How Lodestar actually reads a chart

Every Lodestar reading starts by computing your natal chart with Swiss Ephemeris — the same JPL-derived ephemeris library used by professional astronomers and astro-software going back two decades. The math is deterministic: the same birth date, time, and place always produces byte-identical planetary positions, house cusps, nakshatras, and Vimshottari dasha timings. No hidden randomness, no “house style”, no drift.

• Sidereal Vedic zodiac with the KP-Newcomb ayanamsa, calibrated byte-for-byte against the Astro Murali reference worksheet.
• Placidus & KP cusps to arc-second precision, computed independently of any narrative layer.
• Vimshottari dashas resolved to five levels deep (mahadasha → bhukti → antara → sookshma → prana) with balance-at-birth.
• Reproducibility promise: feed the same birth data twice and you get the same chart, same houses, same significators — forever.

KP is a mid-20th-century South Indian refinement of classical Vedic astrology that shrank every prediction down to a single testable question: at any given moment, will a specific life event fructify or not?It does this by looking at the “sub-lord” of each of the twelve house cusps — a hyper-fine subdivision of each nakshatra star that acts as the deciding vote. If the cusp sub-lord of the 10th house signifies career-friendly houses (2, 6, 10, 11), a promotion is possible. If it signifies denying houses (5, 8, 12 for career), it’s denied outright, no matter how flattering the rest of the chart looks.

Every Lodestar answer is grounded in this framework. When you ask “will I get promoted this year?” the engine computes:

• Which houses that intent activates (2 · 6 · 10 · 11 for career).
• Which planets significate those houses at four levels of KP hierarchy (A / B / C / D).
• The cusp sub-lord verdict for the primary house.
• Whether the ruling planets at the moment of the question overlap the significators.
• Whether the current dasha-bhukti chain agrees.
• Whether transits ratify or block the promise.

These six layers form the “Ruling-Planet Cascade” — a deterministic scorecard that runs before any language model sees your question. When the layers agree, the reading speaks with confidence. When they disagree, the reading is required to hedge.

This is where Lodestar leaves traditional astrology behind and starts behaving like a scientific instrument. Every reading decomposes into a set of discrete symbolic rules — e.g. csl:10:Saturn (“the 10th cusp sub-lord is Saturn”), dasha:Jupiter→Venus (“native is in Jupiter mahadasha, Venus bhukti”), and so on. These rules are logged, and when you later tell us whether the prediction was accurate (via a thumbs-up, thumbs-down, or a follow-up outcome check-in), each rule’s ledger updates.

The ledger uses Wilson lower-bound confidence intervals — the same statistical device Reddit uses to rank comments — so a rule with 8 confirmations out of 10 is trusted more than a rule with 80 confirmations out of 100 only if the sample supports it. Rules with weighted n < 5 stay in a “building” tier and any downstream language model is forced to hedge on them. Rules whose Wilson lower bound exceeds 0.55 are unlocked for decisive language. Rules that repeatedly fail (Wilson upper bound ≤ 0.35) are demoted: the reading will actively caution the user.

Practically: the more people ask Lodestar questions and report back, the sharper the entire engine gets — for everyone. The ledger is the compounding moat.

Beyond per-rule calibration, Lodestar retrieves your “chart twins”: past readings whose rule firings overlap yours by at least 35% (Jaccard similarity), and that carry a verified outcome. We then aggregate their outcomes into an empirical prior with a Wilson lower bound, and hand it to the language model as evidence.

• K-anonymous by design. Precedent is only surfaced when at least 5 verified twins exist — no single user is ever exposed.
• Verified outcomes weighted 3× user thumbs (which are cheap). The strongest signal is a real-world “yes it happened / no it didn’t” answer to a scheduled follow-up.
• Personal calibration overlay. Bayesian shrinkage (κ = 8) blends your own past outcomes with the global rate so your chart’s history counts more as you accumulate readings.

The pipeline above produces a machine-readable chart_slice object: which houses matter for your question, which planets significate them, the cusp sub-lord verdict, the current dasha, the ruling-planet cascade result, the calibrated signal per rule, and the chart-twin precedent. Only then does the language model turn on — and it is prompted to:

• Speak decisively only on rules the ledger has verified.
• Hedge visibly (“this indicator is still under observation”) on rules that haven’t earned confidence.
• Caution the reader on rules the ledger has caught failing.
• Never invent a planetary position, dasha date, or house significator — those come from the engine, verbatim.
• Cite the deterministic slice at the bottom of every reading (the “Why this answer” panel).

We fail over across five providers (Cerebras · Gemini · OpenRouter · Mistral · Cloudflare) with A/B routing, but the model itself is a narrator, not the astrologer. The astrology happened before it ever saw your question.

Every reading includes a “Why this answer” panel showing the exact deterministic slice: which houses we read, which planets we used, which sub-lord verdicts applied, what your current mahadasha and bhukti are, and where the ledger’s confidence stood. If you disagree with a specific sub-claim, thumbs-down it — that weakens the underlying rule for everyone.

For paying customers, we expose the full KP worksheet as a PDF: 9-planet table with 5-level lord chains, 12-bhava cusps with 5-level lord chains, effective links, Old-KP significators, and the Vimshottari tree down to prana. Any independent KP practitioner can verify it against their software of choice.