Spec: Rectification Build — Phase A (time-posterior, sect as marginal)

Status

Draft — the buildable unit for the full posterior

Created

2026-07-11

Owner

Kate

Type

Implementation spec (SDD)

Parent

RECTIFICATION_SPEC.md §3–5 (this is “Phase 3”) · RECTIFICATION_THEORY.md

Precursor

RECTIFICATION_PHASE1_FINDINGS.md — cheap event-character probes topped out at a modest prior (malefic-of-sect: 65%, corr +0.35)

1. Why we’re here

The cheap, near-time-independent probes are exhausted: firdaria timing = null, malefic-of-sect character = a real but modest prior (65%), benefic = null. To do better — and to get time, not just sect — we need the chart itself: sweep candidate birth times and score how well each candidate’s time-dependent timing techniques light the person’s dated events. Birth time is inferred by grid evaluation; sect is read off as the marginal (theory §3–4).

This is the theoretically-grounded core. It is bigger than everything so far.

2. The one thing this phase actually delivers

The pipeline, not any single technique:

candidate grid over the day
  → per candidate: cast chart (full recompute) + evaluate technique activations vs events
  → combine (de-confounded) → un-normalised log-posterior over t
  → normalise → posterior; sect marginal = Σ posterior over day-cells
  → benchmark: sect-marginal accuracy (vs the 65% malefic prior) + time recovery

The first technique to drive it is annual profection — chosen because it is the cheapest time-dependent technique, reuses our significator table, and directly upgrades the failed firdaria experiment (firdaria was ~time-independent; profection’s year-lord depends on the rising sign, so sweeping the time changes the lords and the events can discriminate). The decisive signal is expected from the sharp techniques (directions) that plug into the same pipeline in A2 — but A1 validates the whole machine cheaply.

3. Location & reuse

Standalone (docs/development/specs/rectification/), promote later. Reuses: harness.build_chart (full recompute — profiled at ~7 ms/chart, so a 360-cell × 63-person sweep ≈ a couple of minutes, acceptable), models, significators, sect_signals (the malefic prior, as an optional independent stream), and the benchmark’s Wilson-CI / confusion machinery.

4. The model (concrete)

For a candidate time t on the birth day:

log P(t | events) = log P(t) + Σ_i Σ_j w_j · logL_ij(t)

  • Prior P(t) — uniform over the day for v1 (birth-hour base rates later).

  • L_ij(t) — activation of technique j for event i at chart-cast-t: “how well are event i’s significators lit by technique j around its date.”

  • De-confounding is built in from the start (the firdaria lesson): raw activation is confounded by base rates, so each technique’s contribution is scored as excess over a per-person permutation null (shuffle event type→date), exactly as in sect.py. A technique that lights up for everyone contributes nothing.

  • Sect marginal — P(day) = Σ_{t: sect(t)=day} P(t|events), sect computed per candidate (Phase-0 spec §4.2; never assume the region structure).

5. Milestones

A1 — the pipeline, driven by profection (the proving build)

  • Candidate grid over the day (coarse is fine for profection: the year-lord only changes when the rising sign changes, ~12×/day, so ~20–30 min resolution).

  • Per candidate t, per event i: profected year-lord at age(d_i) = ruler of the sign (rising_sign + whole_years) mod 12; activation = planet_significance of that lord for event i’s type (reuse the table). De-confound vs the permutation null. Combine → posterior over t → sect marginal.

  • Benchmark: sect-marginal accuracy + CI vs (a) chance 50%, (b) majority 54%, (c) the malefic prior 65%. Also report rising-sign recovery (does the posterior mode land near the true ascendant?) as a sanity signal even if sect is weak.

  • Go/continue if the sect marginal is at least on par with the malefic prior and rising-sign recovery beats chance — i.e. the pipeline extracts real time-signal. (Profection is coarse; we do not expect it alone to be decisive.)

A2 — sharp technique: directions to the angles (expected decisive signal)

  • Directed MC/ASC (solar-arc ≈ 1°/yr) or transits reaching natal planets/angles: exact age → minute-level time constraint. Narrow, sharp L_ij(t) bumps.

  • Adds time resolution: report time median-absolute-error, not just sect.

  • The sharp bumps × the coarse profection plateau = the coarse-to-fine behaviour of theory §4.

A3 — combine + calibrate

  • Combine profection + directions (+ optionally the malefic prior as an independent stream) with per-technique weights; calibrate weights + temperature on the corpus (hold out a slice) so the posterior’s credible intervals are honest (theory §7). Escalate combiner complexity only if validation demands (expert weights → logistic → GBT).

6. Data models (extend models.py / a new posterior.py)

@dataclass(frozen=True)
class Candidate:
    minute_of_day: int
    sect: str                       # "day" | "night" (computed)
    log_like: float                 # Σ de-confounded technique activations

@dataclass(frozen=True)
class TimePosterior:
    person: str
    candidates: tuple[Candidate, ...]   # normalised P in .prob
    p_day: float
    p_night: float
    map_minute: int                     # posterior mode
    # (credible interval + per-event/technique contributions added in A2/A3)

7. Testing

  • Determinism — same grid + seed → same posterior.

  • Sanity — feeding a known chart’s events makes the posterior mass fall in the correct sect region and near the true ascendant more often than chance.

  • Marginal correctness — p_day + p_night == 1; p_day = summed day-cells.

  • Standalone tier, not the package suite.

8. Build order

  1. posterior.py — grid + prior + normalise + sect-marginal (technique-agnostic).

  2. Profection activation (A1) + wire into the posterior.

  3. run_posterior_benchmark.py — sect-marginal + rising-sign recovery vs baselines.

  4. Read the number. Pipeline works? → A2 directions. Weak? → the pipeline is still the reusable foundation; go straight to the sharp technique.

  5. A2 directions → time MAE. 6. A3 combine + calibrate.

9. Open questions

  • Q-A1 — grid resolution: coarse (rising-sign, ~24 cells) for profection; fine (~4 min, 360 cells) once sharp techniques enter. Lean: adaptive per technique.

  • Q-A2 — which sharp technique first: solar-arc directions to angles vs transits-to-angles vs ZR-from-Fortune peaks. Decide from the API scout + which carries signal. (ZR-from-Fortune is attractive: it flips by sect and is event-timed.)

  • Q-A3 — is the malefic prior folded in as an independent evidence stream, or kept separate as a reported cross-check? Lean: fold in at A3, weighted.

  • Q-A4 — combiner: start hand-weighted; calibrate on the corpus with a holdout to avoid overfitting n=63.

10. Acceptance criteria

  • posterior.py: grid → posterior → sect marginal, technique-agnostic, tested.

  • Profection activation wired in; benchmark reports sect-marginal + CI + rising-sign recovery vs chance / majority / the malefic prior.

  • A written call: does the time-posterior beat the cheap prior?

  • (A2) sharp technique → time median-abs-error reported.

  • (A3) calibrated posterior; credible-interval coverage ≈ nominal on a holdout.