The method-specific class that connects a paper implementation to common
model construction, training, evaluation, and optimizer hooks.
Auxiliary asset
A non-feature input required by a method, such as a cross-scale map, clinical
report, attribute tensor, or prompt embedding.
Backbone
A registered runtime encoder implementation. This is distinct from the
offline model that produced a cached feature store.
Capability
An operation exposed by an encoder bundle, such as text encoding, soft
prompting, or paired tile-text projection.
Cohort
One annotation-defined collection of slides/cases with a fixed task and
label order.
Config validity
Whether a generated run is internally consistent and satisfies the adapter's
declared contract. Validity does not require files to exist.
Experiment variant
A reportable method configuration with a fixed feature source, resolution
role binding, prompt mode, and recipe. Encoder or projection changes receive
separate variant names.
Feature level
The semantic input type consumed by a method: patch bag, dual-scale bags,
patch sequence, slide embedding, raw tile directory, or composite input.
Feature role
A method-facing name such as bag, low, high, or tiles. Roles are
mapped to independently registered feature sources.
Feature space
The representation produced by one exact encoder/checkpoint. It is tracked
by feature_space_id; tensor width alone does not define it.
Fold
One outer patient/case-disjoint train, validation, and test partition.
Input kind
The physical/semantic storage contract, including patch_bag,
slide_embedding, patch_sequence, or raw_tile_directory.
Native projection
A paired projector released with a model that maps its own visual and text
representations into a shared space. It cannot be borrowed by an unrelated
slide encoder.
Prompt encoder
The runtime text tower used to encode class descriptions or learn soft
context. It may differ from the offline visual feature producer when the
method explicitly defines an alignment.
Prompt profile
The canonical, task-level description source compiled into method-native
prompt assets.
Protocol
The source-of-truth YAML defining cohorts, features, experiments, folds,
shots, prompts, and output conventions.
Readiness
Whether every metadata, split, feature, weight, prompt, and auxiliary file
needed for a valid run currently exists.
Resolution
A declared semantic magnification such as 5x, 10x, or 20x. low and high
are roles and are not hard-coded magnifications.
Run config
One generated YAML passed to an adapter invocation. It records resolved paths
and provenance for one experiment, cohort, and shot level.
Shot
The number of labeled training patients/cases per class, with the same count
used for validation under the standard protocols.
Smoke test
Isolated model construction plus a method-appropriate dummy forward pass. It
validates wiring and finite output shape, not scientific performance.
Upstream reproduction
A run that follows the released task, architecture, features, prompts, and
recipe closely enough to claim reproduction.
Framework extension
A valid adapter run on a new task, prompt source, encoder, projection, or
recipe that was not released upstream.