ImageFirm
Course · Agent engineering · Rev. 2026.10 · Evidence checked 29 Oct 2026

A trading system may improve itself. It may never promote itself.

A redline review of the widely shared “ultimate self-improving trading bot” blueprint, and the governed build that replaces it. What to keep, what fails the moment real money is involved, and the statistical, security and human-authority controls that make continuous improvement safe to run.

12 modules 5 working tools Study time ≈ 5 h Level Advanced Progress 0 of 12
M00

The brief

Orientation · 10 min

The blueprint is a good research loop wired straight into a brokerage account. That single design choice is the whole problem, and this course is about undoing it.

A loop that keeps optimising against its own test results becomes, over enough cycles, a machine for discovering luck. A system that reads the news while holding order authority can be steered by whoever writes the news. Neither failure is fixed by a smarter model. Both are fixed by separation of powers: the part that learns, the part that judges and the part that trades must be different parts, and only a person may move a change from the first to the third.

You will leave with a reference architecture, a ten-stage validation gauntlet, a working deterministic risk gate in Python, a 24-point go-live gate and a twelve-week build plan. Five tools on the page let you test the ideas instead of taking them on trust.

Who this is for

  • Founders and CTOs scoping an AI trading or treasury project.
  • Engineers who can write Python and have run a backtest.
  • Risk, compliance and audit leads who must sign off on one.
You will be able to
  • Name the four design flaws in the blueprint and the control that fixes each.
  • Deflate a backtest for the number of things you tried, and say whether anything survived.
  • Draw the authority boundary between research, validation and execution.
  • Specify a pre-trade risk engine and a three-level kill switch.
  • Defend an LLM-driven research loop against manipulated inputs.
  • Run a go-live review a regulator, an auditor or your own board would accept.

Read this first

This is a course in system design and risk control. It recommends no strategy, instrument, broker or trade, and nothing here is a forecast. ImageFirm is not a licensed investment adviser or broker. Running money or selling signals for anyone other than yourself is a regulated activity in most jurisdictions; module 09 shows where the lines sat on 29 October 2026, and none of it replaces licensed counsel.

Saved in this browser only
M01

Redline review of the blueprint

Audit · 30 min · Tool 01

The blueprint gets the shape of a research loop right: log, analyse, hypothesise, test, compare. It goes wrong in the four places that decide whether capital survives: who holds authority, what counts as evidence, what the system trusts, and what stops it.

Every component in the original drawing is audited below against current evidence. Keep means sound as drawn. Refine means right idea, missing a control. Replace means the design choice itself is unsafe. Add marks components the blueprint leaves out entirely.

Tool 01 · Component audit · filter by verdict
ComponentBlueprint saysVerdictWhyGoverned build
Market data · TradingViewCharts, indicators, real-time and historical data as a bot input.ReplaceTradingView licenses market data for display; its terms list automated trading and automated order generation among prohibited uses and exclude non-display use. [19]Machine-readable data from a vendor or broker licensed for automated use. Keep TradingView as a human research surface.
Market data · MetaTrader 5Forex and CFDs, live or demo.KeepThe official Python package connects to a running MT5 terminal and sends structured trade requests. [20] Prices are your broker's quotes, and demo fills are not live fills.Record broker quotes, spreads and swaps as cost-model truth. Calibrate slippage on live fills only.
News & eventsEconomic calendar, news, sentiment.RefineUntrusted text that steers an LLM. Headline edits invisible to humans changed trading decisions in published tests. [13]Quarantine, Unicode normalisation, hidden-text stripping, source provenance, arrival timestamps.
Alternative dataOn-chain, macro, custom datasets.RefineRevisions, survivorship and licensing silently inflate backtests.Store vintages as first published. Licence terms recorded per dataset.
AI orchestratorPlan, reason, use tools, improve; three model vendors drawn in the box.RefineOrchestrating research is fine. Sending “approved strategies” onward to execution puts the learner in the trading seat. Hard-wiring vendors is a resilience risk.Model-agnostic routing. Orchestrator coordinates research only; no broker credentials or order tools in reach.
Research agentsMarket analyst, strategy researcher, code agent.KeepHypothesis generation, literature synthesis and code drafting are where LLMs earn their place. [11]Sandboxed, allowlisted egress, outputs are artifacts (notes, code, specs), never actions.
Research reviewerValidate logic and find flaws.RefineA reviewer that shares the author's model, prompt and reasoning can share its blind spots, and words are not a test.Independent reviewer (different model or prompt, no access to the author's rationale) plus automated leakage tests.
Approved strategies → executionOrchestrator passes approved strategies to MT5 to place orders.ReplaceThe drawing never says who approves. In a self-improving loop the answer cannot be the loop.Named human promotes a signed version. Every order passes a deterministic risk engine before the broker adapter.
Order manager, broker adapter, execution monitorPosition, modify, cancel; live or paper; fills, slippage, errors.KeepCorrect separation of concerns on the execution side.Idempotent client order IDs, daily reconciliation against broker statements.
Memory & observabilitySignals, orders, fills, PnL, market state.RefineMissing the one record that decides statistical honesty: every experiment ever run.Experiment ledger with trial count; model, prompt, code and data hashes; hash-chained audit log.
Error analysisUnderperformance, drawdown, excessive trades, slippage, regime change, data issues.RefineOmits the failures that look like success: leakage, look-ahead, cost-model error, manipulated inputs, model nondeterminism.Expanded taxonomy (module 10) with an owner per category.
Reasoning engineanalyze_failures() finds patterns in failed trades and proposes changes.RefinePattern-hunting in a few losing trades is how noise becomes a rule. LLMs handle numerical time series poorly, by the account of the organiser of the 2025 live LLM trading contest. [9]The LLM writes hypotheses in words, each with a mechanism. Statistics decide.
Validation & testingBacktest, out-of-sample, transaction costs, risk limits, robustness.RefineConsulted every cycle, the out-of-sample set silently becomes in-sample. [4]Purged cross-validation for development, a locked vault spent rarely, trial-adjusted Sharpe, then paper, shadow and canary.
Strategy update · Approve / RejectApply approved changes with version control (v2.1).ReplaceApproval is automated by the test result, so the system promotes itself.Signed release, evidence pack, human approval, staged capital, rehearsed one-step rollback.
Loop step 5 · CompareKeep if better, otherwise revert.Replace“Better” over a short window is a coin toss, and constant switching is itself overfitting that churns costs.Champion and challenger on a pre-registered horizon, with a minimum track record and a significance test.
Deterministic risk engine & kill switchNot drawn.AddThe reference case: in 2012 Knight Capital sent more than 4 million orders in 45 minutes and eventually lost more than $460 million. [15]Pre-trade limits, price collars, throttles, latched loss breaker, three-level kill switch (module 07).
Human promotion gateNot drawn.AddRegulators expect the management body to stay accountable for decisions made by AI. [22]Named approver, four-eyes deployment, evidence pack (module 09).
Security boundaryNot drawn.AddIt combines private data, untrusted content and outbound channels (the lethal trifecta), then adds the power to trade. [17]No path from untrusted text to an order (module 08).
Cost and capacity model“Transaction costs” as one tick-box.AddAn edge measured before spread, slippage and financing is not yet an edge.Spread, commission, slippage, financing and market impact, calibrated on live fills (module 06).
Regulatory perimeterNot drawn.AddLeverage caps, licensing and AI-marketing rules differ by jurisdiction and change. [24] [28] [29]Perimeter review per jurisdiction, re-run on every material change (module 09).
Incident response & retirementNot drawn.AddPublished edges decay on average; the question is whether you notice yours first. [6]Severity-graded runbook and retirement criteria written before launch (module 10).
Redline · Reasoning engine

patterns = find_patterns(trades) → hypothesis = propose_changes(patterns) → return hypothesis

A hypothesis is not a change. It is a counted trial that must pass a gauntlet the proposer cannot see, and only a person may turn the survivor into a release.

improvement_cycle.py · illustrative
def improvement_cycle(ledger, gauntlet, vault, approver):
    failures   = ledger.failures(since=ledger.last_review)   # evidence, not anecdotes
    hypotheses = research_llm.propose(failures, max_n=5)     # words + mechanism, no orders
    for h in hypotheses:
        trial = ledger.register(h)                           # counted BEFORE it is tested
        result = gauntlet.run(trial, data="development")       # deterministic statistics
        ledger.record(trial, result)
    best = ledger.best_candidate(min_dsr=policy.min_dsr, n_trials=ledger.count())
    if best is None:
        return "no change"                                    # the common, healthy outcome
    verdict = vault.test_once(best)                          # the vault is spent sparingly
    pack = evidence_pack(best, verdict, ledger)
    return approver.review(pack)                            # only a person promotes

Keep the loop. Remove its authority. Count everything it tries.

Saved in this browser only
M02

Why self-improvement overfits

Statistics · 45 min · Tools 02–03

Every idea the loop tests is a lottery ticket. Buy enough tickets and one wins, every time, with no skill involved. This module tells you how many tickets you bought and what a real winner has to look like.

2.1Every test is a trial

Academic finance learned this the hard way. After hundreds of published return factors, Harvey, Liu and Zhu estimated that a newly discovered factor should clear a t-statistic above 3.0, not the conventional 2.0, to be taken seriously. [1] A self-improving loop recreates that “factor zoo” at machine speed: an LLM can propose and backtest dozens of variants overnight. The number of trials is the most important number the blueprint never logs.

Run the machine below. Every strategy in it is pure noise by construction, so any Sharpe ratio you see is luck.

Tool 02 · The overfitting machine

Zero-edge strategies, one lucky winner

Each strategy draws daily returns at random: mean zero, 1% daily volatility. The best in-sample performer is then run for 12 fresh months.

Best in-sample Sharpe
—
Same strategy, next 12 months
—
True edge of every strategy
0.00
Cumulative return of the luckiest zero-edge strategy
In-sample (selected)Next 12 months

Press “Run the machine”. With JavaScript off: the best of 100 noise strategies over 3 years averages a Sharpe near 1.5, and about 1.9 at 1,000 strategies (computed for this course).

2.2The holdout wears out

The blueprint runs out-of-sample testing on every cycle. That feels rigorous and is not. When each analysis choice depends on earlier results from the same data, false discoveries inflate; teams that repeatedly check a fixed holdout overfit to it, so leaderboard scores climb while true performance does not. [4] After enough cycles, the blueprint's out-of-sample set is training data with extra steps.

The fix is a three-way split with different access rules:

  • Development data, used freely, with purged and embargoed cross-validation so overlapping labels cannot leak. [5]
  • The vault, a locked holdout touched once per finalist, every access logged, with a fixed budget of looks per year.
  • Forward time: paper, shadow and canary trading. The only data the loop cannot have seen.

2.3Deflate before you celebrate

The Deflated Sharpe Ratio adjusts an observed Sharpe for selection under multiple testing and for non-normal returns, so a strategy chosen from many trials is judged against the best result luck alone would have produced. [2] The Probability of Backtest Overfitting estimates how often the configuration that ranks best in-sample falls below the median out-of-sample. [3] Both need one input the blueprint does not keep: the honest trial count.

Tool 03 · Deflated Sharpe calculator

Is this Sharpe ratio skill, or the best of many guesses?

Blank assumes the trials had no edge, the conservative default.
Luck benchmark (annualised)
1.46
P(skill > 0), ignoring trials
99.97%
Deflated Sharpe (P beats luck)
82.4%

Not distinguishable from selection luck at the 95% level. Keep researching; do not promote.

Method: Bailey & López de Prado (2014). The benchmark is the expected maximum Sharpe of N trials; the Deflated Sharpe is the probabilistic Sharpe ratio measured against that benchmark, corrected for skew and kurtosis. [2] The 95% line is a common convention, not a law; set your threshold before you look.

The defaults tell the story. An annualised Sharpe of 2.0 over three years of daily data looks overwhelming against zero (about 99.97%). Admit that it was the best of 100 trials and the deflated probability falls to about 82%; at 1,000 trials, about 58%. It stays above 95% only up to about 15 trials. These figures are computed for this course under the no-edge assumption; enter your own.

2.4Real edges decay too

Even genuine signals erode. Across 97 published predictors, McLean and Pontiff found portfolio returns 26% lower out-of-sample and 58% lower after publication. [6] For LLM-read news, Lopez-Lira and Tang report that strategy returns decline as LLM adoption rises, consistent with prices becoming more efficient. [32] So the loop has two jobs, not one: find improvements, and detect decay in what is already live. The honest output of most cycles is no change.

If you cannot say how many things you tried, you cannot say whether anything worked.

Saved in this browser only
M03

The governed architecture

Design · 35 min

Three planes, one gate, one ledger. Learning, judgement and trading live in separate planes, and nothing reaches the trading plane without a person's signature.

Governed architecture: a governance band above three planes (research, validation, execution) and a memory band below.
Governance · human authority A named approver promotes signed versions on an evidence pack. The kill switch is reachable out-of-band by the operator and the approver.

Research plane

May read anything. May touch nothing.

  • LLM research agents: analyst, researcher, coder
  • Independent reviewer on a separate model or prompt
  • Sandboxed backtester on development data only
  • Quarantined readers for news and alternative data

No credentials · no order tools · no vault access

Hypotheses + code, registered in the ledger

Validation plane

A deterministic judge.

  • Purged cross-validation, cost model, robustness suite
  • DSR and PBO at the true trial count
  • Vault holdout on a budget of looks
  • Paper and shadow harness on live data

No LLM verdicts · cannot rewrite ledger history

Evidence pack → human approval → signed release

Execution plane

Trades only what was signed.

  • Strategy runtime, pinned to a signed version
  • Deterministic risk engine, fails closed
  • Broker adapter: the only credential holder
  • Reconciliation and execution monitor

No free-text input · no self-updates · no outbound calls but the broker

Memory & observability · the ledger Every trial counted before it runs; point-in-time data store; model, prompt, code and data hashes per result; hash-chained audit log shipped off-host.
Rule 1

The part that reads untrusted text cannot place an order.

Cuts the trifecta's outbound leg, and the order button with it.

Rule 2

The part that proposes a change cannot approve it.

The loop learns; a person promotes.

Rule 3

Every trial is counted before it runs.

Without the count, no statistic in module 02 means anything.

Mapped to ImageFirm's 8-layer agent architecture
LayerIn this systemNon-negotiable control
1 · Interface & channelsOperator console, alerts, approval screen, kill switch.Kill switch works out-of-band, from a phone, in one action.
2 · Planning & orchestrationResearch orchestrator scheduling agents and backtests.No execution tools registered. Ever.
3 · Memory & contextExperiment ledger, research notes, point-in-time store.Append-only; the trial counter cannot be decremented.
4 · ModelsRouted LLMs for research; frozen statistical or ML models inside strategies.Model, version, training cutoff and prompt pinned per result.
5 · Tools & actionsAllowlisted data fetch, code sandbox, backtester; broker adapter only in execution.Least privilege; no secrets in prompts, repositories or logs.
6 · State & dataPositions, working orders, fills, reference prices.Broker is the source of truth; reconcile at least daily.
7 · Observability, safety & evalsRisk engine, drift monitors, injection red-team evals.Deterministic, fails closed, alerts a human.
8 · DeploymentSigned releases, staged capital, rollback.Four-eyes release; the loop cannot deploy itself.

Can an LLM ever sit inside a live strategy?

Yes, under three conditions. It runs as a frozen, versioned component of a signed release, never as a live agent with tools. Its output is constrained to a schema (a label or a number in a range) that deterministic code validates before use. And its influence on position size is capped by the risk engine, so the worst output it can produce is a bounded, pre-approved loss. A news classifier that emits {"sentiment": -1..1} under those conditions is a feature. An agent that reads the news and calls order_send() is a liability.

Separation of powers is the product. Intelligence is a component.

Saved in this browser only
M04

Data you can trust

Data · 30 min

A backtest is only as honest as its clock. Most fatal bugs are not in the strategy; they are in when the strategy was allowed to know something.

4.1Three timestamps on every record

  • Event time: when it happened in the world.
  • Availability time: when your system could first have read it, after publication, vendor delay and ingestion.
  • Decision time: when the strategy acted.

A decision may only use records whose availability time is earlier than the decision time minus your latency budget. Store all three; most leaks come from keeping only the first.

Data contract · one per dataset
  • Source, licence, and whether automated (non-display) use is permitted
  • Availability-time rule and vendor delay
  • Vintage policy: values as first published, revisions stored separately
  • Universe coverage, including delisted instruments
  • Time zone, session boundaries, corporate-action handling
  • Known gaps, and a named owner
The leaks that make backtests lie
LeakHow it entersTest that catches it
Look-aheadTrading at a bar's close using that bar's close; using a daily high before the day ends.Shift every feature one bar later; an edge that vanishes was never there.
SurvivorshipTesting only instruments that still exist today.Rebuild the universe point-in-time, delisted names included.
RevisionsMacro and fundamental data restated after the fact.Backtest on first-release vintages only.
Label overlapMulti-day targets overlapping across train and test folds.Purging and embargo in cross-validation. [5]
Quote mismatchTesting on one feed, trading on your broker's prices and spreads.Compare backtest fills with live broker fills weekly.
Model memoryAn LLM that was trained on the period you are backtesting.Post-cutoff evaluation only, or a time-consistent model (4.2).

4.2The LLM's memory is a leak

An LLM backtested on news from inside its own training window may already know what happened next. Glasserman and Lin identified two distortions: look-ahead bias, the model knowing how returns followed a headline, and a distraction effect, general knowledge about a named company interfering with the read of the headline itself. In their in-sample tests, headlines with company names removed produced better trading performance than the originals, most of all for large companies the model knows well; out of sample, look-ahead was no longer the concern, though distraction could remain. [7]

He, Lv, Manela and Wu trained models only on text that existed at each point in time, with yearly checkpoints through 2024. In their next-day news-return task, the bias turned out to be modest, and they stress that its size depends on the model and the application. [8] Lopez-Lira and Tang avoid the issue by testing on headlines published after the models' training cutoff. [32] Modest in one study is not zero in yours.

Rule · LLM-dependent signals

Evaluate them only on data after the model's training cutoff, or with a chronologically consistent model. Anonymise entities where that does not destroy the signal. Record the model's cutoff next to every result in the ledger.

4.3Licence before latency

The blueprint's first data box is a charting platform whose terms license market data for display and list automated trading among prohibited uses. [19] Using it as a bot feed risks account termination and leaves your evidence trail built on data you were not entitled to use that way. Check every source for automated-use rights before you check its speed, and keep the answer in the data contract.

Ask of every number: when could I have known this, and am I allowed to use it this way?

Saved in this browser only
M05

Where LLMs earn their place

Agents · 35 min

Use language models for language: hypotheses, code, synthesis, critique. Keep them away from counting, from grading their own work, and from the order button.

Fit by task
TaskFitCondition or reason
Generating hypotheses from literature and failure logsStrongEvery hypothesis is registered as a trial before testing.
Drafting strategy and test codeStrongCode is checkable: sandbox, unit tests, independent review.
Summarising incidents and logs for peopleStrongThe human still reads the source record before deciding.
Turning text into features (sentiment, events)ConditionalFrozen, schema-constrained, validated after the training cutoff (M04).
Inferring patterns from raw numeric time seriesWeakThe 2025 contest organiser's own diagnosis. [9] Use statistical tools.
Grading its own proposalsWeakShared blind spots. The validation plane is deterministic.
Final risk decisionsUnsafeJailbreak success above 95% in a 2026 study (M07). [14]
Placing ordersUnsafeGives untrusted text a path to action (M08). [17]

5.1What live contests show, and what they don't

In nof1's Alpha Arena, frontier models traded real money autonomously. Season 1 ran on the Hyperliquid exchange in crypto perpetuals, with $10,000 per model, from 17–18 October (the two sources differ by a day) to 3 November 2025. [9] [34]

Alpha Arena Season 1 · final balances as reported; returns computed from $10,000
ModelFinal balanceReturn
Qwen3 Max$12,231+22.3%
DeepSeek$10,489+4.9%
Claude Sonnet 4.5$5,799−42.0%
Gemini 2.5 Pro$5,445−45.6%
Grok$4,208−57.9%
GPT-5$4,126−58.7%

Season 1.5 moved to US stocks in late November 2025 and added four modes, one of them with memory and self-learning, the blueprint's central idea. As reported on 8 December 2025, only one model, Grok 4.2, was profitable overall, at about 12%. [10]

How to read this: one run per model, a few weeks, a handful of models, no repetition. That is a single sample path. The leaders changed between seasons, which is consistent with noise, though the asset class, model versions and modes changed too. Treat it as evidence of how fast autonomous agents with order authority can lose capital, not as a ranking of trading skill, and certainly not as a reason to pick a model.

5.2Research frameworks worth studying

TradingAgents mirrors a trading firm: fundamental, sentiment and technical analysts; bull and bear researchers who debate; traders with different risk profiles; and a risk-management team. Its authors report better cumulative return, Sharpe ratio and maximum drawdown than baselines. [12] R&D-Agent(Q) automates the research loop itself: hypotheses drawn from a knowledge base of past results, a code agent that implements them, backtests on Qlib, and a bandit scheduler that decides whether to work on factors or models next. It reports up to roughly twice the annualised return of classical factor libraries on CSI 300 with about 70% fewer factors. [11]

Read both as architecture patterns. Their performance figures are backtests, and module 02 applies to them exactly as it applies to you. R&D-Agent(Q) is, structurally, the blueprint's loop placed where it belongs: in research, with no broker attached.

5.3Same prompt, different answer

Thinking Machines Lab traced most LLM output variation at temperature zero to inference kernels that are not batch-invariant: a request's output can change with the size of the batch it happens to share, which depends on server load. In their test, 1,000 runs of one prompt produced 80 distinct completions on default kernels and a single completion on batch-invariant kernels. Determinism costs speed: in a separate benchmark the deterministic setup took 42 seconds against 26 by default, and 55 before its attention kernel was optimised. [31]

The consequence for audit is direct: you cannot reproduce an LLM-driven decision by running it again. Store the actual output, the full input, the model identifier and every parameter in the ledger. Reproducibility comes from your records, not from the model.

Pattern · the independent reviewer
  • Runs on a different model family, or at minimum a different system prompt.
  • Sees the hypothesis, the code and the results; never the author's reasoning.
  • Must name one way the result could be leakage, one way costs could kill it, and one regime where it should fail. A review that cannot is rejected, not the hypothesis.
  • Its opinion is advisory. The gauntlet in module 06 is binding.

Let the model write the hypothesis. Never let it grade the exam or hold the pen that signs.

Saved in this browser only
M06

The validation gauntlet

Evidence · 40 min

Ten gates in a fixed order. Each produces evidence and each has a fail action agreed in advance. A candidate that skips a gate has not passed it; it has not been tested.

  1. Register the hypothesis

    Before touching data: mechanism (why anyone would pay you for this), expected sign, horizon, universe, capacity estimate and the conditions that would kill it.

    Evidence: ledger trial ID · Fail: no ID, no test

  2. Check the data contract

    Every input passes the module 04 contract and leak tests, with availability timestamps.

    Evidence: contract and dataset hashes · Fail: fix the data, re-register as a new trial

  3. Development test

    Purged, embargoed cross-validation. Combinatorial purged CV gives a distribution of backtest paths instead of one flattering line. [5]

    Evidence: Sharpe distribution, not a point · Fail: close the trial

  4. Cost model

    Spread, commission, slippage, financing or swap, and market impact at your intended size, taken from your broker's actual quotes. Report the break-even cost.

    Evidence: net-of-cost metrics · Fail: edge does not survive costs doubled

  5. Robustness

    Neighbouring parameters, regime splits, resampled histories, and the result with the best few trades removed. A knife-edge optimum is a fitted accident.

    Evidence: stability report · Fail: performance collapses near the chosen parameters

  6. Deflate

    Deflated Sharpe at the ledger's full trial count, and the probability of backtest overfitting. [2] [3]

    Evidence: DSR, PBO · Fail: below the threshold fixed in advance

  7. Open the vault, once

    One look at the locked holdout. The look is logged and spent whatever the result.

    Evidence: vault result, access record · Fail: close the trial; no second look

  8. Paper trading

    Live data, simulated fills, the full pipeline including the risk engine and reconciliation.

    Evidence: paper-versus-backtest tracking · Fail: tracking error beyond tolerance, investigate the pipeline

  9. Shadow

    Runs beside the live champion on the same live data, logging the orders it would have sent without sending them.

    Evidence: decision log against live quotes · Fail: behaviour departs from the specification

  10. Canary

    Real orders on a small allocation with a loss budget in currency fixed before the start, approved by a named person. Scaling happens in steps, each one a new approval.

    Evidence: live fills against the cost model · Fail: budget hit, automatic halt and human review

Example promotion policy · illustrative values
TestExample threshold
Trials countedAll, abandoned ones included
Deflated Sharpe at that count≥ 0.95
Probability of overfittingBelow a ceiling you set first
Cost stressEdge survives 2× costs
Paper and shadowMinimum duration and trade count, fixed first
CanaryLoss budget in currency; latched halt

The values are examples. The discipline is fixing yours before any result is visible.

Replaces “keep if better, otherwise revert”

Champion and challenger

  • The comparison horizon is registered before the challenger runs.
  • Both are measured over the same live period, the challenger in shadow.
  • The challenger must win by more than the noise of that horizon, after switching costs.
  • A minimum track record applies; two good weeks are not one.
  • A person approves the switch, and the old champion stays deployable for rollback.

Write the fail actions before the results exist. Afterwards, everyone negotiates.

Saved in this browser only
M07

The risk engine and the kill switch

Control · 45 min · Reference code

The last line of defence should be the simplest code in the system: short, deterministic, tested, and impossible to argue with.

7.1Why the risk manager cannot be a language model

A 2026 study of multi-agent LLM trading systems attacked each role in turn. Jailbreaking the risk-manager agent, the final decision node, succeeded on 98.9% of attackable days for buy-targeted attacks and 95.5% for sell-targeted ones. The authors conclude that no architecture they tested is inherently robust, and flag the design risk of coupling safety enforcement with final decision authority. The test covered four large-cap stocks and Bitcoin over the first quarter of 2026, without trading costs, so treat the magnitudes as indicative. [14]

Rule

An LLM may advise on risk. Only code may enforce it, and only a person may switch it back on.

Control catalogue
StageControlWhat it stops
Pre-tradeInstrument allowlist; per-order notional capFat fingers and strategies wandering into instruments nobody approved.
Pre-tradePosition and gross-exposure limits that count working ordersMany small orders adding up to one large unapproved bet.
Pre-tradePrice collar against a trusted reference priceOrders far from the market, including prices hallucinated upstream.
Pre-tradeOrder-rate throttle; idempotent client order IDsRunaway loops and duplicate submissions after retries.
Pre-tradeLatched daily-loss breaker; trading-hours windowLosing all day; trading into closed or thin sessions.
At-tradeHeartbeat (dead-man's switch); cancel on disconnectOrders working while the system that placed them is dead.
Post-tradeFill and position reconciliation with the brokerThe system believing it is flat when it is not.
SystemThree-level kill switch; signed-version pinning; four-eyes configuration changesA bad release or a changed limit becoming a loss before anyone looks.

For regulated firms in the EU and UK these mirror the MiFID II algorithmic-trading expectations: kill functionality that cancels all unexecuted orders, pre-trade limits on volume, value and messages, and price collars. The UK regulator flagged limits set significantly above normal trading activity as poor practice, because they make it hard to show the controls work. [21]

7.2Reference implementation

A deterministic gate in standard-library Python. It counts working orders toward exposure, rejects with stable reason codes, latches the loss breaker, and writes a hash-chained audit trail. It is a reference for review, not production code: add your broker adapter, persistence, tests and monitoring.

risk_gate.py · Python 3.10+ · tested for this course
"""
risk_gate.py - deterministic pre-trade gate (reference implementation).

No language model sits on this path. Every order any strategy wants to send
passes check() first; only submit() may reach the broker adapter, and the
adapter is the only object in the system that holds broker credentials.
Requires Python 3.10+.
"""
from __future__ import annotations

import hashlib
import json
import time
from dataclasses import asdict, dataclass, field
from datetime import datetime, timezone


@dataclass(frozen=True)
class Limits:
    allowed_symbols: frozenset[str]
    max_order_notional: float      # per order, account currency
    max_position_notional: float   # per symbol, absolute
    max_gross_exposure: float      # sum of absolute positions
    max_daily_loss: float          # positive number, realised + unrealised
    max_orders_per_minute: int
    price_collar_pct: float        # e.g. 0.01 = reject beyond 1% of reference


@dataclass(frozen=True)
class Order:
    client_id: str                 # idempotency key, unique per intent
    symbol: str
    side: str                      # "buy" | "sell"
    qty: float
    limit_price: float
    strategy_id: str
    strategy_version: str          # signed release tag, never "latest"


@dataclass
class State:
    positions: dict[str, float]          # symbol -> signed notional
    reference_price: dict[str, float]    # symbol -> trusted mid, not LLM output
    pending: dict[str, float] = field(default_factory=dict)  # working orders
    day_pnl: float = 0.0
    halted: bool = False                 # latched: only a human resets it
    sent_ids: set[str] = field(default_factory=set)
    order_times: list[float] = field(default_factory=list)


class Reject(Exception):
    """Raised with a stable reason code; the order never leaves the process."""


def check(o: Order, s: State, lim: Limits, now: float | None = None) -> None:
    now = time.time() if now is None else now
    if s.halted:
        raise Reject("KILL_SWITCH_ACTIVE")
    if s.day_pnl <= -lim.max_daily_loss:
        s.halted = True                                  # trip and stay tripped
        raise Reject("DAILY_LOSS_BREAKER")
    if o.client_id in s.sent_ids:
        raise Reject("DUPLICATE_ORDER")
    if o.symbol not in lim.allowed_symbols:
        raise Reject("SYMBOL_NOT_ALLOWED")
    if o.side not in ("buy", "sell") or o.qty <= 0 or o.limit_price <= 0:
        raise Reject("MALFORMED_ORDER")
    ref = s.reference_price.get(o.symbol)
    if not ref:
        raise Reject("NO_REFERENCE_PRICE")               # fail closed
    if abs(o.limit_price - ref) / ref > lim.price_collar_pct:
        raise Reject("PRICE_COLLAR")
    notional = o.qty * o.limit_price
    if notional > lim.max_order_notional:
        raise Reject("ORDER_TOO_LARGE")
    delta = notional if o.side == "buy" else -notional
    exposure = lambda k: s.positions.get(k, 0.0) + s.pending.get(k, 0.0)
    new_pos = exposure(o.symbol) + delta                 # count working orders too
    if abs(new_pos) > lim.max_position_notional:
        raise Reject("POSITION_LIMIT")
    others = sum(abs(exposure(k)) for k in set(s.positions) | set(s.pending)
                 if k != o.symbol)
    if others + abs(new_pos) > lim.max_gross_exposure:
        raise Reject("GROSS_EXPOSURE_LIMIT")
    recent = [t for t in s.order_times if now - t < 60.0]
    if len(recent) >= lim.max_orders_per_minute:
        raise Reject("ORDER_RATE_THROTTLE")
    s.order_times[:] = recent + [now]


class AuditLog:
    """Append-only JSONL with a hash chain, so silent edits are detectable.
    In production, load the last hash on start-up and ship lines off-host."""

    def __init__(self, path: str, prev: str = "0" * 64) -> None:
        self.path, self.prev = path, prev

    def write(self, event: str, **data) -> None:
        rec = {"ts": datetime.now(timezone.utc).isoformat(), "event": event,
               "prev": self.prev, **data}
        line = json.dumps(rec, sort_keys=True, default=str)
        self.prev = hashlib.sha256(line.encode()).hexdigest()
        with open(self.path, "a", encoding="utf-8") as fh:
            fh.write(line + "\n")


def submit(o: Order, s: State, lim: Limits, broker, log: AuditLog):
    try:
        check(o, s, lim)
    except Reject as r:
        log.write("reject", reason=str(r), **asdict(o))
        return None
    s.sent_ids.add(o.client_id)
    delta = o.qty * o.limit_price * (1 if o.side == "buy" else -1)
    s.pending[o.symbol] = s.pending.get(o.symbol, 0.0) + delta
    log.write("accept", **asdict(o))
    return broker.send(o)          # adapter: the only code holding credentials


def on_fill(s: State, symbol: str, signed_notional: float) -> None:
    """Called by broker reconciliation for every fill (partial fills included).
    Cancels and rejects call it with the unfilled remainder moved out of pending."""
    s.pending[symbol] = s.pending.get(symbol, 0.0) - signed_notional
    s.positions[symbol] = s.positions.get(symbol, 0.0) + signed_notional


def kill(s: State, broker, log: AuditLog, reason: str, who: str) -> None:
    """Level 2 kill: block new orders, cancel all working orders.
    Flattening positions (level 3) is a separate, human-confirmed action."""
    s.halted = True
    broker.cancel_all()
    log.write("kill", reason=reason, by=who)

7.3The kill switch, specified

Level 1 · Pause

Block new orders

Working orders stay. Any monitor or operator may trigger it; automatic triggers allowed.

Level 2 · Cancel

Block new, cancel all working

Operator or automatic breaker. This is the regulatory “kill functionality”.

Level 3 · Flatten

Close positions

Human-confirmed only. Flattening into a disorderly market can be the most expensive action available.

Reset at every level is human-only, with the reason logged. Rehearse all three in paper on a fixed calendar; an untested kill switch is a belief, not a control.

7.4Two incidents every builder should know

Knight Capital · 1 August 2012

A deployment that traded

Code for a function retired in 2003 was never removed. In 2012 the flag that used to switch it on was reused for new code, and a technician copied that code to seven of eight servers, with no second review and no written procedure requiring one. On the eighth server, the flag woke the old code. [35] In the first 45 minutes the router sent more than 4 million orders while trying to fill 212 customer orders and traded more than 397 million shares; Knight eventually lost more than $460 million. The SEC imposed a $12 million penalty under the Market Access Rule. [15]

For a self-improving system: every release is a trade. Delete dead code, never reuse flags, require four eyes on deployment, reconcile in real time, and keep the kill switch one action away.

Flash Crash · 6 May 2010

An algorithm with no sense of price

A sell programme of 75,000 E-mini contracts, about $4.1 billion, ran on an algorithm that kept selling while volume was strong, executing without regard to price or time. It completed in about 20 minutes, faster than the market could absorb it. [16]

For a self-improving system: every execution routine needs price and time constraints, not only a participation rate. A loop in which your own selling raises the volume that drives more of your own selling is a design flaw, not a market event.

Make the risk engine boring enough that nobody can talk it out of anything.

Saved in this browser only
M08

Security: break the lethal trifecta

Security · 35 min

Simon Willison's lethal trifecta is an agent with access to private data, exposure to content an attacker can control, and a way to communicate externally. [17] The blueprint has all three, plus a sharper version of the third: it can place orders.

The three legs, before and after
LegIn the blueprintIn the governed build
Private dataBalances, positions, keys and strategy logic within the orchestrator's reach.Held by the execution plane; research sees anonymised or aggregated state.
Untrusted contentNews, sentiment, web pages and custom datasets fed straight to the orchestrator.Quarantined readers whose output is a validated schema, never free text that can instruct.
Outbound channels and ordersOrchestrator passes strategies to order placement.Only signed strategies, through the risk engine. Research egress is allowlisted.
Evidence · Jan 2026

Invisible headline edits

Researchers altered one stock's headlines on one day in ways humans cannot see. Swapping Latin letters in company names for identical-looking Cyrillic ones broke the link between headline and stock on about 99% of edited headlines and cut cumulative return by about 3.7 points on average, 17.7 at worst. Hidden CSS text flipped sentiment on about 66% of headlines. The system stayed profitable, so the losses would likely go unnoticed. [13]

Evidence · Aug 2026

Multi-agent is not a defence

Across roles and architectures, no design was robust. Majority voting helped only while compromised agents stayed a minority. The authors' conclusion: security has to come from deliberate validation and aggregation, not from adding agents alone. [14]

Evidence · 2026

Grounding helps, partly

Automated misinformation red-teaming of a Bitcoin trading agent reached a 26.67% attack success rate. Letting the agent check text against recent price data cut it to 18.33%: useful, nowhere near sufficient. Simulated environment, one asset. [33]

These are the trading-specific faces of OWASP's LLM01 Prompt Injection and LLM06 Excessive Agency. [18]

Controls · input hygiene
  • Unicode NFKC normalisation and mixed-script detection on every entity name
  • Strip zero-width characters and visually hidden content before any model reads it
  • Provenance on every item: source, signature where available, reputation
  • Two independent sources before a market-moving item can move a position
  • Check text claims against price and volume before acting on them
Controls · authority
  • Research plane holds no credentials and no order tools
  • Model output is structured data validated by code, never concatenated into commands
  • Egress allowlist for every research process
  • Secrets in a vault, short-lived, out of prompts and logs
  • Broker API keys scoped to trading, never withdrawals, IP-restricted where offered
Controls · assurance
  • Injection eval suite, homoglyph and hidden-text cases included, run on every model or prompt change
  • Alert when one source drives a decision swing
  • Signed strategy artifacts; pinned package versions
  • Quarterly external red-team on the research plane

Assume the news was written by your adversary, and design so that it does not matter.

Saved in this browser only
M09

Governance and the regulatory perimeter

Authority · 40 min · Snapshot 29 Oct 2026

Humans retain final authority, and regulators increasingly say so in writing. The perimeter moves; this snapshot is analysis, not legal advice, and every row needs licensed counsel before you rely on it.

Roles and their limits
RoleHeld byDoesCannot
ProposerResearch agentsWrites hypotheses and codeTouch the vault, hold credentials, deploy
ReviewerIndependent agent and a human quantChallenges evidence, names failure modesApprove
ApproverA named person with authority and competencePromotes signed versions, sets limitsApprove their own change
OperatorPerson on dutyMonitors, triggers the kill switchChange limits alone
AuditorInternal or externalReads ledger and audit logWrite anything
Evidence pack · what the approver signs against

Hypothesis registration and trial ID · total trial count · gauntlet results by gate · DSR and PBO · vault result and access record · paper and shadow logs · cost model and break-even · risk limits · reviewer notes · rollback plan · regulatory perimeter note · the approver's decision and reason.

Regulatory snapshot · verified 29 October 2026
WhereInstrumentWhat it means for this build
EU · UKMiFID II algorithmic trading (Art. 17, RTS 6)Investment firms must test algorithms, hold kill functionality and pre-trade controls, and run an annual self-assessment approved by senior management. [21] For a private desk trading through a broker it is the best-practice template; it binds once you need authorisation.
EUAI Act, as amended by Regulation (EU) 2026/1744Published 24 July, in force 27 July 2026. Annex III high-risk obligations move to 2 December 2027 and Annex I to 2 August 2028; Article 50 transparency duties apply from 2 August 2026, with a grace period to 2 December 2026 for machine-readable marking by systems already on the market. [23] Analysis: trading algorithms are not an Annex III use case (credit scoring is), so financial-services law, not the AI Act, sets most duties here.
EUESMA statement on AI in investment services (30 May 2024)The management body stays accountable for AI-made decisions; clients are told how AI is used; models are tested, monitored and documented, staff use of outside AI tools included. [22]
EURetail CFD rules (ESMA 2018, now national)ESMA's 2018 measures set retail leverage from 30:1 on major currency pairs down to 2:1 on crypto, a 50% margin close-out and negative balance protection; national regulators had found 74–89% of retail CFD accounts lose money. [28] The temporary measures expired on 31 July 2019, and most national regulators replaced them with permanent rules at least as strict. [37]
USSEC Rule 15c3-5, market accessBroker-dealers with market access must maintain risk-management controls and supervisory procedures. Knight Capital paid $12 million under it, with inadequate controls for code deployment and testing among the charges. [15]
USAdvisers Act Marketing Rule · “AI washing”In March 2024 the SEC settled with Delphia ($225,000) and Global Predictions ($175,000) over AI claims they could not substantiate. [24] Any “self-improving AI” claim made to investors needs evidence on file.
USSEC predictive data analytics proposalWithdrawn on 12 June 2025. [26] Existing antifraud, fiduciary and marketing duties still apply.
USFINRA 2026 Regulatory Oversight Report (9 Dec 2025)Names agent risks (autonomy and scope creep, auditability, data handling, misaligned reward functions) and encourages firms to consider human-in-the-loop oversight, tracking of agent actions and guardrails. [25]
US banksModel risk guidanceSR 11-7 was rescinded on 17 April 2026. The revised interagency guidance excludes generative and agentic AI from scope, and, as of April 2026, the agencies planned a request for information on AI. [27] There is no ready-made supervisory template for LLM agents; your governance has to stand on its own.
TürkiyeSPK leveraged transactions (Tebliğ III-37.1)Retail leverage capped at 10:1 with a 50,000 TL minimum initial deposit since February 2017. [29] Reported as unchanged in May 2026; [36] confirm with your intermediary.
TürkiyeLaw 7518 (2 July 2024)Crypto-asset service providers brought under SPK authorisation. [30] Analysis: managing money or giving investment advice to others falls under Capital Markets Law licensing; counsel sign-off required.
MercosurNational securities regimesNot verified in this edition. Licensing for third-party funds or signal distribution must be confirmed with local counsel in each country before any offering.
The line that matters everywhere

The moment the system trades for anyone but its owner (clients, investors, often followers copying its signals) you are likely in licensed territory, and what you claim about it becomes regulated too.

Authority needs a name. Claims need evidence. Both need a date.

Saved in this browser only
M10

Operating the system

Operations · 30 min

Launch starts the experiment; it does not end it. Most of the value of the governed build sits in a weekly review that a person actually attends.

Monitor · alert when
SignalAlert when
Live versus backtest trackingOutside the tolerance band set at promotion
Slippage versus cost modelPersistently above the model
Risk-engine rejects by reasonAny spike: upstream bug or attack
Limit utilisationRunning near caps for long periods
ReconciliationAny break
Input integrityMixed-script entities, feed gaps, one source swinging a decision
Model and providerAny change to a pinned version or a deprecation notice
HeartbeatsAny missed beat
Error taxonomy · replaces the blueprint's six boxes
  • Performance: underperformance, drawdown, overtrading
  • Evidence: leakage, look-ahead, survivorship, undercounted trials
  • Cost: spread, slippage and financing drift
  • Input integrity: manipulated or corrupted text and data
  • Model behaviour: nondeterminism, version drift, changed refusals
  • Infrastructure: connectivity, clock skew, broker outages
  • Process: unapproved change, skipped gate, expired review
SEV1

Uncontrolled exposure or loss breaker tripped

Level 2 kill, approver paged, written review before any restart.

SEV2

A control is degraded

Reconciliation break or a monitor down: Level 1 pause until restored.

SEV3

Drift within limits

Logged and taken to the weekly review.

Retirement criteria · written before launch
  • Live performance below the registered floor over the registered horizon
  • Slippage consistently above the cost model
  • The mechanism is invalidated by a change in market structure
  • Capacity exhausted at the approved size

Retire to shadow, never delete. The ledger keeps the history, including the failures.

Weekly review · 45 minutes, a person in the chair
  1. New trials and the running count
  2. Candidates at each gate
  3. Live against expected, per strategy
  4. Incidents and rejects
  5. Limit utilisation
  6. Regulatory or provider changes
  7. Decisions, owners and dates, recorded

A self-improving system needs a self-reviewing owner.

Saved in this browser only
M11

Twelve-week build plan and go-live gate

Capstone · 40 min · Tool 04

Build the governance first and the strategy last. By week twelve the system should be able to tell you, with evidence, that it has nothing worth deploying. That is the most valuable sentence it will ever produce.

Build sequence
WeeksBuildExit criterion
01–02Experiment ledger, point-in-time store, data contracts, licences confirmedEvery dataset has a contract; a dummy trial is counted end to end
03–04Research sandbox: agents, independent reviewer, backtester on development dataA red-team test confirms the research plane cannot reach the broker or the vault
05–06Validation plane: purged CV, broker-quote cost model, robustness, DSR and PBOThe gauntlet rejects a deliberately overfitted strategy
07Risk engine and kill switch; broker adapter as sole credential holderEvery reason code unit-tested; all three kill levels rehearsed in paper
08Input hygiene, schema outputs, egress allowlist, injection eval suiteHomoglyph and hidden-text cases neutralised in evals
09–10Paper, then shadow, for the first candidate that survived the vaultTracking inside tolerance for the pre-set period
11Evidence pack, named approver, perimeter review with counselA signed approval, or a documented no
12Canary with a loss budget; weekly review in the calendarFirst review held; retirement criteria filed
Tool 04 · Go-live gate

Twenty-four checks before real money

Items marked CRITICAL block go-live on their own. Ticks are saved in this browser only.

0 / 24 Blocked · 8 critical items open

Passing the gate earns a canary, not a fortune.

Saved in this browser only
Q

Knowledge check

Tool 05 · 10 questions · instant feedback

Ten questions, one per idea that matters most. Answers explain themselves, so a wrong answer still teaches.

0 of 10 answered
§

Sources

Checked 29 October 2026 · figures as reported by each source

Where a figure in this course is computed rather than quoted (returns from reported balances, Deflated Sharpe examples, simulator output), the text says so. Secondary sources are used where the primary text was not retrievable; read the primary before relying on a number.

  1. Harvey, Liu & Zhu, “…and the Cross-Section of Expected Returns”, NBER Working Paper 20592, 2014. nber.org
  2. Bailey & López de Prado, “The Deflated Sharpe Ratio”, Journal of Portfolio Management 40(5), 2014. ssrn.com
  3. Bailey, Borwein, López de Prado & Zhu, “The Probability of Backtest Overfitting”. ssrn.com
  4. Dwork, Feldman, Hardt, Pitassi, Reingold & Roth, “The reusable holdout”, Science, 2015 (Google Research summary). research.google
  5. Purged cross-validation, embargo and CPCV, after López de Prado, Advances in Financial Machine Learning, 2018. wikipedia.org
  6. McLean & Pontiff, “Does Academic Research Destroy Stock Return Predictability?”, Journal of Finance 71(1), 2016. pdf
  7. Glasserman & Lin, “Assessing Look-Ahead Bias in Stock Return Predictions Generated by GPT Sentiment Analysis”, arXiv:2309.17322, 2023. arxiv.org
  8. He, Lv, Manela & Wu, “Chronologically Consistent Large Language Models”, arXiv:2502.21206. arxiv.org
  9. ForkLog, “Four out of six AI models suffer losses in trading tournament” (Alpha Arena Season 1), Nov 2025. forklog.com
  10. ForkLog, “AI model Grok 4.2 triumphs in trading tournament” (Season 1.5), 8 Dec 2025. forklog.com
  11. Li et al., “R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization”, arXiv:2505.15155. arxiv.org
  12. Xiao, Sun, Luo & Wang, “TradingAgents: Multi-Agents LLM Financial Trading Framework”, arXiv:2412.20138. arxiv.org
  13. Rizvani, Apruzzese & Laskov, “Adversarial News and Lost Profits: Manipulating Headlines in LLM-Driven Algorithmic Trading”, arXiv:2601.13082, Jan 2026. arxiv.org
  14. Na et al., “Poisoning Agentic Alpha: Adversarial Vulnerabilities Across Roles and Architectures in Multi-Agent Trading Systems”, arXiv:2608.24069, Aug 2026. arxiv.org
  15. U.S. SEC, Press Release 2013-222, Knight Capital Market Access Rule charges, Oct 2013. sec.gov
  16. Christian Science Monitor, “Flash crash report: a tale of how not to make a big trade”, on the SEC–CFTC report, Oct 2010. csmonitor.com
  17. Simon Willison, “The lethal trifecta for AI agents”, 16 Jun 2025. simonwillison.net
  18. OWASP Top 10 for LLM Applications 2025 (Invicti summary). invicti.com
  19. TradingView, Terms of Use, Policies and Disclaimers. tradingview.com
  20. MQL5 documentation, MetaTrader 5 Python integration: order_send. mql5.com
  21. UK FCA, “Algorithmic Trading Compliance in Wholesale Markets”, Feb 2018. fca.org.uk
  22. ESMA, Public Statement on the use of AI in the provision of retail investment services, 30 May 2024. esma.europa.eu
  23. Lewis Silkin, “The Digital Omnibus on AI enters into force today”, 27 Jul 2026. lewissilkin.com
  24. Morrison Foerster, “SEC targets ‘AI washing’ with two new settled cases”, Mar 2024. mofo.com
  25. Debevoise & Plimpton, on FINRA's 2026 Regulatory Oversight Report, Dec 2025. debevoise.com
  26. Ropes & Gray, “SEC clears unfinished rulemakings from regulatory agenda”, Jun 2025. ropesgray.com
  27. Orrick, “Agencies overhaul model risk management guidance for banks”, Apr 2026. orrick.com
  28. ESMA, “ESMA agrees to prohibit binary options and restrict CFDs to protect retail investors”, 27 Mar 2018. esma.europa.eu
  29. QNB Invest, notice on the SPK leverage amendment (Tebliğ III-37.1), 10 Feb 2017. qnbinvest.com.tr
  30. SPK, “Kripto Varlık Hizmet Sağlayıcılara İlişkin Duyuru” (Law 7518), 2 Jul 2024. spk.gov.tr
  31. Thinking Machines Lab, “Defeating Nondeterminism in LLM Inference”, Sep 2025. thinkingmachines.ai
  32. Lopez-Lira & Tang, “Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models”, arXiv:2304.07619. arxiv.org
  33. Liu et al., “AutoRedTrader: Autonomous Red Teaming of Trading Agents through Synthetic Misinformation Injection”, arXiv:2605.09185. arxiv.org
  34. Bitcoin.com News, “6 bots with real money: Hyperliquid hosts first-ever AI trading showdown”, Oct 2025. bitcoin.com
  35. J. R. Varma, “SEC order explains Knight Capital systems failure”, 2013, summarising the SEC administrative order. jrvarma.in
  36. Track360, “Forex kaldıraç nedir? SPK sınırları…” (2026 guide), May 2026. track360.io
  37. Mondo Visione, “ESMA ceases renewal of product intervention measures relating to contracts for difference”, Jul 2019. mondovisione.com
Work with ImageFirm

Building an AI system that touches money?

We design the authority boundaries, the evidence trail and the go-live gate before the first line of strategy code is written. Bring one workflow, not a vague AI ambition.

Start an enquiry