1 lightning.finite_horizon.studies.mac_q1_campaign

lightning.finite_horizon.studies.mac_q1_campaign

Frozen Mac Q1 comparisons through the canonical diagnostic experiment.

The campaign changes only TRAIN population and learning rate. Sampling blocks are recovery boundaries, not corpus epochs. Evaluation retains selected and final roles separately; no result here certifies a deployable policy.

1.1 Classes

Name Description
MacQ1CampaignConfig Closed population and unchanged scorer/optimizer template for six fits.

1.1.1 MacQ1CampaignConfig

lightning.finite_horizon.studies.mac_q1_campaign.MacQ1CampaignConfig()

Closed population and unchanged scorer/optimizer template for six fits.

1.1.1.1 Attributes

Name Description
d0 Historical R1 TRAIN population in frozen source order.
d1 Historical-root expansion admitted separately from D2.
validation Unchanged exposed development validation; never called untouched test.
experiment Native Huber volumetric scorer with no auxiliary objectives.
admission_receipt Parent-owned full-array/actor/teacher/source admission receipt.
admission_receipt_sha256 Exact admission bytes bound before any optimizer transaction.
deadline_utc Owned supervisor soft-pauses at a complete sampling-block boundary.
seeds Two registered seed-specific common cold states.
tie_band Original-label practical tie band, not a fitted threshold.
max_allocated_bytes Commissioning memory bound; MPS current/driver counts are distinct.

1.1.1.2 Methods

Name Description
validate_campaign Reject additional interventions or altered comparison support.
1.1.1.2.1 validate_campaign
lightning.finite_horizon.studies.mac_q1_campaign.MacQ1CampaignConfig.validate_campaign(
)

Reject additional interventions or altered comparison support.

1.2 Functions

Name Description
cell_experiment Resolve one registered cell without changing scorer, loss or validation.
evaluate_campaign_cell Cold-read verified final-primary and selected-secondary original H1 rows.
prepare_campaign Freeze full admitted populations, common cold states and TRAIN controls.
run_campaign_cell Charge one canonical fit; preserve failures and never retry automatically.
update_budget Keep 390 factual chain presentations per block under one frozen batch width.
verify_campaign Authenticate producing source, config, cold weights and full populations.

1.2.1 cell_experiment

lightning.finite_horizon.studies.mac_q1_campaign.cell_experiment(
    config,
    cell,
    seed,
    phase,
)

Resolve one registered cell without changing scorer, loss or validation.

1.2.2 evaluate_campaign_cell

lightning.finite_horizon.studies.mac_q1_campaign.evaluate_campaign_cell(
    config_path,
    output,
    *,
    fit_receipt,
    cell,
    seed,
    phase,
    destination,
)

Cold-read verified final-primary and selected-secondary original H1 rows.

1.2.3 prepare_campaign

lightning.finite_horizon.studies.mac_q1_campaign.prepare_campaign(
    config_path,
    output,
)

Freeze full admitted populations, common cold states and TRAIN controls.

Preparation performs no optimizer updates. DataModule admission precedes population hashing and guarantees scene-disjoint role contracts.

1.2.4 run_campaign_cell

lightning.finite_horizon.studies.mac_q1_campaign.run_campaign_cell(
    config_path,
    output,
    *,
    cell,
    seed,
    phase,
    attempt,
    resume_checkpoint=None,
)

Charge one canonical fit; preserve failures and never retry automatically.

1.2.5 update_budget

lightning.finite_horizon.studies.mac_q1_campaign.update_budget(config, phase)

Keep 390 factual chain presentations per block under one frozen batch width.

1.2.6 verify_campaign

lightning.finite_horizon.studies.mac_q1_campaign.verify_campaign(
    config_path,
    output,
    *,
    producer_root=None,
)

Authenticate producing source, config, cold weights and full populations.

Fitting uses the current checkout by default. A newer read-only observer may explicitly name the frozen producing checkout: its package, campaign owner and execution-profile bytes must still match the original protocol. This changes where source is authenticated, never the required digests. The caller separately authenticates its observer implementation; this function does not certify that implementation or authorize a new fit.

1.2.6.1 Parameters

Name Type Description Default
config_path Path Closed campaign configuration whose bytes are sealed. required
output Path Campaign root containing the immutable protocol and controls. required
producer_root Path | None Absolute checkout root for read-only source verification; omitted for all native fitting calls. None

1.2.6.2 Returns

Name Type Description
MacQ1CampaignConfig Admitted configuration and original protocol, preserving producing
dict[str, Any] identities even when the observer implementation differs.
Back to top