Lincei Quant Research Engine

Status: active self-funded capital-first QuantConnect/LEAN + LLM alpha system.

Last aligned: 2026-05-27.

Lincei is a personal autonomous alpha system whose first monetization goal is self-funded capital allocation: running continuously, researching strategies, collecting validation artifacts, and eventually trading the operator's own pre-funded capital only after promotion evidence, pre-trade risk checks, and reconciliation gates pass.

Darwinex/Zero is a later monetization path. It should not drive the first architecture. A Darwinex/Zero track record is useful only after the self-funded capital loop can produce a defensible signal and execution record.

Read SPEC.md first. It is the canonical active spec index. If another document conflicts with SPEC.md, SPEC.md wins.

Direction

• Self-funded capital allocation is the priority.
• QuantConnect Cloud and LEAN are the strategy validation/runtime foundation.
• The Oracle Cloud ARM server is the intended always-on control plane for scheduling research, data, feature, ablation, import, paper trading/shadow trading, reconciliation, and alert jobs.
• LLMs are LLM-derived feature generators, not allocators and not broker operators.
• Parallelization is expected before promotion: corpus ingest, hypothesis extraction, data ingest, feature jobs, LLM jobs, ablations, backtests, and Cloud artifact imports should run concurrently where safe.
• Execution-like work is single-writer: promotion, portfolio target consolidation, risk cuts, paper trading/shadow trading intent, reconciliation, and broker-write pre-trade risk checks must have one canonical state.
• Real broker writes remain blocked until a separate broker-write implementation spec is approved.

Core Hypothesis

The project hypothesis is:

A point-in-time parallel research pipeline that combines durable numeric baselines, ML features, and LLM-derived features can produce after-cost, benchmark-relative returns that survive QuantConnect Cloud validation, current paper trading/shadow trading artifacts, reconciliation, and later self-funded capital execution.

The Alpha Architect corpus review changes the near-term priority:

1. Prove simple liquid baselines first: trend following, defensive allocation, momentum, daily-return features, and cost-aware rebalancing.
2. Add LLM-derived alpha only as typed features and ablations.
3. Treat factor crowding, factor valuation, macro regimes, index effects, and filing language as research hypotheses that need broader data and vintage controls.
4. Revisit Darwinex/Zero after the self-funded capital deployment-grade track record exists.

System Flow

flowchart TB
    subgraph PAR["Parallel research pipeline"]
        CORPUS["Research corpus<br/>articles / papers / notes"]
        HYP["Hypothesis registry"]
        DATA["Market / news / filing / macro ingest"]
        FEAT["Feature jobs<br/>symbol / source / window"]
        LLM["LLM-derived feature jobs"]
        ABL["Ablations<br/>numeric / LLM / combined"]
        BT["Backtest and parameter sweeps"]
        CLOUD["QuantConnect Cloud imports"]
    end

    CORPUS --> HYP
    DATA --> FEAT
    HYP --> ABL
    FEAT --> ABL
    LLM --> ABL
    ABL --> BT
    BT --> CLOUD

    CLOUD --> LEDGER["Promotion ledger<br/>all variants retained"]
    LEDGER --> ALPHA["Approved AlphaDecision"]
    ALPHA --> LEAN["LEAN Insight"]
    LEAN --> TARGET["Portfolio targets"]
    TARGET --> RISK["Risk cuts"]
    RISK --> PAPER["Paper trading/shadow trading intent"]
    PAPER --> RECON["Reconciliation"]
    RECON --> PREFLIGHT["Broker-write pre-trade risk check<br/>blocked until approved spec"]

Loading

Parallel jobs improve research throughput. They do not create multiple execution truths. The system can evaluate many hypotheses at once, but only one consolidated target set can advance toward execution evidence for a given strategy/account/time.

Evidence Ladder

stateDiagram-v2
    [*] --> Hypothesis: research corpus
    Hypothesis --> Baseline: simple numeric strategy
    Baseline --> Ablation: numeric / LLM / combined
    Ablation --> LocalLEAN: local LEAN debug
    LocalLEAN --> CloudImport: QuantConnect Cloud backtest/import
    CloudImport --> PaperShadow: current paper trading/shadow trading
    PaperShadow --> Reconciliation: intended vs observed state
    Reconciliation --> PromotionReview: cost, slippage, tax, bias checks
    PromotionReview --> OwnCapitalCandidate: broker-read-only then write spec
    PromotionReview --> Rejected: failed or overfit evidence
    OwnCapitalCandidate --> DarwinexCandidate: later track-record path

Loading

Promotion requires direct supporting evidence. Unit tests protect schemas and API contracts, but they do not prove strategy performance or broker-write readiness.

Research Corpus

The repo stores an initial Alpha Architect corpus for strategy research:

• references/alphaarchitect/README.md
• references/alphaarchitect/index.json
• references/alphaarchitect/strategy-register.md
• docs/own-capital-alphaarchitect-corpus-review.md

The corpus contains 40 sourced articles with metadata and content hashes. It is a hypothesis library, not trading evidence. A stored article becomes useful only after it is mapped to point-in-time data, implemented as a LEAN strategy or feature, and validated through the evidence ladder.

Current Architecture

Layer	Role
Research corpus	Stores articles, papers, source metadata, content hashes, and hypothesis candidates.
Hypothesis registry	Converts research into testable strategy variants and failure modes.
Feature store	Preserves point-in-time and vintage features with availableAt, source refs, and hashes.
LLM-derived alpha	Extracts structured features from text evidence; never emits broker instructions.
Numeric/ML alpha	Provides simple baselines and model features for ablation.
LEAN / QuantConnect	Owns strategy runtime semantics: Insight, portfolio construction, risk, execution in approved modes.
Control plane	Orchestrates jobs, imports artifacts, records ledgers, reconciliation, pre-trade risk checks, and dashboard state.
Broker boundary	Read-only and pre-trade risk checks first; broker writes require a future approved spec.

Repository Map

Path	Purpose
SPEC.md	Canonical active spec index
AGENTS.md	Agent/contributor rules
terminology.md	Canonical terms
docs/spec/	Normative split spec
docs/own-capital-alphaarchitect-corpus-review.md	Self-funded capital architecture review from research corpus
references/alphaarchitect/	Stored Alpha Architect article corpus and strategy register
config/universes/	Universe manifests and caps
backend/	NestJS control plane and ledgers
engines/lean/aggressive_llm_momentum/	LEAN Algorithm Framework strategy
ml/	Offline feature/model helpers
frontend/	Operational dashboard
scripts/	Operator command wrappers

Implementation Status

Implemented in the current branch:

• durable research hypothesis registry and parallel research job ledger;
• Alpha Architect corpus-to-hypothesis ingestion script with idempotent job records;
• multiple-testing bias promotion check that blocks when only winning/no variant evidence exists;
• point-in-time text evidence ingestion, including Hugging Face FOMC evidence;
• QuantConnect Cloud project/backtest listing and manual Cloud backtest import;
• paginated Cloud insights/orders import with Cloud id preservation;
• paper replay separated from current paper trading/shadow trading readiness;
• backtest-cycle dashboard;
• Alpha Architect corpus with 40 sourced articles and self-funded capital strategy review;
• long-term specs for self-funded capital priority, Darwinex/Zero deferral, and parallel research pipeline.

Not implemented yet:

• broad research universe profiles separate from the current theme universe;
• complete vintage-data store for restatable sources;
• simple trend/momentum/daily-return baselines with promotion evidence;
• broker-read-only reconciliation;
• broker-write adapter;
• Darwinex/Zero execution or track-record adapter.

Setup

git clone alone is not enough. Secrets, local data, LEAN workspace state, generated artifacts, and SQLite databases are intentionally gitignored.

Prerequisites:

Tool	Used for
Bun	Backend, frontend, CLI wrappers
Python 3.10+	ML helpers and Lean CLI venv
Docker or Podman-compatible Docker CLI	Local LEAN container runs
QuantConnect account/API token	Cloud backtests, workspace setup, Object Store
OpenAI-compatible API key	LLM-derived features
Oracle Cloud ARM host	Optional always-on control plane target

Bootstrap:

git clone <repository-url>
cd lincei-quant-research-engine
cp .env.example .env

# Fill QUANTCONNECT_USER_ID, QUANTCONNECT_API_TOKEN, OPENAI_API_KEY as needed.
./scripts/bootstrap-dev.sh

Command Cheat Sheet

Run from repository root unless noted.

# Primary capital evidence loop
bun --cwd=backend run lincei -- capital run --max-backtest-workers 1 --json
bun --cwd=backend run lincei -- capital status --json

# Research and text evidence
bun --cwd=backend run lincei -- research corpus
bun --cwd=backend run lincei -- research selected-run-bias
bun --cwd=backend run lincei -- data semantic-evidence --limit 80
bun --cwd=backend run lincei -- alpha run

# Local LEAN and import
bun --cwd=backend run lincei -- lean backtest aggressive_llm_momentum
bun --cwd=backend run lincei -- lean import latest
./scripts/run-local-strategy-smoke
./scripts/verify-lean-cloud-package aggressive_llm_momentum

# QuantConnect Cloud
bun --cwd=backend run lincei -- qc list-projects
bun --cwd=backend run lincei -- qc list-backtests --project-id <project-id> --limit 10
bun --cwd=backend run lincei -- qc import-backtest --project-id <project-id> --backtest-id <backtest-id>
bun --cwd=backend run lincei -- qc cloud-backtest aggressive_llm_momentum
./scripts/qc-cloud-push aggressive_llm_momentum

# Paper trading, shadow trading, learning, broker-write pre-trade risk check
bun --cwd=backend run lincei -- paper run
bun --cwd=backend run lincei -- paper replay
bun --cwd=backend run lincei -- shadow run
bun --cwd=backend run lincei -- learning run
bun --cwd=backend run lincei -- preflight run

The old ./scripts/* operational wrappers are compatibility surfaces during migration; new operator commands should go through lincei.

Exit code 2 means a policy/account/platform blocked result, not necessarily a crash. Examples include missing QuantConnect Cloud credentials, account-tier blockers, missing current paper trading/shadow trading artifacts, stale targets, data-license blockers, or broker-write pre-trade risk check blockers.

Verification

Use the narrowest command that proves the touched surface:

git diff --check

cd backend && bun run build
cd backend && bun run test

cd frontend && bun run typecheck
cd frontend && bun run build
cd frontend && bun run test:run

.venv-ml/bin/python -m pytest engines/lean/aggressive_llm_momentum/tests
./scripts/verify-lean-cloud-package aggressive_llm_momentum
./scripts/run-local-strategy-smoke
./scripts/run-cloud-quality-backtest
./scripts/run-paper-cycle
./scripts/run-live-shadow
./scripts/run-learning-loop
./scripts/live-preflight

Final reports must separate unit-test evidence, local LEAN evidence, QuantConnect Cloud artifacts, paper trading/shadow trading evidence, reconciliation evidence, and blockers.

Key Rule

Build the self-funded capital loop before anything else:

hypothesis -> baseline -> ablation -> Cloud artifacts -> current paper trading/shadow trading -> reconciliation -> broker-read-only -> broker-write spec

Everything else is supporting infrastructure.