STAGE 1CURRENT
Simulated only
All research and trading decisions remain on paper. No broker connection, no real capital at risk.
01 / DELIVERY ROADMAPBUILT → BUILDING → EXPLORING
We built the agent. Now we’re building the infrastructure.
HWA Engine is being developed in deliberate phases: a Claude-built autonomous research agent that already operates in simulated markets, a broker-connected execution layer under active development, and a management interface only when operational needs justify it.
02 / PHASE 0 — AUTONOMOUS RESEARCH & PAPER TRADINGIMPLEMENTED & TESTED
A working agent that questions its own calls.
Phase 0 was fully implemented with Claude AI and Claude Code, and it has operated simulated research and trading workflows on a scheduled weekday cadence — open, midday, and close — across three paper books. Every documented thesis can be revisited with its original evidence, expectation, and invalidation criteria intact.
- Three $100k paper portfolios — swing, long-term, and meme — independently tracked
- Six real Claude Code skills defining the agent’s entire behavior
- Scheduled autonomous research and trade runs, each leaving a dated journal
- 25 reconciled simulated fills at the 2026-10-08 report date
- One verified reflective-learning chain: review → lesson → later citations
THE REFLECTION LOOP
What every journal entry records.
Thesis, dated sources, falsifiable expectation, invalidation level, simulated fill, review-by date, outcome, and the lesson worth keeping.
01 →
ResearchBuild a thesis from dated market evidence.
02 →
Paper tradeRecord simulated decisions and fills.
03 →
ReviewMeasure both reasoning quality and outcomes.
04 ↺
RememberFeed persistent lessons into the next call.
PROCESS DIAGRAM DERIVED FROM AUTOSTOCKTRACKER METHODOLOGY · NOT A LIVE TRADE OR PERFORMANCE DEMO
PUBLIC CASE STUDY — THE RECORDED XOM CHAIN
A decision is only the beginning.
AN ACTUAL RESEARCH-TO-LESSON CYCLE GENERATED BY THE PHASE 0 SYSTEM
The record contains one complete reflective-learning chain so far, and it is reproduced here exactly as dated — a simulated call on XOM, reviewed against its own thesis, generalized into a persistent lesson, and cited in later research.
- 012026-08-10Simulated thesis & decision
BUY XOM — $6,000 paper fill at $159.79 on a live supply catalyst, entered with a falsifiable expectation (≥$170 within 4 weeks) and an invalidation fixed at entry.
- 022026-09-08Retrospective review
The catalyst played out — crude rose ~12% — yet XOM moved just +0.45%. Verdict: reasoning right, outcome lagging; a timing and magnitude miss, not a broken thesis.
- 032026-09-08Lesson L1 recorded
Written to LESSONS.md: integrated majors lag pure upstream producers on a crude-spike catalyst — with concrete sizing and timeframe guidance for future calls.
- 042026-09-09 →L1 cited in later work
Six subsequent journal entries cite L1 by name during position monitoring. The record shows the lesson informing later research — not yet changing a brand-new entry decision.
THE SYSTEM CAPTURES LESSONS IN PERSISTENT PROJECT MEMORY, AND CLAUDE’S SKILLS DIRECT LATER WORKFLOWS TO READ THEM — THIS IS NOT RETRAINING THE UNDERLYING MODEL’S WEIGHTS
SOURCES IN THE PHASE 0 RECORD: RESEARCH JOURNAL · RETROSPECTIVE REPORT · LESSONS.MD · DATED MONITORING REFERENCES
HOW CLAUDE IS USED
Two distinct, documented roles.
BUILDING THE SOFTWARE
Claude Code implemented the system itself — the research methodology, the six skills, the journals and ledger structure, and the reporting views — in interactive sessions under the founder’s direction.
OPERATING THE RESEARCH
Separately, scheduled Claude workflows run the recurring stock analysis, simulated trade decisions, daily reports, and retrospective reviews. Each run writes its own dated records, which remain the authoritative evidence.
.claude/skills/review-calls/SKILL.md — excerpt, verbatim
# Review calls
The learning half of the loop. A call is **due** when its
Invalidation or Expectation level was hit, its **Review by**
date has passed, or the user asks for a review.
2. **Reconstruct what happened** per due call: price path since
entry vs the Expectation and Invalidation written at entry;
… Compare against the evidence the call originally cited.AN AUTHENTIC EXCERPT FROM A PHASE 0 SKILL DEFINITION — NOT A RECREATED CLAUDE CONVERSATION
FROM EVIDENCE TO ROADMAP
- PHASE 0 · OPERATIONALClaude research agent
Autonomous simulated research, decisions, and review — the system above.
- PHASE 1 · BUILDINGRisk / approval boundary
Independent limits and human authorization between analysis and any order.
- PHASE 1 · BUILDINGRobinhood MCP connector
Broker connectivity under development — not an operational capability.
- PHASE 2 · FUTUREOptional UI
Monitoring and administration, unscheduled; not required by the headless engine.
THE COMPLETED CAPABILITY IS SIMULATED PAPER TRADING. NOTHING HERE IMPLIES BROKERAGE CONNECTIVITY IS OPERATIONAL.
03 / PHASE 1CURRENT ENGINEERING PRIORITY — ACTIVE DEVELOPMENT
Taking research into the execution layer.
Our current focus is Robinhood MCP connectivity for a single initial strategy, without abandoning the three research books. The bridge between AI analysis and brokerage orders must be testable, inspectable, and independently constrained.
DEVELOPMENT STANCE
Broker connection, autonomous execution, and live trading are not represented as production capabilities. All live-account testing is contingent on brokerage permissions, security checks, and appropriate human oversight.
PHASE 1 / IMPLEMENTATION TARGETSCURRENT
A
Single-strategy integration
Choose one research approach for the first validated execution pipeline.
B
Permissioned broker adapter
Connect eligible Robinhood MCP tools; keep account access and credentials outside the public website.
C
Deterministic risk boundary
Reject actions that violate preset limits independently of model output.
D
Audit, reconciliation, and rollback
Track decisions and order state, then verify safe shutdown behavior before any limited live testing.
NO TRADING INTERFACE REQUIRED FOR PHASE 1 · HEADLESS DEVELOPMENT
04 / VALIDATION PATHHOW TESTING PROGRESSES
Live testing is earned, stage by stage.
Each stage must hold up before the next is considered. None of these stages involves customer capital, and the final stage remains contingent on brokerage permissions and appropriate human oversight.
STAGE 2NEXT
Controlled broker dry-run
Exercise MCP connectivity and the risk boundary end-to-end without submitting live orders.
STAGE 3GATED
Limited approved testing
Founder-account testing only — and only where law, platform policy, and required oversight permit it.
05 / PHASE 2FUTURE — UNSCHEDULED
Management interface, only when it matters.
A potential monitoring, configuration, and reporting UX may follow once the execution layer earns it. There is no committed build date — and this website is not that interface. Infrastructure comes first.
06 / WHERE THIS GOESLONG-TERM GOALS
Three goals, stated plainly.
Democratize engineering capacity
Use Claude as an assistant in software development, research synthesis, and iterative agent workflows — while preserving accountability for every decision.
Make reasoning inspectable
Connect hypotheses, data sources, proposed actions, retrospectively reviewed outcomes, and reusable lessons into one traceable record.
Build toward controlled automation
Explore broker-connected execution only after independently validated restrictions, appropriate authorizations, and monitoring are working.
NO CLAIMS OF GUARANTEED MARKET EFFICIENCY, TRADING ALPHA, OR ELIMINATED INVESTMENT RISK — THESE ARE ENGINEERING GOALS, NOT PERFORMANCE PROMISES.
07 / CONTACT
elaina@hwagallery.com ↗Let’s talk systems.
For research collaboration, technical partnerships, or company inquiries, contact Hwa Gallery & Co. directly.