Birth-Time Rectification & Sect Recovery — A User’s Guide¶

This guide is different from the other cookbooks. Rectification in Stellium is essentially one function call — but it comes with more caveats than code, because the honest version of this technique is mostly about knowing what it can’t do. Read the “how not to use it” section; it’s the important one.

TL;DR — the one call¶

from stellium import ChartBuilder, analyze_sect

chart = ChartBuilder.from_notable("Frida Kahlo").calculate()
a = analyze_sect(chart)

print(f"{a.p_day:.2f}  → {a.leans}")   # 0.80  → day

Or drop it straight into a report (renders in every format):

from stellium import ReportBuilder

print(ReportBuilder().from_chart(chart).with_sect_rectification().to_string("markdown"))

That’s the whole API surface. The rest of this document explains what that number means, and — more importantly — what it does not.

What rectification is, and why Stellium refuses to fake it¶

The dream. You have someone’s date and place of birth but not the time. Their Ascendant, houses, Moon, and every timing technique depend on that time. Classical rectification promises to recover it by matching known life events against the chart: find the birth time whose directions/progressions/transits “light up” the right ages.

The reality. We tested this dream as rigorously as we knew how — a 63-person corpus of AA-rated (certificate-verified) birth times with documented life events, pre-registered gates, leave-one-out cross-validation, permutation nulls, confound controls, and an out-of-sample holdout. The full write-up is in RECTIFICATION_REPORT.md. The short version:

The map from birth time → chart → life is many-to-one. Two different birth times can produce lives that fit the events equally well. So the inverse is ill-posed (non-unique) — a textbook ill-posed inverse problem, the same species as deconvolution or CT reconstruction. Minute-level birth time is not recoverable from life events, by any method, because the information isn’t there.

Concretely: individual timing techniques placed the true birth time at only the ~55th–59th percentile of candidates — a whisper above chance — and blind combination of them made it worse, not better (the noise adds faster than the signal). So Stellium deliberately does not ship a chart.rectify() -> a birth time. That would be selling a false precision the evidence says doesn’t exist.

What is recoverable: sect (one bit)¶

One thing survived every test: sect.

Sect is whether you were born by day (Sun above the horizon) or night (Sun below it). It’s the coarsest possible fact about a birth time — a single bit — and the most robust.
Stellium recovers it at ~70% accuracy (leave-one-out cross-validated; 70.6% on a fully out-of-sample modern cohort) from two ingredients:
1. The daylight prior — the fraction of the birth day the Sun is above the horizon (P(day | date, latitude)), pure geometry. Longer day ⇒ more likely born by day.
2. The malefic contrary to sect — a 2,000-year-old doctrine that turned out to leave a real, confound-robust fingerprint: a Mars-flavoured life (accidents, violence, heat) leans day; a Saturn-flavoured one (illness, poverty, slow decline) leans night.

The ceiling is truth resolution, not method. Sect is recoverable, unlike minute-level time — but how well depends on how rich your knowledge of the person is. A one-line biography caps you near 70%. Knowing the person first-hand — their temperament, the texture of their hardships — can push a skilled reader to ~90%. The tool gives you the geometry and the structure; you supply the resolution.

The one call: `analyze_sect`¶

analyze_sect(chart, *, events=None, temperament=None) -> SectAnalysis

chart — any natal chart. Both hypotheses are re-cast from its date + place, so the time you built it with doesn’t matter (noon is fine).
events — a sequence of LifeEvent. For notables in the catalog, they’re looked up automatically by name. Pass () to force a geometry-only read.
temperament — soft character traits; auto-looked-up for notables. Flagged low-confidence (see the guardrails).

What you get back — `SectAnalysis`¶

Field	Meaning
`p_day`	Calibrated probability of a day birth (the headline number).
`leans`, `confidence`	`"day"`/`"night"` and `max(p_day, 1-p_day)`.
`daylight_fraction`	The geometric prior alone (before event evidence).
`day`, `night`	The `Hypothesis` structures each sect implies (sect light, out-of-sect malefic, in-sect benefic, with sign + dignity).
`moon_band`	A warning if the Moon changes sign within the 24 h — i.e. its sign is a band, not a point.
`hardship`, `fortune`	`(Mars, Saturn)` and `(Jupiter, Venus)` event-flavour tallies.
`firdaria`	Firdaria time-lord convergence (`day_hits`/`night_hits`/`favors`).
`malefic_temper`, `sect_light_temper`	Soft temperament signals (see guardrails).
`technique_votes()`	Each technique’s day/night vote, for the convergence tally.

Reading the output¶

For Frida Kahlo, with_sect_rectification() produces (abridged):

Anchor:
  Daylight prior P(day)  0.55   (geometry, the base rate)
  Calibrated  P(day)     0.80   → leans DAY (80% conf)
  ⚠ Moon band: Moon crosses Taurus → Gemini within the 24h — a band, not a point.

                       IF DAY                         IF NIGHT
  Sect light           Sun in Cancer (peregrine)      Moon in Taurus (exalted)
  Out-of-sect malefic  Mars in Capricorn (exalted)    Saturn in Pisces (peregrine)
    → hardship reads   hot, sharp — accidents…        cold, slow — illness…
  In-sect benefic      Jupiter in Cancer (exalted)    Venus in Gemini (peregrine)

  hardship flavour: leans DAY (4.0 / 2.0)     fortune flavour: leans NIGHT (0.6 / 7.6)
  Convergence: DAY 3 · NIGHT 1

Read it top-down: the anchor is the quantitative answer; the day/night table lays out what each hypothesis means so you can check it against what you know; the evidence shows which way the life actually tilts. Note the honesty — Frida’s fortune reads Venus/night (she was an artist) even as her hardship reads Mars/day. A real cross-current, shown rather than smoothed over. Her true birth was by day; the anchor recovers it.

How not to use it — the guardrails¶

This is the part that matters.

p_day is one bit, not a clock. It tells you day vs night, nothing finer. It is not a birth time and cannot be turned into one. If you need a house or an exact Ascendant, this tool can’t give it to you — and neither can anything else, honestly (see the ill-posed section).
It’s an indicator, not an oracle. ~70% on biography-thin data means it’s wrong about one chart in three. Treat a lean as a hypothesis to weigh, never a verdict to act on. Every output says so; believe it.
Don’t feed it the answer. The classifier reads the character of a life, not chart placements that depend on the time you’re trying to find. Keep your inputs (events, temperament) independent of any time you already suspect.
Don’t trust the soft temperament signals on strangers. malefic_temper and sect_light_temper were null on the corpus — a keyword tally can’t tell “hot-tempered” from “the opposite of hot-tempered.” They light up only when you supply real, first-hand knowledge of the person. The output flags them; heed it.
Convergence is counted, never summed. We measured that blindly adding the timing techniques together cancels the signal. The tool shows you how many independent techniques vote each way — it never fuses them into one score, and neither should you.
When the anchor and the timing lenses disagree on sect, trust the anchor. The daylight × malefic anchor is the only cross-validated signal. The timing lenses (directions/transits/profection in the convergence matrix) are whisper-level and routinely point the wrong way. Use them to explore time-within-sect, not to overrule the anchor on sect.
It doesn’t replace your judgment — it structures it. The right mental model: this is a decision-support display that constructs both hypotheses and lays out the evidence, so a knowledgeable human can adjudicate. The one step it will never take for you is the final call. That’s by design.

Supplying your own events¶

For anyone not in the notables catalog, pass events explicitly:

from stellium import ChartBuilder, Native, analyze_sect
from stellium.data import LifeEvent

chart = ChartBuilder.from_native(Native("1974-05-20 12:00:00", "Lima, Peru")).calculate()

events = [
    LifeEvent(date="1998-09", precision="month", type="accident",
              description="serious motorcycle crash, broke leg", significance="major"),
    LifeEvent(date="2005", precision="year", type="career",
              description="promoted, took over the family firm", significance="moderate"),
]
a = analyze_sect(chart, events=events)
print(a.p_day, a.leans)

LifeEvent fields: date ("YYYY", "YYYY-MM", or "YYYY-MM-DD"), precision, type (an event-taxonomy key — accident, health_crisis, bereavement_parent, career, relationship, windfall, …), description (free text; the Mars/Saturn and Jupiter/Venus keyword flavour is read from here), and significance (major/moderate/minor). With no events, pass events=() for a geometry-only read — the daylight prior with no event tempering.

Note on the geometry-only number. With no events, p_day reflects the daylight prior as seen by the trained model, which centres the malefic feature at its corpus mean — so “no evidence” reads slightly day-leaning rather than exactly the raw daylight fraction. That’s expected; the raw geometric prior is in daylight_fraction.

Advanced: the convergence matrix (exploratory)¶

For a fuller picture — and to see why timing can’t pin the time — there’s a heavier two-lens display:

print(ReportBuilder().from_chart(chart).with_sect_convergence_matrix().to_string("markdown"))
# or, as data:
from stellium import convergence_matrix
m = convergence_matrix(chart)

It runs ~3–10 s (it sweeps 96 candidate times with primary directions at each) and shows two lenses over that sweep:

Lens A — the structural band: every genuinely distinct chart across the 24 h (each sect region × Ascendant sign), scored by solar-arc directions, transiting chronocrators to the angles, and annual profection.
Lens B — the event hooks: a histogram of how many events get an apt directed/transiting hit at each candidate time — the times the events themselves “nominate.”

It exists to be read, not obeyed. Its timing rows are whisper-level and it never sums them into a verdict — expect the event-hook histogram to be spread across the whole day (the ill-posed signature), and expect the two lenses to sometimes fight the anchor. When they do, re-read guardrail #6. This is the honest, mechanised form of the “convergence of multiple techniques” that traditional rectifiers rely on — with the quiet truth made visible: convergence displays candidates, it doesn’t prove one.

The data behind it¶

The auto-lookup for notables draws on two curated datasets shipped with Stellium (kept separate from birth data because their provenance is lower):

from stellium import get_notable_life_events, get_notable_temperament

events = get_notable_life_events("Frida Kahlo")     # ~Rodden-B, taxonomy-tagged
traits = get_notable_temperament("Frida Kahlo")     # soft/interpretive (warns on access)

Life events — dated, taxonomy-tagged, gathered from biographies + AstroDataBank via research. Sourced but not certificate-verified — roughly Rodden B grade. Each carries an honest precision and a representative_date helper.
Temperament — soft, interpretive character descriptors, not measurements. The getter emits a DataQualityWarning on access, on purpose.