LLM-Assisted Content Enrichment and Theming¶
Status: Exploratory design; not scheduled and not a description of an online runtime dependency.
Audience: Product, editorial, trust and safety, and lexical-platform reviewers.
This is a living design and idea document for using language models to enrich MiniGames lexical content, support editorial themes, and assist content moderation. It records current capabilities, proposed architecture, decisions, risks, experiments, and open questions. Ideas in this document are not commitments or descriptions of shipped behavior unless explicitly marked as current.
Related architecture:
Product vision¶
The lexical catalog should eventually describe more than a mapping from a term to one definition. It should be able to support games involving meaning, associations, language variation, wordplay, and timely editorial themes. Potential content includes:
- Multiple senses and original glosses for one term.
- Phrases, idioms, collocations, quotations where rights permit, and common speech patterns.
- Morphology, inflections, register, dialect, pronunciation, and parts of speech.
- False friends and false cognates across languages, with precise relationship labels.
- Semantic topics, entities, events, moods, eras, genres, and cultural associations.
- Wordplay attributes such as puns, homophones, hidden words, transformations, and rebus potential.
- Difficulty, familiarity, ambiguity, offensiveness, sensitivity, and age-suitability signals.
- Time-bounded editorial collections for holidays, launches, eclipses, releases, and internet events.
LLMs can help propose and organize this material at scale. They must not become an unversioned, unreviewed source of truth or an online dependency of ordinary puzzle generation.
Current foundation¶
The current catalog already provides useful seams for future enrichment:
RawLexicalEntrymay carry case-insensitive string tags from a source.- Tags from multiple sources are unioned when normalized entries merge.
LexicalEntryretains immutable tags alongside forms, graphemes, scores, and provenance.- Tags are included in the catalog content fingerprint.
- Eligibility profiles may require tags.
- A theme-aware eligibility context can require an entry to contain a matching tag.
- Catalog views and game indexes reference catalog entries rather than copying the lexical corpus.
Current limitations are equally important:
- Production sources provide little thematic metadata.
- Tags are untyped strings with no confidence, evidence, hierarchy, validity interval, or review state.
- Reverse Cross production generation currently rejects theme selection.
- Letter Golf accepts a legacy theme parameter, but its production catalog index is unthemed.
- No LLM classification, editorial review, theme manifest, or enrichment publication pipeline is implemented.
- The current generators do not distinguish theme anchors from neutral fill.
Working principles¶
- Run LLM enrichment offline or asynchronously, never in the latency-critical generation path.
- Treat model output as a proposal requiring deterministic validation and, where risk warrants, editorial approval.
- Keep lexical identity stable. Enrichment should normally refer to catalog entry IDs rather than produce another copy of each term.
- Version every published artifact by content and record its model, prompt, policy, and source lineage.
- Separate stable lexical facts from temporary editorial collections.
- Keep game rules game-specific. Shared enrichment supplies metadata; each game decides how that metadata affects selection, generation, clues, validation, and scoring.
- Preserve replayability. A puzzle must record the exact catalog, profile, enrichment/theme artifact, generator options, and random seed used to create it.
- Do not infer safety from topical relevance. Themed content passes the same moderation and rights gates as ordinary content.
- Never copy protected definitions, clues, articles, scripts, or puzzle content into prompts or outputs without verified rights.
Proposed responsibility flow¶
licensed and first-party lexical sources
|
v
normalization + deterministic moderation + provenance
|
v
immutable WordCatalogSnapshot
|
+--> offline enrichment jobs --> candidate assertions
| |
| v
| validation + review
| |
| v
| versioned enrichment artifact
| |
+---------------------------------+
|
+--> stable game profile/view --> game-specific index
|
+--> theme manifest -----------> themed selection/generation policy
The catalog remains the canonical owner of normalized lexical entries. Enrichment artifacts and theme manifests refer to those entries by stable ID and add independently versioned assertions. This avoids rebuilding or duplicating the entire lexical corpus for each temporary event.
Metadata model direction¶
Simple tags remain useful for coarse, stable labels, but richer content should use typed assertions. A future assertion might include:
{
"entryId": "lex:...",
"kind": "topic",
"value": "astronomy",
"confidence": 0.94,
"evidence": ["source-reference-or-editor-note"],
"model": "provider/model/version",
"promptVersion": "lexical-topic-v3",
"policyVersion": "enrichment-policy-v2",
"reviewState": "approved",
"createdAt": "...",
"validFrom": null,
"validUntil": null
}
Candidate assertion kinds include:
topic: astronomy, cooking, horror, elections, basketball.entity: NASA, a film, a person, a place, a fictional character.event: a launch, eclipse, release, anniversary, or tournament.tone: spooky, celebratory, absurd, nostalgic, formal.register: slang, archaic, technical, regional, informal.sense: a particular meaning of a polysemous term.relation: synonym, antonym, homophone, pun target, associated-with.phrase-type: idiom, proverb, catchphrase, collocation, speech pattern.language-relation: translation, false friend, false cognate, loanword.content-signal: sensitive category, age suitability, ambiguity, moderation concern.game-affordance: good seed word, clueable sense, rebus candidate, transformation candidate.
Typed assertions prevent one flat tag namespace from having to represent semantics, provenance, confidence, moderation, and editorial state at once.
Stable tags versus theme manifests¶
Stable tags are appropriate when the association is durable and broadly reusable:
A theme manifest is better for a curated or time-sensitive experience:
{
"themeId": "total-eclipse-2027",
"version": "theme-manifest-v1:<content-hash>",
"language": "en-US",
"validFrom": "2027-07-01T00:00:00Z",
"validUntil": "2027-08-31T23:59:59Z",
"anchorEntryIds": ["lex:..."],
"relatedEntries": [
{ "entryId": "lex:...", "weight": 0.91, "roles": ["seed", "answer"] }
],
"editorialStatus": "approved",
"catalogVersion": "word-catalog-v1:...",
"enrichmentVersion": "lexical-enrichment-v1:..."
}
This distinction matters operationally. Adding or changing tags embedded in LexicalEntry changes
the catalog content fingerprint and causes dependent views and indexes to rebuild. Publishing a
small theme manifest can instead select and weight existing entry IDs without changing the base
catalog. A manifest must declare the catalog/enrichment versions against which its IDs were
validated.
Potential LLM roles¶
Categorization and association¶
- Propose topic, entity, event, mood, era, genre, and register classifications.
- Rank the strength and obviousness of an association instead of returning only a boolean label.
- Distinguish literal membership from looser cultural association.
- Suggest theme anchor entries and secondary fill candidates.
- Cluster candidate terms and identify outliers that weaken theme coherence.
Lexical expansion¶
- Propose phrases, idioms, inflections, collocations, speech patterns, and emerging terms missing from licensed or first-party sources.
- Find likely source records that normalize to the same entry.
- Suggest language variants and relationships while preserving their distinct forms and meanings.
- Identify false friends, false cognates, loanwords, and translation traps, using precise labels rather than treating every cross-language resemblance as a cognate.
LLM proposals are not automatically admitted. New entries need evidence, rights provenance, normalization, duplicate resolution, moderation, and factual review. Generated definitions and clues must be original and checked against authoritative sources; the system must not reproduce protected dictionary or puzzle text.
Sense and clue enrichment¶
- Propose multiple senses for ambiguous entries.
- Write candidate original glosses at controlled reading levels.
- Suggest clue angles, misdirection opportunities, wordplay, and alternate interpretations.
- Identify whether a clue depends on current events, a particular locale, or ephemeral knowledge.
- Detect when two supposedly different clues collapse to the same sense.
Definitions, senses, and clues should be separate versioned entities. One term can have several senses; one sense can support several clues; one clue can be suitable for only certain games, audiences, dates, or difficulty bands.
Moderation assistance¶
- Propose sensitivity categories and contextual usage warnings.
- Detect sexual, hateful, violent, self-harm, drug, harassment, and personally identifying content.
- Distinguish a benign sense from an unsafe or age-inappropriate sense of the same spelling.
- Flag named people, allegations, medical claims, political persuasion, and other high-risk content for mandatory human review.
- Re-evaluate existing content when moderation policy changes.
An LLM moderation result is a signal, not the sole gate. Deterministic deny rules, source policy, game/audience policy, and human escalation remain independently enforceable.
Editorial theme development¶
- Turn an editor's concept into candidate anchors, related vocabulary, clue angles, and exclusions.
- Follow timely events from an approved factual input feed rather than unconstrained web rumors.
- Score candidate sets for coherence, novelty, diversity, solvability, and sensitivity.
- Explain weak associations so an editor can reject them quickly.
- Produce several theme treatments: literal, accessible, expert, pun-heavy, or mechanically unusual.
Game-specific mechanics¶
Reverse Cross¶
An editorial-style Reverse Cross should usually combine a small number of strongly thematic answers with ordinary crossing fill. Requiring every answer to share a topic would often make a grid impossible or produce poor vocabulary.
Potential mechanics:
- Require one or more anchor answers before filling the remaining slots.
- Prefer related entries using weighted candidate ordering without excluding neutral fill.
- Reserve selected long or symmetric slots for theme entries.
- Score completed fills for theme density and association strength.
- Support reveal entries that connect apparently unrelated anchors.
- Add transformations, rebuses, puns, homophones, hidden phrases, or repeated clue mechanics.
- Record the selected theme manifest and anchor IDs in replay identity.
The positional fill index should remain optimized for (length, position, grapheme) constraints.
Theme membership can be represented by a compact eligible-entry set or weight table layered over
that index. A separate full positional index per short-lived theme would waste memory unless
measurement demonstrates that theme filtering during search is too slow.
Letter Golf¶
Letter Golf should normally use theme metadata to choose or weight the seed-word pool. Once a puzzle exists, the full eligible dictionary should continue validating player words unless the game mode explicitly tells players that only themed answers are accepted.
Potential mechanics:
- Halloween, space, sports, music, or event-themed seed words.
- Daily topical weighting with cooldowns to avoid repeated seeds.
- Theme-specific achievement or bonus words while ordinary words remain legal.
- A constrained mode in which each played word must match an announced relation, with separate validation and clear user-facing rules.
- Semantic progression, such as moving through related senses or categories, for a future variant.
The LetterSignature is one word or puzzle's grapheme-count value. WordSignatureIndex is the
shared searchable structure over all eligible Letter Golf entries. A theme should filter or weight
entry IDs around that index, not create a new signature definition.
Future games¶
Enriched content could support:
- Guessing a word from progressively more specific senses.
- Connecting idioms by shared metaphor or origin.
- Detecting false friends across languages.
- Building phrase chains from collocations or speech patterns.
- Sorting terms by topic, register, era, or semantic relation.
- Games based on alternate definitions, ambiguity, homophones, and transformations.
- Event-driven daily challenges selected from approved manifests.
Each new game should define which assertions it trusts and build only the derived index its own query pattern requires.
Publication and versioning¶
Every enrichment artifact should have a deterministic content version. Its fingerprint should cover canonicalized assertions and the inputs that affect their meaning:
- Artifact schema version.
- Base catalog version.
- Model/provider/version identifier.
- Prompt and toolchain versions.
- Enrichment and moderation policy versions.
- Source/evidence references.
- Review decisions.
- Canonically ordered accepted assertions.
The base catalog fingerprint remains one hash for one complete language-scoped catalog snapshot. Game views have separate fingerprints based on catalog version, profile, and context. Game indexes have cache identities based on their view/catalog inputs and index options. Board dimensions, puzzle seed, and a temporary theme do not inherently create new base catalog fingerprints.
A themed puzzle replay identity should eventually include:
catalog version
game profile ID and version
enrichment artifact version
theme manifest ID and version
game index options version
generator options
random seed
Validation and quality gates¶
An enrichment candidate should pass checks appropriate to its assertion type:
- Schema validation and canonical serialization.
- Entry-ID existence against the declared catalog version.
- Language and grapheme consistency.
- Duplicate and contradiction detection.
- Confidence thresholds calibrated against a reviewed benchmark.
- Deterministic moderation and deny-list checks.
- Evidence availability and source-rights validation.
- Temporal validity checks for event-driven material.
- Human review for high-risk categories and public publication.
- Game simulation to measure solvability and puzzle quality, not only semantic relevance.
Suggested evaluation sets:
- Reviewed positive and negative examples for each topic and relationship type.
- Ambiguous words with multiple benign and sensitive senses.
- False-friend versus true-cognate distinctions.
- Timely themes with near-topic distractors.
- Theme sets that are coherent but impossible to use in a grid.
- Prompt-injection and malformed-source cases.
- Regression sets for previously rejected unsafe, defamatory, or low-quality output.
Useful metrics include precision by assertion type, editor acceptance rate, unsupported-claim rate, moderation false-negative rate, theme coherence, lexical diversity, puzzle generation success rate, generation time, and repeat frequency. Raw lexical content and private puzzle solutions must not be placed in high-cardinality telemetry labels.
Security, privacy, rights, and safety¶
- Send only approved data to a model provider and document retention/training terms.
- Treat source text, retrieved event material, and model output as untrusted input.
- Isolate prompts from instructions embedded in source content.
- Do not send user-private data, unpublished puzzle solutions, credentials, or denied-term lists to external models without an explicit approved design.
- Keep evidence and provenance for factual, cultural, and linguistic assertions.
- Require review for living people, allegations, politics, health, protected classes, tragedy, and other high-impact topics.
- Respect source licenses and attribution. A model transformation does not erase upstream rights.
- Avoid close paraphrases of copyrighted dictionary definitions, clues, articles, scripts, lyrics, or existing puzzles.
- Support takedown, correction, expiry, and rollback by artifact version.
Operational direction¶
The preferred first implementation is a bounded batch job producing reviewable artifacts, not a model call from the API:
- Select a bounded set of catalog entry IDs and an enrichment task.
- Produce structured candidate assertions with model and prompt metadata.
- Validate syntax, IDs, evidence, safety, and confidence.
- Present uncertain or high-impact candidates for editorial review.
- Canonically serialize and content-hash accepted output.
- Publish the artifact atomically.
- Load it alongside a compatible catalog snapshot.
- Build or update only the game-specific derived data required by consumers.
- Measure game outcomes and feed reviewed failures into evaluation sets.
Artifact publication should preserve the last known good version on failure, support audit and rollback, and avoid holding two complete lexical corpora in memory. Expensive builds should use the same single-flight and atomic-publication principles as the catalog and game indexes.
Candidate delivery phases¶
Phase A: vocabulary and contracts¶
- Define controlled assertion kinds and namespaced stable tags.
- Specify enrichment and theme-manifest schemas.
- Define provenance, confidence, review, expiry, and versioning requirements.
- Build small hand-reviewed fixtures for Halloween and astronomy.
Phase B: offline categorization experiment¶
- Classify a bounded sample into a few low-risk topics.
- Compare at least one LLM-assisted approach with deterministic/reference baselines.
- Measure precision, reviewer acceptance, cost, latency, and disagreement.
- Publish no user-facing content until thresholds are agreed.
Phase C: editorial theme manifests¶
- Add an internal workflow for proposing, reviewing, versioning, and expiring manifests.
- Implement theme-aware seed weighting for Letter Golf without changing legal-word validation.
- Record manifest identity in generated puzzles and telemetry.
Phase D: Reverse Cross theme anchors¶
- Add required/preferred theme-entry constraints around the existing positional index.
- Evaluate fill success and quality with neutral crossing fill still available.
- Add deterministic replay and bounded fallback when a theme cannot produce a valid puzzle.
Phase E: richer lexical relationships¶
- Add senses, original glosses, idioms, phrases, false-friend/cognate distinctions, and speech patterns as typed, sourced assertions.
- Expose only reviewed relationship types to games.
- Add game-specific indexes as actual query needs emerge.
Decisions to date¶
- The catalog remains canonical for normalized lexical identity and provenance.
- LLM processing belongs outside the synchronous player request path.
- Model output is proposed enrichment, not authoritative lexical data by default.
- Stable metadata and temporary editorial themes are separate versioned artifacts.
- Games consume shared enrichment but retain their own generation, validation, and indexing rules.
- Reverse Cross themes should use anchors plus neutral fill rather than require every entry to be thematic.
- Letter Golf themes should initially influence seed selection, not reject otherwise legal plays.
- Replay identity must include every enrichment or theme artifact that could affect output.
Idea backlog¶
- Annual seasonal packs with freshness controls and regional variants.
- Event feeds for launches, eclipses, major sporting events, and release calendars.
- Theme intensity settings from subtle to explicit.
- Editor-defined exclusions and anti-associations for misleading near-matches.
- Association explanations shown only to editors, not necessarily players.
- Difficulty inferred separately for recognition, spelling, clue interpretation, and cultural knowledge.
- A novelty score balancing current relevance against overused crossword fill.
- Theme graphs connecting entries through typed relations rather than a single shared label.
- Automatic expiry for memes and event associations, with manual promotion to durable metadata.
- Generation of candidate original clues followed by independent factuality and similarity checks.
- Multilingual themes that intentionally use translations, loanwords, or false friends.
- Speech-pattern games based on fillers, discourse markers, idiomatic constructions, or regional phrasing.
- Controlled community theme proposals that enter the same moderation and review pipeline.
Open questions¶
- Which assertion types are valuable enough to justify first-class schemas rather than tags?
- Which themes require human approval, and which low-risk classifications may be auto-published?
- What evidence is sufficient for cultural association or an emerging term?
- Where should enrichment artifacts and theme manifests be stored and promoted between environments?
- How should corrections and deletions propagate to already-generated daily puzzles?
- What confidence calibration and reviewer agreement are required per assertion type?
- How should regional, dialect, age, and accessibility differences affect theme selection?
- Which game should host the first user-visible experiment, and what is the smallest reversible version of that experience?
- How much neutral fill is acceptable before a Reverse Cross theme feels too weak?
- When does filtering an existing game index become slow enough to justify a theme-specific secondary index?
Change log¶
- 2026-08-28: Initial document. Captured the offline enrichment boundary, stable tags versus theme manifests, game-specific mechanics, moderation and rights constraints, versioning, phased delivery, and the initial opportunity backlog.