← Back to Overview
The Generator · Python · OpenAI & Gemini

The Puzzle Generator

A target-aware Python construction system for crossword fill, editorial validation, publisher clue workflows, repair, tracing, and immutable evidence. Accepted output and human release review remain distinct.

3 RoutesVertex, Gemini, OpenAI Capable
Worst EntryExecutable Quality Contract
5Registered Targets
Human ReviewRequired for Release

Multiple Model Provider Routes

Publisher clue and review code can route through Vertex, Gemini, or OpenAI credentials. Provider models, quotas, behavior, and pricing must be verified at operation time.

AI
OpenAI-compatible client path

Used by configured publisher stages with caching and structured validation. No single model name or cost is guaranteed by this page.

G
Google Vertex and Gemini paths

Credential-based routes exist in the model client. Callers still own prompt, schema, retry, budget, and editorial acceptance behavior.

Model clients and callers include error-handling paths, but quota exhaustion, retry behavior, partial artifacts, and spend controls must be verified for the selected route.

Target Owned Quality Gates

Current targets use executable structural, lexical, theme, worst-entry, and publisher policies. The exact gate and receipt for the selected target are authoritative.

#CheckThresholdWhy It Matters
01Target ContractTarget-specificStructure and policy belong to the selected publisher or product target.
02Grid IntegrityExecutableEntries, crossings, dimensions, and symmetry are checked by code.
03Worst-Entry QualityNo waived bad fillThe primary track rejects crosswordese, partials, and low-quality entries.
04Theme PolicyWhen applicableThemed targets enforce answer, placement, revealer, and evidence requirements.
05Release EvidenceImmutable receiptAcceptance and release decisions remain auditable.

From Inputs to Release Evidence

Generation, acceptance, clueing, export, human review, and deployment are separate stages; a target may not use every stage.

Template Selection 15×15 grid with 180°
rotational symmetry
Grid Filling Algorithmic fill from
configured lexicon sources
Target Quality Gate Worst-entry, structure,
theme and policy checks
AI Clue Generation OpenAI or Gemini,
with clue caching
Evidence & Release Receipt, human review,
then optional deployment

Supporting Metadata and Publication Tools

Supporting modules can classify, title, resume, and publish artifacts. Availability and authority vary by target, and release remains a reviewed operation.

  • Legacy Difficulty Classifier

    Easy / Medium / Hard classification exists for legacy app content. It is not the primary NYT-quality success contract.

  • Optional Puzzle Titling

    Some routes can propose titles through model-assisted stages. Editorial review and target policy determine whether a title is usable.

  • Resume-Aware Runners

    Selected runners record state or inspect artifacts to support resumption. Recovery guarantees must be verified for the exact command and storage path.

  • GCS Publisher

    The publisher writes puzzle objects and an index to configured Google Cloud Storage. Release approval, credentials, bucket state, and client delivery remain separate concerns.

Browse Real Puzzles From the Library

These repository examples demonstrate puzzle structure and presentation. Each artifact needs its own target receipt and provenance before it can be described as currently accepted or publishable.

Interested in the Generator?

The generator is included in the full acquisition. Reach out to discuss terms.

← Return to Overview