Glossary

Use this glossary when architecture, sweep, artifact, and workflow terms start to carry too much repo-specific meaning.

Anchor

The locked comparison run for a sweep. New rows are judged against the anchor unless the queue row explicitly declares a different preserved surface.

Architecture-Screen Surface

The canonical benchmark-facing sandwich surface used for current architecture work. Historical staged surfaces still exist for comparison, but they are no longer the default landing zone for new architecture evidence.

Benchmark Bundle

A pinned benchmark task set used to compare runs consistently. It defines the shared comparison surface for benchmark-facing claims.

Benchmark-Full

A sweep execution policy that continues past screening, runs the pinned benchmark bundle, registers the run, and requires a result_card.md.

Checkpoint

A saved training state that can be evaluated, compared, resumed, or exported.

Control Lane

The frozen PFN-style comparison lane:

• tabfoundry_simple
• tabfoundry_staged with stage=nano_exact

Use it for trust and comparison, not as the default landing zone for new architecture work.

Corpus Ref

A stable identifier for a materialized corpus record. Public training docs prefer data.corpus_ref over a raw manifest path for recurring dagzoo-shaped corpora.

Dagzoo Synthetic-Data Lane

The synthetic-data source lane used when tab-foundry reasons about generated corpora. It is separate from curated real-data benchmark ladders.

Delta

One named change package in the system-delta workflow. A delta is the logical unit a queue row applies against the anchor.

Export Bundle

A packaged inference artifact produced by this repo for downstream runtime use. In the current contract it contains a manifest and weights file.

Hybrid Diagnostic Lane

A staged surface that holds some PFN-adjacent assumptions fixed while isolating one question. Useful for diagnosis, but not automatically benchmark-facing promotion evidence.

Inference Handoff

The boundary where this repo stops at producing validated export artifacts and another repo or runtime system takes over serving or long-lived inference ownership.

Manifest

The concrete data/task description consumed by training, evaluation, and inspection flows.

Manifest-Backed Data

The concrete data surface the repo trains on. Public docs usually refer to it through data.corpus_ref; data.manifest_path remains the lower-level direct manifest override and inspection surface.

Model

The learned tabular network family being trained, evaluated, and compared in the repo.

Model Breadth

Capability expansion beyond the settled classification anchor, such as many-class support, regression, later modalities, inference handoff, or scaling-law breadth.

Prior Training

Legacy exact-prior training against the nanoTabPFN prior dump. In current docs this is the train legacy-prior lane, distinct from the corpus-backed default training surface.

Row

One queue entry in a sweep. A row isolates one declared dimension family and records the description, rationale, artifacts, metrics, interpretation, and next action.

Run Directory

The output directory for one training run. It usually contains histories, summaries, checkpoints, and other artifacts produced during the run.

Screen-Only

A sweep execution policy for diagnostic rows that stop after screening metrics. These rows do not register benchmark runs and do not create a result_card.md.

Sweep

A bounded research campaign with one active anchor, one queue, and one rendered matrix. Completed sweeps remain historical evidence rather than being mutated in place.