Why You Can't Leave It All to AI — The Case for Hybrid Architecture

 

Same Question, Different Answers

"Give me the Four Pillars reading for a man born on May 15, 1990, at 6:00 AM."

Early in the project, I asked AI this question three times. All three responses sounded convincing. The problem was that across those three answers, the Day Pillar came back in two different variations — once as Gap-Jin and twice as Eul-Sa. When the Day Pillar, the very foundation of a Four Pillars (Saju) reading, differs between attempts, every interpretation built on top of it collapses. That single experiment settled the matter: you cannot let AI directly calculate the Four Pillars.

The Danger of the AI Black Box

Why does this happen? Setting up the four pillars is far more than a simple date conversion.

The Year Pillar doesn't change on January 1st. It changes at the Start of Spring (Ipchun), one of the 24 solar terms. In 2024, the Start of Spring fell on February 4th at 5:27 PM. Someone born on February 3, 2024, is in the year 2024 by the Western calendar, but in Saju terms they still belong to the previous year. This solar term boundary must be determined to the minute.

The Month Pillar follows the same logic. Months aren't determined by the Gregorian calendar but by solar terms. From Minor Cold to the day before the Start of Spring is the Ox Month; from the Start of Spring to the day before the Awakening of Insects is the Tiger Month. Twelve of the 24 solar terms — specifically the "junctions" — define the month boundaries, and they too must be calculated to the hour and minute.

The Hour Pillar introduces yet another trap: True Solar Time correction. Korea Standard Time is based on the 135th meridian east, but Seoul's actual longitude is roughly 127 degrees east. Without correcting for this difference, the Hour Pillar can shift. Especially for people born right at the boundary between two-hour periods — say, between the Rat Hour and the Ox Hour — the correction can change the Hour Pillar entirely.

When you ask AI for a Saju reading, there is no way to verify whether the model internally performed the solar term boundary checks and True Solar Time corrections accurately. This is the core problem of the black box: you can't judge whether the result is correct just by looking at the output.

The Opacity of Interpretation

There was something even more concerning than calculation accuracy: the reasoning behind the interpretation was invisible.

Suppose the AI says, "You have a strong Indirect Wealth (Pyeonjae), so you have an aptitude for business." Anyone who knows Saju would immediately ask: Which pillar is the Indirect Wealth in? How many are there? Is it the Favorable Element (Yongsin) or an unfavorable one relative to the Day Master? Does it include Hidden Stems (Jijanggan)?

When you let AI handle the entire reading, there's no way to expose this intermediate reasoning. If the user can't verify why a conclusion was reached, it's no longer Saju analysis — it's fortune-telling. The value of Four Pillars study lies in its systematic analytical process, and that entire process gets hidden.

Professional Saju apps always display the Chart Table (Myeongsikpyo). Four pillars with their Heavenly Stems (Cheongan) and Earthly Branches (Jiji), the Ten Gods (Sipsin) at each position, Hidden Stems, and Five Elements ratios. This data itself provides value to the user. Relying solely on AI makes this "visualization of the analytical process" fundamentally impossible.

The Hybrid Solution

Hybrid architecture solves both problems simultaneously.

The rule engine's domain is clear: Gregorian-to-lunar conversion, solar-term-based Month Pillar determination, True Solar Time correction, Day Pillar calculation, and Hour Pillar calculation. Post-pillar analysis is also the rule engine's job: determining the Five Elements of each Stem and Branch, calculating Ten Gods relationships relative to the Day Master, extracting Hidden Stems, detecting combinations and clashes, arranging the 12 Life Stages, computing Five Elements ratios, and determining the Favorable Element.

All of these follow explicit, deterministic rules. The element of Gap is always Wood. If the Day Master is Gap and the counterpart is Mu, the Ten God is always Indirect Wealth. There's no room for AI judgment here, nor should there be.

AI's domain is equally clear: taking the structured analysis output from the rule engine and synthesizing it into a natural-language interpretation that humans can read. A request like, "The Five Elements ratio is Wood 35%, Fire 10%, Earth 20%, Metal 25%, Water 10%, and the Favorable Element is Fire. Explain the personality traits of this chart."

This separation produces three effects.

First, accuracy is guaranteed. The rule engine's calculations are deterministic — the same input always produces the same output. Results can be cross-verified against traditional almanac sites, and unit tests can automate validation.

Second, transparency is ensured. Users can see the Chart Table, Five Elements chart, and Ten Gods relationships visually. They can verify what data the AI interpretation is based on.

Third, the interpretation remains natural and engaging. If you built interpretations purely from rules, you'd get dry listings like "Indirect Wealth: 2, Direct Wealth: 1, financial fortune: average." AI transforms this data into a contextual narrative.

Drawing the Boundary

The heart of hybrid architecture is ultimately drawing the line between "what to delegate to AI" and "what to keep away from AI."

The criterion is simple: "Is there one correct answer, or many?" The element of Gap being Wood has one correct answer — the rule engine handles it. "What are the personality strengths of this chart?" has many valid answers — AI handles it.

More specifically: calculation goes to code, judgment goes to AI. Classification goes to code, narrative goes to AI. This principle extends well beyond the Saju project to AI collaboration in general.

The trickiest boundary in practice was determining the Favorable Element (Yongsin). The Favorable Element identifies which of the Five Elements the chart most needs, but different schools of thought use different methods. The억부법 (Suppression-Support method) judges based on the Day Master's strength, while the 조후법 (Seasonal Regulation method) judges based on season and temperature. Should this go in the rule engine or be delegated to AI?

The answer was the rule engine. Favorable Element determination follows rules — "if this condition, then this element" — and those rules can be expressed in code. However, to accommodate the different schools, we adopted a flexible structure: "default is the Suppression-Support method; the Seasonal Regulation method can be selected in settings." AI is only asked to interpret what the determined Favorable Element means for this particular chart.

What This Experiment Taught Us

That simple early experiment — asking AI to directly calculate a Saju reading — determined the entire architecture of the project. If we had skipped that experiment and started with the assumption that "AI will figure it out," the project would have crumbled midway.

AI can do an astonishing number of things. But "what AI can do" and "what you should delegate to AI" are different questions. If you ask whether AI can calculate Four Pillars, the answer is usually yes. But if you ask whether you can trust the result, verify it, and show the process to users, the answer changes.

Hybrid architecture leverages the strengths of both AI and code. This isn't a compromise — it's a strategy that maximizes both.

Coming Up Next

One side of the hybrid — the rule engine — computes according to fixed rules. How do you elevate the quality of the other side — the AI interpretation? In the next installment, we'll explore the core of prompt engineering and the dramatic impact of Gungtong Bogam reference data on interpretation quality.