three disciplines, held constant.

there is no second meaning of the word inside numeratio. paper trading, simulated fills, and shadow-mode operation are not live and never appear on the public site under that label. the phrase is reserved for systems with real orders, real capital, and a real audit trail from the venue they trade on.

the status of every strategy is one of five values: roadmap (designed but not yet implemented), setup (under operational hardening, no capital yet), live (in production), paused (capital remains deployed but the strategy is not executing), and retired (decommissioned). the labels are applied with the same meaning across every strategy. a strategy in setup does not get the visual treatment of a strategy in live, even briefly, even when the distinction is subtle.

forward testing — running a strategy with small live capital before committing real size — is part of the live category, not a separate marketing-friendly stage. a strategy in forward testing is labeled as such so the reader can tell it apart from a strategy running at full size.

what this means for the numbers shown anywhere on this site: every figure is derived from realized fills, recorded by the strategy and verified against the venue. there are no projected returns, no expected-value estimates, no curves drawn from historical simulation. if a strategy's track record begins on a given date, that is the date the strategy reached production size; forward-testing periods are preserved internally but are not retroactively counted as track record.

the discipline this rules out is the most common pattern in algorithmic trading offerings: a backtest curve presented as evidence of viability. backtests are useful internally — they inform whether a strategy is worth building. they are not evidence of anything until the system is running. once it is running, the backtest becomes irrelevant; the real curve replaces it.

the separation is structural, not a policy that could be relaxed. each strategy's code, venue account, ledger, and schedule are its own. strategies reuse shared infrastructure — the patterns for alerting, persistence, logging, and time-keeping — but they never share running state: no aggregate ledger, no shared cache, each strategy aware only of its own positions, balance, and history. a change in one strategy's position cannot reach another's, because there is no path for it to travel.

isolation also separates credentials from execution. the keys a strategy needs — venue authentication, signing material, withdrawal authorizations — never sit on the machine that runs it; they live on a separate, hardened server and are fetched at startup, never written to the execution host's disk. so every strategy spans two servers, one holding secrets and one running the strategy, and compromising either does not, on its own, hand over the other. that two-server split is a baseline expectation in any institutional security review — here it is part of the architecture, not a setting that can be switched off.

the payoff is that one strategy cannot harm another: a bug in one's order routing cannot drain another's account, a mis-set risk parameter cannot reach another's position sizing, a drawdown in one cannot raise exposure in another — the strategies do not coordinate, net, or compose. the common alternative is a single engine running everything through shared books and shared state: cheaper to build, and cheaper to lose. running each strategy as its own system costs more engineering, and that cost is paid on purpose — a single engine across strategies has the operational shape of a fund, and numeratio is not building one.

the comparison covers everything. each strategy's own ledger — every fill, every position it believes it holds, every dollar of p&l it computed — is checked against what the venue reports independently. when the two disagree, it surfaces within hours, not at month-end.

the mechanic is unglamorous. once a day, at a fixed time, the strategy pulls the venue's authoritative positions and balances, compares them to its own records, and writes the result to an append-only audit log — each entry chained to the hash of the one before it, so any later alteration is detectable. it runs without an operator; a person is needed only when the comparison flags a mismatch. the mismatches that show up are mundane and fixable — a fill lost to a dropped connection, a fee booked under a different rounding rule, a trade the two sides date a few minutes apart — and the point is to catch them while they are still small, not to pretend they never happen.

a faster check runs between the daily ones. on a short cycle, the strategy re-verifies its live positions against the venue — every position it holds still exists there, nothing exists there it has no record of, sizes and directions agree on both sides — and is suppressed only while a position is legitimately in flux, so it never raises a false alarm mid-trade. a gap between intent and reality is caught in minutes, not at the next day's close.

the record is built to outlast the process that writes it. every reconciliation — its inputs, findings, and resolution — is preserved, and a strategy's full history is mirrored off the trading machine on its own schedule, so it survives the loss of any single host. a strategy can be paused, retired, or migrated and its audit trail travels with it; nothing written is discarded, because a track record is only as credible as the ability to reconstruct it — and that depends on the underlying data still existing.

what this means for the figures shown here: every performance number has been checked against the venue. the strategy's record is the source; the venue's record is the check. a divergence between them is never a number that gets published — it is a signal to repair something first.