1 lightning.finite_horizon.studies.development_screen
lightning.finite_horizon.studies.development_screen
Portable retained native-H1 development screen observations.
The frozen 24-scene cohort uses four held-out scene groups. Within a scene, paired state regrets are averaged; each usable physical scene receives one aggregate vote, including measured-zero and noninformative states. The 0.005 threshold and lowest-index constant tie rule remain frozen. Uniform-tie sensitivity is descriptive and cannot replace that decision.
This module verifies captured assignment/admission, fit/checkpoint lineage and finite paired evaluations. It neither admits current datasets nor fits models. Historical native-H1 loss receipts are explicitly versioned; current Q1 loss projection remains owned by objective.py and commissioning.q1_metrics.
1.1 Attributes
| Name | Description |
|---|---|
| CURRENT_Q1_READOUT_CONTRACT | Current commissioning Q1 state-mean projection; no historical value conversion. |
| HISTORICAL_LOSS_CONTRACT | Native H1 losses captured in diagnostic-fit-v1 receipts; never re-evaluated. |
1.2 Classes
| Name | Description |
|---|---|
| PlannedScene | One source-planned physical scene and its completed or unavailable attempt. |
| ScreenAssignment | Frozen source, split, model, and decision identity for one diagnostic screen. |
| ScreenInput | Exact assignment file and its caller-frozen content digest. |
| ScreenResult | Content-bound terminal receipt for one development-screen attempt. |
1.2.1 PlannedScene
lightning.finite_horizon.studies.development_screen.PlannedScene()One source-planned physical scene and its completed or unavailable attempt.
1.2.1.1 Attributes
| Name | Description |
|---|---|
| scene_id | Stable physical scene identity; the independent aggregation unit. |
| sample_index | Planned VIN source-row index, including unavailable acquisitions. |
| store_id | Frozen rollout store identity for this planned scene. |
| status | Captured availability; a rejected scene is distinct from measured zero. |
| acquisition_status | Original acquisition terminal status, preserved without admission promotion. |
| reason | Captured exclusion or execution failure reason; empty for successful acquisition. |
| acquisition_receipt_path | Captured terminal receipt location; portable readback does not follow this path. |
| acquisition_receipt_sha256 | SHA-256 of the captured terminal acquisition receipt, including unavailable attempts. |
| rollout_manifest_sha256 | Full completed-store manifest SHA-256; None when the acquisition is unavailable. |
| rollout_store_content_sha256 | Completed-store content seal; None when no complete store was captured. |
| owner_sha256 | Immutable completed-store owner digest; None for unavailable acquisitions. |
1.2.2 ScreenAssignment
lightning.finite_horizon.studies.development_screen.ScreenAssignment()Frozen source, split, model, and decision identity for one diagnostic screen.
1.2.2.1 Attributes
| Name | Description |
|---|---|
| schema_version | Exact historical assignment schema identifier. |
| source_plan_path | Original frozen acquisition plan location; retained only as provenance in cold readback. |
| source_plan_sha256 | SHA-256 of the acquisition plan captured before learner outcomes. |
| source_readback_path | Original captured source-readback location; cold reading does not follow this path. |
| source_readback_sha256 | SHA-256 of the captured passing source readback that preceded the screen. |
| actor_store_dir | Original actor-store location; no actor payload is read during retained analysis. |
| actor_manifest_sha256 | Full actor manifest SHA-256 shared by all historical folds. |
| rollout_store_root | Original rollout store root; retained as provenance rather than reopened. |
| evaluator_identity | Frozen scientific evaluator identity shared by admission, fit, and held-out observations. |
| implementation_revision | Captured fit source revision, retained as historical provenance. |
| scenes | All 24 consecutive source intentions, including rejected and failed acquisitions. |
| test_groups | Four disjoint held-out scene groups; each scene appears exactly once. |
| seed | Frozen historical fit seed 7. |
| optimizer_updates | Frozen historical 2000-update budget per scientific fold. |
| cpu_threads | Frozen historical four-thread execution setting. |
| practical_tie_band | Frozen 0.005 native gain margin for practical action regret. |
| minimum_improvement | Frozen 0.005 minimum scene-equal constant-minus-learner improvement. |
| deadline | Aware frozen historical ISO timestamp; cold reading may occur after it expires. |
1.2.3 ScreenInput
lightning.finite_horizon.studies.development_screen.ScreenInput(
assignment_path,
assignment_sha256,
)Exact assignment file and its caller-frozen content digest.
1.2.3.1 Attributes
| Name | Description |
|---|---|
| assignment_path | JSON file naming all planned rows and source identities. |
| assignment_sha256 | SHA-256 of assignment bytes, fixed before learner outcomes. |
1.2.4 ScreenResult
lightning.finite_horizon.studies.development_screen.ScreenResult(
receipt_path,
receipt_sha256,
)Content-bound terminal receipt for one development-screen attempt.
1.2.4.1 Attributes
| Name | Description |
|---|---|
| receipt_path | Atomic top-level receipt covering completed, failed, and unreached folds. |
| receipt_sha256 | SHA-256 of receipt_path, required for every readback. |
1.3 Functions
| Name | Description |
|---|---|
| diagnose_development_screen_uniform_tie | Compute the post-hoc order-neutral constant-control sensitivity. |
| main | Cold verify an explicit manifest and its independently retained seal. |
| read_development_screen | Verify historical retained fit artifacts and paired physical-scene usefulness. |
| read_development_screen_assignment | Read one content-bound assignment for portable retained-result analysis. |
| read_portable_development_screen | Verify an independently anchored portable native-H1 result and full aggregate. |
| reduce_development_screen_evaluations | Validate paired retained H1 rows and recompute the frozen scene-equal decision. |
| seal_development_screen | Copy retained observation artifacts exclusively and return the independent seal. |
| validate_development_screen_retained_fold | Validate one copied fold without Git, checkpoints, or rollout stores. |
1.3.1 diagnose_development_screen_uniform_tie
lightning.finite_horizon.studies.development_screen.diagnose_development_screen_uniform_tie(
assignment,
evaluations,
)Compute the post-hoc order-neutral constant-control sensitivity.
The frozen constant assigns equal scores and therefore selects the lowest admitted candidate index. This diagnostic instead averages practical regret uniformly across every admitted tied action. It does not alter the frozen control, threshold, or development decision.
1.3.2 main
lightning.finite_horizon.studies.development_screen.main()Cold verify an explicit manifest and its independently retained seal.
1.3.3 read_development_screen
lightning.finite_horizon.studies.development_screen.read_development_screen(
source,
audit_path,
result,
)Verify historical retained fit artifacts and paired physical-scene usefulness.
This archived-location interface never reopens datasets or Git. Use the portable seal interface when the captured absolute fit paths are unavailable. Returned native-H1 observations remain historical and grant no fresh fit or data admission. The captured aggregate schema is preserved for comparisons.
1.3.4 read_development_screen_assignment
lightning.finite_horizon.studies.development_screen.read_development_screen_assignment(
source,
)Read one content-bound assignment for portable retained-result analysis.
1.3.5 read_portable_development_screen
lightning.finite_horizon.studies.development_screen.read_portable_development_screen(
manifest_path,
*,
expected_manifest_sha256,
)Verify an independently anchored portable native-H1 result and full aggregate.
All 22 intended files are required; fit checkpoint bytes are hashed without loading models. Captured absolute paths remain provenance and are never followed. Every fold’s learner/constant state, candidate, source, and fit lineage is admitted before the full scene-equal aggregate is recomputed. A hash-rebound readback with any fabricated aggregate field is rejected.
This readback verifies captured admission consistency, rather than reopening canonical datasets or granting current execution admission. Historical native-H1 losses are not rewritten as the current state-balanced Q1 readout.
1.3.6 reduce_development_screen_evaluations
lightning.finite_horizon.studies.development_screen.reduce_development_screen_evaluations(
assignment,
evaluations,
)Validate paired retained H1 rows and recompute the frozen scene-equal decision.
This reduction is independent of Git state, checkpoints, and rollout stores so a publication bundle can cold-read copied evaluation receipts. Callers remain responsible for binding those receipts to their immutable source artifacts.
1.3.7 seal_development_screen
lightning.finite_horizon.studies.development_screen.seal_development_screen(
source,
audit_path,
result,
*,
readback_path,
destination,
)Copy retained observation artifacts exclusively and return the independent seal.
Original bytes/paths/revisions and the captured aggregate remain unchanged. Sealing requires the historical retained fit/checkpoint files to be available; subsequent cold reading needs only this bundle and its independently retained returned manifest hash. Neither operation executes a fit or accesses stores.
1.3.8 validate_development_screen_retained_fold
lightning.finite_horizon.studies.development_screen.validate_development_screen_retained_fold(
assignment,
fold_index,
fit_receipt,
learner,
constant,
)Validate one copied fold without Git, checkpoints, or rollout stores.