Skip to content

ASPECT_BACKEND_STANDARD

TheDaniel166 edited this page Jul 26, 2026 · 6 revisions

Moira Aspect Backend Standard

Governing Principle

The Moira aspect backend is a sovereign computational subsystem. Its definitions, layer boundaries, invariants, failure doctrine, and determinism rules are stated here and are frozen until explicitly superseded by a revision to this document.

This document reflects current implementation truth. Focused verification counts belong in completion receipts rather than frozen doctrine.


Part I — Architecture Standard

1. Authoritative Computational Definitions

1.1 Ecliptic aspect

An ecliptic aspect in Moira is:

A detected angular relationship between two distinct celestial bodies whose angular separation along the ecliptic falls within a declared orb of a canonical aspect angle.

Element Definition
two distinct bodies body1 != body2; no self-aspects
angular separation angular_distance(lon1, lon2) → [0, 180] degrees, folded at 180°
canonical aspect angle An angle from moira.constants.Aspect.ALL
orb abs(separation - angle)
within declared orb orb <= allowed_orb
allowed_orb default_orb * orb_factor, or the caller-supplied override for that angle

The admission test is fully reconstructable from the stored vessel:

orb        == abs(separation - angle)
orb        <= allowed_orb
separation  = angular_distance(lon1, lon2)

1.2 Declination aspect

A declination aspect in Moira is:

A parallel or contra-parallel between two distinct celestial bodies whose signed declinations satisfy one of the two defined relationships within a declared orb.

Type Admission test Orb formula
Parallel Same nonzero hemisphere and abs(dec1 - dec2) <= allowed_orb orb = abs(dec1 - dec2)
Contra-Parallel Opposite nonzero hemispheres and abs(dec1 + dec2) <= allowed_orb orb = abs(dec1 + dec2)

Two points exactly on the equator form one exact Parallel. A single equatorial point has no hemisphere and forms neither relation with a non-equatorial point. Hemisphere qualification precedes the orb formula so both relations cannot be emitted for one pair near 0°.

Detected DeclinationAspect vessels carry no motion data; consequently the legacy generic aspect_motion_state(DeclinationAspect) remains NONE. The separate DeclinationAspectMotionWitness is the first-class temporal product when declination rates are supplied.

1.2.1 Module ownership

moira.declination_aspects governs parallels, contra-parallels, their hemisphere/equator doctrine, policy, result vessels, and signed motion witness. moira.aspects governs ecliptic-longitude and whole-sign relationships and retains compatibility re-exports for the historical declination imports. Shared domain/tier/family classification types live in the private moira._aspect_types taxonomy layer so both engines use identical enum and classification objects without circular imports.

Declination is a second angular coordinate on the celestial sphere. It is not full Cartesian 3D geometry because radial distance is not part of these relationships.

1.2.2 Whole-sign direction and overcoming

moira.aspects also owns Hellenistic aspect direction and whole-sign overcoming. hellenistic_superiority_truth() is the governing atomic receipt for one ordered body pair: HellenisticDirectionTruth preserves directed-arc geometry and sinister/dexter applicability, while HellenisticOvercomingTruth preserves both inclusive sign places and the typed winner-or-neither relation.

Direction is not_evaluable when no aspect angle is supplied, for conjunction or opposition, or at an exact directed conjunction/opposition boundary. The reason is preserved and an evaluable direction cannot be suppressed. Overcoming remains evaluated for finite pair geometry. Sign counting is inclusive: the first body overcomes the second only when the first occupies the tenth sign from the second, which is nine zero-based sign steps forward. Reversing the operands reverses the question. A third-sign/fourth-sign confusion must not be accepted merely because the bodies form an undirected square.

AspectData.direction and overcoming(lon1, lon2) are compatibility projections from that shared raw truth. Detector-produced AspectData also carries the aggregate hellenistic_superiority_truth receipt. No additive dignity or superiority score is admitted by this composition layer.

1.3 Admitted aspect

An aspect is admitted when the admission test passes. Admission is binary: the aspect either qualifies or it does not. There is no partial admission and no confidence score at the detection layer.


2. Layer Structure

The backend is organized into twelve phases. Each phase operates only on outputs produced by phases below it. No phase reaches upward.

Phase  1 — Core aspect detection
Phase  2 — Relational truth preservation
Phase  3 — Classification
Phase  4 — Inspectability
Phase  5 — Doctrine inputs
Phase  6 — Policy surface
Phase  7 — Geometric strength
Phase  8 — Temporal state
Phase  9 — Canonical configuration
Phase 10 — Multi-body pattern layer
Phase 11 — Relational graph / network layer
Phase 12 — Harmonic / family intelligence layer

Layer boundary rule

A function in phase N may consume results from phases 1 through N−1. It may not:

  • re-run position arithmetic
  • re-compute aspect admission
  • alter a vessel produced by an earlier phase in place
  • introduce new doctrine inputs not present in that phase's entry point

3. Delegated Assumptions

The aspect engine delegates to external modules without redefining them:

Concern Delegated to Convention
Angular distance arithmetic moira.coordinates.angular_distance Returns [0, 180], fold at 180°
Canonical aspect definitions moira.constants.Aspect.ALL 22 zodiacal aspects
Default orb table moira.constants.DEFAULT_ORBS {angle: max_orb}
Aspect tier lists moira.constants.ASPECT_TIERS Major / Common Minor / Extended Minor

The aspect backend does not redefine any of these. Changes to these constants propagate automatically.


4. Canonical Aspect Set

The complete set of recognised and detectable aspect types is declared in CANONICAL_ASPECTS — a module-level tuple of 24 names, frozen at import time.

Tier Count Names
Major 5 Conjunction, Sextile, Square, Trine, Opposition
Common Minor 6 Semisextile, Semisquare, Sesquiquadrate, Quincunx, Quintile, Biquintile
Extended Minor 11 Septile, Biseptile, Triseptile, Novile, Binovile, Quadnovile, Decile, Tredecile, Undecile, Quindecile, Vigintile
Declination 2 Parallel, Contra-Parallel

Rules:

  • The 22 zodiacal names correspond 1-to-1 with entries in Aspect.ALL.
  • The 2 declination names are produced exclusively by find_declination_aspects.
  • CANONICAL_ASPECTS carries no detection logic; it is a declaration only.
  • No aspect not in CANONICAL_ASPECTS can be produced by any detection function.

5. Classification

AspectClassification classifies every admitted aspect on three independent axes:

Axis Type Rule
domain AspectDomain ZODIACAL for ecliptic; DECLINATION for parallels
tier AspectTier Derived from AspectDefinition.is_major and membership in Aspect.EXTENDED_MINOR
family AspectFamily Derived from _FAMILY_BY_NAME; maps each aspect name to its harmonic family

Classification is descriptive only. It describes what was detected, not how it should be interpreted. Strength, dignity weighting, and reception scoring are excluded from the classification layer.

Family grouping

Aspects in the same harmonic series share a family:

Family Members
CONJUNCTION Conjunction
OPPOSITION Opposition
SQUARE Square
TRINE Trine
SEXTILE Sextile
SEMISEXTILE Semisextile
SEMISQUARE Semisquare
SESQUIQUADRATE Sesquiquadrate
QUINCUNX Quincunx
QUINTILE Quintile, Biquintile
SEPTILE Septile, Biseptile, Triseptile
NOVILE Novile, Binovile, Quadnovile
DECILE Decile, Tredecile
UNDECILE Undecile
QUINDECILE Quindecile
VIGINTILE Vigintile
DECLINATION Parallel, Contra-Parallel

6. Policy Layer

AspectPolicy encapsulates longitude-aspect doctrine and retains declination_orb as a compatibility adapter for historical callers.

Field Type Default Effect
tier int | None None 0=Major only, 1=Major+Common Minor, 2=All; None defers to include_minor
include_minor bool True Include Common Minor when tier is None
orbs dict[float, float] | None None Custom orb table {angle: max_orb}; overrides orb_factor when set
orb_factor float 1.0 Multiplier on all default orbs; ignored when orbs is set
declination_orb float 1.0 Ceiling for Parallel and Contra-Parallel detection

When a policy argument is passed to a detection function it takes full precedence over any corresponding individual keyword arguments. Individual parameters remain available for backward compatibility.

DEFAULT_POLICY reproduces the historical default behaviour of all four detection functions.

DeclinationAspectPolicy is the governing declination policy:

Field Default Effect
orb 1.0 Admission ceiling for both relationships
exact_tolerance_deg 1e-9 Exactness precedence threshold
rate_tolerance_deg_per_day 1e-12 Relative signed-error rate treated as stalled
hemisphere_policy STRICT_SIGNED Parallel uses the same nonzero hemisphere; Contra-Parallel uses opposite nonzero hemispheres
equator_policy PAIRED_EQUATORIAL_PARALLEL Two equatorial points form one Parallel; one equatorial point is unclassified

Policy validation

AspectPolicy validates its fields at construction:

Condition Raises
orb_factor <= 0 ValueError
declination_orb < 0 ValueError

7. Geometric Strength

AspectStrength is a pure arithmetic exactness summary derived from orb and allowed_orb only.

surplus   = allowed_orb - orb
exactness = 1.0 - orb / allowed_orb
Field Definition
orb Angular deviation from target angle; always non-negative
allowed_orb Orb ceiling applied at admission
surplus allowed_orb - orb; remaining headroom
exactness 1.0 - orb / allowed_orb; 1.0 = exact, 0.0 = at boundary

aspect_strength does not interpret strength. It does not weight by aspect family, body dignity, or orbital speed.

Strength validation

aspect_strength validates its input before computing:

Condition Raises
orb or allowed_orb is non-finite, or orb < 0 ValueError
allowed_orb <= 0 ValueError
orb > allowed_orb ValueError

Whole-sign aspects are categorical rather than orb-admitted. Their strength identity is orb = allowed_orb = surplus = 0 and exactness = 1.


8. Temporal-State Doctrine

MotionState formalises the motion-aware truth already stored in applying and stationary. It maps the complete decision space without ambiguity:

Vessel/domain orb stationary applying → MotionState
DeclinationAspect any — — NONE
Whole-sign AspectData any — — NONE
Zodiacal AspectData <= 1e-9 any any EXACT
Zodiacal AspectData > 1e-9 True any STATIONARY
Zodiacal AspectData > 1e-9 False True APPLYING
Zodiacal AspectData > 1e-9 False False SEPARATING
Zodiacal AspectData > 1e-9 False None INDETERMINATE

Categorical/no-motion domains yield NONE. For zodiacal aspects, EXACT takes precedence over stationary and applying/separating classification.

Temporal consistency rules

  • APPLYING ↔ is_applying is True and is_separating is False
  • SEPARATING ↔ is_separating is True and is_applying is False
  • is_applying and is_separating are never simultaneously True
  • Both are False when applying is None

First-class declination motion

DeclinationAspectMotionWitness classifies a selected relationship from caller-supplied declinations and optional declination rates, all expressed in the same declared frame, timescale, origin, and correction regime.

Relationship Signed error E Relative rate dE/dt
Parallel dec1 - dec2 speed1 - speed2
Contra-Parallel dec1 + dec2 speed1 + speed2

Away from exactness, orb = abs(E) and orb_rate = sign(E) * dE/dt. A negative orb rate is APPLYING; a positive orb rate is SEPARATING. Exactness takes precedence. A relative rate inside the policy tolerance is STATIONARY, while missing or partial rates are INDETERMINATE. An individual zero declination rate does not force STATIONARY when the relationship's relative error is still changing.

The witness is instantaneous (instantaneous_no_event_search). It does not prove that a relationship will perfect before either body reverses in declination; that stronger claim requires an epoch-based exact-root search.


9. Multi-Body Pattern Doctrine

find_patterns operates over an already-admitted list[AspectData]. It does not re-run position arithmetic.

Implemented patterns and their structural requirements:

Kind Bodies Required edges
STELLIUM ≥3, maximal clique Mutual Conjunction between every pair
T_SQUARE exactly 3 One Opposition (A–B) + Square(A–C) + Square(B–C)
GRAND_TRINE exactly 3 Trine(A–B) + Trine(B–C) + Trine(A–C)
GRAND_CROSS exactly 4 Two Oppositions + four Squares (closed cross)
YOD exactly 3 Sextile(B–C) + Quincunx(A–B) + Quincunx(A–C)

Pattern ordering rules

  • Output order: Stellia, T-Squares, Grand Trines, Grand Crosses, Yods.
  • Each pattern kind is emitted at most once per unique body set (frozenset).
  • Stellium: smaller subsets contained within a larger Stellium are suppressed.
  • All other kinds are independent. A Grand Cross may also contain T-Squares; both are reported.
  • Within each kind, patterns are ordered by sorted body-name iteration (outer loops always iterate over sorted(all_bodies)).

Pattern contributing-aspects ordering

The aspects tuple inside each AspectPattern is sorted by (body1, body2, aspect). This ordering is stable and independent of the input list ordering.

Structural aspect counts

Kind len(aspects)
STELLIUM (3-body) 3
STELLIUM (4-body) 6
T_SQUARE 3
GRAND_TRINE 3
GRAND_CROSS 6
YOD 3

10. Relational Graph Doctrine

build_aspect_graph expresses the chart as a deterministic aspect network. Bodies become nodes; each admitted AspectData becomes an edge.

Graph construction rules

  • Every body that appears in at least one aspect gets a node.
  • Bodies supplied via bodies= that have no aspects get degree-0 nodes.
  • nodes is sorted by body name.
  • edges is sorted by (body1, body2, aspect).
  • components is sorted by (min(component), len(component)) ascending.

Node invariants

Invariant Expression
Degree consistency degree == len(edges)
Family count consistency sum(family_counts.values()) == degree
Incidence Every edge in node.edges has body1 == name or body2 == name
Edge ordering edges sorted by (body1, body2, aspect)

family_counts keys are AspectData.aspect strings (e.g. "Trine"), not AspectFamily enum values. This is intentional: it preserves per-name granularity at the graph layer (Trine vs Biquintile both count as QUINTILE family at the harmonic layer, but remain distinct at the graph layer).

Derived properties

Property Definition
hubs Nodes with maximum degree; empty tuple when all nodes are isolated
isolated Nodes with degree 0, sorted by name

11. Harmonic Intelligence Doctrine

aspect_harmonic_profile derives the family distribution of admitted aspects at both the chart level and per body.

Profile construction rules

  • chart covers all aspects in the input list.
  • by_body has one entry per body that appears in at least one aspect.
  • by_body keys are in sorted body-name order.
  • A body with no aspects has no entry in by_body.

Family resolution order (when classification is absent)

  1. a.classification.family when classification is not None (normal case).
  2. _FAMILY_BY_NAME[a.aspect] when classification is None and the name is a known zodiacal name.
  3. AspectFamily.DECLINATION as the fallback for any unrecognised name (covers "Parallel", "Contra-Parallel", or custom names).

AspectFamilyProfile invariants

Invariant Expression
Total sum(counts.values()) == total
Proportions count len(proportions) == len(counts)
Proportions sum abs(sum(proportions.values()) - 1.0) < 1e-9 when total > 0
Dominant membership Every member of dominant is a key in counts
Proportion range All proportions in [0.0, 1.0]
Key ordering counts and proportions keys follow AspectFamily declaration order
Dominant ordering dominant sorted by AspectFamily.value (alphabetical)

12. Non-Goals

The following concerns are explicitly outside the scope of the current aspect backend:

Excluded concern Reason
Interpretation (e.g. "this aspect is challenging") Doctrine-specific; belongs above the engine
Dignity weighting or reception scoring Requires a separate dignity model
Body-specific orb weights Not in current AspectPolicy
Antiscion contacts A separate geometric computation
Cross-chart (synastry) relational policies Multi-chart context not yet in scope
Kite, Mystic Rectangle, Grand Quintile Require oriented topology or 5-body matching
UI rendering or serialization Belongs above the engine
Harmonic chart generation Separate from aspect detection

Part II — Validation Codex

1. Validation Environment

Property Value
Authoritative runtime .venv in the project root
Python version 3.14.x (as resolved by .venv)
Test runner pytest via .venv\Scripts\python.exe -m pytest
Test file tests/unit/test_aspects.py
Baseline Focused aspect test file passes under strict known-issues mode
Acceptable result 0 failures, 0 errors

No test in test_aspects.py may be modified to make the implementation pass. A failing test is always treated as an implementation defect, not a test defect, unless the test itself is proven incorrect.


2. Invariant Register

This register is the normative source of truth for all subsystem invariants. Each invariant is identified by a short code for traceable reference.

INV-TRUTH — Truth preservation

Code Invariant
T-1 Zodiacal orb == abs(separation - angle) to floating-point precision
T-2 Zodiacal orb <= allowed_orb; whole-sign orb == allowed_orb == 0
T-3 orb_surplus == allowed_orb - orb >= 0
T-4 separation is in [0, 180] degrees
T-5 For a Parallel: orb == abs(dec1 - dec2)
T-6 For a Contra-Parallel: orb == abs(dec1 + dec2)
T-7 dec1 and dec2 are in [-90, +90]
T-8 Parallel and Contra-Parallel hemisphere predicates are mutually exclusive
T-9 overcoming(lon1, lon2) is true exactly when (sign1 - sign2) mod 12 == 9
T-10 HellenisticDirectionTruth is evaluated only for a supplied non-conjunction/non-opposition angle away from directed 0°/180° boundaries
T-11 HellenisticOvercomingTruth.relation exactly matches the two inclusive sign places and admits an evaluated neither result
T-12 Detector-produced AspectData.direction is the compatibility projection of its hellenistic_superiority_truth.direction_truth

INV-CLASS — Classification

Code Invariant
C-1 classification.domain is ZODIACAL or WHOLE_SIGN for detector-produced AspectData
C-2 classification.domain == DECLINATION for every DeclinationAspect
C-3 classification.family == _FAMILY_BY_NAME[aspect] for every zodiacal aspect
C-4 classification.family == AspectFamily.DECLINATION for every DeclinationAspect
C-5 classification.tier == MAJOR for every aspect in {"Conjunction","Sextile","Square","Trine","Opposition"}
C-6 Classification is identical for the same aspect name across all calls

INV-STR — Geometric strength

Code Invariant
S-1 0.0 <= orb <= allowed_orb for any orb-admitted vessel passed to aspect_strength
S-2 surplus == allowed_orb - orb
S-3 0.0 <= exactness <= 1.0
S-4 exactness == 1.0 - orb / allowed_orb
S-5 aspect_strength raises ValueError when allowed_orb <= 0
S-6 aspect_strength raises ValueError when orb > allowed_orb
S-7 Whole-sign strength is (orb=0, allowed_orb=0, surplus=0, exactness=1)

INV-MOT — Temporal state

Code Invariant
M-1 APPLYING ↔ is_applying is True and is_separating is False
M-2 SEPARATING ↔ is_separating is True and is_applying is False
M-3 Zodiacal exactness yields EXACT before stationary/applying classification
M-4 Non-exact zodiacal data yields STATIONARY when stationary is True
M-5 A detection-only DeclinationAspect and whole-sign data yield generic MotionState.NONE
M-6 is_applying and is_separating are never simultaneously True
M-7 Non-exact zodiacal data yields INDETERMINATE when motion is unavailable
M-8 Parallel signed error/rate are dec1 - dec2 and speed1 - speed2
M-9 Contra-Parallel signed error/rate are dec1 + dec2 and speed1 + speed2
M-10 Non-exact declination orb_rate < 0 yields APPLYING; orb_rate > 0 yields SEPARATING
M-11 Declination exactness precedes rate classification
M-12 Missing declination rates yield INDETERMINATE unless exact
M-13 Declination STATIONARY means the relative error rate is inside tolerance, not that either individual body has zero rate

INV-PAT — Pattern layer

Code Invariant
P-1 Every body in pattern.bodies appears in at least one aspect in pattern.aspects
P-2 pattern.aspects is sorted by (body1, body2, aspect)
P-3 No pattern body-set is emitted more than once per kind
P-4 Stellium sub-cliques contained in a larger Stellium are suppressed
P-5 T_SQUARE has exactly 3 contributing aspects
P-6 GRAND_TRINE has exactly 3 contributing aspects
P-7 GRAND_CROSS has exactly 6 contributing aspects
P-8 YOD has exactly 3 contributing aspects
P-9 STELLIUM (3-body) has exactly 3; (4-body) has exactly 6 contributing aspects
P-10 Output order: Stellia, T-Squares, Grand Trines, Grand Crosses, Yods
P-11 find_patterns does not mutate the input list

INV-GRAPH — Relational graph

Code Invariant
G-1 degree == len(edges) for every node
G-2 sum(family_counts.values()) == degree for every node
G-3 Every edge in node.edges involves that node as body1 or body2
G-4 node.edges sorted by (body1, body2, aspect)
G-5 graph.nodes sorted by body name
G-6 graph.edges sorted by (body1, body2, aspect)
G-7 graph.components sorted by (min(c), len(c)) ascending
G-8 build_aspect_graph does not mutate the input list

INV-HARM — Harmonic layer

Code Invariant
H-1 sum(counts.values()) == total
H-2 len(proportions) == len(counts)
H-3 abs(sum(proportions.values()) - 1.0) < 1e-9 when total > 0
H-4 Every member of dominant is a key in counts
H-5 All proportions are in [0.0, 1.0]
H-6 chart.total == len(aspects)
H-7 by_body[name].total equals the number of aspects in which name participates
H-8 by_body keys are in sorted body-name order
H-9 aspect_harmonic_profile does not mutate the input list

INV-POL — Policy

Code Invariant
PO-1 AspectPolicy raises ValueError when orb_factor <= 0
PO-2 AspectPolicy raises ValueError when declination_orb < 0
PO-3 DEFAULT_POLICY is a valid, constructable AspectPolicy
PO-4 Invalid tiers, non-boolean flags, and non-finite policy values raise ValueError
PO-5 Caller-supplied orb mappings are defensively copied and exposed read-only

3. Determinism Register

The following ordering guarantees are normative. Any permutation of an input list must produce identical output on all of these:

Context Determinism guarantee
find_aspects result order Sorted by orb ascending
find_declination_aspects result order Sorted by orb ascending
find_patterns — pattern order Stellia, T-Squares, Grand Trines, Grand Crosses, Yods
find_patterns — body-set per pattern frozenset (order-independent identity)
find_patterns — aspects tuple Sorted by (body1, body2, aspect)
build_aspect_graph — nodes Sorted by body name
build_aspect_graph — edges Sorted by (body1, body2, aspect)
build_aspect_graph — components Sorted by (min(c), len(c))
build_aspect_graph — node.edges Sorted by (body1, body2, aspect)
aspect_harmonic_profile — by_body keys Sorted body-name order
aspect_harmonic_profile — counts keys AspectFamily declaration order
aspect_harmonic_profile — dominant Sorted by AspectFamily.value (alphabetical)

4. Failure Doctrine

The following table lists every condition that raises an exception, the exception type, and the diagnostic guarantee:

Function / constructor Condition Exception Diagnostic guarantee
aspect_strength allowed_orb <= 0 ValueError Message includes "allowed_orb" and the offending value
aspect_strength orb > allowed_orb ValueError Message includes "orb" and both values
aspect_strength Non-finite or negative geometry ValueError Message identifies orb or allowed_orb
AspectPolicy orb_factor <= 0 ValueError Message includes "orb_factor"
AspectPolicy declination_orb < 0 ValueError Message includes "declination_orb"
AspectPolicy Invalid tier, flag, mapping, or non-finite value ValueError Message identifies the policy field
Detection entry points Non-finite longitude/declination/speed or out-of-range declination ValueError Message identifies the geometry field

All functions remain pure computations over valid inputs. Collection-oriented detectors return empty results for empty valid mappings; first-class analysis surfaces require at least two included points.

Behaviour on empty input

Function Input Returns
find_aspects {} []
find_declination_aspects {} []
find_patterns [] []
build_aspect_graph [] AspectGraph(nodes=(), edges=(), components=())
aspect_harmonic_profile [] AspectHarmonicProfile(chart=empty, by_body={})

5. No-Mutation Guarantee

Every public function beyond the detection layer accepts its inputs by value and does not mutate them:

Function Guarantee
find_patterns(aspects) Does not alter aspects or any element of it
build_aspect_graph(aspects, ...) Does not alter aspects or any element of it
aspect_harmonic_profile(aspects) Does not alter aspects or any element of it
aspect_strength(aspect) Does not alter aspect
aspect_motion_state(aspect) Does not alter aspect

6. Cross-Layer Consistency Rules

These rules govern the logical relationship between layers. Each rule must hold on any output produced by the detection layer.

Rule Expression
Classification–family a.classification.family == _FAMILY_BY_NAME[a.aspect] for all zodiacal AspectData
Classification–domain Detector-produced AspectData is explicitly ZODIACAL or WHOLE_SIGN
Strength–surplus a.orb_surplus == aspect_strength(a).surplus
Graph–harmonic sum(node.family_counts.values()) == hp.by_body[node.name].total for every node
Harmonic–total hp.chart.total == len(aspects)
Motion–is_applying aspect_motion_state(a) == APPLYING ↔ a.is_applying is True
Motion–is_separating aspect_motion_state(a) == SEPARATING ↔ a.is_separating is True
is_major–is_minor a.is_major != a.is_minor for any classified AspectData

7. Public Surface Register

Complete public surface of moira.aspects as of Phase 12:

Enumerations

Name Values
AspectDomain ZODIACAL, DECLINATION, WHOLE_SIGN
AspectTier MAJOR, COMMON_MINOR, EXTENDED_MINOR
AspectFamily 18 members, including explicit DECLINATION and UNKNOWN families
MotionState APPLYING, EXACT, SEPARATING, STATIONARY, INDETERMINATE, NONE
DeclinationMotionState APPLYING, EXACT, SEPARATING, STATIONARY, INDETERMINATE
AspectPatternKind STELLIUM, T_SQUARE, GRAND_TRINE, GRAND_CROSS, YOD
HellenisticAspectEvaluationStatus EVALUATED, NOT_EVALUABLE
HellenisticOvercomingRelation BODY1_OVERCOMES_BODY2, BODY2_OVERCOMES_BODY1, NEITHER

Frozen dataclasses

Name Fields
AspectClassification domain, tier, family
HellenisticDirectionTruth evaluation status, aspect angle, directed arc, direction, reason
HellenisticOvercomingTruth evaluation status, sign indices, inclusive places, relation, reason
HellenisticSuperiorityTruth ordered body names and longitudes plus direction and overcoming components
AspectPolicy tier, include_minor, orbs, orb_factor, declination_orb
AspectStrength orb, allowed_orb, surplus, exactness
AspectPattern kind, bodies, aspects
AspectGraphNode name, degree, edges, family_counts
AspectGraph nodes, edges, components; properties hubs, isolated
AspectFamilyProfile counts, total, proportions, dominant
AspectHarmonicProfile chart, by_body
DeclinationAspectAnalysis positions, aspects, orb, reference_frame, timescale, provenance; derived counts and normalized mapping
DeclinationAspectPolicy orb, exact/rate tolerances, hemisphere policy, equator policy
DeclinationAspectMotionWitness selected relationship, declinations/rates, signed error, relative rate, orb/orb rate, state, policy and coordinate provenance

Immutable result vessels

Name Rationale
AspectData Frozen after complete construction by a detector
DeclinationAspect Frozen after complete construction by a detector
OutOfBoundsBody Frozen after complete construction by its detector

The vessels are terminal (not designed for subclassing) and document their structural invariants explicitly in their class docstrings.

Module-level constants

Name Type Content
CANONICAL_ASPECTS tuple[str, ...] 24 canonical aspect names
DEFAULT_POLICY AspectPolicy Policy matching historical detection defaults

Detection functions

Name Signature Returns
find_aspects (positions, *, include_minor, tier, orbs, orb_factor, policy) list[AspectData]
aspects_between (body1, lon1, speed1, body2, lon2, speed2, ...) list[AspectData]
aspects_to_point (positions, point_name, point_lon, ...) list[AspectData]
hellenistic_superiority_truth (lon1, lon2, aspect_angle=None, *, body1="body1", body2="body2") HellenisticSuperiorityTruth
find_declination_aspects (declinations, *, orb, policy) list[DeclinationAspect]
declination_aspects_from_declinations (declinations, *, reference_frame, timescale, orb, policy) DeclinationAspectAnalysis
declination_aspect_motion_witness (body1, dec1, body2, dec2, aspect, *, speeds, orb, tolerances, reference_frame, timescale, policy) DeclinationAspectMotionWitness

Derived-layer functions

Name Input Returns
aspect_strength AspectData | DeclinationAspect AspectStrength
aspect_motion_state AspectData | DeclinationAspect MotionState
find_patterns list[AspectData] list[AspectPattern]
build_aspect_graph list[AspectData], bodies=None AspectGraph
aspect_harmonic_profile list[AspectData] AspectHarmonicProfile

8. Validation Baseline

Validation uses the authoritative project .venv and the focused aspect, moiety, signed-motion, synastry, pattern, public-surface, external-reference, and relationship REST slices. Test counts are not frozen doctrine; the completion receipt for a change records the exact commands, outcomes, skips, fixtures, and scope exercised. Tests may change only when an explicitly admitted contract changes, with regression coverage for the replacement law.

Clone this wiki locally