Methodology and accuracy research

How are readings generated, and are they accurate?

The charts on this site are calculated by software using calendar and traditional chart rules. This page documents the methods and corrected errors so they can be checked. A language model then reads the calculated chart and writes the interpretation; a person does not write it word by word. The term “master reading” (師傅解讀) on the Chinese site refers to this model-generated text. We currently have no evidence that establishes the readings as accurate. In the August 2026 chart-swap experiment, a reading based on the participant's own birth chart scored a 62.2% hit rate against statements in that person's life history. Control readings based on three fictional sets of birth details averaged 66.0%. With the same model, the participant's own chart did not score higher. There was only one participant, and the experiment used an older process that has since changed, so this also does not establish that readings are inaccurate. It shows that a reading feeling accurate is not, by itself, evidence.

What we can take responsibility for, and make open to scrutiny, is whether the chart has been calculated correctly. The 16 errors corrected so far, their effects, and their fix dates appear under “Corrections already made.” For the reading itself, our approach is to constrain it: it must use the items already calculated in the chart and must not invent or remove stars, palaces, or lines. Each tool states its school-specific choices. We currently have no data that can answer whether these readings predict the future.

To assess a reading yourself, check whether each statement points back to a position in the chart. If it does not, the same sentence could have been written for another chart. The feedback card after each reading asks whether the reading was accurate. These ratings provide data for later evaluation; the prerequisites for the next controlled experiment appear under “Unresolved questions.”

How the homepage numbers are calculated

Homepage figures are calculated from the database and refreshed every 5 minutes. The period begins with the new site's launch on August 29, 2026. Administrator and test accounts are excluded.

The “Explore your birth chart” area shows only BaZi statistics; the “Ask a question” area shows only Six-line I Ching statistics. Reading counts include follow-up questions of the same type and exclude failed or interrupted readings. Zi Wei, relationship charts, children's charts, daily readings, and old records with no type label are not mixed into these two areas.

“Helpful” is the percentage of submitted helpfulness responses that selected “Yes”; “Partly” does not count as yes. “Reader rating” is the average score readers give, from 1 to 10, after finishing a detailed reading. One person can rate more than one reading, so these are response counts, not counts of independent people.

A service's reading count appears only after 100 readings, and its helpfulness percentage requires 20 responses. The average reader rating appears from the first in-depth rating; the number of ratings is available under “About these figures.” Without any ratings, it says “No ratings yet.” One person can rate different readings, so the count does not represent unique readers. A self-rating is not an objective hit rate. As the chart-swap experiment shows, feeling that a reading is accurate does not establish that the chart or interpretation is correct.

The question this page addresses

“Accuracy” is one of the hardest claims to check on a divination site. When a reader feels that sixty or seventy percent of a reading fits, that feeling can be real—but it can come from three very different sources: the chart calculation contributed something, the wording was broad enough to fit anyone, or the statement came from a prompt-template example and would appear for any chart.

Feeling alone cannot separate those sources; an experiment is needed. Our first experiment, numbered 01 and completed on August 5, 2026, was designed to separate them. Its results were not flattering to us. The design, numbers, and limits are published together below.

The method: chart-swap controls

The Barnum effect is the tendency to treat general personality descriptions as uniquely personal. A recognized way to test how much of a reading comes from this effect is a chart-swap comparison: generate a reading from someone else's chart, then score it against the participant's own life. If the other person's reading scores similarly, that score is not measuring the contribution of the chart.

1. Real-chart group: One participant provided a full life history. The site calculated their chart from their birth details, and three different model generations each produced one reading.
2. Decoy group: Three fictional cases had the same gender, a birth year within three years, the same birthplace, and the same question. Only the birth date and time differed. Readings used the same online calculation process, one model, and identical parameters. Decoys were deliberately matched as closely as possible demographically, to isolate the contribution of the chart itself.
3. Blinding: Model and identity labels were removed from all six readings and replaced with randomized identifiers. An extractor unaware of their identities split them into 309 verifiable life claims and removed divination terminology. The claims were pooled and shuffled before going to the scorer.
4. Scoring: Each claim was checked against the participant's life history and marked as a hit or miss. Each was also rated for generality: how many people would feel that the statement fitted them.

Two strict boundaries were part of the design. The person writing prompts did not score claims. The reading-generation side never had access to the participant's life history; the model received only the fields users would normally supply: birth details, gender, birthplace, and question.

What experiment 01 measured

On a narrow screen, scroll the table sideways to read all seven columns.

The three real-chart readings had a combined hit rate of 74.1%; the three decoys, 66.0%—an apparent difference of 8.1 percentage points. But the real-chart group included three models and the decoy group only one, so writing style was mixed into that difference. More than seventy percent of the claims from models A and B were generic; broader statements were easier to score as hits.

Controlling for the model: the real chart scored 62.2%, and decoys averaged 66.0%. The real chart did not outperform the decoys. This result applies to that version of the process, with n = 1.

For claims that both hit and had low generality—statements that fitted without fitting almost anyone—the three real-chart readings had 0 hits in total; decoys had 3. The full pool was heavily weighted toward generic claims: roughly sixty to seventy percent had high generality, while only about 3% were truly specific.

The red-team review found another concrete problem: example scenarios in the prompt template were repeated almost word for word across the six readings. The same “accurate” statement appeared regardless of the chart because it came from the template, not the chart. This finding directly led us to replace 13 template examples with abstract instructions.

Under that version of the process, the participant's feeling that a reading was “sixty or seventy percent right” could also be matched by scoring someone else's reading against their life: the controls reached 66%.

What these numbers cannot establish

• n = 1. Only one participant's life history was available for comparison. Every conclusion is an observation about that person, not a general rule.
• “No benefit from the chart” is not established; “no benefit detected” is the supported wording. The same-model comparison had only one real-chart reading. For the −3.8 percentage-point difference, the approximate 95% confidence interval ran from −21 to +13 percentage points. This design could not even detect a ten-percentage-point advantage for the real chart. It cannot establish that the benefit is zero; the measure is too coarse.
• Scoring was automated, not performed by the participant. Matching claims to a written life history is different from asking how the participant experiences them. Anything absent from that history was excluded from scoring, removing 15% to 38% of each reading's claims from its denominator. The participant's own blinded, claim-by-claim ratings were still awaiting completion; those are the gold standard.
• Zero highly specific hits is a measurement problem as well as a product problem. Only about 3% of all claims were specific, and more specific claims were more likely to concern something omitted from the life history. At the pooled rate, the chance of zero such hits among the real charts' 106 scored claims was already about one in four. Zero versus the decoys' 3 hits is not statistically distinguishable. The current measure cannot detect a genuinely chart-specific result; that is more accurate than saying such a result is impossible.
• The material came from an older process. Every experimental reading was generated before a guidance-trigger bug was fixed—see correction 03 below. Results for that version cannot be generalized to the current product.
• The experiment evaluated retrospective descriptions only. Future predictions were filtered out before claims were pooled. Assessing them would require a follow-up a year later, even though that is the part readers often care about most.

These limitations remain on the page because a methodology page that lists only favorable numbers would be meaningless.

Red-team review

After completion, the experiment was given to an independent review role with one task: find flaws in the method. The review identified 2 Critical, 5 Major, and 6 Minor issues.

Two main conclusions were judged stronger than the evidence and had to be rewritten. “The chart contributes almost zero benefit” became “no benefit was detected.” The word “entirely” was removed from the claim that the feeling of accuracy came entirely from writing style. The review also rejected the original formulations for product copy or as assumptions in later experiments. This page follows the revised conclusions.

The review identified three design biases that weakened the report's own case. The most serious was converting calendar years into ages when extracting claims: this erased mismatched years in the decoys and favored a result where the real chart did not win. Following the review, we reran the analysis after removing all 83 age-anchored claims. The Barnum comparison gap changed from 8.1% to 8.2%, without changing the direction of the conclusion.

Birthplace, time zones, and apparent solar time

Since October 17, 2026, new charts take the local civil date and clock time at the birthplace as input, together with the time zone, longitude, and historical daylight-saving record. Year pillars, month pillars, and the interval to the start of a ten-year cycle compare the actual instants of birth and solar-term transition. Day pillars, hour pillars, and the lunar date used for Zi Wei use the local apparent solar date and time. If you were born outside Taiwan, do not convert to Taiwan time or subtract daylight saving yourself.

If a clock time occurs twice when daylight saving ends, choose the earlier or later occurrence according to the birth record. A skipped, nonexistent time is rejected. Near the International Date Line, the local civil date is the reference and the solar day closest to the local clock is chosen; this is the site's stated calendar convention. Existing charts and readings keep their original records. The new rules apply when you confirm the details and create a new chart.

An unknown birth time is not replaced with noon. On an ordinary date, only three pillars are calculated; the day pillar provisionally uses the civil date, without a solar-time correction across midnight. Zi Wei and the start time of ten-year cycles are not provided. If a solar-term transition occurs on that local birth date, the year or month pillar cannot be uniquely determined, so a time is required first. Historical time-zone data before 1970 may be incomplete, and the equation of time uses an approximation by day of year. Births near boundaries still need more precise records. For the procedure, formulas, and original sources, see apparent solar time and births outside Taiwan. These improvements make time conversion explicit; they do not validate the accuracy of the reading.

Corrections already made

Correct chart calculation takes priority over prompt wording. These issues were found and fixed during the research process. Dates are the dates the fixes went live.

1. 01The equation-of-time formula used the wrong phase2026-08-05

   The code combined the phase from a solar-declination formula with equation-of-time coefficients. It matched 0/4 reference points in the astronomical almanac; the annual mean error was 13.2 minutes and the maximum was 28.7 minutes. For births in Taipei, an estimated 11% of birth times could be assigned to the wrong hour branch. An incorrect hour branch also shifts the Life Palace and, with it, the entire Zi Wei chart.

2. 02Solar-time correction did not change the date when crossing midnight2026-08-05

   When correction moved 23:50 to 00:10 on the following day, the code wrapped the minutes but left the date unchanged. The day pillar therefore remained on the previous day. This affected relatively few people—roughly those born within twenty minutes of midnight—but when it did, the Day Master changed, invalidating the whole chart and its readings.

3. 03The Zi Wei warning for a Ji transformation in the Well-being Palace never triggered2026-08-05

   The Four Transformations data stored the single character 忌 (Ji), but the comparison expected 化忌 (Hua Ji). They could never match. The rule had existed for more than a year and had triggered zero times: charts that should have received a reminder about mental or emotional well-being never received one.

4. 04Reading guidance was incorrectly nested inside the Zi Wei condition2026-08-05

   All five guidance rules were inside the condition checking whether a Zi Wei chart existed. Users who did not know their birth time therefore missed even the rules based only on BaZi. The inputs with the least information, and the greatest need for cautious reminders, received no warnings at all.

5. 05The Six-line I Ching outer-trigram Na Jia indices were misplaced2026-08-29

   The three outer lines took values from the inner-trigram positions. Across all 64 hexagrams, the earthly branches of all 192 outer lines were wrong. There were 128 incorrect Six Relations assignments (66.7%), and 52 hexagrams incorrectly lacked one relation. A focus relation could therefore appear to be absent when it was actually present, potentially reversing the interpretation. The same fix added the previously uncalculated Self and Other lines. Every hexagram was checked against the Na Jia assignment verses in Bu Shi Zheng Zong (《卜筮正宗》). A regression test independent of external data was also added: exactly ten of the 64 hexagrams must be Six-Clash hexagrams.

6. 06Day Master strength counted Resource hidden stems as roots2026-08-29

   In the traditional framework, having roots distinguishes whether a Day Master can carry Wealth and Officer influences. Roots mean the Day Master's own elemental energy; Resource generates that energy but is not itself a root. Combining them in one score systematically labeled weak charts as strong, leading to narratives such as “pursue wealth” or “you can bear the risk.” Root strength and Resource support are now scored separately, with an explicit flag for the presence of a substantial root.

7. 07Extremely unbalanced charts no longer receive forced favorable-element judgments2026-08-29

   Following patterns can have favorable and unfavorable elements opposite to those in ordinary strength-balancing analysis. Applying the ordinary rule can describe a favorable element as unfavorable. These charts no longer receive lists of favorable and unfavorable elements; the result instead says judgment must be reserved. The aim is not to claim the algorithm can correctly identify every following pattern, but to stop it from forcing an answer where it lacks the ability to give one.

8. 08The “abundant Wealth, weak Day Master” warning was both too broad and incomplete2026-08-29

   Previously, any Wealth influence appearing in the heavenly stems, together with a sufficiently low score, could trigger the strongest warning. Conversely, a typical case with all Wealth hidden in the earthly branches and strengthened by a formation and the season could fail to trigger it. The revised rule uses four conditions and grades the warning according to the strength of the evidence.

9. 09Favorable-element wording was made more qualified2026-08-29

   This site selects favorable elements using only strength balancing. It does not apply pattern-support analysis, seasonal adjustment, the illness-and-remedy approach, or mediation between elements. The previous wording presented one school's answer as an undisputed fact. The output now names the method and, when the strength score is near a switching threshold, marks the chart as borderline and uses qualified language.

10. 10Seed examples were removed from prompt templates2026-08-29

    Thirteen example sentences that were being repeated across different charts were replaced with abstract instructions. This followed directly from the red-team review: the same sentences appeared regardless of the chart because they came from the template rather than the chart.

11. 11An absolute wording rule contradicted the requirement for qualified language2026-08-29

    One global rule prohibited words such as “possibly” and “perhaps,” while another required qualified language for borderline charts. The charts most in need of caution therefore faced the strongest conflict, effectively undoing that caution. The prohibition now has an explicit exception.

12. 12Relationship and individual charts used different element totals2026-08-29

    The same chart could show different Five Element totals in two product areas. Both now use the weighted version: 1.0 for the main hidden stem, 0.5 for the middle hidden stem, and 0.3 for the residual hidden stem.

13. 13Readings were being cut off2026-08-30

    After model reasoning effort was set to its highest level, half the output budget went to hidden reasoning. Visible text was shorter than at either of the other two settings; output stopped mid-sentence and the entire ten-year-cycle section was missing. Returning to the medium setting produced the only configuration that completed the full reading.

14. 14Birthplace names did not match the longitude table, so solar-time correction was skipped2026-08-30

    Profiles stored “New Taipei City” (新北市), while the longitude table used “New Taipei” (新北). The exact-match lookup failed and no correction was applied. Testing found that only 1 of the 22 users who had entered birth details at the time had a matching entry. Changhua, Nantou, and Yunlin counties, which were missing from the longitude table, were added in the same fix.

15. 15Solar-term transitions used dates only, and term times were several minutes off2026-09-25

    The year and month pillars checked only whether the birth date was the date of a solar-term transition. People born earlier than the transition time that day were assigned to the next month; on the Start of Spring, the year pillar was also wrong. The low-precision solar-term formula differed from the Purple Mountain Observatory's published values for 1950–2025 by 3.6 minutes on average and 12.8 minutes at most; 2 solar terms even fell on the wrong date. After switching to a high-precision solar-position algorithm, all differences from the published values were within 1 minute. Year pillars, month pillars, ten-year-cycle start times, and the Six-line I Ching month branch now compare times to the minute. The main affected group is people born on a solar-term transition day: those born before the transition return to the previous month, and at the Start of Spring, the previous year as well.

16. 16Zi Wei principal-star brightness tables did not match the classical source2026-09-25

    The brightness entries for fourteen principal stars in twelve palaces were checked cell by cell against the table in Zi Wei Dou Shu Quan Shu (《紫微斗數全書》). Of 168 cells, 127 were wrong. For example, the Sun in the Zi palace should be “fallen” (陷), but the code labeled it “neutral” (平). The table now follows that source and matches the widely used open-source chart software iztro cell by cell. Brightness labels on every Zi Wei chart, and brightness data sent into readings, change accordingly.

This list also means that, before the errors were found, readings on this site had already been interpreting incorrect charts for some time.

Unresolved questions

• The participant's own blinded ratings are incomplete. The difference between the participant's claim-by-claim ratings and the automated scores is itself data we need to examine.
• The sample is still n = 1. Expanding to 3 to 5 participants requires people willing to provide full life histories and birth details.
• The second experiment has not started. It will examine how strongly a prompt should guide the reading: give a conclusion, point toward one, or give none. Its prerequisites are a finalized removal of seed examples and a new baseline using the corrected process.
• Every output-affecting change breaks comparability. Records from before and after a correction are not mixed in the same statistics, and each boundary is recorded. The Six-line I Ching Na Jia correction, prompt-wording changes, and model-setting changes each have their own boundary. The next experiment's baseline must be collected after the last such change. This slows progress, but comparisons made by mixing versions would be meaningless.
• BaZi calculation still has known omissions. These include the effects of clashes, punishments, combinations, and meetings on roots; three-branch combinations and seasonal assemblies forming patterns; and seasonal adjustment. These cannot be expressed within the current strength-balancing framework. We state that limitation instead of pretending they have been implemented.

What you can check on the site

Reading feedback. A feedback card appears after each completed reading. It asks whether the reading was accurate, rather than whether the reader was satisfied: satisfaction can reflect prose quality and length, precisely the factors experiment 01 aimed to separate. Scores run from 1 to 10, and guests can respond. This is the only ongoing source of accuracy-feedback data. Readings that fail or are stopped do not show the card, so service failures are not mixed into perceived accuracy.

School-specific choices. Traditional chart interpretation has no single agreed answer; schools can interpret the same chart differently. Each tool states the choices used here. Zi Wei's Four Transformations follow the commonly used modern Zhongzhou school. BaZi uses the late-Zi-hour convention: the day pillar does not advance after 23:00. Hidden-stem weights are fixed, without the seasonal day-by-day allocation of ruling hidden stems. A different result elsewhere often reflects a different school rather than a calculation error.

What is and is not calculated. The final sections of the BaZi, Zi Wei Dou Shu, and Six-line I Ching tool pages list what the software actually calculates and what it deliberately leaves out. Readings are constrained not to invent derivations from items that have not been calculated.

References

The definitions of the Barnum effect and cold reading used on this page come from the following literature. Its experimental approaches also informed the design.

• Forer, B. R. (1949). The fallacy of personal validation: A classroom demonstration of gullibility. Journal of Abnormal and Social Psychology, 44(1), 118–123.
• Sundberg, N. D. (1955). The acceptability of “fake” versus “bona fide” personality test interpretations. Journal of Abnormal and Social Psychology, 50, 145–147.
• Dickson, D. H., & Kelly, I. W. (1985). The “Barnum effect” in personality assessment: A review of the literature. Psychological Reports, 57(2), 367–382.
• Hyman, R. (1977). Cold reading: How to convince strangers that you know all about them. The Zetetic (Skeptical Inquirer), 1(2), 18–37.