Idea in. Evidence out.
The pipeline, in plain English
strategy idea → clear rules → recorded market data → trade check → final result check → profit/loss → reproducible report
Two halves that never mix: AI compiles (turns the messy human claim into exact rules) and the engine proves (a fixed, rule-based program with no AI in it produces every number you see on this site).
Clear rules: the decoder
AI reads the strategy as written and asks exact questions: buy at what price? how much? when exactly? Every decision the author never made is filled with a default that is disclosed and counted. The number of guesses is printed on the record. If the rules are hidden, paywalled, or missing, the output is a refusal, with reasons.
Real results across 14 real public strategies: 1 compiled with zero guesses, 6 compiled with counted guesses, 7 refused. Three separate AI models read the clearest source and produced exactly the same rules.
Recorded market data: the snapshot
Every test runs on a saved market-data snapshot with a fingerprint. That locks exactly which data produced the result, so the test can be rerun and checked later. Data that can't be identified can't be replayed. By design. If anyone edits a saved snapshot, verification fails before a report is produced.
Technical name: DatasetVersion. Largest on record: 4,380,961 market events from one month of live prediction-market data, with real final results attached.
Trade check: could each trade actually go through?
A signal is not a trade. For every trade the strategy wants, Proofline checks the order book: was there enough available, at that price, at that moment? Full availability = a verified trade. Not enough = a partial trade, counted as partial. No proof at all = the trade is refused, not imagined.
This check was tested against 352 real trades from a live bot (a "calibration"): assuming a shown price always fills was wrong about 1-in-5 times on Polymarket and 1-in-3 on Kalshi; the strict availability + freshness check was right 93 of 93 times. Known limit, disclosed on records: the simulator does not yet enforce a quote-freshness cutoff.
Fees: counted on every trade
Venue fee schedules are applied to every simulated trade. Small numbers, but they decide marginal strategies: the one-month replay on this site paid $158.43 in fees inside its $1,500.90 loss.
Final result check: settlement
Every position is closed with the market's official final result (win or lose), taken from the venue itself. If a market hasn't resolved, the position stays open and the report says "pending." Proofline never guesses an outcome from price direction. Where two venues disagreed about the same window (it happens 5.4% of the time in the recorded month), both results are kept, separately.
Trade record + report: rerun it and get the same answer
Every signal, trade, fee, and final result lands in a step-by-step trade record (the ledger). The finished report carries fingerprints of the data, the rules, the checking model, and itself. The test that matters: rerun the record. The output must be exactly the same. The 8-hour live paper test on this site was rerun from its own recording and produced the identical trade record.
Plain-English glossary
Every technical term used anywhere on this site, translated.
Fill: Would the trade actually go through?
A fill means the market had enough available at the shown price for the trade to happen.
Depth: Was there enough available to buy or sell?
Depth checks whether the order book had enough shares/contracts at that price.
Depth verified: The book had enough available for the trade.
Proofline checked the live order book and confirmed the simulated trade was possible at that size.
Partial fill: Only part of the trade could go through.
The strategy wanted more than the market had available at that price.
Settlement: Final market result.
Settlement is how the market officially resolved: YES or NO, win or lose.
Replay: Rerun the strategy on recorded market data.
Replay means Proofline ran the same strategy over saved market data to see what would have happened.
Ledger: Trade record.
The ledger is the step-by-step record of signals, trades, fees, and final outcomes.
Hash: Fingerprint.
A hash is a fingerprint of the report/data. If the data changes, the fingerprint changes.
DatasetVersion: Saved market-data snapshot.
A DatasetVersion locks which market data was used so the result can be checked again.
P&L: Profit or loss.
P&L means how much the strategy made or lost after trades and fees.
P&L computable: Profit/loss can be calculated.
Every open trade had a final result, so Proofline can calculate the full outcome.
Oracle: Market result source.
The oracle is the source that decides how the market officially resolves.
Calibration: Checking the simulator against real fills.
Calibration compares Proofline's fill model against real trading logs to see whether the simulator is realistic.
Byte-identical: Exact same result.
When Proofline reran the record, it produced the exact same output.
For technical readers: reproducing a record yourself
The engine is a Python CLI. Given a record's rule and data fingerprints:
# rebuild the data snapshot (its fingerprint must match the record) proofline import-c3 --path <tape> --out <dir> --dataset-id <id> # rerun the pinned rules under the pinned fill model proofline replay --spec spec.json --dataset <dir> --fill-model v0.1 # reports are reproducible to the exact byte; fingerprints must match proofline compare --spec spec.json --dataset <dir> --out-dir <out>
The per-record fingerprints live in each report's "Fingerprints" section.
What the numbers on this site are
Where the data comes from
Proofline uses recorded market data and venue/source result data where available. Some records use historical market data, live paper captures, venue result sources, or other source-backed datasets. Each report explains what was used. Different markets have different data quality, so Proofline labels what can and cannot be verified. When data is missing, the report says so instead of filling the gap.
Proofline tells you what actually happened under the rules and data shown. See the reports →