One operator, one AI partner, a full systematic trading platform
Morpheus42 researches, tests, deploys and monitors automated trading strategies across multiple markets and venues. It treats evidence and risk control as engineering problems.
What it is
Morpheus42 is a privately built platform that covers the whole life of a trading idea: from a written hypothesis, through testing on more than 20 years of minute-level data, to live execution and daily review. One person sets the direction and approves risk. Claude writes, tests, deploys and watches the system alongside them.
It runs around the clock on a small private estate, with a custom execution engine, hundreds of scheduled workers, a live command centre and independent monitoring.
Built with AI, governed by a person
- Engineering: AI sessions write the code and run the tests before every commit.
- Operations: AI reads live state, investigates faults and carries out approved changes through controlled routes.
- Research: parallel agents run studies, while separate agents check them. Independent review catches what the author missed.
- Memory: hundreds of structured notes carry decisions, closed questions and known traps from one session to the next.
- Human gates: anything that adds risk needs the operator. AI safeguards stop other risky actions and hand back the exact step for a person to take.
Inside the platform, models are used to read and check, not to predict prices. AI audits review each day's changes against a catalogue of known research errors, and a finding only counts if it can be quoted from the code.
Research that is built to say no
Most backtests flatter their author. This process is designed to do the opposite.
- Pre-registered: the rule, sample and pass mark are written and fixed before any result is seen.
- One look: held-out data is read once, recorded and never reused.
- Real costs: spreads, commissions and financing are measured per account and instrument from live data, then refreshed nightly.
- Causal fills: orders fill in the order prices actually traded. A rule that only works with hindsight is thrown out.
- Multiple-testing control: results are corrected across every test run, with returns clustered by day.
- Honest nulls: every result states the smallest edge the test could have detected, so "nothing found" is never overstated.
- A permanent registry: more than 570 ideas are logged with their evidence and status, so nothing gets re-tested by accident.
The outcome so far: from about 900,000 tests across more than 180 instruments, one edge has cleared every gate. That ratio is the point. A process that rejects almost everything is one whose rare passes are worth attention.
Risk and change control
| Control | What it does |
|---|---|
| Governance as code | About 2,000 live settings, each change stamped with an owner, reason, evidence and review date |
| Layered limits | Trade, day, account and fleet limits stop trading well before any external limit is reached |
| Fail-closed design | If data is stale or a check cannot run, the system stops trading rather than guessing |
| Forward testing | New strategies earn real-fill evidence at small size, separate from the main book |
| Audit trail | Operator actions are recorded in a tamper-evident chain, and results cannot be silently altered |
| Reversible by default | Every deployment leaves a backup and a one-line rollback, and temporary changes carry a dated revert |
Engineering discipline
- Commit gates: self-tests, the relevant test suite and a front-end lint run before anything is accepted.
- Verified releases: a matching file hash proves a file landed. A build stamp from the running system proves it is actually loaded.
- Measured clocks: time offsets are measured per venue, including daylight-saving changes, because clock errors are among the most common research traps.
- Monitoring that fails loudly: every scheduled job is checked against its expected cadence, and alerts are tested with real sends.
- Whole-system audits: data, code, interface and documentation are audited independently, and the findings are tracked until closed.
- Status anywhere: the fleet can be checked by voice, phone or watch through read-only channels that cannot touch a trade.
What the audits catch
A few real examples of faults that looked healthy on the surface:
| Fault | Effect |
|---|---|
| Account number truncated to 32 bits | Per-account settings were silently ignored |
| Significance test inverted | Whole research labs could never pass anything |
| Orders filled in signal order, not traded order | An apparent edge disappeared once fills were made causal |
| A single daylight-saving rule assumed for every venue | Valid entries were refused after a clock change |
| A stale temporary file renamed over live settings | A change quietly reverted, and a parity check caught it |
Architecture
Select any part of the system to see what it does, or follow one of the two tours. The first takes an idea from a written hypothesis to live trading. The second follows what happens when something goes wrong.
Why it matters
Morpheus42 shows what one person can run with AI when the discipline is built into the system: the research rigour, operational control and auditability usually associated with a full team.