AI Bazi: Testing Consistency Through Repetition

AI八字 is a shorthand for AI-assisted Bazi calculation and interpretation. The term is written both as AI八字 and 八字ai in Taiwan and Hong Kong, where both forms have measurable search volume and point to the same practical concern: using language models to produce or explain a Four Pillars chart. The core question is not whether a model can generate a plausible reading, but whether it can produce a stable one. A low-cost way to begin testing that stability is repetition: take the same chart and the same question, then ask three times in three separate new conversations.

同一份命盘问三次

The simplest repeat test is to keep the input fixed. The same birth data should be used, the same sex, the same location convention, and the same question. Each attempt should be made in a new conversation so that the model does not carry over its previous answer as context. The question should be copied exactly each time. If the model is asked to produce the Eight Characters, the first thing to compare is whether the four pillars match across attempts. If the model is asked to interpret a given chart, the chart should be supplied in the prompt rather than left for the model to derive, so that the test isolates the reading layer.

Asking once shows what the model can produce. Asking three times shows whether the production is repeatable. A single answer may look fluent and internally coherent. Three answers placed side by side reveal whether the tool is giving one answer or three different ones dressed in similar language. This is the cheapest possible audit: it requires no external dataset, no expert panel, and no ground-truth database. It only requires the discipline to keep the prompt identical and the sessions separate.

三次答案不同说明了什么

The three attempts can fail in two distinct places. If the eight characters differ across the three runs, the problem is at the chart-construction layer. The model has not reliably converted the same birth input into the same 干支 pillars. That is a determinism failure. Bazi calculation is rule-based: a given date and time should map to a fixed set of pillars. If a language model returns different pillars for the same input, it is not performing calculation. It is generating something that resembles calculation.

If the eight characters are the same across all three attempts but the conclusions differ, the problem is at the interpretation layer. The chart is stable, but the reading is not. The model may emphasize different elements, choose different 用神, or shift its tone between attempts. A user receiving only one of those answers would have no way of knowing that the other two exist. The instability is invisible in a single session.

If both layers are stable across the three runs, the tool is at least self-consistent. That is not a guarantee of correctness. It means only that the system produces the same output for the same input under the same conditions. For a practitioner or a product builder, that is the minimum condition for treating the output as material rather than noise.

一致性与正确性的区别

Consistency and correctness are not the same property. A tool can be perfectly consistent and still wrong. If a model always assigns the same wrong 日柱 for a given date, it passes the three-run test and fails the truth test. Consistency means the answer does not vary. Correctness means the answer matches the established calculation rules or, in the case of interpretation, aligns with a defensible reading tradition.

Consistency is a necessary condition for correctness, not a sufficient one. A correct tool must be consistent, because a correct answer cannot be correct on one run and incorrect on the next for identical input. But a consistent tool can be consistently incorrect. This distinction matters because many demonstrations of AI Bazi focus on fluency. A reading may sound authoritative and still be unrepeatable. The three-run test separates the question of whether the output is stable from the question of whether it is true. Stability can be measured cheaply. Truth requires reference charts or expert review.

The reverse is also true: an inconsistent tool is necessarily wrong on at least some runs. If the same input yields different 干支 across attempts, those outputs cannot all be correct. Inconsistency is itself a conclusion. It tells the user that the system is not reliable for calculation, regardless of how plausible any single answer appears. That is a stronger claim than saying the tool may be wrong. It is wrong by construction, because at least one of the differing outputs must fail to match the deterministic result.

怎么把一致性测出来

The measurement procedure is straightforward. Define one prompt that includes the birth data and the request. Run it three times in separate sessions. Record the eight characters verbatim. Then record the main interpretive claims: the strength of the day master, the favorable and unfavorable elements, the broad life-area comments, and any specific statements about timing. Compare the three transcripts side by side.

For the calculation layer, the test passes if the four pillars are identical across all three runs. Any divergence in even one character is a failure at the calculation layer. There is no partial credit for getting three of four pillars right, because a user cannot safely use a chart that may contain a silent error in the remaining pillar.

For the interpretation layer, the comparison is less binary. Identical wording is not required. A language model samples from a probability distribution on every generation, so different phrasing is normal. The question is whether the substantive claims agree. If one run says the chart is weak water, another says strong metal, and the third says balanced earth, the reading is unstable even if each sentence is grammatical. If all three runs agree on the day-master assessment and the broad direction of favorable elements, while differing in wording and minor emphasis, the interpretation layer can be considered stable enough for further checking.

Deterministic calculation consistency comes from the algorithm. The same input processed by the same rules produces the same output. That is a property of ordinary software. A language model does not work this way. Its generation mechanism includes sampling, so the same prompt can produce different tokens on different runs. Different wording is expected and acceptable. Different 干支 is not, because the 干支 are not a matter of phrasing. They are the output of a finite, rule-governed mapping. If the model cannot reproduce them, it is not doing the mapping.

把不稳定的部分标出来

Once the three runs are compared, the unstable parts should be marked. This is the smallest amount of work needed to turn raw model output into usable reference material. A paragraph that says the same thing three times can be kept with a note that it passed the repeat test. A sentence that changes across runs should be tagged as unstable. If the chart pillars differ, the entire output is unstable at the calculation layer and should not be used as a chart. If only certain interpretive claims differ, those claims can be marked as low-confidence.

Marking instability does not require resolving it. The point is to prevent a reader from treating a single sampled answer as a fixed result. A user who sees that the model gave three different 用神 across three runs knows that the 用神 line is not reliable from that tool. A user who sees that the pillars were identical knows that the model at least reproduced the chart consistently, even if the chart itself has not been verified against a reference.

For a knowledge base or a product, unstable material should either be removed or displayed with a warning. Stable material can be retained, but with the understanding that stability is only the first filter. It says nothing about whether the reading matches any school of Bazi practice. The practical workflow is: run the three-way test, mark the unstable parts, discard calculation-level failures, and only then spend effort checking the stable remainder against external references.

The AI八字 space contains both deterministic calculators and language-model interpreters. The two should not be confused. A deterministic calculator is testable by comparing its output with a known chart. A language model is testable first by repetition, because its output is sampled. The three-run test is the cheapest way to apply that test. It does not establish that the tool is good. It establishes whether the tool is even a candidate for being good.