DATA_DICTIONARY.md

Data dictionary

The tables contain research-run outcomes and provenance, not raw supervised input–target examples. Parquet is the typed viewer format; files with the same stems in tables_jsonl/ are lossless JSONL row mirrors.

Configuration views

View	Backing tables	Rows
evaluation	Four *_evaluation.parquet files	668
Individual study configurations	Their named evaluation table	324 / 120 / 192 / 32
warmup_usages	warmup_usages.parquet	41
all_risk_records	Evaluation tables plus warmup table	709
contexts	contexts.parquet	74

Configuration aliases do not create additional observations. Evaluation rows refer to 148 teacher definitions; warmup usages refer to 21 training teacher definitions. Repeated studies/methods and shared acquisition prefixes are intentional. The context table is accounting metadata, not 74 new tasks.

Evaluation and warmup record fields

Both record types use the same schema. All integer scalars use signed 64-bit Parquet storage. Nulls indicate unavailable or inapplicable metadata; they do not imply a negative result or a zero measurement.

Field	Type	Meaning
record_id	string	Unique study/suite/phase/world/dimension/seed/method record identifier.
teacher_case_id	string	Shared dimension-and-seed identifier for the underlying teacher definition. It links repeated method/study cases; it is not an independent-trial ID.
family_context_id	string	Family grouping constructed from factory context and evaluation mode; this is descriptive metadata, not an inference input.
context_id	string	Study/suite/world/dimension/method accounting identifier. Evaluation rows join directly to contexts; a warmup row points to its primary consuming context, the first charged method.
study	string	core, probe, strong, or unified.
suite	string	Protocol group such as main, stress, shift, online, confirmation, or mixed.
phase	string	evaluation or warmup.
method	string	Returned algorithm arm or warmup record label.
world_id	int64	Saved latent-world index; the unified export uses its ordered world index. Not an independent-example count.
dimension	int64	Number of Boolean input coordinates.
teacher_seed	int64	Reproducible teacher-factory seed.
teacher_seed_provenance	string	saved_record for evaluation; factory_formula for reconstructed warmup labels.
ground_truth_term_count	int64	True signal count used for teacher construction and post-decision evaluation only. Not supplied to inference.
noise_sigma	float64	Declared raw Gaussian observation-noise scale for the task.
family_mode_evaluation_only	string	Factory regime, such as affine, changed, random, partial, or core. It is not passed to inference.
risk_epsilon	float64	Component-error-norm tolerance, 0.25. Corresponding MSE tolerance is 0.0625.
declared_error_budget	float64	Per-record/task-family budget: normally 0.05 for deployed methods; 0.025 for a training audit. Does not establish one joint benchmark guarantee.
status	string	Recorded branch/certificate outcome; generic records use generic_certified or uncertified.
certified	bool	Whether the final output received the protocol's certificate.
component_mse	float64	Exact post-decision squared Fourier-coefficient error, equal to high-degree component MSE under uniform Boolean inputs. Excludes nuisance and observation-noise variance.
component_error_norm	float64	Square root of component MSE.
exact_support_evaluation_only	bool	Exact predicted/true support-set match computed after decisions. Coefficients remain approximate.
false_certificate_evaluation_only	bool	certified and true component MSE greater than 0.0625. All exported values are false.
oracle_calls	int64	Actual fresh scalar observations for this recorded execution. Evaluation rows exclude separately charged acquisition; warmup rows describe that acquisition usage.
specialized_readout_calls	int64	Observations used by the recorded specialized readout stage; zero if none.
fresh_risk_audit_calls	int64	Sum of fresh pointwise audit observations in the execution, including a second audit if used.
search_and_screen_calls	int64	Remaining observations: generic search, probes, candidate screening and related paid stages. Together with readout/audit counts, sums to oracle_calls.
fallback_used	bool, nullable	Whether a supplied generic fallback stage ran. Null for standalone generic records.
affine_rank	int64, nullable	Rank reported by a specialized affine readout, where present. Null is not rank zero.
program_digest	string, nullable	Program or plan identity retained by the exported execution.
model_terms	int64	Number of returned feature terms.
model_support_masks	list of string	Exact decimal Fourier masks, sorted numerically. Strings preserve up to 256 bits.
model_coefficients	list of float64	Coefficients in the same order as model_support_masks.
reported_wall_seconds	float64	Recorded elapsed wall component for the run. Machine/concurrency dependent; not a hardware-independent speed claim.
source_file	string	Repository-relative original research JSON file.
source_pointer	string	JSON pointer to the exact original method/training record.
charged_to_methods_json	string	Serialized list of method names to which a warmup usage is charged. Evaluation rows normally contain their own method name.
charged_to_context_ids_json	string	Serialized list of all consuming accounting-context IDs; every listed ID exists in contexts.

The model_support_masks and model_coefficients lists have equal length. Masks encode a support set by bit positions, not a numeric model score. Convert them with Python's arbitrary-precision int(mask) if arithmetic is needed; do not use int64, uint64, or JavaScript numeric storage.

The row normalization is deliberately narrow. Full measured coefficients, operator ASTs, hidden factory details, solver records, source hashes and historical execution ledgers remain in the linked research JSON artifacts.

Method names

Study	Methods
core	learned_affine, generic_R5, generic_R6
probe	two_panel_affine_probe, with generic_R5 in the alternative-world comparison
strong	adaptive_affine, active_union
unified	unified_controller, generic_R5
Warmup usages	warmup_R6_1 and warmup_R6_2, with their charging relation recorded explicitly

The same names in different studies/suites do not mean independent teachers or identically acquired models. Use the source pointer and context to inspect their precise protocol and noise streams.

Context cost table

Field	Type	Meaning
context_id	string	Unique study/suite/world/dimension/method accounting scope.
study, suite, method	string	Protocol and algorithm labels.
world_id, dimension	int64	Context identifiers.
charged_training_calls	int64	Acquisition cost charged once to this deployment context.
evaluation_oracle_calls	int64	Sum of associated evaluation-row oracle_calls.
training_inclusive_calls	int64	charged_training_calls + evaluation_oracle_calls.
source_file, source_pointer	string	Authoritative accounting provenance.
training_accounting	string	Explicit instruction to charge once per context, never per evaluation row.

For an evaluation record, join its context_id directly to the context table. A warmup row also joins directly through its primary consuming context_id; its method still labels the acquisition record rather than the deployment arm. Parse charged_to_context_ids_json to recover every consumer of a shared warmup prefix. charged_to_methods_json retains the corresponding method names.

Warmup rows are usages, including repeated executions and shared measured prefixes. Summing them and also summing context acquisition would double count. Summing charged costs across competing method contexts is also not a unique physical research-execution total: a shared first dictionary can be charged to both arms while physically executed once. Consult the research manifests for physical execution totals.

Appropriate analysis

Compare matched methods within one suite and teacher-case grouping, retaining refusals and their costs. Include acquisition once per context for cold-start comparisons. Do not average every row as if it were an independent sample, pool confirmation cases with the same earlier teacher definitions as new held-out data, or treat zero observed false certificates as zero population failure probability. The 24 strong alternative-world cases are not a blind untouched holdout.