WHICH COMMODITY
MOVED THE MOST?

Cotton. Wheat. Corn. Three consecutive entries. Three commodities chosen after they had already moved 10–13% in a single week.

The thesis was right each time. Supply tight, demand holding, price action confirmed. But correct thesis doesn't guarantee correct entry. Every position was a chase — the opportunity was real, the timing was late.

The one trade that actually worked was the opposite: a short on natural gas, entered as the market was weakening. Contrarian, not momentum.

The structural problem. The scoring system rewards recent price moves. A commodity up 12% on the week scores high — that reads as “strong momentum, high conviction.” But from an entry standpoint, a 12% move is evidence the opportunity already happened. The people who made money did it two weeks ago.

Underneath this, the Bayesian learning engine was supposed to correct for repeated mistakes. It tracks outcomes per commodity and adjusts confidence weights over time. The problem: when it was built, it only got wired into the puzzle prediction — the game AI’s commodity picks. The trade entry scorer never connected to it. Every trade decision was made as if it were the first trade ever.

What changed. A momentum penalty now applies at the scoring stage. Commodities that moved more than 10% in the prior week get docked before conviction is calculated. The Bayesian adjustments now feed into trade decisions, not just puzzle picks. And trade outcomes — wins and losses — now feed back into a separate learning state that accumulates over time.

The wheat and corn positions are still open. Whether they close as wins or losses doesn’t change the analysis. The entry pattern was wrong on its own terms regardless of outcome. The fix is in.

This entry exists because the pattern was worth naming before the outcome was known. That’s the rule: write it down when you see it, not after you find out if you were right.

Four trades. Four failures the system didn’t catch on its own.

Natural gas: the stop logic was direction-blind. A SHORT position’s stop was checked the wrong way in the code. The system executed exactly what it was programmed to do, logged a WIN, and moved on. It had no way to know it was wrong — it can’t read its own code.

Rough rice: entered with no executable micro contract. The contract catalog was supposed to prevent that entry. It didn’t. Catalog validation failure.

Wheat: approaching its stop for two days. An alerts table exists. Alerts were being written to it. Nobody was reading it.

Corn: entered after a 12.9% weekly run. High conviction. Same pattern as cotton, same pattern as wheat. The scoring system had no mechanism to penalize momentum chasing. Nothing noticed the repeat.

The honest answer to the question was: mostly no. Not yet. Each failure was a different layer of blindness — a code bug, a validation gap, an unread table, a missing rule. The common thread: the system executes decisions but doesn’t reflect on them. Each daily run starts fresh. There’s no persistent thread that looks across trades and asks “wait, I keep doing this.”

What changed today. A portfolio health audit now runs before every new entry decision. Five checks: stop proximity, win rate, momentum pattern, contract validity, undelivered alerts. If anything is CRITICAL, new entries are blocked. The check that should have been running from the start.

The Bayesian learning engine — which was supposed to adjust trade scoring based on outcomes — was only connected to the puzzle game, not to the trade entry scorer. That bridge is now built. Trade outcomes feed back. The momentum penalty is now hardcoded into scoring. The two-path state system that was causing trades to disappear from the journal is fixed.

Right now, wheat is 1.65% above its stop. The audit fires CRITICAL on that tomorrow morning. No new positions open until it resolves.

The one thing the system still can’t do: notice its own code is wrong. That class of failure requires a human. Everything else is now a check.

Entry 001 identified a pattern. Entry 002 identified the failures underneath it. Both were correct diagnoses. Neither of them named the deeper problem.

The momentum penalty — the fix from Entry 001 — docked high-conviction scores on commodities that had moved more than 10% in a week. It was a real improvement. It wasn't enough. The reason: adjusting a score that was built on price movement doesn't fix a process that starts with price movement. It penalizes the most egregious version of the same mistake. The question being asked was still wrong.

The question the old system answered: What has moved the most this week?

The question a working research process answers: What has a real catalyst that the market hasn't priced yet?

Those are not variations of the same question. They are different questions that produce different starting points, different data sources, and a different track record.

How rice was found. The one profitable trade on record — rough rice, June 2026, +$278 — was found by a different process entirely. Not a price scan. A news search. India's monsoon was running 40% below normal. El Niño had just been declared. Export ban risk was rising. The rough rice price at that moment: down 5.6% on the month. A real catalyst, a quiet market. That was the entry.

The old system, presented with the same data at the same moment, would have scored rough rice near zero. It was going the wrong direction. Nothing about it looked like momentum.

The structural inversion. The old system was built to find things that had already moved, then justify those moves with news. The research process would search for context to explain a price spike. The price spike was the starting point. Everything else was rationalization.

A working process runs in the opposite direction. The catalyst is the starting point. The price is the check. If the catalyst is real and the price hasn't moved — that's the setup. If the price has already moved, the opportunity has already happened for someone else.

What changed. The research system now runs two discovery channels in parallel. The first monitors market anomalies — unusual price action, curve shifts, positioning extremes. The second sweeps external sources daily: USDA, EIA, NOAA, major commodity newswires. When a catalyst is found and the market hasn't reacted, that's the highest-priority case in the system. When the market moves with no catalyst in sight, the first job is to find out why.

One other structural change: the research analysis is now explicitly prohibited from using price action as evidence that a catalyst matters. The instruction is built directly into the analysis: "If your evidence that this catalyst matters is that the price moved, you have no catalyst." This closes the loop that produced the wheat and corn entries — where the price move became evidence the thesis was right, and the thesis became evidence the trade was good.

Whether this produces better trades remains to be seen. The process is right. The outcomes come next.

Corn moved 24% in four days. The system watched every tick. The decision each day: WATCH. WATCH. WATCH. Never enter.

I called this discipline. It wasn't. There's a difference between a disciplined trader and a fearful one. A disciplined trader has rules and enters when they're met. A fearful trader finds reasons not to enter. We built the second one.

The momentum penalty — added after two losing trades — was working exactly as designed. A commodity up more than 10% in a week got docked 2–3 points before conviction was calculated. The design was correct for what it was: a mechanism to avoid chasing. What it actually did was make the system better at saying no, without making it any better at finding real setups. Those are different problems. We solved the wrong one.

The deeper failure. The momentum penalty was a symptom. The root cause was underneath it. The research system — every thesis it has ever written — was reverse-engineered from a price number. It receives a price, identifies that the price moved, then searches for a story that explains the movement. That is not research. That is rationalization dressed as analysis.

A working process runs in the opposite direction. Find the catalyst. Check if price has absorbed it. If it hasn't — that's the setup. The rice trade, the one that actually worked, was found that way: not by looking at price, but by looking at monsoon data and asking whether the market had noticed yet.

That process was never encoded into the system. We built everything else.

What the architecture review found. A full audit of the codebase — run by a separate model with no prior context — confirmed the diagnosis and found two additional problems. First: the trading system was coupled to the puzzle game. Candidates were only evaluated on days those commodities rotated into the puzzle. Corn might be evaluated once every four or five days regardless of what was happening in the market. Second: a half-built research system was sitting dormant in the codebase. It had an event calendar, a news scanner, and a thesis agent. None of it was connected to anything. It had been built and abandoned.

What changed today. The old price-first scanner was disabled. A new research agent now runs every morning at 6:40 AM, before anything else. It searches the web for each of the four tradeable commodities, checks whether any known events are in the next 48 hours, and logs whether each piece of news has already been absorbed by price. It doesn't trade. It watches and records. That's Phase 2 of a four-phase rebuild.

Tomorrow is the first real test. WASDE — the USDA's monthly supply and demand report — drops at noon. The morning brief already has pre-event context: an independent yield tour puts corn production 4 bushels per acre below the analyst consensus. The market is quiet heading into the report. Whether the brief flags that gap correctly, and whether it would have generated a valid pre-event thesis, is what we'll grade tomorrow afternoon.

The process is changing. The outcomes come next.