01 / PLATFORM ARCHITECTUREHWA ENGINE

Autonomy built on research, grounded in controls.

HWA Engine is a headless backend — a service, not a customer trading UI. It extends Claude-driven research with an explicit boundary between analysis, authorization, and broker execution: AI can propose; independent software and human authorization control what is permitted.

01 / INPUTEXISTS

Market data

Price context, fundamentals, research sources, and defined hypotheses.

02 / REASONINGEXISTS

Claude research

Structured analysis, falsifiable calls, reviews, and persistent lessons.

03 / CONTROLPLANNED

Risk engine

Independent deterministic limits: position bounds, permissions, exposure checks, shutdown.

04 / AUTHORIZATIONPLANNED

Human authorization

Required review and approval boundaries before any action can reach a broker.

05 / CONNECTIONIN DEV

Robinhood MCP

Broker connectivity and controlled order workflows where supported.

06 / RECORDFOUNDATION

Audit & monitoring

Research records exist; broker-order reconciliation and alerts are Phase 1 goals.

IMPLEMENTED FOUNDATIONPHASE 1 WORKFuture-state Phase 1 architecture, in development. The displayed elements have different implementation states — see each label.

02 / REASONING LAYERPHASE 0 — IMPLEMENTED & TESTED

A headless research system that shows its work.

The engine’s foundation is AutoStockTracker, a Claude Code research workspace. Every simulated call is recorded with its thesis, dated evidence, expected outcome, and the price level that would prove it wrong — then revisited in retrospective reviews.

Three independently tracked paper books, each starting from a simulated $100k
Valuation-first analysis with falsifiable expectations and invalidation levels
Simulated fills and portfolio journals — no real-money trades in Phase 0
Persistent lessons read before subsequent calls: reflection, not retraining

03 / AUTONOMOUS CLAUDE WORKFLOWSPHASE 0 — IMPLEMENTED & TESTED

Claude operates the research loop — it didn’t just build the website.

Six Claude Code skills define everything the Phase 0 agent does, executed on a scheduled weekday cadence: research at the open, midday, and close. Every run leaves its journals, fills, and reports in the record. Claude the development assistant wrote the system; Claude the agent runs it.

/analyze
Valuation-first deep dive on one ticker, ending in a verdict per research book.
/market-scan
Screener-sourced undervalued candidates, market pulse, and sentiment sweep.
/trade-run
Fills pending orders, sweeps invalidations and targets, executes simulated trades, and writes the day’s journal.
/meme-run
Sentiment-driven meme-book judgments — no valuation gate, tight stops.
/daily-report
Close-of-day report: book status, allocation charts, valuation table, tomorrow’s plan.
/review-calls
Retrospective: judges past calls against their original theses and extracts durable lessons.

Phase 0 operations page: the six Claude Code skill definitions rendered from the repository, the scheduled weekday routine cadence, and the run-coverage record including its known gaps. — How the agent runs — the six Claude Code skill definitions, the scheduled weekday cadence, and the run-coverage record with its gaps stated plainly. VIEW FULL SIZE ↗

04 / THE LEARNING LOOPPHASE 0 — IMPLEMENTED & TESTED

Lessons that outlive the trade.

After each review, generalizable observations are written to a persistent lessons file that is read before subsequent calls. Earlier reasoning errors stay visible to future decision cycles — a documented feedback loop between past reviews and new analysis.

WHAT THIS IS — AND ISN’T

This is journal-based reflection: recorded lessons informing future prompts and methods. It is not model fine-tuning, self-training, or persistent weight updates, and it is not presented as a proven improvement in trading returns.

HOW OUR REVIEW PROCESS WORKS

Every reviewed call is judged on two independent axes: was the reasoning sound, and was the outcome favorable?

SOUND REASONINGFAVORABLE OUTCOME

Confirm

The thesis worked for the stated reasons. Keep the approach and note what made it repeatable.

SOUND REASONINGUNFAVORABLE OUTCOME

Examine

Discipline held but the market disagreed. Look for the missing variable before changing the process.

FLAWED REASONINGFAVORABLE OUTCOME

Flag

A profitable accident. The gain is not allowed to validate the process that produced it.

FLAWED REASONINGUNFAVORABLE OUTCOME

Correct

The clearest signal. The error is generalized into a lesson for future decision cycles.

EXPLANATORY PROCESS MATRIX · NOT A TALLY OF RESULTS OR A PERFORMANCE CLAIM

05 / THE SIMULATED RECORDDATED EVIDENCE · PAPER TRADING ONLY

The record, drawn from its own files.

Every chart below is generated from the Phase 0 repository’s records — journal/history.csv, journal/ledger.md, and the dated portfolio.json snapshots. All values are simulated paper trading from a $100k-per-book start. Gaps in the record are shown as gaps, never reconstructed.

SIMULATED NET WORTH PER PAPER BOOKSNAPSHOTS 2026-08-10 → 2026-09-17 · SOURCE: journal/history.csv

SWING · $99,142 AT 09-17
LONG-TERM · $100,777 AT 09-17
MEME · $100,437 AT 09-17

Simulated net worth per $100k paper book, from journal/history.csv close-run snapshots, 2026-08-10 → 2026-09-17. Rows exist only for days a close run recorded a snapshot; gaps (including 2026-08-13 → 08-21) and the end of the series reflect missed scheduled runs and are shown, not reconstructed. Not a live performance feed and not a returns claim.

VIEW AS DATA TABLE

DATE	SWING	LONG-TERM	MEME
2026-08-10	$100,000	$100,000	$100,000
2026-08-11	$99,998	$99,989	$100,000
2026-08-12	$99,997	$100,092	$100,000
2026-08-24	$100,149	$100,973	$100,000
2026-08-28	$99,875	$100,788	$99,995
2026-09-01	$100,175	$100,698	$99,995
2026-09-03	$100,120	$100,979	$100,310
2026-09-08	$99,957	$101,107	$100,310
2026-09-09	$100,077	$101,005	$100,267
2026-09-14	$99,901	$100,146	$100,267
2026-09-16	$99,151	$100,051	$100,318
2026-09-17	$99,142	$100,777	$100,437

SIMULATED FILLS BY MONTH & BOOKAS OF 2026-10-08 · SOURCE: journal/ledger.md

SWING
LONG-TERM
MEME

Simulated fills recorded in journal/ledger.md by calendar month and book — 25 fills total as of 2026-10-08 (October covers the month to date). Every row links to the dated journal entry holding the full written call.

VIEW AS DATA TABLE

MONTH	SWING	LONG-TERM	MEME
AUG 2026	1	3	2
SEP 2026	4	0	13
OCT 2026	0	0	2

LONG-TERM BOOK ALLOCATION — AT ENTRY COSTSNAPSHOT 2026-10-08 · SOURCE: portfolio.json

CASH$78,000 · 78%
TGT$8,000 · 8%
MU$7,000 · 7%
AER$7,000 · 7%

Long-term paper book allocation at entry cost, from the dated portfolio.json snapshot (2026-10-08). Positions are shown at recorded cost, not marked to market. Simulated capital only.

Phase 0 simulated portfolio page with per-book net-worth chart including a labeled snapshot gap, current holdings at cost, and as-of dates. — Supporting evidence: the Phase 0 portfolio report — per-book paper net worth with the post-09-17 snapshot gap shaded and labeled rather than filled in. VIEW FULL SIZE ↗

06 / DATA INTEGRITYHOW THE RECORD STAYS HONEST

Research you can re-check.

A research system is only as credible as its records. Phase 0 is built so that every claim about the past can be traced to what was written down at the time — including the parts that limit what the results can prove.

Structured, reproducible theses

Each call records its sources, expectation, and invalidation level, so it can be re-read and re-judged later without hindsight edits.

Decision time vs. fill time

The journal distinguishes when a decision was made from when a simulated fill was recorded, keeping the sequence honest.

Known data limitations

Phase 0 methodology accepts delayed quoted prices. Simulated results are not equated with executable market performance.

The boundary of paper trading

Paper fills ignore liquidity, slippage, and execution friction. That boundary is stated, not hidden.

07 / CONTROL LAYER

PHASE 1 — IN DEVELOPMENT

Independent software decides what is permitted.

Between Claude’s proposals and any brokerage order sits a deterministic policy layer that validates every action on its own terms. These are the target safeguards being engineered — not completed capabilities.

DEVELOPMENT STANCE

Broker connection, autonomous execution, and live trading are not represented as production capabilities. Integration targets Robinhood’s supported Trading MCP, validated in simulation and dry-run first; any limited live testing is contingent on brokerage permissions, security checks, and appropriate human oversight.

01Permitted instruments and account scope
02Position and exposure limits
03Drawdown controls and spending caps
04Explicit trade authorization where required
05Default-deny behavior on failures
06Emergency kill switch
07Order logs and reconciliation
08Secure broker credential handling

08 / STATUS & RESPONSIBILITYWHAT IS REAL TODAY

Exactly where everything stands.

Implementation status of HWA Engine capabilities by development phase
CAPABILITY	PHASE 0	PHASE 1	FUTURE
Claude Code research workflows	Implemented	Extending	—
Three paper portfolios & journals	Implemented	Continuing	—
Persistent retrospective lessons	Implemented	Continuing	—
Broker MCP / order workflows	Not live	In development	—
Independent risk control engine	Not in production	In development	—
Customer-facing interface	No	No	Optional, unscheduled

RESPONSIBLE AI & FINANCIAL SAFETY — OUR COMMITMENT, NOT A CERTIFICATION

Phase 0 is exclusively paper trading, research, and evaluation; there are no public client accounts. Phase 1 is being designed to separate Claude-generated analysis and proposals from execution permissions: independent deterministic limits, least-privilege broker credentials, audit logs, alerts, reconciliation, and an emergency stop — each described as planned, building, or verified according to its actual state. Where Anthropic’s Usage Policy requires it, personalized investment recommendations generated by Claude will receive meaningful review by a qualified human before any action — a deterministic risk engine alone is not a substitute for that review. Autonomous execution of fixed, pre-authorized human-defined rules is a distinct workflow and will be evaluated against all applicable policies and law before any real-capital use. This reflects our own engineering commitments, not a partnership with or approval by Anthropic.

09 / CONTACT

Let’s talk systems.

For research collaboration, technical partnerships, or company inquiries, contact Hwa Gallery & Co. directly.

elaina@hwagallery.com ↗