1 lightning.finite_horizon.diagnostic_cycle

lightning.finite_horizon.diagnostic_cycle

Finite, receipt-backed QH development diagnostic over one frozen N60 cohort.

This module selects cells and calls the canonical diagnostic experiment. It does not implement a second learner or scorer. A profile run uses training scenes only; the full matrix requires an admitted original-label evaluator disposition and a clean, exact source revision.

1.1 Classes

Name	Description
QhLearningCycleConfig	Frozen finite cells, original-store paths and evaluator admission.

1.1.1 QhLearningCycleConfig

lightning.finite_horizon.diagnostic_cycle.QhLearningCycleConfig()

Frozen finite cells, original-store paths and evaluator admission.

1.1.1.1 Attributes

Name	Description
cohort_manifest	Tracked cohort assignment containing exactly eight stores and four folds.
cohort_sha256	SHA-256 digest binding cohort_manifest bytes.
rollout_store_root	Root containing immutable rollout directories named by cohort store identity.
actor_store_dir	Immutable offline actor store shared by every diagnostic cell.
source_campaign_plan	Frozen generation plan required to admit the original-label evaluator.
source_campaign_plan_sha256	SHA-256 digest of source_campaign_plan, required for a full matrix.
evaluator_disposition	Human-reviewed scope of the retained campaign evaluator.
evaluator_identity	Content identity reproduced from campaign scorer and task-sampler settings.
required_source_revision	Exact clean Git revision required for a full diagnostic matrix.
seeds	Three frozen training seeds applied to every learned matrix cell.
max_optimizer_updates	Exact optimizer-step budget for each scientific matrix fit.
profile_optimizer_updates	Smaller optimizer-step budget used only by the training-only profile.
accelerator	One frozen execution device choice shared by every learned matrix cell.
max_epochs_safety	Epoch ceiling that prevents an update-bounded fit from running indefinitely.
validation_every_n_epochs	Epoch interval for full-matrix fits; short profiles validate every epoch.
hidden_dim	Frozen hidden width shared by geometry and root-moment scorers.
dropout	Frozen dropout probability; zero removes stochastic dropout from this comparison.
batch_chains	Number of complete rollout chains in each optimization batch.
learning_rate	AdamW learning rate shared by every learned cell.
weight_decay	AdamW weight-decay coefficient shared by every learned cell.
practical_tie_band	Absolute H1 reward difference treated as practically tied.
write_metric_history	Opt in future learned cells to bounded JSONL telemetry; omitted from legacy assignments when false.

1.2 Functions

Name	Description
main	Run a training-only CPU profile or an admitted full diagnostic matrix.
read_qh_diagnostic_cohort	Read one content-bound cohort through the cycle’s canonical topology contract.
run_cycle	Run only frozen cells, preserving every completed or failed attempt.

1.2.1 main

lightning.finite_horizon.diagnostic_cycle.main()

Run a training-only CPU profile or an admitted full diagnostic matrix.

1.2.2 read_qh_diagnostic_cohort

lightning.finite_horizon.diagnostic_cycle.read_qh_diagnostic_cohort(
    path,
    *,
    expected_sha256,
)

Read one content-bound cohort through the cycle’s canonical topology contract.

1.2.3 run_cycle

lightning.finite_horizon.diagnostic_cycle.run_cycle(
    config,
    *,
    output_dir,
    full_matrix,
)

Run only frozen cells, preserving every completed or failed attempt.