Claude Bazi: How a General Chat Model Outputs Changsheng States

Claude算八字 refers to the use of the Claude general-purpose conversational model to produce a Chinese-metaphysics natal chart reading. Because Claude generates text token by token, the twelve-stage longevity states on a chart are likewise written out as generated language rather than retrieved from a fixed lookup table. This distinction matters when the output is checked against the actual sexagenary pillars.

十二长生是一张查得到的对照表

The twelve longevity stages are 长生，沐浴，冠带，临官，帝旺，衰，病，死，墓，绝，胎, and 养. They are arranged according to the sequence of a given heavenly stem across the twelve earthly branches. The table is fixed: for each stem, the stage at each branch follows a set cyclic order. A reader can therefore take any stem and any branch, locate the intersection on the table, and read off the named stage. The table does not vary by school, and the names themselves are not generated by the model. What the model writes may or may not correspond to that table.

Because the sequence is cyclic, some stages recur at predictable intervals. The stage at one branch determines the stage at every other branch for the same stem, provided the direction of movement is observed. This is why a single mislabeled stage in a generated reading is often easy to spot: it breaks the cycle that the table establishes.

日主在各支的状态与各柱自坐的分别

A natal chart contains two distinct sets of longevity states. The first set is the state of the day master in each of the four earthly branches of the chart. The day master is the heavenly stem of the day pillar. Its state is read against the branch of the year pillar, the branch of the month pillar, the branch of the day pillar, and the branch of the hour pillar.

The second set is the state of each pillar’s own heavenly stem in its own earthly branch. This is often called 自坐, or self-seated state. The year pillar has one self-seated state, the month pillar another, the day pillar another, and the hour pillar another. The day pillar’s self-seated state is not the same thing as the day master’s state in the day branch, even though both involve the day stem and the day branch. They are separate items because the first set asks about the day master across all four branches, while the second set asks about each stem only within its own pillar.

A chart therefore carries eight named states in total: four for the day master across the branches, and four for the self-seated positions. The terms are drawn from the same list of twelve, but the two groups answer different questions.

让工具把两组状态分开报

The site engine outputs both items for each pillar at the same time. A properly separated report gives, for each pillar, the day master’s state in that pillar’s branch and the pillar stem’s self-seated state in its own branch. In practice, this means the output should not present one undifferentiated list of four states and imply that it covers the whole chart.

When the two groups are reported separately, the user can verify each group against the twelve-stage table independently. The day-master group is checked by holding the day stem fixed and reading across the four branch positions. The self-seated group is checked by taking each pillar’s own stem and reading it against that same pillar’s branch. The two checks use the same table but different stem-branch pairs.

A clear format helps. One workable layout is to list the four pillars first, then list the day master’s stage at each branch, then list each pillar’s self-seated stage. The order of the four pillars should remain year, month, day, hour, so that the reader can compare the two lists position by position without reordering them.

把两组混成一组是什么样的错

The common confusion is to report the day master’s state and the pillar stem’s self-seated state as though they were one group. For example, an output may give a single set of four stage names and present them as the longevity states of the chart. On the surface the names look plausible, because they are all drawn from the twelve-stage vocabulary and appear in the expected positions.

The error becomes visible when the items are checked one by one against the table. If the day master is 甲 and the year branch is 亥, the table gives 长生 for the day master at that branch. But if the year pillar’s own stem is 辛, the self-seated state of 辛 at 亥 is not 长生; it is 沐浴. A generated line that assigns only one stage to that position cannot be correct for both readings at once. When the output does not say which of the two it is reporting, the reader may take a self-seated state as a day-master state, or the reverse.

Another form of the same error is to report the day master’s state at the day branch and then reuse it as the day pillar’s self-seated state. This happens when the model treats the day pillar as a special case and merges the two readings. The two values may coincide for some stems and branches, but they are not the same item and do not always agree.

标注层出错时要作废哪几句

This layer of the reading belongs to what can be called the annotation layer, along with hidden stems and the ten gods. The four pillars themselves are the structural layer. If the four pillars are correct, that does not by itself show that the annotations are correct. The two layers need to be judged separately. A chart can have the right year, month, day, and hour pillars while carrying wrong longevity states, wrong hidden stems, or wrong ten-god labels.

The verification method is to fix the day master first, then check each pillar in turn against the twelve-stage table. That produces four day-master states. Then take each pillar’s own stem and check it against that pillar’s branch. That produces four self-seated states. The full check covers eight items across the four pillars. In most cases this can be completed within ten minutes, because the table is short and the pairs are few.

When an error is found in the annotation layer, there is no need to redo the four pillars. The structural layer remains as it was. However, any interpretive sentences that relied on the mistaken annotation should be set aside together with it. If a statement about the day master’s condition in a given branch was based on a wrong stage name, that statement loses its support. If a statement about a pillar’s character was based on a wrong self-seated state, that statement also loses its support. The rest of the reading may stand if it does not depend on the wrong item.

The practical rule is narrow. Discard the wrong annotation and the conclusions drawn from that annotation; do not discard the chart itself. This keeps the correction limited to the layer where the mistake occurred. It also prevents a single generated error from cascading into a needless reworking of the entire reading.