Every time the engine changes direction on an asset, it's logged here as a trade —
server-timestamped, append-only, never edited. No cherry-picking, no backfilling, no deleting
the embarrassing ones. Judge the engine on what it actually said, when it said it.
How the crypto signal track record is measured
Two engines run side by side — Tactical (short-term breakout longs) and
Careful (confirmed shorts) — the two books that make up the live Hybrid engine (v2.4, Jul 8 2026). Every call is scored as a real trade: from the moment the
engine takes a stance until it flips that asset — entry to exit, one position per asset at a time. For a LONG, up is right;
for a SHORT, down is right; CASH is judged by what it kept you out of. A trade that nets under ±1% over its hold is flat — noise, not a win or a loss. Calls the engine still holds show "open · live" (unrealized — not counted until they close).
What these numbers are: the signal flips are real and server-timestamped, but every return shown is
hypothetical — computed by applying the stop rule below to market prices, with no real money traded.
Hypothetical results have inherent limitations (fills are assumed at the stop price; no fees, slippage, or
funding) and do not represent actual trading. Past results, real or simulated, don't predict future ones.
Engine v2.1 — June 16, 2026. Two upgrades went live: real on-chain social (LunarCrush) + whale exchange-flow (Whale Alert) data across 28 coins, and a validated discipline fix — shorts now fire only on strong, confirmed downtrends at half size (the prior engine shorted too aggressively and got whipsawed; backtesting proved the conservative version cut drawdown hard). Nothing is ever deleted — every prior call stays logged below the line. Forward-tested performance is tracked from here.
Engine v2.2 — July 2, 2026: stop-out cooldown. Once a call exits at its initial −4% stop, the engine cannot re-take the same asset & direction for 48 hours — no revenge-trading. Diagnosed from this very record: nearly every loss was a −4% day-one stop, and roughly two-thirds were the engine re-entering an identical just-stopped trade (replaying the record with the cooldown removes 30 of 47 losses at the cost of one winner). Deterministic from public data, so any follower can reproduce every suppression. Applies forward only — nothing below is touched. Engine v2.3 — July 2, 2026: symmetric long gate. Longs now require a strong, confirmed uptrend — the exact same evidentiary bar shorts have faced since v2.1 (strong trend + real order-flow). The record showed the asymmetry plainly: shorts won 68%, longs won 14%, because the old loose gate bought weak bear-market bounces. Replaying the engine's full logged tick history (23k ticks) with symmetric gates: longs go from net-negative to net-positive and the whole book improves. Forward only — the past record stays as-is. Engine v2.5 — July 9, 2026: no shorting into a breakout. A confirmed short call is now vetoed to CASH whenever the short-term engine (the 20-day breakout) is pointing up — you don't short into strength. Diagnosed from the record: the losing shorts were the ones fired while price was breaking out the other way. Replaying the tick history, this rule removes only losing shorts (it cut two, both losers; zero winners) and lifts net return. Small in count but sound in principle, and it matters most in exactly the rally-into-chop regime that's been squeezing shorts. Forward only — nothing below is touched. Engine v2.6 — reviewed Aug 6, 2026: NOT deployed. We tested "hysteresis exits" (hold a short until the signal meaningfully reverses, instead of exiting the moment it grazes the gate) in a private shadow book for 16 days against criteria we wrote down before the test: cut re-entry churn by ≥30% and match the live book's net under the same grading. It crushed the first bar — 65% fewer round-trips, holds 12× longer — and failed the one that matters: graded exactly the way this page grades everything (the followable stop model), it returned −2.0% vs the live book's +5.7% over the same window, because positions held through noise ran into the −4% hard stop 61% of the time instead of 22%. A change that looks like a 65% improvement on the headline metric and loses money under honest grading is a change you don't ship. The shadow book keeps running; if a larger sample reverses this verdict, you'll read it here first. Nothing changed in the live engine. Data integrity — TON (disclosed Aug 6, 2026). Our price venue delisted TONUSDT on Jun 30, 2026, and kept serving its final candles — so TON's signal inputs silently froze, and the TON calls logged Jun 30–Jul 25 reflect a market our data source had stopped tracking (its settled −0.25% result stands as written; nothing is ever regraded). Found in our own Aug 6 audit. Fix shipped the same day: a staleness guard now treats any feed whose newest candle is older than 48h as failed — a dead market reads n/a and can never hold or open a position again. TON stays visible in the app but emits no fresh signals until it has a live venue.
Bug fix — v2.2 cooldown ledger freshness (Aug 6, 2026). The same audit found the 48h no-revenge-trading rule promised above had a plumbing bug: the stop-out ledger it reads was refreshed only after stances were computed, so a fresh stop could be invisible at decision time. Fixed forward — the ledger now refreshes before each decision cycle, so the rule enforces exactly as described. No engine rule changed and no historical row was touched.
This is the engine's current stance on every asset — not the trade log. Most of the time the honest call is cash: it only takes a position when trend, order-flow and valuation line up. A quiet asset isn't a dead feed — it's the engine watching and waiting for a real edge.
Performance grading closed calls…
The legacy result is locked. The public +167.9% value and its 313 settled calls are preserved exactly as the last published snapshot under the former methodology; they are never regraded. The total above starts from that fixed base, then adds only positions entered after the profit-lock-v2 cutoff. Forward calls use one non-overlapping position per asset from entry until the earlier of signal flip, stop, or time cap. The forward total counts the live hybrid books — tactical longs and careful shorts — so each signal is counted once: since v2.4 the careful book's longs come from the same 20-day breakout the tactical book trades, so one breakout is logged in both books, and totalling every row would count that trade twice. Both rows stay visible and served; only the total is de-duplicated. Their outcomes are written once to a server-owned append-only settlement ledger. Only fully closed UTC bars may settle a stop, so a still-forming bar can never rewrite a result — hourly bars since Jul 30, 2026: under the original daily bars, a run shorter than a full day had no eligible bar at all, so its stop could never engage and it settled at the next signal flip regardless of what price did in between. The 13 forward settlements recorded Jul 28–29 under that daily rule stand exactly as written (append-only means our own corrections don't get to rewrite history either). Historical calls below the cutoff display under the locked legacy stop-model — the same methodology as the frozen base — and the public API's settlement ledger contains forward positions only, so the served data always sums to the forward number shown here. Profit-lock-v2 uses a −4% emergency stop, +1.5% arm, +1% floor, 1.5% trail, 7-day unarmed-loss cap, and 14-day armed-winner cap. Results inside ±1% are flat. All displayed returns are gross modeled percentages before fees, slippage, sizing, or compounding. The locked base and forward add-on remain separately visible so neither can silently replace the other.
loading…
This record only covers periods when the engine was running. It is a research log, not
investment advice and not a performance guarantee — see the Terms & Disclaimer.
Past signals, real or logged, don't predict future results.