IQ IqraLearning

Fintech

Trading Desk | Fintech AI Research

Eleven-agent paper desk where RISK outranks the judge.

Atif Jaffery · AI Architect — committee design, verdict rules, and directed implementation

11 agents · RISK outranks JUDGE

01The situation

In 60 seconds

Business problem

Operators want expert-grade equity research without one model’s “tip” and without real brokerage risk. A single persuasive narrative hides disagreement and invites overconfidence.

Unacceptable failure

A position entered without a complete level card, or one that RISK would have vetoed.

Design move

Five playbooks debate in two rounds. JUDGE applies deterministic verdict rules. RISK, in pure code, sizes or vetoes, and outranks JUDGE.

What it shows

Every verdict carries a level card or a logged HOLD, and dissent stays in an immutable ledger.

The problem

Single-model tips hide disagreement and invite overconfidence. An operator needs entry, stop and invalidation levels before any position, and a backtest is worthless if the data leaks the future.

Who feels it

  • Operators using the paper-trading Command Deck
  • Anyone reviewing the desk’s weekly scoring against SPY

What’s at stake

  • Overconfident entries from one persuasive narrative.
  • Positions with no stop and no invalidation.
  • Backtests corrupted by lookahead data.

Constraints

  • Paper trading only. Recommendation-only. Not financial advice.
  • Point-in-time data only.
  • No live brokerage.

Failures this system must never produce

  1. F1

    An entry without a complete level card

    No position without entry, stop, targets, invalidation and a time stop. A stop hit is a mechanical exit.

  2. F2

    Persuasion beating risk

    Language can’t override sizing, caps or the circuit breaker.

  3. F3

    Lookahead in the data

    A fact that isn’t point-in-time stamped is corruption, not signal.

02The approach

How AI solves it

Each failure gets one mechanism, and each mechanism has proof further down this page.

Failure Mechanism The model’s role Proof
F1
Mechanism

JUDGE tallies two debate rounds with deterministic verdict rules and thresholds set by market regime. A tie is HOLD, and no data means no vote.

The model’s role

LLMs write the arguments. The tally is rules.

Explorer →
F2
Mechanism

RISK sizes, caps and runs the circuit breaker after JUDGE, and can veto. Goals change position size, never the entry bar.

The model’s role

None. RISK and TUDOR are pure code.

Guardrails →
F3
Mechanism

JANITOR cleans, validates and point-in-time stamps every feed and quarantines bad data. HAWK watches for invalidation.

The model’s role

Deterministic, with a thin LLM inside JANITOR.

Explorer →

Where the model works, and where it doesn’t

Model-assisted

  • WARRENBuffett-style fundamentals from EDGAR. LLM-heavy.
  • STEVIECatalysts, Form 4 and sentiment. LLM-heavy.
  • STAN and NANCYData first, with a thin LLM on top.
  • JUDGE (arguments)Writes the debate summary; rules decide the verdict.
  • RECEPTIONThe only user-facing language layer.

Rule-bound or human

  • TUDORTrend, moving averages and asymmetry. No LLM.
  • RISKSizing, caps and circuit breaker. No LLM. Outranks JUDGE.
  • JUDGE (tally)Verdict thresholds by regime, covered by tally tests.
  • JANITOR and HAWKData hygiene and market watch, deterministic.
  • SCRIBEImmutable ledger and contested-HOLD tracking.
  • OperatorMakes paper-trading decisions.

Agents and who holds authority

Five playbooks argue; only rules decide. JUDGE tallies the debate, and RISK, in pure code, can veto it.

Click any card for its contract: inputs, outputs, decision authority, and guardrails.

LLM model-assisted Code deterministic Human person in the loop
Cleans · Watches

Data

Nothing reaches the committee without a point-in-time stamp.

Argues

Committee

Two rounds of debate. STAN gates the regime instead of voting.

Decides

Verdict

Rules turn the debate into ENTER or HOLD, with a level card.

Vetoes · Acts

Authority

Risk math outranks the verdict; a person makes the paper trade.

Explains · Records

Front & record

Plain English for the operator; an immutable log for review.

Guardrails & trust

  • Paper trading only; recommendation-only; not financial advice.
  • No position without a complete level card; stop hit = mechanical exit.
  • Tie = HOLD; no data = no vote.
  • RISK outranks JUDGE; goals change size, never the entry quality bar.
  • Point-in-time data only—lookahead treated as corruption.

03See it work

See it run

A 30-second walkthrough of one ticker through the committee, and a live paper-trading demo.

Live demo · cloud

Coming soon

Put one ticker through the committee

  • Ask about one ticker and watch both debate rounds
  • See STAN’s regime gate and the vote tally
  • Read the level card: entry, stop, targets, invalidation, time stop
  • Watch RISK resize or veto
  • Find the contested HOLD in the SCRIBE ledger
Open live demo · soon Book a walkthrough
  1. 0:00–0:06 Intake

    RECEPTION turns the operator’s goal into a risk envelope.

  2. 0:06–0:14 Debate

    Five playbooks argue in two rounds.

  3. 0:14–0:22 Verdict

    JUDGE tallies by regime and drafts a level card.

  4. 0:22–0:30 Risk

    RISK sizes or vetoes, and SCRIBE logs the result.

Interactive architecture

Walk a real request through agents, services, stores, and human gates. Click any node for its contract.

04The result

What changes for the operator

Disagreement becomes a research tool

Contested HOLDs are logged instead of forced into a trade.

Plain-English recommendations with risk framing

Every ENTER comes with a level card, or it’s a HOLD (F1).

Every call is on the record

SCRIBE keeps an immutable ledger, scored weekly against SPY.

How I check it behaves

Paper-trading research results. These are not financial performance claims.

Check What it counts Target Last run
Entries without a level card F1 Positions opened without entry, stop, targets, invalidation and time stop 0 Run pending
Ties resolved as HOLD F1 Tied tallies that became anything other than HOLD 0 Run pending
RISK vetoes honored F2 Vetoed verdicts that still entered a position 0 Run pending
Lookahead leaks F3 Facts used before their point-in-time stamp 0 Run pending

05Looking back

Architecture decisions

  • Parallel agents and consensus, not a cascade F1

    So one agent’s error can’t flow through unchecked.

    Cost acceptedMore model calls per ticker.

  • TUDOR and RISK as pure code F2

    So the critical math can’t hallucinate.

    Cost acceptedThose rules change only through code review.

  • Surface contested HOLDs

    So disagreement is visible instead of forced into action.

    Cost acceptedMore HOLDs and fewer trades.

Known limits and next step

Research · disclosed

Live market validation is still in progress. This is a research desk, and every result so far is a paper result.

The committee design and tally rules are built and tested; how well the playbooks perform is still being measured.

Next step

  1. Publish the weekly agent scoring against SPY with its date range.
  2. Keep live brokerage out of scope.

Also disclosed

  • Not a broker. Not financial advice.

Project demo