CDAtlas

Evidence, not just links

See what constrained decoding actually specifies, guarantees, and costs.

A machine-readable map of structured generation across formal languages, probabilistic inference, serving systems, benchmarks, and runtime enforcement.

Atlas v0.2 · transparent evidence status · no copied abstracts

01SpecifyWhat property?
02CompileWhat representation?
03AllocateHow does search change?
04EnforceWhere and when?
05EvaluateWhich guarantee or cost?
—catalog records
—peer reviewed
—full texts reviewed
—linked artifacts

Interactive catalog

Filter the field by the property you care about.

Loading…

Evidence map

What the current literature reports—and leaves blank.

README snapshot ↗

Charts use only fully coded records. Discovery records remain searchable but are excluded from this synthesis.

Constraint specification

Enforcement placement

Guarantees

Systems measurements

How to read the Atlas

Four separations prevent most category errors.

01

Validity ≠ quality

A parseable string can still be semantically wrong. We code structural and task outcomes separately.

02

Soundness ≠ coverage

Emitting only grammatical strings does not prove that every intended string remains token-reachable.

03

Local mask ≠ conditional sample

Stepwise renormalization is not automatically the base sequence distribution conditioned on validity.

04

Proposal validity ≠ safe commit

Structured output and runtime authorization operate on different properties and lifecycle boundaries.