Oracle is a three-layer system: scripture-grounded computation, structured verdict synthesis, and AI narration. Each layer has a clear boundary — and the AI is never allowed to guess chart data.
1
COMPUTATION
187 engines · scriptural formulas
Each engine implements a specific classical formula with verse-level citations. Shadbala strength calculations, Vimshottari dasha timing, Ashtakavarga transit scores, yoga identification, dosha detection — all computed from planetary positions using the Swiss Ephemeris. No heuristics. No training data. Pure astronomical math applied to classical rules.
2
VERDICT SYNTHESIS
Structured 5-layer hierarchy
Engine outputs pass through a strict 5-layer hierarchy: structural placement, planetary strength, modification rules (doshas, yogas, cancellations), dasha timing, and cross-scripture confirmation. The result is a tier-banded verdict per life domain — weak, moderate, strong, blessed, or obstacles — informed by scripture but optimised for clarity. No single score pretends to compress a life into a number.
3
NARRATION
AI language, not AI astrology
GPT receives the structured verdict, the raw engine data, and strict anti-hallucination system prompts. Its job is narration only — translating tier bands and engine outputs into natural language. It cannot invent chart details, override engine verdicts, or claim certainty the engines didn't provide. If the engine says "uncertain," the narration says "uncertain."
17 CLASSICAL SCRIPTURES
Every engine traces to a verse.
No engine in Oracle operates without a scriptural citation. When you see a claim, you can ask Oracle what verse it comes from — and it will tell you.
Brihat Parashara Hora ShastraThe foundational text
Brihat JatakaVarahamihira · ~6th century
PhaladeepikaMantreswara · classical
SaravaliKalyana Varma · 10th century
Jataka ParijataVaidyanatha · 15th century
Uttara KalamritaKalidasa · classical
Jaimini SutrasMaharshi Jaimini
Laghu ParashariDasha principles
Sarvarth ChintamaniVenkatesha · 16th century
Krishnamurti Paddhati (KP)K.S. Krishnamurti · 20th
Lal KitabPt. Roop Chand Joshi · 1941
Prashna MargaHorary classics
Tajika NeelakanthiAnnual charts
Sudarshan ChakraTri-ascendant analysis
Hora RatnamBalabhadra · 17th century
Chamatkara ChintamaniBhatta Narayana
DharmasindhuMuhurta classical
What the AI cannot do
The AI narration layer has hard constraints baked into its system prompt. These aren't suggestions — they are structural:
- Cannot invent planetary positions. Every position it references must appear in the engine output.
- Cannot invent yogas or doshas. It names only what the engines explicitly detected.
- Cannot claim higher certainty than the engines. If the engine returned "weak indication," the narration cannot upgrade it to "certain."
- Cannot predict prohibited categories. Death timing, lottery numbers, medical diagnoses — these are blocked at the narration layer regardless of what the user asks.
- Cannot override scripture. If the engine applied a Kakshya Hrasa reduction or a cancellation rule, the narration must reflect it.
⊙ Why this matters
Most "AI astrology" products let the AI free-style the entire reading. The AI hallucinates planetary positions, invents yogas that aren't there, and fabricates certainty it doesn't have. Oracle's architecture makes that impossible by design. The engine does the astrology. The AI writes the sentences.
What makes it hard
The real work wasn't the AI layer — it was the 187 engines. Each one required: reading the relevant scripture(s), mapping ambiguous classical language to precise mathematical rules, handling cancellation conditions, applying strength reductions like Kakshya Hrasa, reconciling conflicts between scriptures, and validating against documented charts. Some engines took a week. A few took a month.
The 5-layer verdict synthesis came next. Raw engine outputs can conflict — one yoga says "blessed marriage," another says "delayed marriage." The verdict layer resolves these using scripture-grounded precedence rules, not averages. Averaging classical astrology produces mush; stacking it with clear precedence produces truth.
What it took to ship
Oracle has been in development since early 2026. One solo developer (Delhi). Over 132 master tasks in the current rebuild alone. 750+ validation events tested across 157 charts. Every scripture re-read. Every engine refactored at least once. Zero hardcoded text — every sentence the user reads is dynamically generated from chart data.
And we're still improving. The methodology page shows exactly where we stand and what we're working on next.
See the numbers →