1 lightning.finite_horizon.studies.continuing_evidence
lightning.finite_horizon.studies.continuing_evidence
Render source-bound continuing-QH receipts without fitting or reacquiring data.
Q1 rows pair by physical state and candidate labels; uncertainty resamples physical scenes. QH residuals remain checkpoint-dependent diagnostics, while feasibility scores predict the recorded operational action mask. Neither metric family establishes endpoint policy benefit.
1.1 Classes
| Name | Description |
|---|---|
| ContinuingCellDisposition | Charged-attempt disposition; missing completion never becomes a result. |
| ContinuingCloudBinding | Startup run identity and failure-time controls authenticated by receipts. |
| ContinuingHorizonSupport | Exact retained loss numerator and query support for one prediction horizon. |
| ContinuingStudy | Admitted continuing-QH snapshot retaining failed and unlaunched cells. |
| ContinuingTrainingObservation | One sparse completed-update observation identified by its original event. |
| ContinuingValidationCheck | Domain reduction of one retained check, without policy-value interpretation. |
| InterruptedQhStudy | Fixed-weight failed-attempt diagnostics, distinct from completed fits. |
1.1.1 ContinuingCellDisposition
lightning.finite_horizon.studies.continuing_evidence.ContinuingCellDisposition(
cell,
disposition,
attempt_charged,
completed_result_admitted,
)Charged-attempt disposition; missing completion never becomes a result.
1.1.2 ContinuingCloudBinding
lightning.finite_horizon.studies.continuing_evidence.ContinuingCloudBinding(
cell,
run_id,
entity,
project,
terminal_completed_updates,
module_config_json,
startup_sha256,
failure_sha256,
)Startup run identity and failure-time controls authenticated by receipts.
1.1.2.1 Attributes
| Name | Description |
|---|---|
| cell | Failed continuing-study cell whose startup cloud identity and terminal local history were |
| run_id | WandB run identity from the retained startup receipt, required to match the run segment of the |
| entity | WandB entity namespace parsed from the authenticated startup URL; it locates the run without |
| project | WandB project namespace parsed from the same startup URL, paired with entity and run_id. |
| terminal_completed_updates | Successful-update count in the unique admitted failed-fit terminal event; bounds cloud diagnostic |
| module_config_json | Canonical detached JSON of module controls retained in the failure receipt; readback uses these |
| startup_sha256 | SHA-256 of exact retained startup-cloud-readback bytes binding the launch and cloud run identity. |
| failure_sha256 | SHA-256 of exact retained failure-receipt bytes binding the failure-time controls and disposition. |
1.1.3 ContinuingHorizonSupport
lightning.finite_horizon.studies.continuing_evidence.ContinuingHorizonSupport(
horizon,
loss_sum,
state_count,
)Exact retained loss numerator and query support for one prediction horizon.
1.1.3.1 Attributes
| Name | Description |
|---|---|
| mean_huber_loss | Normalize by recorded state queries; unsupported horizons remain missing. |
1.1.4 ContinuingStudy
lightning.finite_horizon.studies.continuing_evidence.ContinuingStudy(
snapshot_json,
_histories=(),
cloud_bindings=(),
)Admitted continuing-QH snapshot retaining failed and unlaunched cells.
1.1.4.1 Attributes
| Name | Description |
|---|---|
| cloud_bindings | Admitted failed-cell cloud/run/control bindings in registered cell order. Empty by default means no |
| provenance | Expose admitted producer bindings independently of renderer-owned mappings. |
| cell_dispositions | Preserve launch/journal charge facts and failed or unlaunched status. |
| histories | Return isolated validated telemetry, including failed/partial attempts. |
| training_observations | Retain sparse train events without smoothing or imputing absent losses. |
| validation_checks | Project admitted telemetry into immutable, support-aware domain rows. |
1.1.5 ContinuingTrainingObservation
lightning.finite_horizon.studies.continuing_evidence.ContinuingTrainingObservation(
cell,
seed,
sequence,
optimizer_update,
optimization_loss,
loss_available,
loss_semantics,
)One sparse completed-update observation identified by its original event.
1.1.5.1 Attributes
| Name | Description |
|---|---|
| cell | Registered continuing-study cell identity linking this row to its retained fit history, including |
| seed | Fit seed from the admitted history identity; repeated observations from this seed are dependent |
| sequence | Original zero-based history-event sequence, preserving observation identity when optimizer-update |
| optimizer_update | Recorded successful-update coordinate for this sparse train event; absent intermediate events are |
| optimization_loss | Finite manual-optimization combined objective for the recorded completed update, or None when the |
| loss_available | Explicit retained-loss availability bit, paired with optimization_loss; false denotes missing |
| loss_semantics | Producer label manual_optimization_objective_for_completed_update, distinguishing this sparse batch |
1.1.6 ContinuingValidationCheck
lightning.finite_horizon.studies.continuing_evidence.ContinuingValidationCheck(
cell,
seed,
optimizer_update,
moving_target_qh_equal_active_horizon_mean,
fixed_label_h1_huber_mean,
h1_state_count,
recursive_state_count,
active_horizon_count,
aggregation_semantics,
horizon_support,
)Domain reduction of one retained check, without policy-value interpretation.
Fixed-label H1 loss is a mean over its recorded support. The mixed QH loss retains the producer’s equal-active-horizon reduction and moving targets. Query counts describe support, not independent physical scenes.
1.1.7 InterruptedQhStudy
lightning.finite_horizon.studies.continuing_evidence.InterruptedQhStudy(
snapshot_json,
)Fixed-weight failed-attempt diagnostics, distinct from completed fits.
1.1.7.1 Attributes
| Name | Description |
|---|---|
| snapshot_json | Immutable admitted populations, paired scene summaries and raw support. |
| snapshot | Return detached reductions without granting fit/selection authority. |
| provenance | Expose original source provenance for canonical reporting identity. |
1.2 Functions
| Name | Description |
|---|---|
| admit_archived_continuing_study | Admit relocation using preserved execution identity and current verified bytes. |
| admit_archived_interrupted_evaluations | Admit externally pinned cold/partial validation receipts and pair scenes. |
| admit_continuing_study | Admit continuing receipts without model/data reconstruction or fitting. |
| analyze | Validate all available cell receipts and compute paired scene summaries. |
| sha256_file | Return the digest of exact retained receipt bytes. |
| summarize_feasibility | Summarize candidate-level operational-mask scores with explicit support. |
| summarize_horizon_rows | Validate held-out QH residual support and summarize scene-clustered errors. |
1.2.1 admit_archived_continuing_study
lightning.finite_horizon.studies.continuing_evidence.admit_archived_continuing_study(
archive,
study_prefix,
base_prefix,
)Admit relocation using preserved execution identity and current verified bytes.
1.2.2 admit_archived_interrupted_evaluations
lightning.finite_horizon.studies.continuing_evidence.admit_archived_interrupted_evaluations(
archive,
study_prefix,
)Admit externally pinned cold/partial validation receipts and pair scenes.
This consumes evaluated receipts, never checkpoint bytes or a fabricated completed fit. It verifies retained source/support identities and original labels before reducing dependent states within recipe, then recipes within physical scene. Bootstrap draws resample scenes with weights/selection held fixed. Candidate-pooled operational calibration remains separately named.
1.2.3 admit_continuing_study
lightning.finite_horizon.studies.continuing_evidence.admit_continuing_study(
study,
base_study,
)Admit continuing receipts without model/data reconstruction or fitting.
1.2.4 analyze
lightning.finite_horizon.studies.continuing_evidence.analyze(
study,
base_study,
*,
archive=None,
study_prefix=None,
)Validate all available cell receipts and compute paired scene summaries.
1.2.5 sha256_file
lightning.finite_horizon.studies.continuing_evidence.sha256_file(path)Return the digest of exact retained receipt bytes.
1.2.6 summarize_feasibility
lightning.finite_horizon.studies.continuing_evidence.summarize_feasibility(rows)Summarize candidate-level operational-mask scores with explicit support.
Calibration bins are equal-width over [0, 1), except the last includes 1. Invalid recall and false-valid rate use the invalid-action count as their shared denominator; no positive/negative class is imputed when absent.
1.2.7 summarize_horizon_rows
lightning.finite_horizon.studies.continuing_evidence.summarize_horizon_rows(
horizon_rows,
query_rows,
population,
expected_h1_rows,
)Validate held-out QH residual support and summarize scene-clustered errors.