Why We Don't Let AI Write the Reading
Same floor plan, same birth date — the answer must be identical today and tomorrow.
1. Where the problem starts
Hand a floor plan to a general-purpose language model and ask how the feng shui reads, and it will produce a fluent, confident answer. Ask twice about the same plan and the answer may differ; rephrase the question and the conclusion may reverse. It is generating text, not computing a result — its objective is plausibility, not reproducibility. For a product measured on accuracy, that is fatal.
2. Where we draw the line
The model does exactly one job: read the image. It identifies the compass bearing, the direction the entrance faces, which objects are present, and where they sit in frame. These are perception tasks — the model is good at them and is allowed to be uncertain. We require structured coordinates, never conclusions.
3. Why judgement must be rules
Once coordinates are in hand, everything else is code. Bearings map deterministically to the eight palaces. Each auspicious or adverse judgement is evaluated against documented rules. Remedies are looked up from an object-to-direction table. The same input always yields the same output — which is the precondition for talking about accuracy at all, and what lets a user compare two runs and check them.
4. Marking uncertainty honestly
If a photo contains no compass and the user supplies no facing direction, the code does not invent a bearing. It states plainly that there is no directional basis this time, and says how far the conclusion can be trusted. We would rather give a conservative answer with a basis than a polished answer that was made up.
Frequently asked
So what do you use AI for?
Image recognition only: reading the compass bearing and locating objects in a photo. Results are handed to the rule engine as structured coordinates. The model takes no part in any judgement.
What if the recognition is wrong?
Every reading passes through a confirmation step that lists the bearing and detected objects, and lets you correct them before the report is generated. Corrections are logged and used to improve recognition.
Where do the rules come from?
From published classical sources and established conventions: Ba Zhai palace mapping, Jin Suo Yu Guan sand-and-water rules, annual Flying Stars, and Four Pillars stem-branch calculation. The rule set is versioned and auditable.
Want to test this? Analyse the same floor plan twice and compare the scores and remedies — they will match exactly.