Phase 1 findings — sect classifier v0 (firdaria × significators)

Date

2026-07-11

Verdict

NO-GO — firdaria-lord matching does not recover sect

Code

docs/development/specs/rectification/ (sect.py, run_benchmark.py)

Spec

RECTIFICATION_PHASE0_SPEC.md §5–6

The experiment

Blank each of the 63 corpus people’s birth time; predict day vs night from their dated events alone, by matching the active firdaria time-lord (day-order vs night-order) to each event’s natural planetary significators (theory §6). This is the pre-registered proving experiment: does the approach beat chance?

Result

Variant

Accuracy

Notes

Raw day-vs-night sums

49.2%

strong night bias (predicted night 43/63)

Permutation-null de-confounded

41.3%

bias gone, balanced confusion, still chance

Best achievable (any threshold, any ablation)

54.0%

= majority class (“always day”)

  • 95% CI (de-confounded) = [30%, 54%] — straddles 50%.

  • corr(continuous signal, true sect) = −0.17 to −0.22 across all variants — slightly negative, ≈ 1.7 SE from zero at n=63, i.e. indistinguishable from no signal.

  • Ablations (major-lord-only, major-events-only, ± sub-lord) all land at the majority-class 54%. The null is robust.

Pre-registered gate (acc ≥ 65% AND CI-low > majority): NO-GO.

What we learned (the diagnosis)

  1. The raw comparison is confounded. The night firdaria order front-loads Saturn/Jupiter/Mars in the event-dense early decades, and those are the generically-significant planets (Saturn signifies 8/15 event types). So the night ordering scores higher for anyone, regardless of true sect — a nuisance factor, not sect signal. A per-person permutation null removes it cleanly.

  2. De-confounded, there is simply no signal. Firdaria periods are long (7–13 yr), so a documented life samples only ~5–7 distinct major lords — very few “measurements” of a type↔lord association that is itself weak. Firdaria lacks the resolution to fix sect.

  3. This is a clean, cheap result: the shortcut of reading sect off firdaria alone — appealing because it’s near time-independent — does not work. Better to know before building a posterior on it.

Options from here (a genuine fork)

  • A — Sharp techniques, not coarse firdaria. The theory’s real claim is sect as the marginal of the full event-posterior. Firdaria was a hopeful shortcut; the signal may live in the sharp, angle-tied techniques (directions/transits to the angles, ZR peaks, profected-year lords) that the shortcut skipped. This is bigger (needs the candidate grid + more engines) but is the theoretically grounded path.

  • B — The sect-dependent dignity/lots channel (spec §5.1 fallback): benefic/ malefic of the sect, the sect light, and the Lot of Fortune/Spirit formula flip. Fuzzier and more temperament-linked (leans on the soft channel we deferred).

  • C — Reconsider whether events-only can fix sect at all, vs. needing the temperament/chart-coherence signal a human rectifier actually uses.

The machinery (loader, harness, significators, benchmark, contrastive report) is built, tested, and reusable for whichever direction we take.


Phase 1B — malefic-contrary-to-sect probe (option B) — SIGNAL

probe_malefic_sect.py. Doctrine: the malefic out of sect is the sharper destroyer — Mars out of sect by day, Saturn out of sect by night — so a life’s misfortunes should carry the flavour of its contrary-sect malefic. Prediction: Mars-flavoured hardship → day, Saturn-flavoured → night. Nearly chart-free (reads the character of the misfortunes via a priori Mars/Saturn keyword lists), and a completely different mechanism from firdaria.

metric

value

accuracy

65.1% (41/63)

decided-only (drop 14 keyword-silent ties)

65.3% (32/49)

corr(mars−saturn, day)

+0.346 — ≈ 2.9 SE from 0, p ≈ 0.005

95% CI

[52.8%, 75.7%] (lower bound just under the 54% majority)

confusion

both diagonals dominant (day 22/12, night 10/19)

Pre-registered gate: still NO-GO (CI-low 52.8% < majority 54%) — but this is a genuine, directionally-correct effect, not the firdaria null. The extremes are convincing: strongest-Saturn lives (Ted Bundy, Hannah Arendt, Ali, Obama) classify night; strongest-Mars (Hemingway, Frida Kahlo, JFK, van Gogh, Plath) classify day.

Interpretation: events do carry sect information — via the malefic-of-sect character of misfortune, not via the firdaria time-lord sequence. Modest (65%), needs development, and the keyword lists carry some researcher DOF (set a priori, but worth a holdout).

Phase 1C — benefic-of-sect mirror + combination — benefic is NULL

sect_signals.py, probe_sect_combined.py. Symmetric hypothesis: Jupiter is benefic of sect by day, Venus by night → Jupiter-flavoured fortune → day, Venus-flavoured → night.

signal

accuracy

corr(score, day)

malefic-only

65.1%

+0.346

benefic-only

52.4%

+0.021 (null)

combined

52.4%

+0.169 (diluted)

The benefic mirror carries no signal, and combining it dilutes the malefic (+0.35 → +0.17). Diagnosis: fortune-character is domain-linked, not sect-linked — artists accumulate Venus-flavoured fortune (art/exhibitions), scientists and politicians Jupiter-flavoured (awards/office), independent of sect — so the benefic channel injects profession noise. Misfortune-character (sudden/violent vs slow/chronic) is far more universal, which is why the malefic side works and the benefic doesn’t.

Conclusion: keep malefic-only (corr +0.346, p ≈ 0.005) as a real but modest sect prior; drop the benefic. The cheap event-character approach tops out ≈ 65% — useful as one evidence stream, not a decisive classifier.

Where things stand (post-B)

  • Two mechanisms tried from events alone: firdaria timing = null; malefic-of-sect character = real but modest (65%, p≈0.005, doesn’t clear the strict gate); benefic-of-sect = null (domain confound).

  • Events do carry sect signal, but weakly and only through misfortune character. A decisive sect read likely needs the chart itself (option A: the sharp, angle-tied techniques → full posterior, sect as the marginal), which is the larger build we deferred.


Phase A — time-posterior + the daylight prior — first GO

posterior.py, profection.py, sect_classifier.py, run_sect_benchmark.py.

Built the technique-agnostic pipeline (candidate grid → per-candidate likelihood → posterior → sect marginal) and drove it with annual profection (year-lord depends on the rising sign → time-dependent, unlike firdaria). Then the control that reframed everything.

Profection is null (and harmful). Rising-sign recovery is at chance (exact 9.5% vs 8.3%; mean sign-distance 2.95 vs 3.0), and the profection posterior’s sect marginal (65%) is worse than a uniform posterior. Coarse lord-matching carries no time signal — consistent with the firdaria null.

The control that mattered — the daylight prior. A uniform posterior (zero technique signal) sets p_day = the daylight fraction of the birth day (longer day ⇒ more likely born by day = P(day | date, lat)). Alone it scores 68.3%, corr +0.40 — beating every event-based signal. The strongest sect predictor is a free geometric prior that uses no events at all.

But events add real independent value. The malefic-of-sect signal is only mildly correlated with the daylight prior (+0.19), and its partial correlation with sect controlling for daylight is +0.30 — genuinely independent evidence. Combining them (informed prior × event likelihood — the theory’s exact structure) via a 2-feature logistic:

model

accuracy

majority baseline

54.0%

malefic-alone

65.1%

daylight-alone

68.3%

daylight + malefic (LOO-CV)

69.8%, CI [57.6%, 79.8%]

daylight + malefic (in-sample)

73.0%

Pre-registered gate (LOO acc ≥ 65% AND CI-low > majority): GO. Cross-validated, so not overfit; fitted weights daylight +0.80 / malefic +0.72 (both real).

Takeaways:

  1. A calibrated sect classifier at ~70% (LOO) exists — daylight prior + the malefic-of-sect event evidence. First result to clear the bar.

  2. Most of the strength is the free geometric prior; the event evidence adds a real but modest independent boost. This is honest and defensible, not a dramatic rectification win.

  3. Profection (coarse timing) is null — the remaining upside for time (not just sect) is the sharp angle-tied techniques (directions to angles, ZR-from-Fortune peaks), still untested. Those are the real rectification signal if it exists; the pipeline (posterior.py) is built and ready to host them.


Phase A2 — primary directions (the sharp technique) — null for time

directions.py, run_directions_benchmark.py. For each candidate time, direct the seven planets to the four angles (DirectionsEngine) — hits land at ages that move strongly with birth time — and score whether directed hits whose promissor signifies an event fall near that event’s age (de-confounded).

metric

directions

chance

time median |Δ| (MAP vs truth)

406 min

~360 min

mean posterior mass ±90 min of truth

0.138

0.125

mean posterior mass ±180 min of truth

0.258

0.25

sect marginal

68.3%

(= daylight leaking through)

Directions does not localize time — mass-near-truth is at chance and median error is worse than chance. Best-case retest (day-precision events only, tight 0.5-yr orb): median |Δ| 320 min, mass ±90 min 0.132 — still chance. The sect 68% is just the daylight fraction showing through a near-uniform posterior, redundant with the classifier.

Verdict on the timing-technique family (firdaria + profection + directions)

All three are null. Automated rectification by “sweep the birth time, match events to timing-technique activations via a significator table” does not carry signal — coarse or sharp. Expert rectification works because a human selects meaningful event↔direction correspondences with judgment; the blind, all-events, significator-table version loses exactly that.

What actually works, and all that works:

  1. the daylight-fraction geometric prior (68%, no events), and

  2. the malefic-of-sect event character (adds independent value → 70% LOO).

Neither is a timing technique. Minute-level time rectification is not achievable with this automated approach; the deliverable is the ~70% sect classifier (to be strengthened by an external birth-hour prior — research pending). This is the central, rigorously-established result of the investigation.


Temperament / personality — null for sect (probe_temperament_sect.py)

Tested the soft channel now that full charts are available. Sect-light doctrine: Sun-led (Solar) character → day, Moon-led (Lunar) → night; Solar/Lunar keywords a priori.

value

corr(temperament, sect)

+0.03 (null)

corr(temperament, malefic)

−0.03 (independent, but of noise)

partial corr(temperament, sect | daylight, malefic)

+0.11 (n.s. at n=63)

LOO: daylight+malefic

69.8%

LOO: daylight+malefic+temperament

68.3% (no gain — slightly worse)

Temperament is null and adds nothing (it slightly hurts LOO). This also settles the chart-contextualised version without building it: “does temperament match the chart’s sect-role assignment” decomposes into the sect light (Solar/Lunar — just tested null) and the benefic of sect (Jovial/Venusian — = the null benefic-events probe). Both building blocks are null.

Why temperament fails where malefic-of-sect works: the malefic signal rides a specific, clean axis — the character of misfortune (violent/Mars vs chronic/Saturn) maps directly onto the out-of-sect-malefic doctrine. General temperament is swamped by the far larger non-sect determinants (Sun sign, Moon sign, dominant planet, Ascendant); the sect overlay is too faint to detect against that. Fittingly, sect shows up in the character of harm, not in general personality — consistent with the tradition emphasising sect most for the malefics’ operation.


Natal dignities / placements — null for sect (probe_dignity_sect.py)

Tested planetary sign-dignities two ways: (1) a diurnal-vs-nocturnal dignity balance (new channel), and (2) dignity-weighting the malefic (enrich what works — scale hardship by the natal condition of Mars vs Saturn).

model

LOO-CV

daylight + malefic

69.8%

daylight + malefic + dignity-balance

68.3% (worse)

daylight + malefic(dignity-weighted)

65.1% (worse)

Both fail. The dignity-balance’s in-sample partial corr looks sizable (−0.32) but does not survive LOO — textbook noise-as-signal, and a reminder of why every signal is cross-validated. Expected: sect is a horizon fact, ~orthogonal to which signs the planets occupy. Dignity-weighting the malefic degrades it — natal malefic condition is noise for the misfortune-character signal.

Signals tested — the full map

channel

signal

result

geometry

daylight fraction (prior)

+0.40 — works (68%)

event character

malefic-of-sect (misfortune flavour)

+0.35 — works, independent (→70%)

event character

benefic-of-sect (fortune flavour)

null (domain confound)

timing

firdaria time-lord × significators

null

timing

annual profection (rising-sign lords)

null

timing

primary directions to angles

null (sect and time)

temperament

sect-light (Solar/Lunar)

null

natal dignity

diurnal/nocturnal balance; dignity-weighted malefic

null

Bottom line. Feature space for an automated, corpus-scale rectifier is thoroughly explored. Only daylight prior × malefic-of-sect survives (LOO 69.8%, GO). Minute-level time is null across every timing technique. Further feature-hunting on n=63 has negative expected value (multiple-comparisons risk — the dignity −0.32 that vanished under LOO is the warning).


External birth-hour prior — no improvement (historical cohort) (birth_hours.py)

Replaced the uniform-birth-time assumption with an external, cited hourly distribution (birth_times_research.md; population vital-stats, no chart samples). Distribution A spontaneous_prewar (nocturnal, peak 04:00) as primary, an era-aware blend toward the modern daytime curve for later births.

prior (+ malefic)

LOO-CV

uniform daylight

69.8%

spontaneous_prewar

66.7%

era-blend

68.3%

corr(prior, sect): uniform 0.398, spontaneous 0.408, era-blend 0.399. The birth-hour prior does not improve sect prediction — equal-or-worse in LOO. Reason: the daylight fraction already captures the date/latitude effect; the hourly curve is broad (peak:trough ~1.9:1) and applied identically to all, so it shifts P(day) ~monotonically (absorbed by the logistic) rather than adding discrimination.

Modern cohort test (resolves the caveat). Added 20 post-1970 AA/A births (rectification-modern-cohort.yaml) to test the medicalized daytime curve on the era it’s meant for. It still doesn’t help:

prior

modern AA (n=17) acc

corr

uniform daylight

70.6%

+0.251

modern daytime

64.7%

+0.245

spontaneous

47.1%

+0.245

The modern prior shifts every P(day) up ~0.08 (a near-uniform monotonic shift), so the correlation barely moves (0.245–0.251) — it only slides the 0.5 threshold and mis-flips borderline night births to day. The lesson: a better prior (marginal P(day)) is not a better classifier (discrimination). The birth-hour curve improves calibration of the average but adds no ranking signal; the daylight fraction already owns the person-to-person variation (date + latitude).

Bonus — out-of-sample validation. These 20 were never in the 63, yet the daylight prior alone scores 70.6% on the 17 AA, matching its ~68% on the historical set. The classifier generalises cleanly across eras. The birth-hour lever is now definitively closed.


Selection-bias / confound check on the malefic signal — robust

The corpus is famous + AA-recorded (Western, 19th–20th c., dramatic documented lives). The one positive event signal (malefic-of-sect, +0.35) is the finding most vulnerable to that. Tested the two concrete confounds — profession and sex:

malefic-sect correlation, controlling for

value

nothing (raw)

+0.346

gender

+0.349 (gender explains 0% of sect variance)

category

+0.355

category + gender

+0.361 (both explain 5% of sect variance)

Completely stable (strengthens slightly). Women vs men are day-born at 52% vs 55% — sect is birth time of day, mechanistically independent of who you are, so neither profession nor sex can confound it. The “criminals are violent and happened to be day-born” artifact is ruled out.

What remains (untestable): biographical-emphasis bias (biographers front-page violent events, under-report chronic decline) and whether the signal generalises beyond famous, well-documented lives. These can’t be tested without a non-famous documented-event corpus, which is structurally impossible — validating rectification requires known-time + documented events = famous + AA + Western + recent. That is an inherent ceiling on the whole enterprise, not a fixable flaw. The ~70% classifier is validated on this population, not claimed for humanity.