The Lab assistant
Ask the Lab in plain words — what the assistant can and cannot do, Standard and Deep, credits per question, chats at once per plan, stopping, and how its searches stay honest.
What this is for. The assistant is a research partner inside the Strategy Lab. Describe an idea — "I'm new, I have $100, find something simple to test" — and it reads the data, screens wallets, proposes angles, saves strategies, runs backtests, starts paper runs (after asking you) and reports what the tests measured. It is research, never a recommendation to buy or sell. It is part of every paid plan from Pro Trader (plan table).
Simulated performance
Simulated. Replayed on past trades. Not a forecast or advice. Backtests replay the past at prices a follower could have paid; paper trading scores live signals on real order books without placing orders. Neither is a forecast or advice, and both carry a measured optimism.
These results are based on simulated or hypothetical performance results that have certain inherent limitations. Unlike the results shown in an actual performance record, these results do not represent actual trading. Also, because these trades have not actually been executed, these results may have under- or over-compensated for the impact, if any, of certain market factors, such as lack of liquidity. Simulated or hypothetical trading programs in general are also subject to the fact that they are designed with the benefit of hindsight. No representation is being made that any account will or is likely to achieve profits or losses similar to these being shown.
Ask a question
Open polyrank.app/lab and press New chat in the rail. Start from one of the three starters (I'm new, A market I follow, I know what I want), a template, or type your own question in Ask about traders, markets or your strategies….
Pick the mode beside Send (Mode for this chat): Standard or Deep. The line under the box says which mode the chat starts in and what a typical question costs.
Send. The assistant's steps appear in plain words as it works (Read the templates, Saved a new strategy, Ran a backtest, …); each answer ends with its cost in credits.
Stop halts an answer and keeps what was already found (You stopped the assistant. What it finished is kept.). Delete this chat removes a conversation; chats are listed under Chats in the rail, with a dot while one is answering.
Standard and Deep
| Mode | Is | Costs |
|---|---|---|
| Standard (default) | Quick answers | ≈ 13 credits a typical question |
| Deep | Slower, more thorough | ≈ 2× the credits — Lab Pro and up |
A chat keeps the mode it started in; start a new chat to switch. A new chat starts in the mode you picked last. Polyrank does not name the model in the product; the privacy policy names the provider and what it receives (the conversation, never your name, email or wallet).
Credits
Each question is charged in Lab AI credits — the sum of the model calls it made. A question needs a balance of at least 25 credits to start; the plan's monthly allowance is spent first, then top-ups. The mode menu shows Next question ≈ N credits and your balance. Backtests, paper runs and searches the assistant runs do not spend credits; they count against the plan's Lab limits like any other.
Chats at once
Each plan answers a set number of chats at the same time (queued or working) — the Chats answering
at once row of the plan table; comped Desk seats answer 6 and Enterprise 8. Asking
in one more chat while that many are answering is refused until one finishes: "2 of 2 conversations
are answering; wait for one to finish." By API, 429 lab_agent_busy with running and limit. A
chat whose answer is out while a backtest it started still runs does not count.
Long backtests, the queue and the first answer
- A long backtest does not hold the answer. The assistant waits a few seconds for a backtest it started; a longer one keeps running and the answer goes out at once with a Backtest running card. When the run ends, its result card is posted to the same chat by itself (no credits).
- Waiting your turn. While a question waits for the assistant, the chat says Queued, n ahead.
By API,
GET /v1/lab/agent/conversations/{id}carriesqueue_position(questions of every user ahead of it) andawaiting_backtests(results still to come). - The first answer is quick. A chat's first answer makes a limited number of tool calls (more in Deep mode); when it reaches them it answers with what it found, says what it did not check yet and offers to go deeper.
From code
With an API key, the assistant can be asked outside the app: POST /v1/lab/agent/conversations starts a chat, …/{id}/messages replies and …/{id}/stop stops an
answer — each spends the account's AI credits like a question in the app. The MCP server
exposes the same as polyrank_lab_ask, polyrank_lab_reply, polyrank_lab_stop,
polyrank_lab_list_chats and polyrank_lab_get_chat.
What it can do
The assistant works through tools, each shown as a plain-words step:
- Wallets and metrics — the metric catalog, rank or screen wallets as of a date, read a wallet's history month by month.
- Strategies — list templates and your strategies, create and update them, archive one (it asks first), list their runs.
- Runs — start and read backtests; start a paper run (it asks first) and read paper runs; the catalog of own-rule conditions.
- Your account — your follows, watchlists, alerts and plan.
- Data — the research tables described in Research and data, read-only, and a market's details and live order book from Polymarket.
A backtest it starts goes to the same queue as yours, under your plan's limits.
What it cannot do
It cannot place orders, move funds or turn on live trading. It does not know the future: every backtest or paper figure in an answer is simulated and marked as such, with the simulated-results statement one click away. Its answers are research, not advice.
Searches
When an idea has more than three variants, the assistant pre-registers a search instead of
trying variants one by one: what varies, what stays fixed, the budget in backtests and minutes, and a
window whose most recent part stays unseen. Above three backtests it asks for your go-ahead (a search
is capped at 60). It reads the chosen variant on the unseen hold-out once per strategy, then
reports how many variants were tried, the chosen variant against its control, the best of N
(in-sample), the deflated statistic and the overfit risk, and says plainly when nothing stands out
from luck. The search card shows its state (waiting for your go, running in the background,
reading the hold-out, finished). A search spends backtests, not credits. By API:
GET /v1/lab/agent/searches/{id}.
Related
- Lab AI credits — allowance, top-ups and your daily ceiling
- Strategies · Backtesting · Paper trading
- Research and data — what it can query
Your own rules
Strategies that follow no wallet — conditions on price, spread, depth, moves, volume and timing, checked on the recorded order books, with their baseline and the sealed final test. Lab Pro and up.
Research and data
What history the Lab's backtests, rules and assistant can read — fills since 2022, recorded order books since July 2026, daily metric snapshots — the time ranges, and the known gaps.