The kiln that built VEGA

How a solo trader's struggle with discipline became a governance platform for regulated industries.

VEGA did not start as a platform idea. It started as a trading bot that kept breaking itself.

In 2023, I began building an automated cryptocurrency trading system. One-hour candles, TradingView indicators, a VPS on Oracle Cloud running 24/7. The technical part was straightforward. The hard part was everything around it.

The bot had a 48% win rate. That means it loses more often than it wins. The math still works because winners are larger than losers. But here is what nobody tells you about a 48% win rate: you will regularly have 5 or 6 consecutive losses. The probability is about 2%. It will happen every few months, like clockwork.

And every time it happens, the same voice appears: change the parameters.

Chapter 1

The optimization trap

My parameters had survived Walk-Forward Analysis across 4 years of data, 25 coins, 35,000+ price bars. They were validated. But after 4 losing trades in a row, none of that felt real. The urge to "fix" things was overwhelming.

I gave in. Multiple times. Every time I changed parameters after a short losing streak, I destroyed the statistical validation that made the strategy viable in the first place. I was not improving the system. I was fitting it to noise.

The worst version of this happened in early 2026. I added four filter modules over two months. Each one looked good in isolation. When I ran them together, three out of six pairs cancelled each other out. The baseline without any of them performed better. Months of work for negative value.

Chapter 2

Three AIs that cannot see each other

The solution came from an unlikely place: adversarial review.

I started sending every proposed change to three different AI models, each with a fixed role and no visibility into what the others said. The protocol was simple: a skeptic looks for statistical flaws, an optimist looks for engineering problems, and an arbiter synthesizes and casts the deciding vote.

I called it KGR: Knowledge Guard Review. The first rule was non-negotiable: verified beats claimed. If any reviewer asserts something about the code, the arbiter checks the production server with grep before accepting it. Logical reasoning is a hypothesis. Code on disk is a verdict.

In its first month, KGR caught a resource allocation conflict between concurrent decision modules. Two independent subsystems had each built their own sizing logic. Neither knew the other existed. Both could claim the same resources simultaneously. A path to catastrophic failure that no single reviewer would have found.

No single reviewer would have caught it. The skeptic flagged the statistical anomaly in position sizes. The optimist traced it to the architecture. The arbiter confirmed both codepaths existed on the production server. Three perspectives, one truth.

Chapter 3

Six rules written in losses

Over 420+ versions of the bot, a set of immutable rules crystallized. Not from theory, but from specific, documented failures. Each rule has an evidence file with the exact trade, the exact mistake, and the exact loss.

These rules are not guidelines. They are circuit breakers. When I proposed changing parameters after four losses in May 2026, the system blocked me with my own words from three weeks earlier. The bot had logged the date, the evidence, and the exact emotional state I was in when I wrote the rule. It was arguing with a version of myself that was thinking more clearly.

Chapter 4

The hospital test

In June 2026, a colleague at a public-sector institution asked for help. They had approximately 80 internal regulatory documents: governance protocols, standard operating procedures, compliance records. Some of them might reference outdated legislation. Checking manually would take weeks.

I ran the same KGR protocol on the documents. Three AI reviewers, each examining independently, adversarial verification of every legal reference.

The result: over 100 compliance findings across multiple departments. One finding was critical: a governance body had expanded its formal remit beyond its original legislative mandate. The relevant authority belonged to a national-level commission, not the institution itself. It was not a typo. It was a jurisdictional anomaly with real legal consequences, buried deep in the document matrix.

That moment clarified something. The KGR protocol was not a trading tool. It was a document governance tool that happened to be born in trading.

Chapter 5

The kiln metaphor

There is a moment in ceramics when the sculpture moves from clay to kiln. In clay, you can reshape endlessly. In the kiln, you cannot open the door. If you do, the piece cracks from thermal shock.

I recognized this pattern in my own work. The transition from backtesting (clay) to live trading (kiln) required a different psychology. The artist in me wanted to keep reshaping. The engineer needed to close the door and wait.

In June 2026, the bot went into a data-collection freeze. No parameter changes. No new modules. Just collect data and wait for September. During that freeze, the methodology kept working on other problems. The same adversarial review that stopped me from destroying my trading validation was now catching legal anomalies in hospital documents.

Chapter 6

From bot to platform

VEGA is the generalization of everything that was built to keep one trading bot honest. Every architectural decision traces back to a real failure:

The module protection system (four tiers from immutable to community-contributed) mirrors the bot architecture where some modules are locked and others are experimental. The document pipeline (upload, OCR, verification, AI review, approved) is the same pipeline that processes trading signals, adapted for regulatory documents. The legislative similarity groups that organize EU countries by legal tradition came from the compliance audit that revealed how differently Slovenia and Germany regulate the same medical procedure.

Chapter 7

The meta-proof

In July 2026, the VEGA Platform Specification went through its own KGR review. Round 80. Three independent reviewers examined the specification that defines how reviews should work.

They found 8 issues. A missing liability disclaimer. Insufficient GDPR data-retention language. A group name with political connotations. An apparent paradox in how "immutable" modules could receive security patches.

All 8 were fixed. All 3 reviewers ratified unanimously. The specification was improved by the process it describes.

VEGA was not designed on a whiteboard. It was extracted from 420 versions of a trading bot, 80 adversarial review rounds, a compliance audit of approximately 80 institutional documents, 6 rules written in real losses, and one sculptor's metaphor about not opening the kiln.

The bot is still in the kiln. The methodology is already outside it.