Glossary

Core terms

Term	Meaning
Limen	Vaquum's Bitcoin research engine for Python and Compiled experiments, analysis, backtesting and reusable Sensors and Cohorts.
Python Research	Caller-owned data, preparation, models, backtesting and selection in ordinary Python; no Manifest or built-in-model whitelist is required.
Compiled Research	A YAML experiment executed through the CLI, with its configured validation, preparation and evidence contracts.
YAML manifest	A Compiled experiment definition validated, compiled and executed by the `limen` CLI.
`Manifest`	The abstract experiment-pipeline interface. Runnable pipelines use `MLManifest` or `RuleBasedManifest`.
`MLManifest`	Manifest implementation for trainable reference architectures.
`RuleBasedManifest`	Manifest implementation for predicate-driven, non-learned strategies.
Single-File Decoder	A Python experiment unit that exposes a parameter surface and either a manifest or custom prep/model functions. Previously called Single File Model.
SFD	Short form of Single-File Decoder. Previously called SFM.
Foundational SFD	A packaged SFD under `limen.sfd.foundational_sfd`; most have a matching YAML template.
Reference Architecture	A class-based model implementation and function wrapper used by foundational manifests.
Universal Experiment Loop	The experiment engine that executes parameter rounds.
UEL	Short form of Universal Experiment Loop.
round / permutation	One concrete parameter combination executed by UEL.
experiment artifact	A persisted run output such as `metadata.json`, `results.csv`, `round_data.jsonl.gz`, `checkpoint.json`, or `audit.jsonl`.
`Trainer`	Reconstructs and replays selected rounds from a saved YAML experiment, validates metrics, and returns Sensors.
`Sensor`	A native predictor wrapping fitted state and inference behavior. Python Sensors preserve caller-owned preparation, cutoffs and output meanings; Compiled Sensors carry their reconstructed Manifest and recorded identities.
`Cohort`	Compatible Sensors composed with an aggregation rule, timestamp checks and unavailable-result handling. Recorded Compiled membership can additionally bind artifact identities.
selector	For artifact-backed Compiled composition, a callable choosing string permutation IDs. Python callers can select their fitted models directly before constructing Sensors.
`HistoricalData`	Stateful data-access class for Limen's supported market-data sources and files.
reducer	A `PruningStrategy` that analyzes completed rounds and returns declarative search interventions.
`MSQ`	Mutable Search Queue inside UEL, used by CLI and Python search-strategy runs to select candidates and apply optional feedback.
`Log`	Analysis wrapper over an experiment log and retained round artifacts.
benchmark	Outcome analysis of retained research results. Confusion metrics and advisory walk-forward selection reports have different evidence requirements; neither creates statistical acceptance or an independent leaderboard.
backtest	Evaluation under an explicit trading policy. Limen offers a snapshot ledger and an event execution engine; Python callers can supply their own backtester. State the execution assumptions and measured units.
short/flat snapshot	Explicit binary short-return approximation through `short_flat_strategy`; Backtest defines fills, costs and exclusions.

Observation and decision boundaries

Term	Meaning
fitting cutoff	Latest observation permitted to affect fitted model or preparation state. A later inference call does not extend it.
selection cutoff	Latest observation permitted to affect the choice of candidate, threshold or membership. It can be later than fitting.
qualification	An explicit check of whether a candidate or result meets declared conditions. Passing does not make reused validation observations untouched.
final holdout	Observations excluded before research and used after the candidate is frozen. The native reservation records local exposure history; caller-owned exclusion records only what the caller actually enforces.
availability time	Earliest time an observation or prediction can be used. A bar's timestamp alone does not prove that its closing value was already known.
reconstruction	A validated refit and replay of recorded native behavior. It is different from retaining the original evaluated Python object and from restoring arbitrary model weights.

Implementation names

limen.experiment.manifest_core contains the manifest implementations, and limen.experiment.experiment_core contains Universal Experiment Loop. These are module names, not a general *_core naming promise for other domains.

Core terms

Observation and decision boundaries

Implementation names

Read next