1 lightning.finite_horizon.scene_sampler

lightning.finite_horizon.scene_sampler

Frozen scene-balanced whole-chain exposure for prospective Q1 experiments.

Sampling balances physical scenes, not correlated tasks or states. Within each scene the admitted chains form a seeded cyclic roster. A complete schedule is sealed before training; epochs denote recovery blocks, never corpus passes.

1.1 Classes

Name Description
QhSceneBalancedSampler Expose an immutable finite schedule through native Torch sampling.
QhSceneSamplingConfig Declare finite TRAIN-only exposure; module seed controls every draw.
QhSceneSamplingReceipt Detached JSON-compatible TRAIN schedule and seal returned after construction.
QhSceneSamplingRow Root/task-qualified identity of one admitted TRAIN chain position.

1.1.1 QhSceneBalancedSampler

lightning.finite_horizon.scene_sampler.QhSceneBalancedSampler(
    rows,
    *,
    config,
    seed,
    corpus_identity_hash,
)

Expose an immutable finite schedule through native Torch sampling.

Scene order depends only on seed and sorted scene IDs, so adding chains to the same scenes preserves paired physical-scene exposure. Chain traversal depends on qualified identity within each scene. Admission rejects schedules that cannot visit every TRAIN chain. set_epoch reproduces whole-block recovery; mid-block exact continuation is outside this contract.

1.1.1.1 Attributes

Name Description
receipt Return a detached JSON-compatible seal of identities and all draws.
epoch Return the current frozen draw block for read-only acquisition audit.

1.1.1.2 Methods

Name Description
set_epoch Select a declared recovery block, rejecting over-budget exposure.
1.1.1.2.1 set_epoch
lightning.finite_horizon.scene_sampler.QhSceneBalancedSampler.set_epoch(epoch)

Select a declared recovery block, rejecting over-budget exposure.

1.1.2 QhSceneSamplingConfig

lightning.finite_horizon.scene_sampler.QhSceneSamplingConfig()

Declare finite TRAIN-only exposure; module seed controls every draw.

1.1.2.1 Attributes

Name Description
protocol Closed whole-chain sampling identity; default training remains unchanged.
blocks Number of complete validation/recovery blocks in the frozen schedule.
draws_per_block Chain presentations per block; must divide exactly into training batches.

1.1.3 QhSceneSamplingReceipt

lightning.finite_horizon.scene_sampler.QhSceneSamplingReceipt()

Detached JSON-compatible TRAIN schedule and seal returned after construction.

The sampler validates unique qualified rows and full schedule coverage; diagnostic-history admission independently replays persisted receipts. An annotation alone does not authenticate a decoded JSON payload.

1.1.3.1 Attributes

Name Description
schedule_hash Stable hash of all schedule fields before this hash field is appended.

1.1.4 QhSceneSamplingRow

lightning.finite_horizon.scene_sampler.QhSceneSamplingRow(
    scene_id,
    source_manifest_hash,
    source_sample_index,
    target_task_identity,
    store_index,
    rollout_row_id,
)

Root/task-qualified identity of one admitted TRAIN chain position.

1.1.4.1 Attributes

Name Description
scene_id Physical scene used as the independent exposure and evaluation unit.
source_manifest_hash Immutable actor-root manifest binding; separate from rollout origin.
source_sample_index Persisted actor source-row index within the bound root manifest.
target_task_identity Source-qualified target geometry identity; repeated tasks retain distinct chains.
store_index Position in the frozen ordered rollout-store population.
rollout_row_id Native factual chain row within its immutable rollout store.
Back to top