How the pieces fit together
The same names show up on several pages, so here is the map in plain words. A bot is what trades. Models and skills are parts you equip on a bot. The Quant Lab tests and searches on paper. Head-to-head compares two of your paper bots. Build it Better rebuilds your bot. Every statement below was checked against the platform's code and its live catalog on 28 October 2026.
The picture
Models and skills are equipped on a bot. A bot trades on paper (play money) or live (your own funds). The Quant Lab search replays model rules on past crypto prices and starts the winners as paper bots. Head-to-head scores two paper sides over the same window. Build it Better rebuilds your bot and hands it back as a paper clone.
1. Model, skill and bot: what is the difference?
A bot (also called an agent) is the thing that trades. The cards on /perpetual/agents are bots, and when you clone one, your copy on /my-agents is your bot.
A model (algorithm) is an entry in the Algorithms, Data & Models catalog on /adm. There are 20: ten biomimetic models (HELIX, PRION, ENZYME, SYNAPSE, CORTEX, TELOMERE, IMMUNE, CIRCADIAN, MITOSIS, CRISPR) and ten probability models (RUPTURE, ORACLE, VERDICT, SIGNAL, MIRAGE, ANCHOR, DRIFT, RECOIL, ENDGAME, FLUX). PRION, SYNAPSE and TELOMERE are paid; the rest are free. A bot can carry up to 5 models.
A skill is an entry in the Skills & Tools catalog on /skills-n-tools. There are 29, in groups: indicators (MACD Crossover, RSI Momentum ...), sentiment, data feeds (Binance WebSocket ...), filters (Volume Threshold ...), strategy modifiers and a few paid or coming-soon ones. A bot has 3 primary and 3 secondary skill slots.
So yes: a bot has models and skills equipped, and they are two separate catalogs with separate slots. A skill is never a model. MACD Crossover is a skill. The name MACD also appears elsewhere: the HELIX model's rule uses a MACD crossover, and MACD Momentum is one of the Backtested Strategy bots. Those are three different things that share an indicator.
What drives a running bot's orders today is the bot it was cloned from (its strategy) and your settings (stop-loss, take-profit, position size, markets). The platform marks every equipped model as backtest only: models are used when the bot is backtested and by the Quant Lab search, and they do not yet change a live order.
Where to look: /adm/skills-n-tools/perpetual/agents/my-agents
2. What does the Quant Lab search run over?
Its own list: the 16 model rules that have a backtest setup. That is the ten biomimetic models from /adm plus six strategy rules that belong to Backtested Strategy bots on /perpetual/agents: cryptobull_blue and cryptobull_red (CryptoBull Bot Blue and Red), hybrid2_blue and hybrid2_red (Hybrid 2 Blue and Red), stockbull_blue (StockBot Blue) and stockbear_red (StockBot Red, the short-selling one).
It does not search skills, the ten probability models (RUPTURE, ENDGAME, ANCHOR ...), your own bots, or the perpetual-market agents such as Sync Master. Three more strategies (cross-exchange arbitrage, funding-rate carry, DEX order-book replay) have a setup but run on spreads, funding or DEX data instead of candles, so they cannot be searched yet.
For each model rule and each crypto market with at least 1,200 hourly bars (about 50 days), it tries up to 27 settings: stop-loss, take-profit and a stricter, equal or looser entry threshold. It picks the best on the earlier 60% of the history, skips 24 hours, then checks the picks on the later part it never saw. A result only counts as a win if it clears a bar that gets higher the more combinations have ever been tried, so most are thrown out.
A winner is started as a paper bot: a paper clone with the winning model equipped. That bot's live entries are decided by direction mapping (Delta Hunter), not by the rule that won; the winning stop-loss and take-profit become a whole-bot pause. Step 5 of the worked example below shows this on a real run. The 9 models it cannot start as free paper bots are PRION, SYNAPSE and TELOMERE (paid models; the lab never buys) and the six strategy rules above. Those six are not entries in the /adm catalog at all, so they cannot be equipped on a bot; a win for one of them is reported only. A win that needs a changed entry threshold is also reported only, because a paper bot cannot run a changed entry rule yet.
Where to look: Quant Lab (Experiments)/adm/perpetual/agents
3. Backtested Strategy versus real trades: which can be optimized?
A Backtested Strategy card shows a modeled result: its rules were replayed on past prices (the period is on the bot's page, for example 30 days or 6 months). No orders were placed to make that number.
A bot with wins and losses, such as Sync Master or Delta Hunter, shows a record: each win or loss is a bet the agent actually placed on our perpetual markets. Your own clones on /my-agents also have records: every trade they close on paper or live is counted.
The Quant Lab has two parts. Its Bots section runs experiments: a paper clone of a Backtested Strategy bot, or one of your own paper bots, judged on its own paper trades against your rule (minimum trades, minimum return, maximum drawdown, several windows in a row). When one fails, the lab can propose one change and try again, if your rule allows it. Its Hero-agent search optimizes model rules, not bots, as described above.
What cannot go in: a live bot (the lab tests paper bots only) and the perpetual-market agent cards themselves. A paper clone you made of any bot can go in.
Where to look: /perpetual/agentsQuant Lab (Experiments)
4. Where is the Auto-optimize ready tag?
On /my-agents, and only while you are signed in. It is not on /perpetual/agents. It went live on 27 October 2026.
It appears next to your own active cloned bots that have at least one equipped model the search can replay (one of the ten biomimetic models). A bot that carries only probability models such as RUPTURE, ENDGAME or ANCHOR, or no model at all, does not get the tag.
It means the Quant Lab search can look for better stop-loss, take-profit and entry settings for that bot's model. It does not change your bot by itself. The Auto-optimize button next to it starts the full search (not only for that bot) from your company, and the results appear on the Experiments page. It needs the Pro plan (Max 5x and Max 20x include it) and a company. Paper only: going live always waits for your approval.
Where to look: /my-agentsQuant Lab (Experiments)
5. Head-to-head: what is the Cymetica-improved version?
/h2h runs two of your sides over the same window and scores them from their real paper trades: return, win rate, drawdown and an equity curve. The original side is your own bot trading through the API or MCP in paper mode, or a paper clone you own. The improved side is a paper clone you own.
The page calls the improved side the Cymetica rebuild, and its first step is Build it Better: you ask for your bot to be rebuilt, and the rebuild arrives on your account as a paper clone. So the intended improved version is a Build it Better rebuild of your own bot, not a Quant Lab search result.
Head-to-head itself does not change or optimize any bot. You may put any paper clone you own on the improved side, including one the Quant Lab started. The Quant Lab uses the same scoring for its paired experiments. A verdict appears once each side has closed 3 trades.
Where to look: /h2h/build-it-better
6. One worked example: a real search, start to finish
One real Quant Lab search followed from start to finish, in the five steps a newcomer asked for. The run is search no. 3 of the lab that asked for this example: the daily search of 27 October 2026 (started 13:29 UTC, finished 14:01 UTC), which tried 1,728 combinations on 12 markets, shortlisted 5 and found 0 winners. Every number below is read from that run's stored rows or recomputed with the code's own formula; the only invented thing is the winner in step 5, and it is labelled.
Step 1. The rule: stockbull_blue
stockbull_blue is the StockBot (Blue) rule. It reads hourly candles of one market (open, high, low, close, volume) and computes four indicators on the closes: EMA 9, EMA 21, RSI 14 and MACD 12/26/9. It buys (opens a long position worth 10% of its capital, up to 5 open at once) when, on a closed bar, EMA 9 is above EMA 21 and RSI 14 is below 65 and the MACD histogram is above 0; the order fills at the next bar's open. It sells (closes every open position) when EMA 9 is below EMA 21 and RSI 14 is above 50. A single position also closes when it is down by the stop-loss or up by the take-profit from its entry price. Each trade pays a 0.1% fee per side and 5 basis points of slippage. Its own stop-loss is 5%; it has no take-profit of its own.
All 16 rules the search replays, one line each, read from each rule's code rather than from its catalog description. Every rule also closes a position at its stop-loss or take-profit. The six Blue/Red rules belong to Backtested Strategy bots; a Red rule is the same pair of conditions traded short: it opens a short on the sell condition and covers on the buy condition.
| Rule | Opens | Closes | Reads | Size | Own stop-loss |
|---|---|---|---|---|---|
| HELIX | a long when the MACD histogram is above 0 and EMA 9 is above EMA 21 | when the MACD histogram is below 0 and EMA 9 is below EMA 21 | hourly candles; MACD 12/26/9, EMA 9, EMA 21 | 10% of capital, up to 3 open | 5% |
| ENZYME | a long when RSI 14 is below 30 | when RSI 14 is above 70 | hourly candles; Bollinger bands 20/2 (not used), RSI 14 | 8% of capital, up to 3 open | 4% |
| CORTEX | a long when RSI 14 is below 40 and the MACD histogram is above 0 | when RSI 14 is above 65 and the MACD histogram is below 0 | hourly candles; RSI 14, MACD 12/26/9, SMA 50 (not used) | 6% of capital, up to 2 open | 6% |
| IMMUNE | a long when RSI 14 is below 35 and volume divided by its 20-bar average is above 1.5 | when RSI 14 is above 65 and volume divided by its 20-bar average is above 1.3 | hourly candles; RSI 14, 20-bar volume average, SMA 20 (not used) | 7% of capital, up to 2 open | 5% |
| CIRCADIAN | a long when RSI 14 is below 35 and the Bollinger band width (upper minus lower, in price) is below 500 | when RSI 14 is above 70 | hourly candles; ATR 14 (not used), Bollinger bands 20/2, RSI 14 | 8% of capital, up to 2 open | 7% |
| MITOSIS | a long when EMA 9 is above EMA 21 | when EMA 9 is below EMA 21 | hourly candles; EMA 9, EMA 21, ATR 14 (not used) | 5% of capital, up to 4 open | 4% |
| CRISPR | a long when RSI 14 is below 35 and the MACD histogram is above 0 and volume divided by its 20-bar average is above 1.2 | when RSI 14 is above 70 and the MACD histogram is below 0 | hourly candles; RSI 14, MACD 12/26/9, Bollinger bands 20/2 (not used), 20-bar volume average | 12% of capital, up to 2 open | 3% |
| PRION | a long when RSI 14 is below 30 | when RSI 14 is above 75 | hourly candles; Bollinger bands 20/2 (not used), RSI 14, SMA 50 (not used) | 10% of capital, up to 3 open | 5% |
| SYNAPSE | a long when EMA 21 is above EMA 50 and RSI 14 is below 45 | when EMA 21 is below EMA 50 and RSI 14 is above 55 | hourly candles; EMA 21, EMA 50, RSI 14, ATR 14 (not used) | 6% of capital, up to 3 open | 6% |
| TELOMERE | a long when RSI 14 is below 30 and EMA 9 is above SMA 20 | when RSI 14 is above 70 and EMA 9 is below SMA 20 | hourly candles; RSI 14, ATR 14 (not used), SMA 20, EMA 9 | 5% of capital, up to 2 open | 4% |
| stockbull_blue (StockBot (Blue)) | a long when EMA 9 is above EMA 21 and RSI 14 is below 65 and the MACD histogram is above 0 | when EMA 9 is below EMA 21 and RSI 14 is above 50 | hourly candles; EMA 9, EMA 21, RSI 14, MACD 12/26/9 | 10% of capital, up to 5 open | 5% |
| stockbear_red (StockBot (Red)) | a short when EMA 9 is below EMA 21 and RSI 14 is above 50 | covers when EMA 9 is above EMA 21 and RSI 14 is below 65 and the MACD histogram is above 0 | hourly candles; EMA 9, EMA 21, RSI 14, MACD 12/26/9 | 10% of capital, up to 5 open | 5% |
| cryptobull_blue (CryptoBull Bot (Blue)) | a long when RSI 14 is below 30 and volume divided by its 20-bar average is above 1.5 | when RSI 14 is above 55 | hourly candles; Bollinger bands 20/2 (not used), RSI 14, 20-bar volume average, EMA 9 (not used) | 8% of capital, up to 3 open | 4% |
| cryptobull_red (CryptoBull Bot (Red)) | a short when RSI 14 is above 55 | covers when RSI 14 is below 30 and volume divided by its 20-bar average is above 1.5 | hourly candles; Bollinger bands 20/2 (not used), RSI 14, 20-bar volume average, EMA 9 (not used) | 8% of capital, up to 3 open | 4% |
| hybrid2_blue (Hybrid 2 (Blue)) | a long when RSI 14 is below 40 and the MACD histogram is above 0 | when RSI 14 is above 65 and the MACD histogram is below 0 | hourly candles; RSI 14, MACD 12/26/9, SMA 20 (not used), ATR 14 (not used) | 10% of capital, up to 5 open | 5% |
| hybrid2_red (Hybrid 2 (Red)) | a short when RSI 14 is above 65 and the MACD histogram is below 0 | covers when RSI 14 is below 40 and the MACD histogram is above 0 | hourly candles; RSI 14, MACD 12/26/9, SMA 20 (not used), ATR 14 (not used) | 10% of capital, up to 5 open | 5% |
Reads = the indicators the rule computes on hourly candles; one marked 'not used' is computed but named in neither condition, so it never changes a decision. Size = the share of capital per position and how many positions may be open at once. Own stop-loss = the rule's default, which the search varies (0.6x, 1x, 1.6x).
Step 2. One market and the settings it is tried with
The market is DOT (Polkadot against USDT, hourly bars). The run searched 12 markets: ADA, ATOM, AVAX, BNB, BTC, DOGE, DOT, ETH, LINK, LTC, SOL and XRP; every rule is tried on every market. For stockbull_blue on DOT the grid was 9 settings: stop-loss 3%, 5% or 8% (0.6x, 1x and 1.6x of the rule's own 5%) times take-profit 8%, 15% or 30%. Position size, the number of open positions and the entry conditions were the rule's own.
This run came one day before the entry threshold joined the grid (28 October 2026). A search started today tries each of those 9 settings with the rule's entry as it is, stricter and looser (for stockbull_blue: RSI 14 below 65, below 60, below 70), so up to 27 settings per rule and market; a rule whose entry has no movable number (HELIX, MITOSIS) still gets 9. A full search today is 4,752 combinations instead of 1,728.
Step 3. The search period and the result of each setting
The run pinned 90 days of DOT hourly bars, from 29 July 2026 14:00 UTC to 27 October 2026 13:00 UTC: 2,160 bars (the run stores the first and last bar of its window; the count was re-read from the same window). The search period is the earlier 60%: 1,296 bars, 29 July 14:00 to 21 September 2026 13:00 UTC. The next 24 bars (21 September 14:00 to 22 September 13:00 UTC) are skipped so indicator warm-up cannot leak across the cut. The search sees nothing after that. The 9 settings on the search period:
Each setting is ranked by its return minus half its maximum drawdown. The best, stop-loss 3% / take-profit 30%, scores 29.60 − (5.13 / 2) = 27.04 and was shortlisted: at most one setting per rule-and-market pair goes forward, at most 5 overall, each with at least 8 trades and a gain. The other 8 were thrown out as 'not shortlisted on the search period'. All 9 made money on the search period, which is exactly why a held-back check is needed.
| Stop-loss | Take-profit | Search-period return | Trades | Max drawdown | Win rate | Verdict |
|---|---|---|---|---|---|---|
| 3% | 8% | +12.79% | 108 | 5.65% | 48.1% | not shortlisted |
| 3% | 15% | +21.34% | 99 | 5.13% | 48.5% | not shortlisted |
| 3% | 30% | +29.60% | 94 | 5.13% | 46.8% | shortlisted |
| 5% | 8% | +11.56% | 108 | 6.07% | 48.1% | not shortlisted |
| 5% | 15% | +20.39% | 99 | 5.17% | 48.5% | not shortlisted |
| 5% | 30% | +28.59% | 94 | 5.2% | 46.8% | not shortlisted |
| 8% | 8% | +10.15% | 108 | 6.08% | 48.1% | not shortlisted |
| 8% | 15% | +18.70% | 99 | 5.63% | 48.5% | not shortlisted |
| 8% | 30% | +26.78% | 94 | 5.63% | 46.8% | not shortlisted |
Step 4. The held-back period, the bar and the verdict
The held-back period is the rest: 840 bars, 22 September 2026 14:00 UTC to 27 October 2026 13:00 UTC. Only the 5 shortlisted combinations of the whole run were replayed there: stockbull_blue on DOT, LINK, AVAX and SOL, and HELIX on DOT.
The bar. N is the number of combinations this lab has ever tried: 1,728 (all of them in this, its first, search). The test is one-sided at 5%, split over N: z is the normal quantile at 1 − 0.05 / 1,728 = 4.0213. A held-back result must beat z × sd × √trades, written as a percentage of the $1,000 starting capital and never less than 1%, where sd is the standard deviation of that replay's own trade results. For stockbull_blue on DOT (SL 3 / TP 30): 59 held-back trades, sd ≈ $3.23 per trade, so the bar is 4.0213 × 3.23 × √59 / 1,000 = 9.97%. The more combinations a lab has tried, the higher z and the bar: at today's 4,752 combinations z is 4.2535.
The result: stockbull_blue on DOT lost 3.09% on the held-back period over 59 trades (maximum drawdown 7.99%; both halves of the period lost, −$23.12 and −$2.98). Verdict: lost. It never got as far as the bar, and a loss is never called a win. The other four shortlisted combinations lost too, so the run ended with 0 winners and started no paper bot. That is the usual outcome: the setting that looks best on the past tends not to repeat, and the method is built to say so.
| Rule | Market | Setting | Search-period return | Held-back result | Trades | Bar | Verdict |
|---|---|---|---|---|---|---|---|
| stockbull_blue | DOT | SL 3 / TP 30 | +29.60% | -3.09% | 59 | 9.97% | lost |
| helix | DOT | SL 3 / TP 30 | +16.39% | -1.94% | 43 | 8.62% | lost |
| stockbull_blue | LINK | SL 8 / TP 30 | +23.25% | -12.30% | 74 | 10.77% | lost |
| stockbull_blue | AVAX | SL 8 / TP 30 | +20.90% | -6.87% | 74 | 5.84% | lost |
| stockbull_blue | SOL | SL 8 / TP 30 | +20.24% | -10.01% | 90 | 8.24% | lost |
Step 5. What happens to a winner
No combination in this run won, so the winner here is invented and labelled as such. Invented winner A: stockbull_blue on DOT, stop-loss 3% / take-profit 30%, gaining 12% on 59 held-back trades with both halves positive, above its 9.97% bar. The lab writes 'kept' on the row with the reason 'gained 12.00% on the held-back period (59 trades), above the bar of 9.97% for 1,728 combinations tried', then checks whether it can start a paper bot. For stockbull_blue it cannot: the rule is not an entry in the /adm catalog, so it is not free to equip, and the row gets 'not started as a paper bot: this bot model is not free to equip'. The win is reported in the lab's results and log, and no bot of any kind is created.
Invented winner B: the same numbers for HELIX on DOT (HELIX on DOT was in this run's shortlist). HELIX is a free /adm model, so the lab starts a paper bot: a paper clone of the Backtested Strategy bot CryptoBull Bot (Blue) (for a short rule it would be CryptoBull Bot (Red)), titled 'helix · DOT · SL 3 / TP 30', with stop-loss 3%, take-profit 30% and market DOT written onto it and HELIX equipped in model slot 1. It is funded from your paper balance with the capital in your lab rule and switched on; its card carries the note 'Chosen by the search: gain +12.00% on the held-back period, 59 trades, 1,728 combinations tried (bar 9.97%)'.
Which rule does that paper bot trade? The direct answer
Not the rule that won. A paper bot made from a Backtested Strategy bot gets its live entries from a direction mapping, not from the template or the equipped model: an upside bot resolves to DELTA (Delta Hunter), a downside bot to FADE. Delta Hunter looks at the market's last 5-minute epoch: after a move larger than 0.5% it bets on a reversal, otherwise it follows the move. It reads no EMA, RSI or MACD. The equipped HELIX is marked backtest only and does not steer an order. The same is true of a model you equip from /adm: today it changes backtests and the search, never a live entry. So the bot started from invented winner B is a Delta Hunter bot on DOT, which is exactly what the Strategy tab of your own clone says.
The winning stop-loss and take-profit are applied, but with a different meaning. In the search they were per-position exits: one position closes at −3% or +30% from its entry. On the paper bot they are the bot-level pause the Strategy tab shows: when the bot's overall result since it was started reaches −3% it stops trading with the reason stop-loss, and at +30% with the reason take-profit. Nothing on the bot closes one position at −3%.
In one sentence: a Quant Lab win is a backtest finding about a rule; the paper bot it starts trades Delta Hunter with the winning numbers used as a whole-bot stop, and a win for one of the six Blue/Red rules starts no bot at all. We state this as a gap, not a feature.
Needs a build: why the paper bot does not trade the winning rule yet
The platform already has a bot that runs an indicator rule live: the blank bot ('Blank bot (your rule only)' when you clone) evaluates the backtester's own rule grammar (EMA, RSI, MACD, Bollinger, volume) on closed candles of any chart interval, 1 hour included, with per-position stop-loss and take-profit, and fills paper trades at the live spot price. What is missing is the bridge. The Quant Lab's winner path creates a CryptoBull clone and equips a model; it would instead have to create a blank paper bot, write the winning rule onto it (its indicators, its buy and sell conditions, the 1-hour interval, the market, and the stop-loss and take-profit as per-position exits), teach the lab's scoring to read that bot's positions, and keep short rules reported-only until the blank bot can short on the spot lane. Until that is built, every bot the search starts is a Delta Hunter bot, and this page says so.
Where to look:Quant Lab (Experiments)/adm/my-agents
Glossary
| Term | Where it appears | What it is |
|---|---|---|
| Model (algorithm) | /adm | A catalog entry you equip on a bot: 10 biomimetic and 10 probability models, up to 5 per bot. Used in backtests and by the Quant Lab search. |
| Skill | /skills-n-tools | A catalog entry you equip on a bot: 29 indicators, data feeds, filters and modifiers, in 3 primary and 3 secondary slots. |
| Bot / agent | /perpetual/agents, /my-agents | The thing that trades. Your clone is your bot. |
| Backtested Strategy | /perpetual/agents | A Cymetica bot whose number is a replay on past prices, not trades. |
| Perpetual-market agent | /perpetual/agents | A platform agent (Sync Master, Delta Hunter ...) whose wins and losses are bets it placed. |
| Paper / live | /my-agents | Paper trades with play money. Live trades with your own funds. |
| Quant Lab | /company/app/experiments | Your company's testing lab: Bots (paper experiments) and the Hero-agent search. |
| Hero-agent search | /company/app/experiments | A backtest search over 16 model rules on hourly crypto prices, checked on a held-back period. |
| Auto-optimize ready | /my-agents | A tag on your own cloned bot whose equipped model the search can replay. |
| Head-to-head | /h2h | Two of your paper sides, same window, scored from their real paper trades. |
| Build it Better | /build-it-better | Cymetica rebuilds your bot; the rebuild arrives as a paper clone. |
What we could not confirm
Whether an equipped skill changes a running bot's orders: we did not find a trading path that reads equipped skills (Memecoin Mining runs as its own feature), so we make no claim either way.
The pairing of the stockbear_red rule with the StockBot (Red) card is by description (both are the short-selling stock bot); the card's own web address says stockbull-red.
For AI agents
A plain-text copy of this page is at /how-it-fits.txt and the same summary is in /llms.txt.