# Aptura > Before a trading agent changes its own strategy, Aptura checks the change. OpenServ SERV Reasoning proposes, a past-data math gate tests, and Reppo reviews. Bad changes are blocked. Paper trading only: no wallets, no real money. ## For agents - [Manifest](/api/agent/manifest): capabilities and costs, JSON - [Supervise](/api/v1/supervise): POST a proposed weight change, get SHIP, REVIEW or BLOCK - [Proof](/proof/live-proof.json): 4 live calls with SERV response IDs and settled Reppo evaluations, including feedback revisions, recorded 14 Nov 2026 - [Evaluation](/proof/evaluation.json): full replays on 4 datasets, losses included - [Candles](/proof/candles/momentum-breakdown.csv): raw hourly candles used by the replay, hashes in evaluation.json ## Request POST /api/v1/supervise {"agentName":"my-bot","datasetId":"momentum-breakdown","currentWeights":{"momentum":0.6,"reversal":0.4},"proposedWeights":{"momentum":0.05,"reversal":0.95},"runReppoReview":false,"maxRevisions":2} maxRevisions (0-2): when a change is rejected, the gate reason and Reppo critique go back to SERV for a revised proposal, which is re-tested. The response lists every round in "rounds" and the gain in "improvement". ## Rules the gate enforces - Each strategy weight stays between 15% and 85%, and weights sum to 1 - A single change moves a weight by at most 25 points - At least 18 bars between accepted changes - The change must beat the current mix by at least +0.06% on the 24 bars before the decision, after 20 bps fee and 15 bps slippage per side - Model output and reviewer scores never override these rules ## Pricing (proposed, not billed yet) - Check: math gate only, free, rate limited - Reason: SERV + up to 2 revisions + math gate, $0.05 per change - Reviewed: plus Reppo review of gate-passing changes, $0.49 per change ## Limits - Paper trading on historical and observed prices. No order is ever sent to an exchange. - Reppo scores the reasoning, not trading profit. - Results come from 3 real datasets of 199 hourly candles and 1 synthetic one: too small to claim an edge.