1 lightning.finite_horizon.studies

lightning.finite_horizon.studies

Frozen scorer study protocols, admitted evidence, analysis and archive ownership.

Runtime composition remains in finite_horizon.experiment. Import each study owner directly; this namespace introduces no registry or reexports.

1.1 Package guide

This package owns frozen scorer-study protocols, schema-specific evidence admission and scientific reductions. The runtime composition interface remains QhExperiment; current optimization lives in objective.py and module.py. The study namespace has no registry or reexport facade: callers import the owner they use.

1.2 From bytes to scientific claims

archive.py authenticates original producer/configuration and result bytes against an externally retained manifest digest. Transport verification establishes integrity. inventory.py projects curated attempts, controls, statuses and provenance from receipts. Scientific admission belongs to the schema-specific readers, which verify population, support, pairing and checkpoint roles. A catalogued attempt can be failed, interrupted or unlaunched.

flowchart LR
  B["Sealed bytes and trusted digest"] --> A["Archive verification"]
  A --> I["Receipt inventory and execution status"]
  A --> R["Schema-specific scientific admission"]
  R --> S["Admitted study and reductions"]
  S --> P["Reporting snapshots and figures"]

flowchart LR
  B["Sealed bytes and trusted digest"] --> A["Archive verification"]
  A --> I["Receipt inventory and execution status"]
  A --> R["Schema-specific scientific admission"]
  R --> S["Admitted study and reductions"]
  S --> P["Reporting snapshots and figures"]

The source owners are:

• architecture_study.py and architecture_results.py: frozen architecture protocol, canonical diagnostic-fit dispatch and bounded CUDA measurements in the former, pure scene-equal reductions/support identities and paired comparisons in the latter;
• first_model_evidence.py: captured Q1 inputs, supported original-label evaluation, TRAIN-fitted controls and readiness admission;
• cuda_q1_posthoc.py: closed cuda-q1-first-v1 admission from independently retained S01 metadata/checkpoint pins; validates native terminal history, selected/final roles and original-label support without deserializing checkpoints. Runtime inference separately verifies actual checkpoint tensors;
• first_model_analysis.py: exploratory post-hoc Q1 reductions over admitted inputs, with physical-scene observations and intervals;
• development_screen.py and development_ranking.py: cold retained-screen admission and physical-scene candidate-order diagnostics; the screen module also owns the manifest-verification command;
• continuing_evidence.py: continuing-QH terminal dispositions, horizon support, residuals and operational-feasibility summaries.

1.3 Interpretation and failure contracts

Physical scenes are statistical units. States, recipes, candidate rows and repeated horizon queries share evidence; they are not independent scene replicates. Preserve recipe pairing, seed, checkpoint role and original support when comparing results. Original-label Q1 metrics have fixed labels; QH residuals describe the fitted moving-target objective. Operational feasibility is auxiliary to the hard action mask. These quantities answer different scientific questions.

Exploratory post-hoc analysis keeps fitted weights and frozen checkpoint selection fixed; it does not revise readiness gates. Failed and incomplete attempts remain visible with their dispositions. Missing telemetry is distinct from absent scientific support. history.py supplies optional telemetry and common bindings; the readers retain their own evaluator normalization and terminal rules.

Historical source/configuration bytes and hashes retain their execution identity. Current code or Python-path changes alter package fingerprints and may invalidate frozen resume checks. Portable readback and report rendering establish neither training-resume compatibility nor held-out policy effectiveness.

1.4 Presentation and evidence locations

aria_nbv.reporting consumes admitted products, builds snapshots and renders/exports results. Campaign-specific progression stays with the owning campaign. Living missions/reports live in their owning .omx campaign; immutable scientific exports live under experiments/. Configuration inputs and brief purpose/link READMEs live under .configs/. Figure registration is discovery, separate from scientific admission and evaluated thesis inclusion.

The consolidated implementation retains the historical larger Q1 and interrupted continuing-QH lineage. Its publication evidence is under experiments/2026-10-research/scorer.

1.5 Verification

Focused tests cover archive integrity, inventory dispositions, admitted products and post-hoc reductions, architecture results and reporting. Readback requires the correct historical payloads and externally trusted digest; local success does not establish hosted CI or thesis admission.

1.6 Prospective matched continuing-QH evidence

matched_evidence.py admits the four-cell 2026-10-09 score-only/joint protocol separately from historical continuing-QH. It requires current fit-v3/training-v6 receipts, 40 exact TRAIN passes (6,160 successful whole-chain updates), same-seed cold/exposure identity, and all 640 ordered original-label validation states. Each successful update must retain all 16 dense H1 states. Canonical coverage keeps 16 state records per chain and 40 H1 observations per state; the total dense candidate exposure must equal 40 passes over the frozen TRAIN support. Recursive queries retain exactly one selected candidate per supported state. Final fixed-budget checkpoints are primary; validation-selected checkpoints are separate secondary evidence. Operational teacher calibration is distinct from H1 utility and moving TD residuals.

Portable publication uses the existing authenticated archive. The selected study prefix contains the original config.toml, protocol.json, protocol-sha256.json, cell launch/terminal/fit/exposure/query-coverage receipts, and evaluations.json. The prospective protocol also seals ordered materialized operational teacher support directly from actor masks, including invalid real candidates and excluding padding. Evaluation must reproduce its digest and counts; agreement between two arms alone cannot authenticate classification labels or their denominator. The typed MatchedEvaluationIndex binds the constant receipt and each evaluated cell’s final/selected H1/joint receipts by relative path and SHA256; it names the independent evaluator package digest. Checkpoint payloads are excluded. Actual runtime evaluation admits the checkpoint bytes; cold report readback authenticates retained observations and metadata, without reopening the old execution paths or asserting resume compatibility.

Failed, interrupted, unlaunched, completed-but-evaluation-pending, evaluation-failed and unsupported H1 dispositions remain visible. No arm comparison is manufactured from missing support. Original candidate/label seals, operational teacher masks, phase-zero horizon queries and physical scene/recipe composition must align before paired reduction. State means reduce within recipe and then equally across recipes and physical scenes. Each seed has its own scene-bootstrap interval conditional on fixed weights; two seeds do not establish seed-population uncertainty. R0/crop-R1 pairs share physical scenes but have different target tasks, so recipe effects are descriptive. MatchedStudy.provenance retains analysis_source_sha256, the deterministic Python package-tree digest of the executing reductions. Its conservative scope includes unrelated package edits and remains distinct from the historical fit source_package_digest and evaluator_source_package_digest. The retained bootstrap_method is percentile_scene_bootstrap_95, with bootstrap_seed=20261009, bootstrap_draws=10000 and bootstrap_unit=physical_scene. reduction records equal states within recipe, equal recipes within physical scene and equal physical scenes; conditioning retains the fixed-weight/checkpoint limitations. These settings govern computation and provenance at the same source owner. Focused proof lives in test_qh_matched_evidence.py.

admit_archived_interrupted_evaluations in continuing_evidence.py admits exactly cold.json and interrupted.json from an independently pinned analysis archive. The original failed continuing-QH seed41 attempt remains failed: cold update 0 and historically selected partial update 770 are never completed or final checkpoints. Admission matches source/configuration/population seals, ordered manifests and R0/cropR1 recipe identity, physical states, dense original labels and materialized operational-teacher candidates before calculation. InterruptedQhStudy preserves original and evaluator/method provenance separately. Candidate means reduce within state, then states within recipe, recipes within physical scene and scenes equally. Its exploratory paired intervals use 20,000 shared physical-scene draws, seed 20261009, holding weights and checkpoint selection fixed. Candidate-pooled feasibility and calibration bins remain separate descriptive products. Within-state AUROC compares operational-teacher positive/negative probabilities with ties weighted one half; missing classes retain explicit unavailable support. Its scene contrast uses the same paired draws and higher values indicate better discrimination. Threshold summaries use 0.5 without validation fitting. No geometric-safety, policy-value or joint-versus-score-only superiority claim follows. Regression/cold-readback proof: tests/reporting/test_interrupted_qh_report.py.

Interrupted H1 utility diagnostics remain separate from the established regret, calibration and teacher-discrimination results. Signed bias uses original FP32 prediction-minus-label residuals. Retained informative-pair counts are checked against the producer’s FP32 arithmetic; strict sign agreement gives tied predictions zero correctly ordered pairs. Within-state Pearson correlation is descriptive, with constant labels/predictions and absent informative pairs explicitly unavailable. No unsupported state is dropped to manufacture a full-population interval. States reduce within recipe chains, chains equally within recipes, recipes equally within physical scenes, then scenes equally. Shared paired scene draws hold weights and historical selection fixed. Recording-state summaries carry actual candidate, state, chain, scene-recipe and physical-scene support; they are correlated descriptive strata. Positive pair-correctness contrasts favour interrupted weights, while signed bias is a calibration diagnostic: a common state offset cannot change argmax. The reporting section exposes additive utility_* tables, utility_contrasts and recording_state_diagnostics result IDs through the existing ReportRequest.

Recording-state figures retain paired physical-scene trajectories alongside equal-scene means, with scene/recipe/chain/candidate and metric-specific support in their source table. Missing scene metrics remain missing and plotted gaps are never connected. These observations show scene heterogeneity without pretending that recording states or candidate rows are independent bootstrap units.