Engine Progress Report — Verification, Progress, and a Null Result We Publish Anyway
July 19, 2026 · Our principle has not changed: every claim can be checked, and every experimental result — good or bad — is published as-is.
100%
Lu, Quan and Ji stars of the Ziwei Four Transformations — all ten heavenly stems consistent with an independent library
320/320
Decade and annual cycle transformations — all rule checks passed
1.6×
Heavyweight timing markers — odds of a major event vs baseline (honest lift, not recall)
100% vs 31%
Engine-anchored vs free-form AI — fidelity of core favourable element (yong shen) judgements
The Ziwei Layer Passes Independent Cross-Verification
VERIFIEDThe Four Transformations (sihua) system of Ziwei Doushu (Lu, Quan, Ke, Ji) carries long-standing differences between schools, and it is the layer where AI-driven Chinese astrology most easily goes quietly wrong. We did something rarely done in this field: we cross-checked two completely independently developed open-source chart casting implementations against each other. For the birth-year transformations, all ten heavenly stems were compared cell by cell — the positions of the Lu, Quan and Ji stars agree in every case. The only divergence concentrates on the Ke star under three stems, a documented difference between the Sanhe and Zhongzhou schools, not an error. We have formally frozen our adoption of the Sanhe system, written that school choice into the engine's single source of truth, and every future change is re-verified automatically.
One layer up, the decade and annual cycle transformations: across 320 randomly generated birth charts, the direction of decade progression (forward for yang-male/yin-female charts, reverse for yin-male/yang-female), the five-element bureau starting age, palace stems, and the palaces the transformation stars land in all passed at 100%, cross-checked against the independent library. In other words: from chart casting through birth-year transformations to decade and annual transformations, the entire computation chain is now verified.
Timing Engine: We Revised Our Own Headline Number Down
VERIFIEDWe used to state here that “94.5% of real events fall in the years our engine flagged.” After a self-audit, we are correcting that publicly. The figure is true but misleading: the engine flags roughly 91% of all years as carrying a signal, so almost any year “hits” — its precision (flagged as a signal year, and a major event actually occurs) is only 5%, essentially equal to the 4.9% natural base rate. The recall looks impressive but carries almost no discriminating power. The real signal lives in the rare, heavyweight markers (three-harmony frames, decade-year coincidence, month-pillar clashes): when those fire, the odds of a major life event that year are about 1.6× baseline — modest, but real. We chose to revise our own headline down, because an honest 1.6× beats an inflated 94.5%. This is what “publish the null anyway” looks like applied to ourselves.
Product Progress (Live)
Conversations Have Memory, Identities Never Mix
A new conversation identity system: every conversation is anchored to one explicit subject (yourself, a family member, a partner), with birth charts, Q&A history and past-event memory all isolated per subject — when you read for a family member, the answers always come from their chart, never mixed up. Follow-up dialogue has also been upgraded: it builds on what came before instead of repeating what has already been said.
Not Sure What to Ask? The Engine Opens the Conversation
“Ask another question” now suggests questions based on your own birth chart — organised by angles such as career, timing, relationships and self, each offering a quick form and a deep form, showing you how to phrase a question to reach the most substantial answer.
Birth Charts Self-Repair Automatically
The engine now automatically scans every 15 minutes and repairs any incomplete birth chart data — this class of problem has gone from “fixed when discovered” to “exists for at most 15 minutes”.
English Edition — a New Archetypal Voice
A redesigned English reading voice for Western readers (The Mountain, Growth Allies, Life Chapter…), with a more distinctive style — and not one grounding rule relaxed: years are stated only when the engine has actually computed them, and dates and events are never invented.
The Engine Opens as AI Infrastructure (MCP)
We have opened our deterministic computation layer as a public MCP service — any AI assistant can call our engine directly to cast charts, check dates and read annual cycles, instead of guessing on its own. Free-form AI judging the core favourable element (yong shen) is only 31% faithful; anchored to our engine, it is 100%. That is the reason for opening it.
What's Next
Verifying the self-transformation and flying-transformation layers against classical case records (a hand-built gold standard), letting the blind-test feedback flywheel keep accumulating real outcome data, and extending the MCP tools around actual user needs (compatibility, date selection range scans). The same principle throughout: verify first, then claim.
All figures in this report come from a reproducible test harness; methods and raw results are archived in the engineering repository. Verification methods: cross-comparison of independent implementations, property-based testing, and pre-registered experimental design (placebo controls + outcome shuffles + multiple-comparison correction).