moira

Physical Heliacal Visibility Phase 1 Radiance/Response Checkpoint

Date: 2026-07-30

Status: adaptive radiance, response integration, interpolation, and storage gate passed

Runtime boundary: external research evidence only; not an engine table, loader, API, or release

Governing plan: PHYSICAL_HELIACAL_VISIBILITY_IMPLEMENTATION_PLAN.md

Predecessor: PHYSICAL_HELIACAL_VISIBILITY_PHASE1_ALTITUDE_PRESSURE_INTERPOLATION_CHECKPOINT_2026-07-30.md

Compact source-owned receipt: tests/artifacts/visibility_reference_lab/phase1_radiance_response_checkpoint_2026-07-30.json

Outcome

The final v9 artifact closes the remaining scientific-table questions for the first physical-visibility data pack:

an adaptive, independently seeded MYSTIC radiance design;
source-locked CIE photopic and scotopic response integration;
untouched off-grid response holdouts;
a dense direct-extinction interpolation surface with untouched midpoint holdouts;
separate solver, interpolation, and storage-error receipts;
a fail-closed deep-twilight law; and
an explicit boundary between shipped response products and a retained monochromatic diagnostic.

The same immutable artifact passed independent reconstruction under WSL Python and the repository’s Windows Python environment.

Spectral Reference Selection

An initial 550 nm reference produced an imbalanced scotopic error. The failed artifact was rejected before admission. A separate training-only diagnostic then evaluated 507, 519, 531, 543, 550, and 555 nm over eight fixed seeds at the exposed training point.

The 531 nm candidate minimized the larger of the photopic and scotopic relative standard errors:

Response	Relative standard error at 531 nm
Photopic	0.0185032
Scotopic	0.0710570

The diagnostic did not execute any holdout and did not alter the fixed response-error ceilings. The final artifact uses 531 nm consistently for spectral importance sampling, shape normalization, and the independent absolute anchor.

Admitted Response Surface

The response training grid is:

Coordinate	Nodes
Solar-center altitude	-9, -6, -3, 0 degrees
Target true altitude	0.25, 5, 20, 45 degrees
Relative solar azimuth	0, 60, 120, 180 degrees

Nine complementary Latin-square midpoint combinations were never used to build or tune the 64-node table. Trilinear interpolation is performed in log10 response luminance.

Quantity	Maximum error	95th percentile	Fixed ceilings
Photopic response	0.354270 mag	0.287715 mag	0.5 / 0.3 mag
Scotopic response	0.252966 mag	0.244817 mag	0.5 / 0.3 mag

No response threshold was relaxed after holdout execution.

The full response calculations use REPTRAN-fine ALIS from 380 through 780 nm and the exact source identities for the CIE photopic and scotopic tables. Solver uncertainty is preserved per table cell. Across the completed artifact, the maximum response relative standard errors are 0.209430 photopic and 0.232138 scotopic.

Direct Extinction and Storage

The direct-extinction surface contains 57 target-altitude training nodes from 0.25 through 45 degrees and 400 one-nanometre spectral bins. Fifty-six untouched midpoint altitudes produce 22,400 independently evaluated holdout bins.

Quantity	Measured	Fixed ceiling
Maximum extinction error	0.0212954 mag	0.05 mag
95th-percentile extinction error	0.00279149 mag	0.02 mag
Mean extinction error	0.000881366 mag	reported, not gated

Interpolation is linear in extinction magnitude over log10(target_true_altitude_deg + 0.25). Extrapolation is prohibited.

Little-endian IEEE-754 binary32 was selected for the data pack. Its maximum measured storage error is 9.487461198887104e-7 mag against a fixed 1e-5-mag ceiling. A separately evaluated quantized candidate also passed its own ceiling but was not selected; the standard-library-readable binary32 format is simpler and retains materially more margin.

Monochromatic Diagnostic Boundary

The reconstructed 531 nm surface remains in the error envelope as an intermediate diagnostic. It is not:

shipped as a runtime table;
used by runtime interpolation; or
an admission gate for the response-integrated products.

Its maximum comparison error is 0.544179 mag and its 95th-percentile error is 0.423551 mag. Reporting those values is intentional. Hiding them or treating the monochromatic surface as an admitted proxy for the integrated responses would be a false claim.

Deep-Twilight Law

The modeled solar-twilight table ends at a solar-center altitude of -9 degrees. A request below that boundary is typed not_evaluable with reason:

solar_twilight_below_data_pack_domain

A Monte Carlo non-detection is never converted into physical zero. Measured directional or total-background routes are independent and remain available. A future lower-Sun expansion requires new versioned data-pack evidence, not an undocumented extrapolation.

Immutable Receipt

Field	Value
Artifact directory	`phase1-radiance-response-2026-07-30-v9`
Runs	662
Files	5,621
Bytes	317,168,353
Generation fingerprint	`aef8bdc07948ff5367dba1834baea708dfea0bc0dffb6898899dabc6f231c8c0`
Manifest SHA-256	`6bb91212d1d54762af8276ea066b4c6d5f4df837d84a46057f57b35924bae12f`
Summary SHA-256	`829774108d1c1337399784c20254c1d6cc5fc071d532450f4afb74431d870104`
Linux independent validation	Passed
Windows independent validation	Passed

Eight rejected radiance designs remain represented by compact source-owned failure receipts. They preserve convergence, zero-normalization, spectral balance, exposed-holdout, interpolation, and cross-platform validation failures without allowing any rejected artifact to become an admitted baseline.

Explicitly Unchanged

no moira runtime or native code changed;
no public Python, facade, serializer, REST, or OpenAPI contract changed;
no installed dependency changed;
no CIE, libRadtran, or REPTRAN source file entered the engine wheel;
no automatic download or calculation-time network path was introduced; and
limiting-magnitude and event-time error propagation remain owned by Phases 2 and 3 respectively.

The next and final Phase 1 gate is compilation and dual-platform validation of the separate immutable data pack.

This site is open source. Improve this page.