Talos

The creed

Four tenets. Each one is a line of code, not a slogan — the enforcement is named beside it.

Stay your blade from the flesh of bad setups. The pre-trade gate refuses. size_position() returns a reason rather than a guess when the stop is too wide for the risk cap, and can_place() blocks a trade that would breach the daily loss limit or the trailing drawdown — before the order is built.
Take profit for the innocent. Every order is a bracket. build_order() attaches a stop-loss and a take-profit in ticks at construction time and refuses to build an order that carries more risk than the cap. There is no naked entry to forget to protect.
We work in the dark to protect the capital. Dry-run is the default and every decision is written to an audit trail. A live order needs dry_run == False and is_armed() == True; arming expires after 24 hours and does not survive a reboot.
Nothing is permissible. Everything must be checked. Seventy-seven checks pass before anything is trusted, and the caps live in a frozen dataclass rather than a config file. The dry path enforces the identical gate — only the send differs — so the logic is proven before a dollar is at stake.

Discipline is our edge. Execution is our duty.

We do not predict. We execute.
We do not hope. We verify.

The doctrine

Most trading systems are built strategy-first: find an edge, then bolt on risk management once real money is already moving. Talos was built the other way round. The safety layer was finished, tested and left disarmed — because nothing in the research programme has passed a gate yet, and arming without one is how accounts die.

Five invariants, each enforced in code rather than configuration:

Caps in code: Risk limits live in a frozen dataclass, not a config file. A limit you can edit at 2am is not a limit.
Two keys: A live order requires dry_run == False and is_armed() == True. The holstered default keeps everything dry regardless of arming state.
Auto-disarm: Arming expires (24h default) and does not survive a reboot. There is no persistent armed state to forget about.
Kill file: A file on disk halts placement unconditionally, checked every tick. Stopping must be easier than starting.
Operator only: The assistant never arms this system. Not a preference — a standing rule, and the reason the count above still reads zero.

What the pre-trade gate checks

Every order passes one function before it can exist. It enforces per-order and per-day caps, one-shot-per-signal, and — since 30 July 2026 — the account limits that actually end funded accounts.

The defect that motivated this

The gate used to ask “have I already breached the daily loss stop?” It never asked “would this trade breach it?” It was reactive: it noticed a breach after the fact. Worse, it counted realised P&L only, while a prop account’s daily limit includes unrealised — and it had no concept of the trailing drawdown at all.

It now refuses a trade that would breach either limit, before the order is built.

A worked example, from a real logged signal

One of the research streams fires with stops roughly 373 index points wide on MNQ. That is $746 of risk per contract — 75% of a $1,000 daily loss limit on a single contract. Two such losses in one day would breach the account outright.

Eleven months of expectancy analysis never surfaced that, because expectancy and survival are different questions. The instrument below answers the second one.

The instrument

Three prop-firm rules are misunderstood often enough to be worth stating plainly. All three are implemented here exactly as the firms actually apply them:

The trailing drawdown follows your peak end-of-day balance — not intraday highs — and locks once it reaches your starting balance.
The daily loss limit is measured from the day’s opening balance and includes unrealised P&L.
The consistency requirement is a rising target, max($3,000, 2 × best day) — not a fixed bar.

This runs entirely in your browser. Nothing is transmitted, nothing is stored anywhere but this device, and it makes no prediction of any kind — it answers only what a trade costs you against the rules.

Pre-trade guard

Account size Peak EOD balance Balance at day open Realised today Open unrealised

Daily loss limit—

—of $1,000

Trailing drawdown · TRAILING—

—buffer $2,000

Instrument Entry Stop Contracts

—

Enter a trade

Set an entry and a stop above.

Stop distance	—
Risk per contract	—
Total risk if stopped	—
Share of the daily limit	—
Maximum size within both limits	—
Consecutive losses before the day ends	—

Where Talos sits

Talos is the last step in a chain, not a standalone product. It is deliberately the least autonomous agent in the constellation: everything upstream decides whether to act, and Talos only decides how — and refuses when the answer is that it shouldn't.

Augur: Reads the evidence. Every strategy is pre-registered with a kill criterion before the first data point, and nothing reaches Talos until a forward gate passes. The evidence ledger is why the order count here is still zero — it records what has been falsified, not what has been promised.
Mercury: Runs the commercial side. Trading capital comes from work that has actually been paid for, which is the honest sequence: fund the account from revenue, not from hope.
Aion: The long-cycle view. Talos operates on the timescale of a single trade; Aion holds the one where a six-year drawdown is a normal feature rather than an emergency.
Vulcan: Builds the artefacts — this page and its emblem among them.
Sovereign-Minds: The constellation these agents belong to, and the discipline they share.

Status — holstered

Increments one through five are complete: safety core, dry-run loop with an audit trail, broker transport, arming with a kill file, and drift monitoring. 77 checks pass. No real broker API has ever been contacted.

Arming waits on a forward gate whose decision rule was frozen before any outcome was visible, and which reads in late September at the earliest. Until then there is nothing to arm for — and the honest reason the order count is zero is not restraint, it is that no strategy has earned it.

THE EVIDENCE FOR THAT CLAIM — INCLUDING EVERY FALSIFIED STRATEGY — LIVES IN THE AUGUR EVIDENCE LEDGER →

A trading system that has never traded is an odd thing to publish. It is here because the engineering claim is the point: the safety layer exists, it is tested, and it is disarmed — and that order of operations is the whole argument.