About Deep Oracle
Every claim in a Deep Oracle reading traces back to a named classical source and a deterministic calculation. If we can't show you where a conclusion comes from, it doesn't belong in the reading.
What is Deep Oracle
Deep Oracle is an AI-powered reading platform covering three distinct traditions: BaZi (Four Pillars of Destiny), Western natal astrology, and Tarot. We practice these systems. We built Deep Oracle because the AI tools we tried couldn't tell us why they said what they said — and when we checked the classical sources, their answers often didn't match. Our architecture puts ground truth first. A deterministic engine computes the structure of your reading — stems and branches, pattern classification, useful god, and pillar interactions for BaZi; planetary positions to arcsecond precision via the Swiss Ephemeris for Western; cryptographically sound random draws with elemental-dignity relationships computed before interpretation for Tarot. Only after the structure is computed does the AI write the interpretation, and it writes against that structure rather than inventing conclusions from its training data.
Our three systems
BaZi (Four Pillars of Destiny)
We ground our BaZi analysis in three Qing-era classics. Zi Ping Zhen Quan (子平真詮), attributed to Qianlong-era jinshi Shen Xiaozhan, anchors our pattern theory (格局). Di Tian Sui (滴天髓) — a text of disputed authorship traditionally associated with Liu Ji, which we cite through the commentary tradition that grew around it — informs our reading of stem-branch essential nature. Qiong Tong Bao Jian (窮通寶鑑), as systematized by Qing-era Yu Chuntai from the earlier folk manuscript Lan Jiang Wang, drives our seasonal useful-god analysis. Three supplementary texts — San Ming Tong Hui (attributed to Ming scholar Wan Minying), Shen Feng Tong Kao, and the Song-compiled Yuan Hai Zi Ping — inform the broader framework. Our engine integrates 38 spirit stars, 120+ seasonal rules, and 11 types of branch interaction, with a validator that refuses output when the calculation and the classical framework disagree.
Western astrology
Our natal chart engine uses the Swiss Ephemeris for planetary positions and applies equation-of-time and longitude correction for true local time. We draw on Ptolemy's Tetrabiblos (2nd century CE) and William Lilly's Christian Astrology (1647) for technical foundation; on Dane Rudhyar's The Astrology of Personality (1936) and Liz Greene's body of work for the depth-psychology tradition; and on Chris Brennan's Hellenistic Astrology: The Study of Fate and Fortune (2017) and Demetra George's Ancient Astrology in Theory and Practice (2019–2022) for the contemporary Hellenistic revival. We draw on both Hellenistic and modern frames, matching the framing to what the chart asks of us.
Tarot
Deep Oracle's tarot readings build on the Rider–Waite–Smith deck (1909, William Rider & Son — based on Arthur Edward Waite's instructions and illustrated by Pamela Colman Smith, whose name modern editions increasingly give equal billing with Waite's). We incorporate elemental dignity — the interaction of fire, water, air, and earth between adjacent cards — as a structural technique rarely implemented in AI tarot tools. Interpretive depth draws on Mary K. Greer's work on reversed cards and autobiographical reading and Rachel Pollack's Seventy-Eight Degrees of Wisdom (1980); symbolic framing from the Aleister Crowley / Lady Frieda Harris Thoth tradition (painted 1938–1943, companion Book of Thoth 1944) informs readings where it clarifies rather than mystifies.
Methodology
Traditional metaphysical systems demand precision before interpretation. A single off-by-one hour in BaZi changes the hour pillar entirely; an arcminute error in planetary position can shift a tight aspect in or out of orb. We handle this the same way across all three systems: deterministic calculation first, AI interpretation second, against the computed structure.
For BaZi, our true-solar-time correction applies both longitude adjustment and the equation of time — the seasonal variation in solar noon — to calibrate to minute precision. Pattern classification, useful god selection, spirit-star presence, and branch interaction types are all computed by rule-based engines written and validated by practicing BaZi readers. The AI does not infer these; it reads them. For Western, the Swiss Ephemeris provides planetary positions to arcsecond precision; house cusps, aspects, and essential dignities are computed, not sampled. For Tarot, card draws use cryptographically secure randomness, and elemental-dignity relationships between adjacent cards are calculated before the reading begins.
This matters because of how large language models fail. The common failure mode — sometimes called hallucination — is the model generating a plausible-sounding answer with no grounding in the actual input. A reading that invents a day-master strength or fabricates a classical citation reads fluent but is worthless. Grounding the model against pre-computed ground truth, and prompting it against that structure rather than the raw birth data, is the architectural difference between a reading we stand behind and one that merely sounds convincing.
Classical foundations
Three primary BaZi texts anchor our analysis; three more inform it. The primary three cover the core analytical frame — Zi Ping Zhen Quan for pattern theory (格局), Di Tian Sui for the essential nature of stems and branches (體性), and Qiong Tong Bao Jian for seasonal useful-god selection (調候). The informing three — San Ming Tong Hui, Shen Feng Tong Kao, and Yuan Hai Zi Ping — contribute comprehensive-system reference, annual-forecasting technique, and foundational stem-branch logic respectively.
A note on authorship. The BaZi canon is a centuries-long conversation across authors whose attribution is sometimes secure and sometimes not. Zi Ping Zhen Quan comes down to us with Qing-era jinshi Shen Xiaozhan as attributed author; Xu Leyue's Republican-era commentary is widely read but sinological scholarship notes it sometimes departs from Shen's intent — so we consult the base text and the commentary tradition, rather than treating the commentary as transparent. Di Tian Sui is what modern sinological discussion calls "a text of disputed authorship"; we cite it by the commentary tradition that grew around it rather than by single authorship. Qiong Tong Bao Jian in its extant form is the product of Qing-era Yu Chuntai's systematization of an earlier Lan Jiang Wang manuscript, which itself survives with significant textual gaps. We treat these texts as what they are: conversations across centuries, not monoliths.
For Western astrology, our interpretive tradition spans nineteen centuries: Ptolemy's Tetrabiblos and William Lilly's Christian Astrology for technical foundation; Rudhyar and Greene for the 20th-century depth-psychology turn; and the contemporary Hellenistic revival (Brennan, George, and colleagues) for the recovery of time-lord, lot, and sect techniques that fell out of use in the modern era.
For Tarot, the Rider–Waite–Smith deck is the baseline, with Mary K. Greer's reversed-card work and Rachel Pollack's Seventy-Eight Degrees of Wisdom as the modern synthesizing voices.
Our editorial standards
The commitments that distinguish our readings from plausible-sounding generated content:
What we will do.
- Cite classical sources by name, per conclusion, where a classical source supports the conclusion. If there is no classical grounding for a claim, the reading says so.
- Mark time uncertainty explicitly. If your birth time is imprecise, the hour pillar is flagged in BaZi; the Ascendant is flagged in Western. Conclusions that depend on an uncertain data point carry a lower confidence label.
- Show the calculation. Every reading includes the computed chart structure — stems, branches, hidden stems, ten-gods, or planetary positions, aspects, houses, dignities — so a practitioner can audit the interpretation against the data.
- Mark confidence honestly when signal is thin. Every section of every reading carries a confidence label — high, moderate, or low. If a specific question doesn't have enough signal in the chart to answer with conviction, the reading marks that conclusion low-confidence rather than inventing something to fill the space.
What we won't do.
- We won't predict specific dates of death, illness, marriage, or other major life events. Structural tendencies, yes; dates, no.
- We won't diagnose medical conditions or recommend treatment. When chart indicators touch health, we flag them as patterns worth discussing with a qualified professional, not as conclusions.
- We won't provide financial advice specific enough to act on. Wealth-section analysis describes structural tendencies, not stock picks or investment timing.
- We won't generate content the calculation engine can't support. If our deterministic stack cannot compute a particular interaction, the interpretation does not get to speculate about it.
- We won't invent information when your birth data is incomplete. If you do not know your birth time, we give you what the day, month, and year pillars can say, and mark the rest as unknown.
Model Changelog
Our AI analysis engine iterates continuously — every significant quality change bumps the version. Readings generated under older versions remain valid, but regenerating usually produces a deeper result.
Age-Aware Reading Framing
Readings for subjects under 18 automatically shift tone to future-outlook, with advice targeting parents
- New buildAgeFramingNote(): under-13 uses full future-outlook framing, marriage/career/wealth presented as adult trends
- Ages 13-17 get lighter framing emphasizing education and development
- Age 18+ completely unaffected — zero change to adult readings
- All BaZi and Western astrology prompt builders now inject age framing
Timing Ratings + Life Level Injection
Deterministic timing ratings and life level assessment injected as prompt ground truth — readings no longer rely on model inference
- Timing ratings (4-domain scores + overall verdict) injected as prompt ground truth
- Life level assessment (5-factor composite: pattern/yongshen/conflicts/flow/luck) injected into prompts
- Past validation and timing sections must align with computed ratings — no contradictions allowed
- Overview section must align with computed life level verdict
- All 8 deterministic engines fully injected — entire analytical pipeline pre-computed
Balanced Strength + Family Analysis Engine
Added "balanced" strength verdict, deterministic family analysis engine pre-computes family member star relationships
- Strength verdict now includes "balanced" option between strong and weak
- Element analysis accepts "balanced" as valid strength value
- Family analysis engine: gender-aware ten-god assignment, per-member star strength and yongshen relationship pre-computed
- Family relations section grounded in _v22_familyAnalysis as source of truth
Precise Pattern Naming + QTBJ School
Deterministic pattern naming algorithm, new QTBJ school with independent analytical pipeline
- Deterministic pattern naming via decision tree (month branch → ten god → pattern name)
- Dynamic ten-god lookup table injection eliminates element relationship hallucinations
- Independent QTBJ school prompt with seasonal-ecological analytical framework
- 4 corrections to QTBJ 120-entry prescription table
Year Pillar Fix + Natural Language
Fixed year GanZhi hallucinations in past validation, improved prose naturalness
- Year GanZhi lookup table: complete mapping from birth year to current+5
- Token budget optimization: 50K→29K input tokens, removed 6 redundant fields
- Natural language patch: diverse bridge patterns replacing mechanical transitions
- Strict 3-item cap on practical advice per section
6 Analytical Modules + Temporal Context
Added pre-computed branch interactions, spirit stars, nayin modules with dynamic temporal context
- Pre-computed branch interactions: 6 clashes, 6 harmonies, 3-harmony, 3-meetings, punishments, harms
- 17 spirit stars with per-pillar detection and timing activation
- Complete 60-cycle nayin mapping
- 120-entry twelve life stages lookup
- 120-entry QTBJ seasonal prescription reference
- Luck pillar + annual pillar interaction pre-computation
- Dynamic temporal context injection
- Persona weighting (paused, pending refinement)
Two-Phase Analytical Framework
Two-phase analysis: Phase 1 forces analytical foundation verification before Phase 2 prose
- Two-phase analytical process (analyticalFoundation → 7 reading sections)
- 5-step pattern verification protocol with hidden stem audit
- Stem combination (合化) detection and transformation
- Normal pattern quality assessment (成格/破格)
- Multi-model evaluation framework (6 models × 10 charts)
- School-specific chart data assembly
Contact
For questions about methodology, content corrections, or feedback on specific readings, write to support@deeporacle.ai. We read and respond personally.