Jyotiṣa Cognitive Reasoning System

This system treats two very different kinds of accuracy separately. Chart facts (the actual astronomical positions and derived values) come from a local ephemeris and a named, versioned CalculationProfile: the same birth data and profile always produce the same ChartSnapshot, hashable and reproducible. Jyotiṣa knowledge (what those positions are traditionally held to mean) follows a different path entirely: editioned source texts flow through reviewed claims into executable rules, and a user-facing answer may only cite a published RuleSetRelease, never an LLM's general training-data sense of what a text "probably" says.

The project is explicit that LLMs know a syllabus of Jyotiṣa, not a critical edition, and treats them accordingly: usable as clerks and adversarial reviewers, never as the source of authority (śāstra). Several rules are enforced as non-negotiable throughout the codebase: no silent convention-mixing (every result belongs to a named CalculationProfile), a source claim, a commentator's interpretation, an implementation rule, and a system inference are kept as distinct types rather than flattened together, contradiction between sources is preserved as data rather than resolved by fiat, and the system has to answer usefully even with the LLM layer switched off entirely. The deterministic calculation core is Python, with a separate knowledge pipeline package and machine-readable schemas defining the contracts between them.