Paste AI Answers Back: Verify Against Charts

Backreading is a verification method for AI-generated bazi interpretations. It means taking the text produced by a model, breaking it back into checkable elements, and comparing those elements one by one against the natal chart. The purpose is not to assess literary quality, but to catch fabricated stems, branches, Ten Gods, or time sequences before they enter your notes.

回读的三个步骤

The method has three passes. Each pass looks for one type of checkable item: ganzhi, Ten Gods, and time points. The order matters because each layer depends on the previous one. If the stems and branches are wrong, the Ten Gods derived from them cannot be trusted. If the Ten Gods are wrong, any discussion of favorable and unfavorable elements is unreliable. Time points come last because luck pillars and annual pillars are the easiest to verify against the engine output.

A backread is not a proofread. In proofreading, you look for grammar, punctuation, and awkward wording. In backreading, you ignore style and look only for elements that either match the chart or do not. A sentence may be beautifully written and still be entirely invented. The backread catches the invention.

步骤一：抽出所有干支

The first pass extracts every occurrence of the ten heavenly stems and twelve earthly branches from the text. This includes stems and branches used inside larger terms, such as a pillar written as 甲子 or a year written as 丙午. Each one is pulled out and checked against the natal chart.

The question for each stem or branch is narrow: does it appear on the chart in the same position the text claims? For example, if the text says the day master is 甲, the engine output must show 甲 in the day stem position. If the text mentions 子 as the month branch, the engine must show 子 in the month branch position. If the text describes a combination involving 寅 and 午, both must exist somewhere on the chart or in the referenced luck pillar.

This pass also catches a common failure mode: the model writes about a chart as if it contains a stem or branch that the user never supplied. Sometimes the invented element is plausible, sometimes it is random. Either way, it fails the check. The correct action is not to fix the one character and move on. The whole segment containing it is set aside.

步骤二：抽出所有十神

The second pass extracts every Ten God name: 比肩，劫财，食神，伤官，偏财，正财，七杀，正官，偏印, and 正印. Each one is checked against the same position on the chart.

A Ten God is not a free-floating label. It is always relative to the day master and always attached to a specific stem or branch. When the text says the month stem is 正官, the engine output must show 正官 in the month stem position. When the text says the year branch contains 偏财, the engine must show 偏财 in the year branch position. Hidden stems inside branches are checked the same way: if the text says the day branch hidden stem is 食神, the engine output must agree.

The most useful check is positional identity. It is not enough that 正印 appears somewhere on the chart. The text must place it exactly where the engine places it. A model can list the correct Ten Gods for a chart but assign them to the wrong pillars, which changes the meaning of the reading. Positional errors are treated the same as invented Ten Gods.

步骤三：抽出所有时间点

The third pass extracts every time point: the start and end years of luck pillars, and the stem-branch pairs of annual pillars. Each one is checked against the sequence shown by the engine.

Luck pillar years are a common source of silent error. A text may describe a 正财 luck pillar as beginning in one year when the engine shows another. The difference can be small, sometimes only a year or two, but it shifts every subsequent pillar. The backread does not ask whether the text’s version is plausible. It asks only whether it matches the engine output.

Annual pillars are checked the same way. If the text says a certain year is 壬寅, the engine sequence must show 壬寅 for that year. If the text says an event falls in 乙巳, the engine must show 乙巳 for the year in question. Where the text gives a range, each endpoint is checked individually.

出现命盘上没有的字怎么办

When the text contains a stem, branch, Ten God, or time point that is not on the chart, the model has supplied content on its own. The correct response is to reject the entire segment, not to correct the single character.

The reason is structural. A fabricated element rarely appears in isolation. It usually enters the text as part of a chain of reasoning. The model may invent a branch to complete a combination, then use that combination to justify a Ten God relationship, then use that relationship to explain a life event. If you only fix the invented branch, the rest of the chain remains in place, now supported by nothing. The sentences that depended on the fabricated element become orphans, and the reader cannot tell which parts of the paragraph still follow from the chart.

Rejecting the whole segment is faster than trying to repair it. You are not obligated to salvage the paragraph. You can ask the model to rewrite the segment without the unsupported element, or you can discard it and write your own summary from the chart. The backread only identifies the failure; it does not attempt to rehabilitate the text.

这套回读大约需要多久

For a medium-length interpretation, a backread takes about five to ten minutes. This is less time than a full careful reread, because the backread is not trying to understand everything the text says. It is scanning for a fixed list of checkable items.

The first pass is the longest, because stems and branches appear frequently and hide inside larger terms. The Ten God pass is quicker once you know the chart. The time point pass is usually the fastest, because luck pillar and annual pillar sequences are compact and easy to scan.

The engine output used by this site is deterministic. The four pillars, Ten Gods, and luck pillar sequence are computed from fixed rules, so they can serve directly as the comparison table for a backread. There is no need to recalculate anything by hand. Keep the engine output open beside the text and move down the list.