The All-Seeing Trading Intelligence Platform
Automated, multi-asset trading across stocks, ETFs, forex, commodities, indices and crypto — unified by a closed-loop intelligence engine that generates signals, gates risk across themes, routes each order through one of 8 broker adapters, grades every outcome, and feeds the result back into the next decision. Built end-to-end. 17713 tests green.
Sauron isn't a signal feed or a bot framework — it's the synthesis of both, plus everything that connects them. 36 evaluators fuse price, news, macro, COT, options flow and 13F filings into composite signals. A multi-dimensional risk gate prevents stacked exposure across asset classes. 8 broker adapters stand behind one interface, and the router picks one venue for each class. Every trade is graded, every gate decision audited, every dollar tracked. Then the results feed back into how the next decision gets made.
AuditLogEntry captures every trade open,
close, and gate reject. Tamper detection via verify-chain. FIFO/LIFO/HIFO tax-lot
bookkeeping with Form-8949-ready CSV export. Daily pg_dump with rolling retention.
The opportunity scanner fuses 36 evaluator kinds into multi-modal setups.
Each setup is admin-editable, weighted, and emits an
OpportunityFlag only when the composite score crosses your threshold.
Every flag becomes a Signal that gets graded against its actual outcome —
so the platform learns which evaluators carry weight in which regime.
Specialised agents powered by Claude, each on the model tier its work calls for, from the fastest to the frontier, with Brier-score calibration tracking. Every agent prediction is scored against real outcomes — bad agents lose trust automatically. AI never sends an order. With one switch on, it may argue against one: before a real-money bot order leaves, the Executioner says why it fails and the Champion why it works, and after a first stretch in shadow the Executioner may cut the size or refuse the entry — never raise it.
Beside the trade debate's two voices sit eight more, the Emperor in the chair. With their own switches on, they sit on a real-money bot order once it has left, and on a real-money ticket when its owner asks; each is graded on what the market did next, and its record is its only weight.
Walk-forward backtester replays historical Signals through your
AssetBotConfig — SL/TP %, cooldowns, sizing — and walks PriceData bars
bar-by-bar to find each exit. With realistic transaction costs + slippage and an optional
train/test split, you see how the bot would have done, not just whether the
rule was profitable in theory. A config that wins on train but loses on test is overfit
and not shippable. The Eye tells you that before any capital moves.
Every trade flows through five stages. Each stage's output is logged, audited, and fed back to make the next decision smarter. No black box — every step is inspectable from the Eye.
Closed-loop means that, with the ladder's switch on, a real-money rule whose recent expectancy falls under half of its stage's baseline drops a rung, one measured as losing drops straight to paper, and every move is recorded with its reason, whether or not anyone is watching.
Drawn from the schedule the platform ships with, read when this page was served: 102 scheduled tasks on one beat, plus the price streams where they are enabled. Markets come in on the left, orders and words go out on the right, and the loop turns around one heartbeat. Hover a step, or tap it on a phone, to read what it runs and how often. Times are UTC.
Synthetic trades stream in from the left. Each carries a random USD-beta,
equity-beta and sector load. Drop the caps and watch which trades the gate refuses
in real time. This is the same logic gate_new_entry() runs against your
actual bots once you're inside.
Each row is a trade the gate just judged. Green = passed,
red = refused, with the
exact cap that tripped. Same logic as production
gate_new_entry().
The Operations Center is the home screen: what the system is doing right now. Live signals arrive on a rail that follows you across every page, fills and gate rejections land as banners the moment they happen, and the headband along the bottom keeps portfolio, exposure, bot state and drawdown in view without a click. Everything below is one page.
One bot per asset class, each with its own entry logic, sizing, cooldowns and skip reasons — every one born switched off, on the paper venue. A rule keeps its full size on real money only once the proving ground has passed it on that class and side, or its own graded record has; one measured as losing there is refused, and everything else enters at a quarter. A stage set by hand without that evidence is labelled as such.
▼ and back down again: with the ladder's switch on, a measured loser drops to paper and a decayed rule one rung. 61 strategies are on the ladder now.
A trade is surrounded on every side: checks that can stop it leaving, care while it is open, guards that watch the machine itself, and an alarm that rings when something is wrong. Most of it ships switched off, and none of the 13 switches the platform marks as real-money decisions is turned on by an all-on button: each is turned on alone, by a person, with the trading PIN, and turning one off stays free.
A personality is not a new engine. It is a coherent preset of the knobs that already decide how a bot behaves — the bar it reads, the volatility its stop and target are cut from, how long a thesis may live, how many bets run at once, how much one stop-out costs. Those knobs used to be set one at a time, and nothing in the platform said which combination was a trading style and which was a typo. Three of them are named now: each graded over a window that matches its own holding period, each with its own band in the account allocator.
No personality borrows to fund a position. None of the three sets a leverage or a margin knob — the preset is timeframes, stop and target multiples, a hold ceiling, a concurrency limit and a risk fraction, and that is the whole list. What the short-term one actually gets is how much position one fixed risk budget buys when the stop is tight, plus how many bets may run at once. Both are cash. Where leverage exists at all it is the broker's: with the leverage switch on, a live ticket offers the multipliers the venue lists, the highest by default, and with its switch on the attack mode picks, for each entry, the highest one the venue lists that its stop allows. It changes the money the broker locks, never the loss at the stop. And a personality moves no capital by itself: it changes no account share, enables no config, and applying one to a bot that is already trading live re-sizes real risk — so that one asks for your PIN.
The entry path is cut in two: every bot proposes before anything executes. Turn the fleet pass on — it ships off, like every switch here that can reach a real order — and one layer finally sees the whole tick at once. The desk ranks the candidates each bot has already cleared through its own gates, grades what that rule has really paid across five lanes of evidence, and prices every candidate by its marginal risk — what it adds to a book that already holds correlated positions, not what it would cost standing alone. Then it fills a risk budget the drawdown governor is allowed to cut. Its only power over a size is to shrink one.
Displaced is not discarded. The desk walks the bars the entry it refused would have lived through, prices the exit it would have found, and books the R it would have paid — so the desk is graded against the fleet it overruled, with the tie-breaks in that counterfactual deliberately set against the desk. A correlation it could not measure enters the arithmetic as zero and is written down as unmeasured, never passed off as measured.
Rules opt in by declaring a parameter schema. When the nightly investigator confirms decay on one of them, mutation proposals are generated that same night — creation is a reflex to evidence, not a calendar appointment. Each candidate perturbs one to three parameters inside their declared bounds, and every one of them has to prove itself on data the parent never saw before an operator is even asked.
Above the signal engine sits a layer whose only job is to understand. It reads the platform's own state on a loop, writes down what it believes, invites its own agents to attack those beliefs, keeps what survives, and forgets the rest. Every claim carries a deadline, and every deadline is graded.
Switched on — it ships off, and costs a frontier-model pass when it runs — Horizon writes, once a month, a five-to-ten-year view of every sector this platform can trade: what structurally changes, what it would take to be wrong, and which risks are worth guarding against years before they arrive. Every other agent on this platform looks hours to weeks ahead. This one is held to exactly the same bar as the rest: prose with no gradable claim in it does not count, so each thesis ends in direction calls at six and twelve months, resolved against the first bar at or after the deadline by the same calibration ledger that scores every other agent.
A conversation with the platform about the platform. It reads Sauron's own accumulated state — the Mind's recent reports, the knowledge graph, the hypotheses, the latest briefing, the earnings reviews, the recent audit decisions and the open demotions — and answers in plain English. It reads no trade on its own: asked from a seat, it is given that seat's facts beside the question. It has no tools to change anything: the research agent can only look. Press Ctrl+K on any page, or open the full conversation from AI Agents.
Every broker speaks the same duck-typed interface (ping · ticker · klines · order_book ·
market_order · account · balance).
The router consults the flagged venues in one fixed order — Saxo, then eToro, then IBKR —
and a class nobody has flagged goes to its default adapter when that adapter has keys; otherwise nothing real is sent.
Encrypted credentials at rest (Fernet). Paper-mode default.
The audit log is append-only and hash-chained. Every trade open, trade close,
and gate rejection produces a row whose payload_hash =
sha256(prev_hash · kind · canonical(data)). Mutating any past row breaks the
chain — verify exposes the tamper.
It sits beside FIFO/LIFO/HIFO tax-lot bookkeeping and a Form-8949 CSV export.
save() refuses to update existing rows; delete() raisesverify_chain() walks forward, recomputes, flags any breaksA password gets you a session. It does not get you these: unticking the demo box, arming a real bot, turning on a switch that moves real money, placing or closing a real order by hand and applying a capital plan each sit behind a second factor you hold, and a screen left unattended stops being a screen anyone can use.
Once you are signed in, every door that asks for that PIN counts wrong answers on one counter: the fifth wrong answer where money opens or moves ends the session, while a door that only stops, a close or the kill switch, never ends it and always obeys the right PIN.
One rule runs under all of it. Every proposal is a gradable call with a deadline on it. Every plan is scored against the counterfactual — what actually happened, or what the trade it displaced would have paid. Inside the platform, anything unmeasured renders as an em dash, never as a zero dressed up as data; on this public page a counter that cannot be read shows 0, by design, so the front door never fails. And every switch that can reach a real order ships off; the 13 switches the platform marks as real-money decisions are never turned on by an all-on button, and each waits for a person to turn it on alone.
So here is the honest position, which is the better story anyway. The desk places no orders. It ranks, and its only power over a size is to shrink it; the allocator moves no share by itself. Both are built to run on every tick and every schedule, behind switches that ship off like every other switch here that can reach the money, and both are graded against what would have happened without them. Where one of those switches has since been turned on, it was turned on by a person who had read the grades — not because a landing page said it was ready.
Battle-tested OSS. Multi-stage Docker build, non-root runtime, ASGI worker
for Channels. daphne serving HTTP and WebSockets;
Sentry-instrumented; Celery beat for the recurring tasks; Channels for the per-user Eye push;
a pg_dump nightly with rolling retention. The platform looks after itself.
Each evaluator pulls from real data feeds (SEC, CFTC, FRED, news APIs, broker depth, etc.)
and outputs a normalised score the scanner can compose. Add a new evaluator with one
register_kind() call — the scanner picks it up automatically.
The last steps the Book records, newest first. Every one is dated, and every one has its own line in the Book.
On the bench, not yet landed: The Orcs' shadow minds. What is being built now →
Sauron starts in paper mode. Wire up your brokers when you've watched it run. Flip the orchestrator on when you're ready. A rule trades real money at full size only once the proving ground or its own graded record has earned it. Cautious by default — aggressive when you ask.
mode="live" per config.
Everything visible without a single broker key.