GOCHARA_WINDOWS_STANDARD

Bounded Gochar windows standard

Date: 11 October 2026. Owner: moira/gochara_windows.py. The implementation and its acceptance evidence are recorded in the validation receipt. This is the window extension of VED-017, alongside the existing dated snapshot and snapshot doctrine.

Admitted product

Return the time partition of the existing Phaladeepika/Sastri seven-classical Gochar profile relative to the natal Moon. Every request searches Sun, Moon, Mars, Mercury, Jupiter, Venus and Saturn, including potential blockers that are absent from the requested presentation subjects. The existing evaluator owns favorable houses, historical indications, all directed Vedha witnesses, exemptions, baseline-only omission and raw BAV testimony.

The twelve named ayanamsas are Lahiri, Fagan-Bradley, Krishnamurti, Raman, Yukteshwar, Djwhal Khul, True Chitrapaksha, De Luce, Aldebaran (15 Tau), True Revati, True Pushya and True Mula. Each is evaluated at the required TT epoch. The five live-star systems resolve the sovereign anchor record directly; missing anchors never fall back to a polynomial. The frame/anchor model is the existing dated Gochar model, not a new interpretation of these traditions.

Optional raw natal BAV derives once from the same seven natal signs and Lagna, using the existing unreduced Raman encoding. Birth location is required exactly when this mode is selected. State rows select the subject's own BAV by absolute transit sign. No threshold, score, probability or strength override is added.

VED-018/019 remain separate: alternative source tables, nodes, other reference points, counter-Vedha, thirds/halves activation, Murthi, dasha/dignity modifiers, nakshatra/chakra techniques and remedies are not selected by this product. The research checkpoint preserves their source differences and locators.

Engine and REST

from moira import Moira, GocharaWindowSearchPolicy

m = Moira()
result = m.gochara_windows(
    2451545.0, 2460967.0, 2460968.0,
    subjects=("Jupiter",),
    search_policy=GocharaWindowSearchPolicy(require_resolved=True),
)
assert result.coverage.domain_accounted
assert len(result.coverage.participants) == 7
assert any(e.planet == "Jupiter" and e.direction == 1 for e in result.events)
assert result.resolve_sample(0).transit == result.samples[0].epoch

gochara_windows(natal_jd_ut1, start_jd_ut1, end_jd_ut1, *, birth_location=None, policy=None, search_policy=None, subjects=None, reader=None) is the standalone entry. gochara_windows_for_datetimes(natal_dt, start_dt, end_dt, **kwargs) accepts three aware civil instants. Both are also instance-reader-bound Moira methods. gochara_window_capabilities() is available through the curated root, facade and Vedic surfaces and as a static facade method.

Method	Route	Contract
POST	/v1/gochara/windows/from-epochs	Three UT1 JDs, existing date policy, separate search policy, optional birth location and subjects.
POST	/v1/gochara/windows/from-datetimes	Three offset-aware ISO datetime strings; convert each exactly once using the existing UTC/UT1 owner.
GET	/v1/gochara/windows/capabilities	Profiles, systems, nominal domain, limits, endpoint and uncertainty semantics; no kernel/anchor probe.

Example epoch request:

{
  "natal_jd_ut1": 2451545.0,
  "start_jd_ut1": 2460967.0,
  "end_jd_ut1": 2460968.0,
  "subjects": ["Jupiter"],
  "policy": {"ayanamsa_system": "Lahiri"},
  "search_policy": {"boundary_tolerance_seconds": 1.0, "require_resolved": true}
}

Unknown fields, boolean/nonfinite numbers, numeric strings, invalid/duplicate subjects, contradictory BAV policy, zero/reversed duration and naive or numeric civil instants fail before astronomical evaluation. Result schema is moira.gochara.windows.v1. Pydantic declares every nested field, validates the result, and REST sends the engine's exact canonical JSON bytes. The base engine uses standard-library serialization and does not import Pydantic.

The result's selected options include forecast.dated_events; the dated and supplied-position snapshot policies do not select that option. Catalogue admission describes this bounded window product, not an expansion of snapshot claims. Eight Gochar routes are now discoverable under classical-vedic.

Domain, clocks and resources

Nominal transit certification is UT1 JD [2415020.5, 2488434.5), corresponding to 1900-01-01 through, but excluding, 2101-01-01. The actual interval must also fit converted TT frame support, required retarded SPK epochs, reader identity and anchor support. The clock adapter derives safe UT1/TT intersection endpoints and returns them in resource receipts. Near either nominal edge, nominal UT1 admission alone does not guarantee support. Natal epochs retain the dated compositor's point-domain rules and may precede 1900.

One request pins a borrowed pool's reader order, owns its point/interval caches and work meter, and restores reader context on every exit. It never closes a caller-owned reader. In particular, prewarmed serving-reader point caches are not inherited by the proxy: their stored readers could bypass request metering. Clock/EOP and anchor identities are checked for change before returning. Receipts retain clock/anchor input hashes, the anchor record, serving time identity, SPK source hashes and cache occupancy/hit/miss measurements.

Partition, evidence and numerical meaning

The requested interval is closed. Each instant belongs to exactly one region:

Kind	Meaning
certain	Seven constant signs and their discrete doctrine state, justified by exclusion evidence; one individually dated astronomical sample.
boundary_band	Closed event-time enclosure with accepted classification/order and requested width. It has no asserted constant state.
unresolved_band	Closed interval where count, contact, order, width, source continuity or another named dependency remains unresolved. Reasons and individual events survive merging.

Neighboring certain regions exclude the endpoints owned by a closed band. Request-edge clipping is retained. Overlapping distinct events share a band and preserve event_order_unresolved; their order is not guessed. Subject views retain references to this atomic partition, including blocker-only bands; this admission does not provide a separate coalescing/stitching API.

An open gap with no representable binary64 interior epoch is retained with no_representable_interior_witness; it cannot acquire a sample on an excluded band endpoint. A certain region touching a request edge may instead use that owned endpoint as its witness. Individual event evidence survives either case.

The existing point rule remains authoritative: exactly 30 degrees is Taurus, including on a retrograde pass. A tangent may have different singleton and neighboring states. The solver conservatively retains possible contacts and does not manufacture an exact singleton from a rounded time. The event schema can distinguish contact/model-seam records; this implementation establishes crossings where proved and otherwise retains possible contacts or reasoned unresolved seam regions. It does not assert exact contact existence.

All twelve sign rays for all seven bodies are searched. Interval value exclusion and necessary Newton contraction account for omitted domains; adaptive subdivision retains possible roots and source/record seams. The initial one-day subdivision seed is not a sampling-based completeness proof. Periodic targets are rebased after both subdivision and coarse-to-precise evaluation so a changed angular lift cannot erase a crossing.

The arithmetic certificate is conditional on RN binary64 arithmetic with gradual underflow and transcendental error at most four ulps. Differential arithmetic encloses a family of continuous surrogate branches, each holding its rounding parameters fixed over the complete interval. Light-time stop branches are retained as a family. Clock/record seams disable continuous contraction where continuity cannot be established.

A crossing count of one means one crossing per continuous surrogate branch under this declared model. It is not a count of sub-ulp sign flips of the discontinuous rounded program, nor a proof of exact physical event time. Endpoint brackets use the full interval's center and derivative enclosures; fresh point calls cannot shrink away its branch/rounding uncertainty to claim existence. Definite derivative sign proves direction and uniqueness within each branch; the retained time band covers the branch family.

Existence/count, direction, event order and time width are separate facts. A proved crossing can still be qualified by boundary_width_exceeds_requested_tolerance. Ordinary mode preserves it in an unresolved band; strict mode rejects the result. A one-second numerical target does not imply one-second agreement with external astronomy.

coverage.domain_accounted reports full partition ownership. coverage.exclusion_certified means discarded domains have the declared exclusion evidence; it does not mean no uncertainty exists. coverage.events_resolved additionally requires no unresolved region. Participant and doctrine coverage are separate fields. Degrees, clocks and ayanamsa values remain time-dependent and appear only in dated samples.

Frozen result construction rechecks types, IDs, ordering, references, source testimony, event evidence, sample signs, exclusions and partition ownership. Natal integrity is checked without repeating BAV. These are consistency checks, not cryptographic authentication of someone else's fabricated kernel result.

Work, output and failure law

The default boundary target is 1 second; admitted targets are 0.01–60 seconds, subject to the representability floor. Callers can lower each positive integer limit, but cannot exceed these finite ceilings:

Limit	Maximum/default
Requested days	366
Evaluations	2,000,000
Reader calls	20,000,000
Root iterations	500,000
Subdivisions	500,000
Retained event candidates	8,192
Cells	16,384
Dated samples	16,384
Evidence records/candidates	500,000
Canonical UTF-8 JSON bytes	32,000,000

The evaluation counter includes isolation visits and evaluated astronomy; reader calls count actual serving operations. Cached work is not recharged. Candidate counts can exceed merged events. Evidence includes retained unresolved candidates as well as exclusion records. Output records are charged before retention, then the complete compact UTF-8 encoding is checked, including civil receipts. Its own byte count is omitted from work to avoid self-reference.

These ceilings retain headroom above the measured annual corpus (about 40,000 evaluations, 210,000 reader calls, 210 samples and 5 MB), including tighter targets and difficult geometry; they are not a latency SLA. Every request is still bounded by the earliest exhausted cap. The exact benchmark matrix and peak resident-memory measurements are in the validation receipt. Annual calls are synchronous work taking minutes on the measured host. Proxy/client timeouts do not cancel a running Python worker; it finishes within its work limits and restores its context. No asynchronous job service is implied.

Outcome	HTTP	Code
Malformed input, policy contradiction or nonunique Lagna	422	validation_error
Outside serving ephemeris coverage	422	gochara_date_outside_coverage
Outside effective certification domain	422	gochara_window_outside_certified_domain
Duration/work/output exhaustion	422	gochara_window_budget_exceeded
Unresolved result requested in strict mode	422	gochara_window_unresolved
Required kernel/time/anchor resource unavailable	503	gochara_resource_unavailable

Gochar window error details have a dedicated typed schema retaining numeric counts/bounds and reason lists. A budget failure discards the incomplete result; no partial 200 or silent truncation is permitted. Ordinary unresolved results return 200 with their evidence. Unexpected failures retain the existing sanitized server-error behavior.

Ownership and preservation

_event_search owns shared intervals, roundoff, differentials, frame/clock and serving-astronomy enclosures, isolation, budgets and reader loans. Original marriage module paths alias these owners for import/monkeypatch compatibility; marriage policy, horizons, history and election composition stay in their existing owners. No native formula, dependency, source table, snapshot default, website or Urania flow changes in this admission. The focused validation includes shared-owner regressions and native-record parity; it is not a claim that the whole repository was tested or that astrology's predictive validity was established. Publication, version release and deployment are separate.