'in-play' stocks — the ones actually moving on that day's news, not the same fixed list every day (adds about +$800); and take only trades that go WITH the stock's bigger daily trend (adds about +$200). That's about a 16x jump in profit — by trading LESS but BETTER (32,000 trades down to 2,200). On real money, expect roughly +$1,200 to +$2,000 per 100 trades — the fresh-data tests landed between +0.12R (about +$1,200) and +0.20R (about +$2,000); call it about +$1,500. NOT the +$3,700 we got on the practice data. The 4 plays that make money on fresh data: revbreak_long, trend_ride, open_drive, vwap_reject_short. Full breakdown on the 90-Day Backtest page.

The recipe — profit climbs as each discipline rule is added

per 100 trades, each risking $100 · green = practice data · blue = honest fresh-data estimate (~+$1,500)
No rules — trade everything+$230
+ Only the morning+$1,630
+ Only the 4 best plays+$2,730
+ Only in-play (SIP) names+$3,530
+ Only with the trend+$3,700
Honest live estimate+$1,500
How to read this — a few quick words:
 • "R" = how much you make compared to what you risk. +0.20R means you make 20¢ for every $1 you risked. Negative = you lose.
 • "Practice data vs fresh data" = we tune an idea on older data (practice), then test it on newer data it has never seen (fresh). An idea only counts if it makes money on FRESH data — otherwise it just memorized the past.
 • "VWAP" = the day's average-price line — the typical price so far. Stocks often bounce off it.
 • "Gapped" = a stock opened way higher or lower than it closed yesterday (a price jump).
 • The play names (revbreak_long, trend_ride…) are each explained with a picture on the Plays page.
 • Why some rows look like they contradict: two phrases mean different things. "Ride" — riding a trend play to its 20-minute line WORKS; but "riding" a bounce-back play's leftover, or holding for days as a swing, FAILED (same word, opposite trades). "The plays" — all 9 plays with no rules LOSE, but the 4 best plays WITH the rules WIN. When a row says one and another says the other, they're about different things — not a contradiction.

✓ What works

✗ What didn't

◇ Open / next

Full log — every question we tested, and what we found

The big question behind all of this: can we turn J3P into a system that reliably makes money, with discipline? Each row is one smaller question we tested to get there — read the question first, then the answer.
DatePhaseThe question → what we triedWhat we found (the answer)Verdict
2026-06-27early (2-week data)
Built the 7 plays + a scale-out exit
Q: Do our setups make money on their own, before we add any rules?
What we tried: Built the seven candidate setups we started with (each is a specific chart pattern we buy or sell on) and an exit rule that takes profit in stages as the trade moves our way. Tested on about 2 weeks of price data recorded once per minute. Note: only 4 of these setups survived later testing; the rest were dropped.
What we found: No money-making advantage on their own yet. When we sorted the setups by our own quality score, the cleanest, highest-quality ones won about 69% of the time while the sloppy low-quality ones lost. So setup quality matters. But the goal of this step was just to build the testing machinery, not to prove the strategy makes money.TOOLING
2026-07-02early (2-week data)
Widened to 28 stocks + a fresh-data test
Q: Do those early edges survive when we test on more stocks and on data we've never seen?
What we tried: Added more stocks and, for the first time, held back some data to test on (the 'fresh data' check).
What we found: No play made money on fresh data — the small-sample 'edges' were just memorizing. Diagnosis: this was a DATA problem (2 weeks is too little), not necessarily bad plays.DIDN'T
2026-07-04deep data
Got 2 years of real data (Massive)
Q: Can we get enough trustworthy price history to test this the right way?
What we tried: Hooked up a deep price-history source and cross-checked it against our other one.
What we found: It matches to the penny — trustworthy. Now we finally have enough history to test properly.WORKS
2026-07-04deep data
Ran all the plays on real data
Q: Was 'no edge' the real answer, or just a quirk of our tiny 2-week data?
What we tried: Ran every play on a month of real data across 8 big stocks.
What we found: Loses money on their own: about -$600 for every 100 trades (each trade risking $100), across roughly 1,800 trades in the test month. So the earlier 'no money-making advantage' finding was real, not just a quirk of the tiny 2-week data source.DIDN'T
2026-07-04deep data
Which 'line up the trends' filters help?
Q: Does only taking trades that agree with a bigger trend help us?
What we tried: Tested three different ways of gauging a stock's trend direction and only taking trades that agree with it: the day-over-day (daily) trend; a 2-day rolling average price line; and shorter trends measured over the last 30, 60, or 100 minutes of trading.
What we found: Only the DAILY trend helps (a nice boost). The shorter intraday trends and the 2-day line don't — they hurt or do nothing.PARTIAL
2026-07-04deep data
What time of day is best?
Q: Does the time of day you trade change whether you make money?
What we tried: Split trades into morning / midday / afternoon buckets to see which time of day performs best, and also tested filters that only take trades when certain average-price lines are tilting up (or down).
What we found: MORNING is clearly best; afternoon is the worst (a money-loser). Adding filters based on whether the day's average-price line or the 50-day average-price line is tilting up or down made results worse. Correction: the real problem is simply the afternoon, not any specific cutoff time like 12:30.PARTIAL
2026-07-04deep data
Combine morning + daily trend
Q: If we stack the two rules that helped, do we finally turn profitable?
What we tried: Stacked the two rules that helped.
What we found: Cut the loss roughly in half: from about -$600 down to about -$300 for every 100 trades, on roughly one-third as many trades. Better, but still losing money at this stage.PARTIAL
2026-07-04deep data
New play: buy the reversal (revbreak_long)
Q: Does the afternoon-reversal idea (buy a falling stock that turns back up) actually make money?
What we tried: The idea: a falling stock breaks its downtrend, pulls back to re-test that broken trend line, then resumes upward. Buy that turn.
What we found: At first it looked like an afternoon loser — but that was a testing mistake: our entry rule triggered too easily, the exit rule was wrong, and our price data didn't cover the late afternoon. Once fixed, the big afternoon reversal winners were real (each winner made about 3 to 6.5 times the dollars it risked). It became one of our 4 best plays.CORRECTED
2026-07-04deep data
Ride the trend to exit (instead of a fixed target)
Q: For trend plays, is it better to ride the trend out than to grab a small fixed target early?
What we tried: The insight: trend plays should stay in the trade and ride the move until the price falls back through its trend line (a line that tracks the stock's recent average price) — rather than grabbing a small fixed target early.
What we found: On the practice data it clearly helped (turned losers into small winners). An early isolated test (using a fixed stock list, with no stock-picking) looked shaky on fresh data, which made us cautious. But a clean head-to-head test on 2026-07-06 (see below) later proved it: for trend plays, riding the trend beats grabbing a fixed target at the price band (a target level set a fixed distance from the day's average-price line) by about 7x on fresh data. It's the right exit for trend plays.WORKS
2026-07-04deep data
The big fresh-data test (3 months)
Q: With our best rules on, does ANY play still make money on data it's never seen?
What we tried: Took the disciplined version (morning + daily trend + the right exit for each play) and tested it on held-back data.
What we found: Most plays were fool's gold — they looked good on the practice (old, tuned-on) data but LOST money on fresh (never-seen) data. The ONE play that made money on BOTH data sets: the short play (vwap_reject_short — shorting a stock whose bounce back up fails). This was the first play with a real, repeatable edge.WORKS
2026-07-04deep data
Does picking in-play stocks matter?
Q: Does it matter WHICH stocks we trade — the day's in-play names vs the same fixed list every day?
What we tried: Up to this point the system had been trading a fixed list of 8 big stocks every day, instead of picking the specific stocks that were actively 'in play' (moving on that day's news). We compared the two approaches: fixed list vs. day's in-play names.
What we found: YES, big time. Trading the day's in-play stocks turned a loss into break-even, just by choosing better stocks. Stocks with big news that day actually move in a direction you can trade.WORKS
2026-07-05deep data
The 90-day test — the big one
Q: Put the WHOLE recipe together over 90 days — does J3P finally make money on fresh data?
What we tried: Rebuilt 90 days of daily in-play stocks and ran the full recipe (in-play + morning + daily trend + the right exit for each play), then checked it on held-back fresh data.
What we found: FIRST TIME IT MAKES MONEY ON FRESH DATA. Practice +0.36 -> fresh +0.12 (still positive). Four plays hold up on both: revbreak_long, trend_ride, open_drive, and the short vwap_reject_short. Trading the in-play stocks is what made it survive.WORKS
2026-07-05mirror tests
Test the long AND short of everything
Q: If our LONG plays work, do their mirror-image SHORT versions work too?
What we tried: We tested whether the opposite (mirror) version of each winning BUY play also works. Important: each mirror is a genuine SELL (short) setup with its OWN trigger — it waits for the exact flipped pattern to appear (for example, an UPtrend that breaks DOWN and then fails a retest), NOT just selling at the moment the buy play fires. We built the true opposite of each winning buy play and checked whether the short made money too.
What we found: They DON'T. They only 'won' on fresh data because the market happened to be falling that month — lost in the earlier period. That's a market-mood fluke, not a real edge. Not added.DIDN'T
2026-07-05exit design
Ride-the-runner exit tweak
Q: On plays that normally sell out at the average-price (VWAP) line, can we make MORE by selling only PART of the position there and holding the rest longer to catch a bigger move?
What we tried: We tested this: on plays that normally sell out at the average-price (VWAP) line, sell only part of the position there and hold the remaining shares longer, exiting them only when the price falls back through the line the trend has been following (a moving-average line). The goal is to capture the occasional much-bigger move.
What we found: At first it looked like a winner (about +$1,000 profit per 100 trades, each risking $100). But five separate re-checks proved the win was FAKE — it came from luckily dividing the practice/fresh data at a favorable point, a software bug that let the held-back shares never register a loss, and a handful of lucky trades carrying the whole result. On a clean test it loses money. We shelved it. The big move it tries to catch is real, but too rare to trade reliably.DIDN'T
2026-07-05stock picking
Which stocks make the plays pay?
Q: What is it about a stock that makes our setups win more — and is it different for longs vs shorts?
What we tried: We matched each stock's traits — did it open far from yesterday's close (gapped), how much trading volume, how far it had already moved that day, and how volatile the day was — to how well the plays performed. We split the analysis between BUY plays (betting the stock rises) and SHORT plays (betting it falls).
What we found: Don't chase what already flew: a quiet-open stock having a wild day pays about 2x; a stock that already gapped or ran hard pays about a quarter. And longs want WILD days while shorts want STEADY days.WORKS
2026-07-06mirror tests
Backside LONG — the mirror we'd skipped
Q: Does the reverse of our backside short play work? The backside short bets a stock will fall; the reverse would be BUYING a stock that has already dropped hard down to its average-price line, hoping it bounces back up. Does that buy version make money?
What we tried: Earlier we had tested reversing our winning BUY plays into SHORT plays (bets the stock falls). We had never tested the opposite direction on the backside play: turning its SHORT into a BUY. We built that BUY version (buy a stock that dropped hard down to its average-price line) and tested it.
What we found: Doesn't work — it's trying to catch a stock still dropping fast (a 'falling knife'), which usually keeps falling. It wins only 28% of the time and loses money on fresh data. Betting that a stock which shot UP will fall back works; buying a stock that already dropped hard hoping it bounces does not. Not added.DIDN'T
2026-07-06exit design
Band exit vs riding-the-trend — the clean test (James's call)
Q: Does taking a small fixed profit at a set price level near the average-price line ('the band') limit how much we make, compared with holding the trade longer and only exiting when the price falls back through the line the trend follows (a 20-bar moving-average line)?
What we tried: We ran the SAME set of trend-play trades three different ways, changing ONLY how each trade is exited: (1) take a small fixed profit at a set level near the average-price line ('the band'), (2) sell in steps at several band levels ('band ladder'), or (3) hold until the price falls back through its 20-bar moving-average line. We tested over 90 days and checked it on fresh (never-seen) data.
What we found: Holding the trade longer clearly wins, and it holds up on fresh data. Taking the small fixed profit at the band makes only about +$200 profit per 100 trades (each risking $100) on fresh data; holding until the price breaks back through its 20-bar moving-average line makes about +$1,500 per 100 — roughly 7x more. This is true for every trend play; two plays (trend_ride and open_drive) actually LOSE money when capped at the band but make good money when held longer. Taking the small profit early was leaving most of the winning move behind. NOTE: this applies only to trend-following plays. Plays that bet a price snaps back to its average (mean-reversion plays) should still take the profit at the band — there is no ongoing trend to hold for. This is different from the earlier rejected experiment (log[13]), which wrongly tried to add trend-holding to a mean-reversion play.WORKS
2026-07-06putting it together
The Cockpit + real news + plain definitions
Q: Can we put everything on ONE screen a trader can actually use to make trade decisions?
What we tried: We pulled everything into one trader screen: a GO / WAIT / SKIP recommendation for each stock with the reasons shown, the real news for each stock, and a plain-English explanation of every number. Along the way we found three data inputs that were arriving wrong or missing, and adjusted the screen to handle them correctly.
What we found: One screen reads the live in-play list and tells you what to do and why. Not yet automatic or alerting (that's the next open item).TOOLING
2026-07-06exit design
Partial at the band vs ride the whole thing
Q: On a trend play, if we take a partial profit at the band and ride the rest, do we make more?
What we tried: Tested the same trend trades three more ways: take 1/3, 1/2, or 2/3 off at the fixed band target and ride the remainder to the trend-line (EMA) break, stop moved to breakeven.
What we found: No — it makes LESS, not more. The band exit is nearly worthless on a trend play, so every share taken off early is a share not riding the trend. Fresh-data profit per 100 trades ($100 risk): ride it all +$1,471; 1/3 partial +$879; 1/2 +$707; 2/3 +$534; cap-all-at-band +$200. The ONLY thing a partial buys is a higher win rate (46% vs 42%) and a smoother ride. So: ride it all for max profit, or a small 1/3 partial if you want it easier to hold.PARTIAL
2026-07-06exit design
Swing exit — hold for days, scale out at the 5/10/20-day averages
Q: What if we turned these into swing trades and exited thirds at the daily 5/10/20-day moving averages?
What we tried: Held the trend entries for days and scaled out 1/3 at each daily average, on the cached daily bars (first pass).
What we found: Looks amazing but it's a MIRAGE. The average was huge, but the TYPICAL (median) trade LOSES, only ~24-34% win, and 67-85% stop out in the first day or two (the intraday stop is far too tight to hold overnight). The giant average is carried by a few rare multi-week runners. With a sensible swing stop it goes NEGATIVE on fresh data (-$2,271 per 100). A lottery, not an edge. PARKED until we can test it properly (real swing entries, wider stops, position-sizing, more history).DIDN'T
2026-07-06harden
Walk-forward — is +0.12R one lucky window?
Q: Does the 4-play stack hold across many time periods, or did it just get one good stretch?
What we tried: The model is fixed (no fitting), so we split the 90 days into 6 consecutive blocks — each genuinely never-before-seen — and checked every block.
What we found: HOLDS. 5 of 6 blocks were profitable, and the pure forward (2nd half) made +0.20R (+$1,957 per 100 trades). Not one lucky window. One yellow flag: the most recent block (late June) was slightly negative (-0.04R, 31% win) — worth watching as we go live.WORKS
2026-07-06real trades
Would the daily-trend rule have helped my real IBIT trades?
Q: On my actual IBIT trades (account 14859), would 'only trade with the daily trend' have kept the good ones and avoided the bad stretch?
What we tried: Pulled every IBIT fill in late-March and May-15-to-Jun-13, paired each day with IBIT's daily trend (above/below its 20-day line).
What we found: YES on direction. You were LONG both windows. Late March: IBIT mostly ABOVE its 20-day -> the rule KEEPS those (choppy, ~break-even). May 15-Jun 13: IBIT was BELOW its 20-day EVERY one of the 16 days (a steady 45->35, -22% drop) -> the rule AVOIDS the whole long campaign, and pointed to SHORT (the winning side). HONEST CAVEAT: the exact dollar P&L is not trustworthy yet — you held positions overnight, so the quick fill math is distorted (a -$8,766/+$8,657 pair gave it away). Direction/trend finding is solid; the true dollars-saved needs proper round-trip reconciliation (next).PARTIAL
2026-07-06go-live
Made it tradeable for tomorrow morning
Q: Can we actually trade the 4 plays tomorrow, hands-on, on the right stocks?
What we tried: Finished the live scanner: fires all 4 validated plays (added open_drive) with a big GO banner + the morning + daily-trend gates; made the daily-trend check pull fresh data; sped it up 60x (guarded a stray backtest). Deployed it to a stable folder with a one-click launcher + a scheduled task that auto-starts 9:25 AM weekdays.
What we found: Live at /system/J3P_LIVE.html. Each morning: the SIP list is built pre-market (7:00 ET), the scanner watches those names 9:25-11:15 and flashes GO when a validated play fires. James just keeps the box on + logged in and watches the page; GO -> check the chart -> trade small by hand. LAB — not live-proven; expect live ~ the +0.12R test.TOOLING
2026-07-06trend playbook
Organized the IBIT autopsy + phase notes into a reusable Trending-Stock Playbook
Q: When a stock starts to trend (the next IBIT, the JNJ we missed) — how do we enter, size, and manage it, every time?
What we tried: Pulled the rules out of the BTC->IBIT MTF-VP autopsy (the deep study of the real IBIT swing) and James's phase-study notes, and organized them into one page: CONTEXT (price vs the 200-day line + its slope) -> GRADE the entry (A/B/C/Pass) -> SIZE by grade -> MANAGE (mandatory profit-take at the 200-day test; never hold for a target inside overhead supply; otherwise scale out on the 5/10/20-day staircase). Added a live board reading JNJ, BBIO, BTDR, ALOY through the same lens.
What we found: /system/J3P_TREND_PLAYBOOK.html. The IBIT numbers make the case: the trade had +5.5R available; trimming at the 200-day = +5.22R, holding for the big target = +0.93R. JNJ today: right CONTEXT (above a rising 200-day, at highs) but ~20% extended -> the playbook says WAIT for a pullback with a defined stop, don't chase. CORRECTION logged: the earlier IBIT filter test should have started from this autopsy (the source of truth for that trade) instead of re-deriving a cruder version.TOOLING
2026-07-07management plans
Management-plans menu + per-play management cards + watcher spec
Q: How does James decide + remember which exit/management style to use on a given trade — and how do we later enforce it live?
What we tried: Added the side-by-side management-plans menu (Hold-for-target / Momo-burst+ceiling-trim / Staircase / Intraday EMA-ride) to the Trend Playbook, a Plan-A/Plan-B management card to each of the 9 plays (exact trims, stop moves, invalidation, tested numbers), and a review-first SPEC for a Management Watcher.
What we found: Playbook copy DONE + reviewable (J3P_TREND_PLAYBOOK + J3P_PLAYS). Watcher SPEC written (J3P_WATCHER_SPEC): 4 alerts (pullback-zone, ceiling/200-DMA TRIM-NOW, staircase-leg sell-1/3, selected-management-line) + staged build (daily read-only -> plan-per-position -> SIM hotkey enforcement). GATE: nothing to production until James reviews the copy + spec. Practice via the VCET replay viewer is the parallel step.TOOLING
2026-07-07management plans
Watcher Phase A (live) + Replay Practice (step 2)
Q: How does James remember the plan live, and how does he drill it into muscle memory?
What we tried: Built + scheduled the Management Watcher Phase A (read-only daily alerts: watchlist pullback-zones + open-position 200-day/staircase exits, 8 AM weekdays). Built the Replay Practice page: reveals the IBIT daily chart bar-by-bar, forces the call at each decision (enter/trim/add/exit), scores vs the playbook.
What we found: J3P_WATCHER.html (JNJ watching → pullback alert when it dips to the rising 20/50-day) + J3P_REPLAY_PRACTICE.html (on-plan path +4.47R vs +0.93R held-to-failure; the 200-day trim is the pivotal call). Both read-only/LAB. Sequence now: playbook copy ✓, practice ✓, watcher Phase A ✓; remaining = watcher Phase B (plan-per-position UI) + Phase C (SIM hotkey — separate sign-off) + more replay tickers.TOOLING
2026-07-07management plans
Replay Practice — 3 trades, 3 lessons (JNJ + BTDR added)
Q: Does the playbook hold across DIFFERENT contexts, and can James drill each one?
What we tried: Turned the replay into a multi-trade practice with an index. Added JNJ (A-context above a rising 200-day) and BTDR (a +45% run that round-tripped) beside IBIT (B under a falling 200-day). Each reveals the real daily chart bar-by-bar and forces the call.
What we found: Three distinct lessons, same discipline: IBIT = DEFENSE (trim at the ceiling, on-plan +4.47R vs +0.93R held) · JNJ = OFFENSE (don't chase; buy the pullback to the rising 20-day @ $223, ride to $263 = +8.51R) · BTDR = PROTECT A RUNNER (staircase the 5/10/20-day breaks = +0.69R vs -0.35R held-to-round-trip; peak was +2.49R). CONTEXT (price vs the 200-day + slope) picks which plan. Index /system/J3P_REPLAY_PRACTICE.html.TOOLING
2026-07-07management plans
Intraday EMA-ride replay (GOOG) + adversarial review of the whole replay suite
Q: Does the practice suite drill the RIGHT lesson — honestly, with no look-ahead or over-stated edge?
What we tried: Added the 4th replay: an intraday 2-min GOOG day-trade (revbreak_long, ride-the-20-EMA vs cap-the-band), picked by ranking all 19,980 durable-play trades and choosing a clean, believable example (not the +19R thin outliers). Then ran a 4-dimension adversarial verification (R-math / look-ahead / honesty / data) across all 4 replays.
What we found: The review CONFIRMED 11 defects; all fixed: ENGINE — added stop enforcement (hold through the stop = −1R) + re-graded the scorecard to headline the ON-PLAN COUNT (not a green R) with a 'the trade bailed you out, not process' caveat, killing the never-lose pedagogy; made 'add' a real (costly) action. IBIT — reframed 'tagging the 200-day' to the overhead supply it actually stalled at (its own s200 sits ~$6 above price). JNJ — re-anchored to the real 6/04 RECLAIM (was entering below the 20-day) and stripped outcome-leaking prompts; 6/18 is now an honest first-close-below → trim. GOOG — now teaches the ODDS (riding beats capping on avg +0.09R→+0.41R but wins only ~45% of trades; +7.4R is a ~98th-pct tail), not the outlier. Re-verify: 39 fixes confirmed, 0 high/medium, 2 cosmetic nits fixed. Lesson banked: a single hand-picked winner drills the wrong magnitude unless the aggregate distribution is on the page.TOOLING
2026-07-08verify
Re-check: are the OPPOSITE (short-mirror) trades profitable? (James's challenge)
Q: The shorts of revbreak/trend_ride/band_break (and the long of the short) looked profitable in a backtest — is that real?
What we tried: Re-ran ALL four mirrors (revbreak_short, trend_ride_short, open_drive_short, and band_break_short — never tested before) on BOTH universes: the look-ahead same-day SIP list AND the honest prior-day list. Same gates (morning + daily-trend-down), same 20-EMA exit. Also re-checked vwap_pull_in (the long opposite of the short).
What we found: NOT REAL — James's skepticism confirmed. Every mirror LOSES in the earlier (rising-market) half and only 'wins' in the later half, when the market itself was falling: revbreak_short train −0.31 → test +0.47; trend_ride_short −0.50 → +0.24; band_break_short −0.33 → +0.20 (honest universe); open_drive_short 0% win. That pattern = the falling market paying ALL shorts (market mood), not a repeatable setup — a real edge must make money in BOTH halves. vwap_pull_in = noise (+0.05, t<1, July −0.75). The one that superficially passed both halves (band_break_short on the look-ahead list) collapses on the honest pre-market list. VERDICT: no mirror survives; do not add any of them.DIDN'T
2026-07-11verify
James's play: earnings gap-down + VWAP reclaim → prior-day VWAP ('1000 times in a row')
Q: When a stock gaps down (earnings) below VWAP and reclaims session VWAP, does it reliably run to the PRIOR day's VWAP?
What we tried: Built the detector (gap_reclaim.py: gap ≤−2% vs prior close, open below prior-day VWAP, first morning close back above session VWAP; stop under low-of-day; target = prior-day VWAP) and backtested all 2,278 cached in-play name-days. Also measured the RAW magnet claim with NO stop at all.
What we found: The magnet claim does NOT hold on our data: even with no stop, price touches the prior-day VWAP the same day only 26% of the time (TRAIN 32% / TEST 16%) — and the bigger (more earnings-like) the gap, the LESS it gets there (2–3% gaps: 49% · 3–5%: 23% · >5%: 13%, because the target is further away). As a trade (stop under LOD): TEST −0.485R, 24% win, 68% stopped. THE REAL KERNEL: the reclaim itself does travel — median best reach +1.32R after the trigger — so a modest-target version may have life, but 'goes to prior-day VWAP' is not it. CAVEATS: our proxy is any ≥2% gap-down on an in-play name, not confirmed earnings; and 90 days ≈ 333 setups — a '1000-row' streak would need a very different (likely narrower) definition. DEPLOYED to the live scanner as a LAB play (earn_gap_reclaim, morning-gated, trend-gate exempt) so the live tracker measures James's actual version forward.DIDN'T
2026-07-11verify
Gap-reclaim CORRECTED to James's real spec (v2) + the earnings scan SIP was missing
Q: With the RIGHT spec — gap down vs the CLOSE, break AND HOLD session VWAP, target = the 2-DAY VWAP, real earnings flag — does the play hold up?
What we tried: James corrected three things in my first test: no prior-day-VWAP-open condition, the reclaim must HOLD (not just cross), and the target is the 2-day anchored VWAP (avwap2 — the multi-day line), not the prior day's final VWAP. Rebuilt the detector to that spec, pulled REAL earnings (kti news earnings: prior-day after-hours + same-day pre-market) for all 87 days, re-ran everything. Also built the LIVE earnings scan he asked for: liquid earnings-morning names now auto-append to the scanner watchlist daily.
What we found: Better, but still not a near-certainty on our data: touching the 2-day VWAP after a held reclaim = 43% of all gap-downs (vs 26% under my wrong spec) and 34% on the earnings-driven subset; as a trade ~breakeven (−0.01R all / −0.07R earnings). HONEST LIMIT: only 35 earnings setups were testable because our bar cache covers SIP names only — which is EXACTLY James's point (SIP misses earnings names). So the historical test is inconclusive on his real universe; the live scanner now watches earnings gappers every morning and the tracker will measure the true rate forward. His spec corrections all made the play better (hold-confirmation cut stop-outs 62%→50%; the 2-day VWAP is a nearer target, 2.5R vs 3.7R away).PARTIAL
2026-07-11verify
FULL-universe earnings gap-reclaim (Massive unlimited) — the first play that passes every honesty check
Q: On ALL 686 earnings gap-downs (not just the 35 in the SIP cache), does James's play hold up?
What we tried: Massive upgraded to unlimited calls → fetched 2-min bars for the complete earnings universe (686 earnings gap-downs ≥2% + liquid, 87 days) and ran spec v2 (reclaim + HOLD session VWAP → 2-day AVWAP target, stop under LOD).
What we found: IT WORKS — and unlike everything else it is NOT tail-driven: 513 of 686 fired the reclaim+hold (75%); avgR +0.124 with a POSITIVE MEDIAN (+0.068) and 53% win (t=2.55); BOTH halves positive (train +0.09 / test +0.26); positive 4 of 5 months INCLUDING the falling June (+0.18) — so not market-mood. THE DIAL IS GAP SIZE: >5% gaps = +0.29R avg, +0.25 median, 57% win, t=3.8 (grade A); 3–5% ≈ flat (B); 2–3% NEGATIVE (C — skip). The '1000-in-a-row to the target' is still not literal (45% touch the 2-day AVWAP) — the play wins because losers are small (only 35% stop) and winners at big gaps outweigh them. Caveats: 2-min close fills (earnings-day slippage is real); pre-market-knowable inputs ONLY (earnings+gap+liquidity — no look-ahead); July n tiny. LIVE: earnings names now on the SIP page (badge + per-name modal + own section), the scanner watches them, and every fire alerts Discord/SMS/banner with the gap grade.WORKS
2026-07-12verify
Pipeline replay — the deployed system replayed over 13 earnings days (James's test)
Q: Forget the concept backtest — does the system AS DEPLOYED (wire → liquid top-10 → detector → gates → tracker) find the setups and execute?
What we tried: Replayed the exact live code path day-by-day 6/23–7/10: rebuilt each morning's earnings watchlist as the scanner would have, ran the gap-reclaim play, scored with the live tracker's math. Bars from Massive (paid). Also: charts now carry prior-day history (MAs work) with ET axes; the live page updates in place (charts survive refresh); results are labeled SIMULATED; the drill page is banner-labeled.
What we found: THE SYSTEM WORKS AS DEPLOYED on this sample: 13 signals over 13 days (~1/day), 46% win, avgR +0.35, +4.6R total — and the A-grades (gap ≥5%) carried it: 5 signals, 60% win, avgR +0.77, +3.9R. Zero-signal days were the system being picky (no gap-down reclaimed and held), not broken. Matches the 686-setup backtest's gap-size finding through the REAL funnel. Still a small sample and still simulated fills — week one = A-grades only, tiny, or watch-only; the live scorecard is the final judge.WORKS
2026-07-13verify
James caught the earnings wire lying — strict re-run cuts the earnings edge to ~flat
Q: AMZN 'reported' on 7/10 but doesn't report until 7/30 — how much of the earnings play was built on false reporters?
What we tried: James spot-checked the SIP earnings badges against real earnings dates and caught the wire attributing analyst previews ('Chevron Poised to Beat — UBS Says'), dividend notices, and roundup stories to companies that never reported. Built a strict confirmed-reporter filter (must have parsed actual numbers or a results-style headline; previews/dividends/roundups rejected) — 7/10: 40 wire records → 4 real reporters. Applied it to the SIP page, the /sip/ strip, the scanner watchlist, and re-ran the full backtest + the 13-day pipeline replay.
What we found: THE EDGE SHRANK BADLY: full universe 686→616 candidates, avgR +0.124→+0.062 (t=1.34, no longer significant); A-grade (gap≥5%) +0.29→+0.13 (t=1.95, borderline). The 13-day pipeline replay FLIPPED: +4.6R → −1.8R (A-grades +0.1R ≈ flat). Translation: much of the apparent earnings edge came from FALSE-earnings names — stocks gapping down on other catalysts. Notably the loose (any gap-down) version outperformed the strict earnings-only version, echoing the earlier control finding: the mechanism may be GAP-RECLAIMS generally, not earnings specifically — worth its own test. RECOMMENDATION CHANGED: watch-only on the earnings play until the live tracker builds evidence; do NOT size it. All surfaces now show confirmed reporters only (Delta stayed; AMZN/CVX/CMS/PSX/RKT/CSX gone).CORRECTED
Living lab notebook — appended as we test. Generated from experiments.json. LAB only; nothing promoted to the Pine.