Probing AI Bazi: Breaking Down and Verifying Specific Claims

追问AI八字 is a method for interrogating an AI-generated bazi reading, not for forcing the model to change its answer. The purpose is to take a block of interpretive text and break it back down into verifiable components: which stems and branches are being used, what rule is being invoked, whether a contrary reading exists, and what information is missing. In this entry, the term itself and the headings are given in Chinese, while the explanation is in English.

四句追问各自逼出什么

The four questions each target a different layer of the interpretation. The first question targets the location of the evidence: which characters in the natal chart are actually being referenced. The second targets the rule itself: what is the complete formulation of the principle being applied. The third targets the counter-case: whether the same chart supports an opposite reading, and under what conditions. The fourth targets the gap: what specific materials are still missing before the conclusion could be firmer.

The aim of the four questions is not to make the model retract a statement. It is to turn a piece of fluent prose into something that can be checked. A reading that cannot survive these questions has not been disproven; it has simply failed to show that it rests on identifiable components.

第一句：这条结论用到命盘里哪几个字

The first question asks for the evidentiary position: which characters in the chart does this conclusion actually use. If the model cannot name specific 干支, then the conclusion has not landed on the chart. It may be a general statement about a ten-god type, a seasonal pattern, or a textbook description that floats above the actual pillars.

A usable answer names the stems and branches involved. For example, it might say that the conclusion rests on the 日主 being 甲, with a 申 branch in the month pillar, or that it uses the presence of 庚 in the hour stem in relation to the day master. The exact characters matter because they are what make the statement checkable. If the model answers with a paraphrase such as “the metal element is strong,” without saying whether this refers to 庚，辛，申，酉, or a combination, the first question has not been answered.

The first question also exposes when a reading is importing assumptions from outside the chart. If the conclusion concerns a clash or combination, the answer must show which two characters are clashing or combining. If it concerns a void branch, the answer must name the branch and the pillar in which it appears. Without the specific 干支, there is no way to verify whether the stated interaction exists in the chart at all.

第二句：这条规则的完整表述是什么

The second question asks for the rule in full. It is not enough for the model to restate the conclusion in different words. The rule must be given as a formulation that can be checked against classical or modern sources. A rule can be verified; a paraphrase cannot.

For instance, if the reading says that a certain branch combination transforms into fire, the second question asks what the complete conditions for that transformation are. Does it require the month branch to be in the relevant season? Does it require the absence of a certain stem? Does it require the transformed element to appear as a heavenly stem in the chart? The model must state the conditions, not merely the outcome.

The wording of the rule matters because different traditions state the same principle with different boundaries. A rule taken from one school may not apply in another. The second question therefore asks not only what the rule says, but whether it is being quoted in full or only in the part that supports the current reading. If the model gives a truncated version, the missing clause may be precisely the one that would limit the conclusion.

第三句：有没有相反的读法

The third question asks for the opposing case. Given the same chart, is there a reading that points the other way, and what would that reading require? This is not a request for the model to invent a contradiction. It is a request to test whether the original conclusion is the only one the chart can support.

A chart may contain a combination that appears to strengthen one element while a clash in another pillar weakens it. The same 干支 may be read as auspicious in one system and inauspicious in another. The third question asks the model to state the conditions under which the opposite conclusion would hold. If the model cannot formulate such conditions, then either the original conclusion is unusually robust or the model has not considered the chart thoroughly.

The value of this question lies in forcing the interpretation to state its own boundary. A conclusion that acknowledges a contrary reading and explains why the contrary reading does not apply is stronger than one that simply ignores it. A conclusion that cannot acknowledge any contrary possibility may be operating from a single template rather than from the chart itself.

第四句：缺哪些资料才能更确定

The fourth question asks about the gap. What materials are still missing before the conclusion could be firmer? An honest answer names specific missing items. It does not continue to elaborate the original reading.

The missing items may include the exact time of birth for the hour pillar, the geographic location for solar time correction, the sex of the person for the direction of the 大运, or the specific lineage of the method being applied. They may include a full birth date when only the year and month are known, or clarity about whether the chart is being read for a past event or a future decision.

A model that answers this question with more interpretation has not answered it. The question asks for what is absent, not for what can be said despite the absence. The ability to name the missing item is itself a sign that the model distinguishes between what the chart shows and what it is inferring.

四句问完之后的处理

After the four questions have been asked, the user separates what can be tied to specific 干支 and specific rules from what cannot. The part that survives this process is kept as the core of the reading. The part that cannot be tied to identifiable characters or verifiable rules is downgraded to reference material. It may still be suggestive, but it is not treated as a finding.

If the model changes its conclusion under questioning, that is not necessarily a failure of the method. It is information. A change indicates that the original conclusion did not rest on stable grounds. The original statement may have been generated from a general pattern rather than from the chart, and the questioning exposed that. The change itself is useful because it tells the user how much weight to give the initial output.

The calculation engine of this site does not change under questioning. The same input produces the same result every time. The 干支 output, the pillars, the ten gods, and the computed cycles are deterministic. What is being questioned is not the calculation but the interpretive layer built on top of it. The four questions do not alter the chart; they test whether the prose about the chart can be reduced to the chart itself.