Physical Heliacal-Visibility Validation and Admission

Date: 2026-08-11 Status: local Phase 7 engine-release gate closed; no publication or deployment Model: clear_sky_naked_eye_point_source_v1

Scope

This document is the current validation authority for Moira's additive physical point-source assessment and event model. It does not replace the separate legacy arcus/limiting-magnitude, Yallop lunar-crescent, or Krisciunas-Schaefer legacy-moonlight validation families.

The machine-checked inventories are:

• physical capability matrix;
• physical API inventory;
• tests/fixtures/physical_visibility_phase7_evidence_registry.json; and
• scripts/generate_physical_visibility_inventory.py --check.

Each admitted claim below names its evidence class. A test snapshot or another engine's output is never promoted to scientific authority merely because it agrees.

Admission summary

Surface	Status	Named evidence classes	Qualification
Physical single-epoch assessment	admitted, opt-in	primary_source_equation_validation, independent_libradtran_holdouts, modern_era_observational_comparison, property_and_invariant_testing, separated_numerical_tolerances	Clear sky, unresolved steady point source, admitted observer protocol and exact pack domain.
Four-phase physical event search	admitted, opt-in	external_ephemeris_event_goldens, property_and_invariant_testing, separated_numerical_tolerances	Mars/Jupiter/Saturn/Sirius only; explicit first/last ownership and certified crossings.
Python/facade/REST/OpenAPI projection	admitted	public_contract_and_openapi_parity, legacy_regression_fixtures	Additive physical routes; legacy defaults and shapes remain exact.
External data pack 1.2.0	admitted immutable resource	independent_libradtran_holdouts, release_artifact_and_offline_install	Separate CC BY-SA 4.0 artifact; never embedded or downloaded.
Two private Phase 6 kernels	admitted implementation detail	native_python_differential	Numerical differential/performance evidence only; no policy or doctrine ownership.
Historical apparition/Sothic cases	legacy corroboration only	historical_event_corroboration	Bounded windows or historical uncertainty; not exact physical-event goldens.
Site-specific experimental moonlight	quarantined research	experimental_site_specific_moonlight_quarantine	No runtime, API, packaged resource, release gate, or active-roadmap dependency.

Scientific evidence

Primary equations

tests/unit/test_visibility_physics.py and tests/unit/test_visibility_spectral.py retain source-level evaluations for Kasten-Young relative air mass, the named Schaefer legacy component family, Crumey/Blackwell point-source equations, CIE MES2 response composition, and Tousey-Koomen Table I comparison values. Their purpose is equation and component validation; they do not claim a universal atmosphere or observer.

Independent libRadtran evidence

The versioned pack was produced in an external offline libRadtran 2.0.6 and REPTRAN reference laboratory. Holdout coordinates and random seeds were kept disjoint from training or representation selection. The admitted checkpoints include:

• untouched radiance-response holdouts;
• untouched named-spectral direct-transmission holdouts;
• altitude, pressure, and joint interpolation holdouts; and
• Windows/Linux byte and validator parity.

The generator and runtime remain separate. libRadtran, REPTRAN data, and CIE source tables do not enter the Python distribution or normal execution.

Observational comparison

The admitted modern-era observational comparison is the Tousey-Koomen 1953 Table I twilight series. It is source bounded and enforces the declared point-source/field relationship. It is not described as a contemporary multi-site observer-population campaign.

Astronomical and event goldens

The Phase 3 source-owned golden binds:

tests/golden/physical_visibility_phase3_events.json
pack manifest:
  cf93433a9f66a5ea92832271ce3c4b023fcc8693164803539a9f1be85b17468c

The independent oracle uses checksum-locked Horizons geometry for Jupiter and the Sun, and Hipparcos/BSC5/CALSPEC/Astropy/PyERFA/IERS sources for Sirius. It does not import Moira's event solver. The admitted residuals were:

Case	Engine versus independent event	Oracle tolerance
Jupiter morning first rising	1.2063503265 seconds	60 seconds
Sirius morning first rising	2.5644600391 seconds	60 seconds

These cases validate ephemeris geometry, target identity, event ownership, crossing semantics, and exact pack use. They are not observed first-visibility dates and do not quantify weather or observer-population uncertainty.

The later Phase 7 broad oracle is the complete release-readiness gate for the public four-target/four-phase product. It covers the full 16-cell matrix with 12 independently reconstructed event times and four explicitly typed no-event/domain engine regressions. All 12 timed cells pass the fixed 60-second external-grid limit; the largest absolute engine/oracle difference is 5.8789461851119995 seconds. The current integrated public-engine replay passes all 16 cells against content-admitted DE441 and the exact 1.2.0 pack.

External acquisition is bound to the separately preserved 5,311-byte source specification scripts/visibility_reference_lab/physical_visibility_phase7_source_acquisition_spec_v1.json, SHA-256 ba061013c6e6258475baab4442b072c82c887aabc840cc19f5b0126542eb9323. Its acquisition-driving fields are identical to the later admitted matrix. The 47 external source files are preserved in the verified external archive named by tests/artifacts/release/physical_visibility_phase7_source_recovery_2026-08-11.json. The archive is not embedded in the repository or Python distribution.

The engine replay and published runtime require neither NumPy nor Astropy. The independent Sirius reconstruction uses the declared development-only Astropy/ERFA toolchain and may load NumPy transitively inside that isolated authority-validation environment.

Historical cases

Published planetary apparition windows and the Sirius/Sothic historical case remain in the legacy validation family with their declared windows and delegation tolerances. Phase 7 classifies them as corroboration only. It does not fit the physical model to them, reinterpret them as exact timestamps, or use engine agreement to erase calendrical, observational, site, or doctrine uncertainty.

Separated tolerance law

The pack's independent holdouts produced these numerical bounds:

Error owner	Maximum	p95 or receipt law
Photopic response interpolation	0.354270 mag	0.287715 mag
Scotopic response interpolation	0.252966 mag	0.244817 mag
Direct-extinction interpolation	0.0212954 mag	0.00279149 mag
Binary32 storage	9.487461198887104e-7 mag	retained separately
Source solver / Monte Carlo	per response cell	maximum contributing corner

The runtime propagates the admitted numerical envelope through the photopic and scotopic backgrounds and threshold to separate limiting-magnitude and visibility-margin bounds. Input/scenario and model uncertainties that lack a defensible numeric bound remain named and unquantified; they are not assigned zero.

Event search separately records:

• scan step and recursion law;
• root-time tolerance;
• root-margin tolerance or non-applicability at a horizon boundary;
• actual residual and bracket;
• the Lipschitz zero-enclosure certificate and unresolved interval count; and
• deterministic event-time sensitivity under admitted atmosphere scenarios.

The numerical search tolerances do not become claims about human-observer or weather accuracy. Moira publishes no aggregate tolerance and no probabilistic confidence score.

Property, invariant, and failure evidence

The focused property/invariant family covers:

• no extrapolation at pack axes or target photometry domains;
• exact regular-file inventory, checksums, compatibility, and symlink refusal;
• multiple hidden crossings, same-sign endpoints, tangencies, and near-zero intervals;
• fail-closed uncertified crossings;
• first/last guard-day ownership;
• horizon-owned versus margin-owned boundaries;
• polar day/night, circumpolar, never-rising, and missing-boundary states;
• deterministic repeated and concurrent execution;
• target-catalog and profile drift;
• no generic stellar color-index guessing; and
• no confidence or synthetic score fabrication.

not_evaluable is not treated as false visibility. not_found is used only when a fully evaluable search contains no owned phase transition.

Contract and compatibility evidence

Thirty-six physical symbols preserve object identity across moira.heliacal, the curated moira root, moira.facade, and moira.sky.visibility. Matching Moira methods forward the full policy. The two dedicated REST operations preserve every engine policy/receipt field and accept no client filesystem path.

The legacy assessment, tonight, and generalized heliacal event request and response schemas remain exact. The current generated inventory is checked against create_app().openapi() rather than copied from prose.

Native differential evidence

Phase 6 admitted exactly two private kernels. The boundary-inclusive differential contained 14,205 response cases, 2,001 direct-extinction cases, and eight-thread determinism. Maximum observed absolute differences were:

Quantity	Maximum	Admission tolerance
S/P ratio	8.881784197001252e-16	5e-15
Response weight	1.734723475976807e-18	5e-18
Response normalization	1.1102230246251565e-16	2e-15
Direct extinction	0	2e-15
Direct transmission	0	2e-15

This is implementation evidence, not a second scientific oracle. Python's numerical paths remain executable differential references.

Release artifact and offline evidence

scripts/validate_physical_visibility_phase7_release.py performs the release gate from a copied current source snapshot:

1. build a wheel and sdist with build isolation disabled;
2. inspect both for the release identity, compatibility contracts, and notice;
3. reject any embedded external .f32le payload;
4. clean-install the source-built wheel with --no-index --no-deps;
5. rebuild a wheel from the extracted sdist;
6. clean-install the sdist-built wheel;
7. import the native backend and both public physical functions; and
8. load the exact external pack under MOIRA_NO_DOWNLOAD=1 and an active socket-deny guard.

The external pack release tool produces a deterministic 14-file archive:

moira-physical-heliacal-visibility-1.2.0.tar.gz
SHA-256:
  0d2c98d0717c45416ad0f8f3e0b72ca28d3975f4f6c8080112ceb2bef8327d71

The hardened final run is recorded in tests/artifacts/release/physical_visibility_phase7_release_validation_2026-08-09.json. That receipt proves only the exact local artifacts it names; it does not claim that they were uploaded, tagged, published, installed downstream, or deployed.

Closed exclusions and downstream boundary

Clouds, live weather, optical aids, extended objects, the Sun, lunar crescents through this model, global site transfer, generic stellar profiles, observer population probability, and synthetic confidence remain outside the model.

The Jones/Paranal experiment is quarantined and does not enter the assessment, event, compatibility, or release contract. Website documentation remains downstream of an actual engine and data-artifact publication gate.