Anthropic's Claude (the Opus and Sonnet tiers, depending on the draft phase). Not ChatGPT, not Gemini, not Grok. Claude was chosen because in our own testing it produces less hallucinated chart data and more grounded prose for long-form reflective writing. We may add other models if they out-perform on specific tasks; we will name them here if/when we do.

The model sees, for each reading draft:

• The chart math output (planet positions, aspects, house cusps, computed deterministically by an astronomy library, not by the AI). See the published chart math for a worked-example chart.
• Your free-text "why interested" and any context you submitted at intake
• A library of practitioner-authored reference passages keyed to chart configurations
• The practitioner's voice samples (5 to 10 prior readings on similar configurations, redacted of identifying info)

It is never given:

• Your email, name, or birth location precise to street level
• Any history of previous readings you've had elsewhere
• Any third-party data we don't have ourselves (no social media scraping; no astrology-database lookups)
• Other House Call users' charts or readings
• Your account password (never; it's bcrypt-hashed before storage)

The model is given the chart math, your context, and the practitioner's reference material for the tradition in question, and the practitioner works through each chart configuration with it directly, not by handing off a chart and reviewing a finished draft afterward. The practitioner is directing the process throughout: steering what the instrument surfaces, cross-referencing it against the chart math as they go, and adding the structural insights and connective judgment the model doesn't reach.

For some sections (the "short answer" callouts, the practical "what to do at the next gate" lists), the practitioner writes from scratch and uses the model only as a sounding board, not a drafter.

It does. Routinely. Specifically, these are the failure modes the practitioner is actively watching for throughout the process, not catching after the fact:

• Hallucinated transits. The model occasionally claims a transit is happening when it isn't. The practitioner cross-checks every named transit against the chart math output as it comes up.
• Generic horoscope drift. The model sometimes slides into "Capricorns are practical and ambitious", true and useless. The practitioner cuts these on sight.
• False certainty. The model writes confident prose even when the chart is ambiguous. The practitioner downgrades to "the chart suggests" or "one possible reading" when warranted.
• Wrong sign. Rare but it has happened, model writes about Saturn in Capricorn for a chart where Saturn is actually in Sagittarius. Always caught by the practitioner checking against the chart math.

We log these patterns and tune the instrument's configuration over time.

Email the practitioner. Every reading carries a feedback channel; we credit any error you flag with a revision and a follow-up section addressing the underlying chart configuration in greater depth. Errors are not embarrassing, they're how the practice gets sharper.

Honestly? Speed and depth. A practitioner writing every paragraph from scratch produces ~3,000 words a day before quality starts to drop. Directing a purpose-built instrument through the chart math and reference material lets the practitioner ship a 3,000 to 3,500 word cross-checked, chart-specific reading, plus its answers, inside the 5 to 21 day turnaround window at a price a reader pays once. Without it, the same reading could not be priced the way it is.

The trade is real: the instrument speeds the mechanical work of surfacing and cross-referencing material, and the practitioner uses the recovered hours on the structural synthesis and the "short answer" callouts that make a reading actually decision-useful. If the trade ever stops working, if AI quality regresses, or if the model's hallucination rate climbs, we'll change it and name it here.