Last Updated: 2026-10-26
Risk: This is a design for offline research, not a tested bot, live-trading instruction or earnings forecast. Automation can repeat errors faster than a person. Trading can lose the entire stake.
“A bot made money” is not enough to reproduce a strategy. You need the decision rule, the information available at that time, the resulting orders and fills, costs, open inventory and failures. Neither the previously repeated 0x8dxd story nor the “1,322% agent” claim supplies that record here. They are not performance evidence for any strategy below.
Identify the software before evaluating its result
Save an official product or repository URL, owner, exact version or commit, license, documented features, pricing terms and the date checked. A familiar name or integration logo does not establish that a particular bot exists with the features an article attributes to it. If that record is missing, mark the product unverified rather than inventing a ranked review.
A test report should identify the software configuration and supply raw records, not just a percentage. For example, 127 wins out of 217 completed trials is about 58.5%; 128 is about 59.0%. A claimed 58.7% needs a different, explicitly defined denominator or weighting. Win count still says nothing about net profit without each trial's stake, price, payoff and costs. Separate unresolved trials instead of silently treating them as wins or dropping them.
Compare the work required as well as the output. A human may interpret an ambiguous rule; software may apply a defined rule consistently to more records. Either can make systematic errors. Automation still requires monitoring and maintenance, and public wallet activity alone does not reliably classify a trader as human or bot. No survival-rate or profitability winner follows without a reproducible, comparable cohort.
Hypothetical operating budget: $40 hosting + $60 model calls + $20 data = $120 per month before any trading costs or human time. That bill exists even if no hypothesis shows an edge. Record recurring and one-off costs separately, and compare with a read-only manual or paper workflow before buying a tool. This is a budget example, not a claim about a named product's price or a return target.
Give each strategy the right test
Scroll sideways to see every column. Keyboard: focus the table region and use the arrow keys.
| Strategy hypothesis | Record in the paper test | Essential failure case |
|---|---|---|
| Quote both sides and capture a spread | Queue assumptions, every fill/cancel, signed inventory, marked inventory value and costs | Only the stale side fills after news; spread revenue is smaller than inventory loss |
| Buy a complete outcome basket below its payout | Exact rules, equal outcome quantities, executable depth, costs and unmatched legs | One leg fills and another moves or never fills |
| React to a scheduled source release | Source arrival time, parsed content, quote time, rule matching and decision time | Stale or misparsed data sends the wrong decision; a later timestamp leaks into the test |
| Copy selected wallet activity | Verified address, observed fill time, available copy price, cash flows and unknown hedges | The copied wallet's apparent profit is unrelated to the observed trade or no longer available |
These experiments cannot share one invented “typical annual return” table. Their data, execution assumptions and risks differ.
Market making: cash received is not profit by itself
Suppose a paper strategy buys 100 YES tokens at $0.48. It then sells 40 at $0.52 before news changes the executable bid for the remaining 60 to $0.30. It has paid $48, received $20.80 and holds inventory marked at $18. Total marked value is $38.80, a $9.20 loss before costs, despite selling 40 tokens above their purchase price.
Selling tokens reduces inventory; it does not make the seller newly long. Track the signed inventory after every fill. A test that records only spread earned, while ignoring unsold tokens and the prices available to exit them, does not measure the strategy's result.
Complete baskets need matching payouts
For one fictional exhaustive binary contract, 100 YES at $0.46 plus 100 NO at $0.51 costs $97 and pays $100 under either ordinary binary outcome. Add $1 of total costs and the modeled surplus is $2. If only 60 NO fill, one outcome pays just $60. The apparent $3 gap did not remove execution risk.
Overlapping monthly events, a national result versus a state result, or a monthly-average rate future versus one policy decision do not automatically form a complete basket. Enumerate joint outcomes and exceptional settlement terms. The payoff worksheet supplies the accounting test; the provider's resolution documentation does not replace the actual contract's rules.
Model missing data and restarts
A paper harness should explicitly reject stale quotes and expired source observations. Record duplicate events, partial fills, cancellations and unknown order outcomes. After a simulated restart, reconstruct state from recorded events; do not assume every request failed or every previous order vanished. A failed data feed is not permission to keep using its last signal indefinitely.
Keep this harness unable to submit real orders. A simulated halt should persist across a restart, and resuming should require a recorded reason. Test a lost response after an apparent fill, a duplicated event, an incomplete book and a delayed resolution. State which cases the harness does not model; a successful simulation is not production readiness.
Review all results, not just the impressive run
Separate realized cash flows, marked open inventory, transaction costs, recurring infrastructure costs and deposits/withdrawals. Define the sample period and market-selection rule before inspecting outcomes. Reserve later observations for evaluation; don't retune on every loss and report the same history as independent evidence.
A useful report says which hypothesis survived which tests, where data were missing and which assumptions dominate the result. “No demonstrated net edge” is a valid result. A high win rate alone can hide rare large losses, and forecast accuracy alone does not establish executable profit.
For source handling and explicit insufficient-information outputs, use the paper AI research workflow. Access must comply with current platform restrictions. AITR may earn commissions from marked referral links elsewhere; this page does not recommend a bot or claim tested returns.