Skip to content

Reverse Cross

Status: Implemented product specification through the branch snapshot described in the Reverse Cross roadmap. Feature availability in production must be verified operationally.

Audience: Product, design, engineering, QA, and support.

Technical references: Reverse Cross generation ADR, Reverse Cross Evaluator Design, and Data Layout and Persistence.

Version history

  • 0.1: Initial version (JS)
  • 0.2: Product model, shared-content direction, Daily/Practice intent, and roadmap
  • 0.3: Bounded Phase 3 generation, positional indexing, replay identity, and fixture fallback
  • 0.4: Persistent Daily/Practice lifecycle, authoritative timing, and production player interaction
  • 0.5: Interruption-tolerant Daily pause, duration statistics, streaks, and deterministic replay
  • 0.6: Feature-gated God View diagnostics, bounded solution analysis, and Practice generation controls
  • 0.7: Strict generated Practice support and validation for 6x6 and 8x8 boards
  • 0.8: Daily generation rollout semantics, shared flag gating, and immutable persisted Dailies

Overview

Reverse Cross gives the player a bank of entries and a blank crossword-shaped grid. The player reconstructs the completed grid by placing every supplied entry Across or Down so that all crossings agree.

Despite sharing a product suite with Letter Golf, Reverse Cross has no connection to the sport of golf. It is a timed reconstruction puzzle: faster completed times are better.

The production player supports persistent Daily and Practice puzzles, explicit server Start, revision-aware autosave and resume, pause, and server-authoritative completion timing. The backend contains bounded on-demand generation behind the default-off ReverseCrossGeneration feature flag; while that flag is absent, disabled, or fails evaluation, it continues rotating among validated 5x5, 7x7, and 9x9 fixtures. Those sizes are presets, not a permanent square-grid constraint. The player supports rectangular puzzles within the configured row and column limits. Daily generation adds a second gate: a missing Daily date uses the production-default provider only when ReverseCrossDailyGeneration and ReverseCrossGeneration both evaluate true for the stable synthetic system identity system:reverse-cross-daily. If either flag is disabled or fails evaluation, or bounded generation cannot produce a puzzle, the missing date is persisted from fixtures instead. Existing persisted Daily puzzle documents are immutable cache hits and are never rewritten by later flag changes.

The implementation plan and current phase status live in the Reverse Cross roadmap.

Core rules

  • The public puzzle contains a grid mask and a bank of entries, but not the completed grid or answer placements.
  • A playable cell accepts one grapheme. A blocked cell accepts none.
  • Each supplied entry must be placed exactly once in one contiguous Across or Down slot.
  • An entry must exactly fill its slot. It cannot run through a blocked cell, extend beyond a boundary, or begin or end in the middle of a longer open run.
  • Graphemes at crossings must agree.
  • The puzzle is complete when every bank entry is used exactly once, every maximal slot is filled, and all crossings agree. A rule-valid arrangement need not match the generator's private witness.

The client and public contract key each bank entry by a stable entry ID, so duplicate display strings, if ever deliberately allowed, remain distinct.

Terminology

Term Meaning
Grid mask The public arrangement of playable and blocked cells.
Slot A maximal contiguous Across or Down run of playable cells.
Entry One item from the supplied word bank. It may eventually be a word, proper noun, abbreviation, or normalized phrase.
Grid form The normalized grapheme sequence placed into cells, such as ONTHEDL.
Display form The player-facing form, such as ON THE DL, when it differs from the grid form.
Solution witness A private completed grid and placement set proving the generated puzzle is solvable; it is not the only accepted arrangement.
Fixture A hand-authored or legacy hardcoded puzzle used before or alongside generated puzzles.

Puzzle modes

Daily

Daily gives every player the same puzzle for the Pacific-time calendar day. It uses the suite-wide puzzle-key convention:

reversecross|global|{yyyy-MM-dd}

The Daily puzzle document is read before any source decision. If that key already exists, the stored definition is returned unchanged to every player, including after later feature-flag changes. For a missing date, Daily evaluates ReverseCrossDailyGeneration, ReverseCrossGeneration, and ReverseCrossStructuralQuality with system:reverse-cross-daily, not the requesting player's identity. This prevents caller-specific Daily source or quality selection. A user-only allowlist will not enable Daily generation; operators should use global All On/Off changes for Daily rollout. A disabled or failed Daily gate, disabled or failed master generation gate, bounded generation outcome, catalog error, or concurrency rejection serves a fixture for the missing date.

The 14-day pre-generation horizon means enabling Daily generation after the horizon is already full of fixtures will naturally produce generated Dailies only as new dates roll into the horizon, roughly two weeks later. Operators can accelerate that only by manually removing reviewed future fixture puzzle documents so they are missing again; routine flag changes do not overwrite them.

Fetching a Daily must not start its clock. The player first sees a short ready screen and explicitly chooses Start; that server-recorded action establishes the competitive start. Progress autosaves and can be resumed on another visit by the same durable user identity.

A started Daily can be paused. Pause waits for pending board changes to persist, closes the current server-owned active segment, and hides the board and entry bank until the player explicitly resumes. When the page becomes hidden, the client obscures puzzle content immediately and makes a best-effort pause request after pending persistence settles. Returning to the page reconciles authoritative play state and never resumes automatically. Browser visibility is only a trigger; server lifecycle timestamps remain authoritative. The web client cannot prevent operating-system screenshots or other external capture.

The play contract exposes canPause so a future Strict or Tournament timing policy can disable pause without changing the normal Daily lifecycle.

Practice

Practice creates a fresh puzzle on demand with a key such as:

reversecross|practice|{shortid}

Practice supports save, resume, abandon, start-another, and explicit pause/resume flows with the same save-before-pause and hidden-content behavior as Daily. Practice completions do not affect Daily streaks, group standings, or Daily personal records. Abandon is a terminal lifecycle action rather than a scoring action. It is available from the paused Practice screen; creating another Practice remains the ordinary way to replace an active puzzle.

Practice may use validated fixtures while generated serving remains disabled or as its bounded fallback. The mode contract does not depend on how a puzzle was produced.

The ordinary New Practice control remains available to every Practice player and follows the production-default provider policy. Targeted playtesters can configure the next Practice request to use Production default, Fixture, or Generated, with an optional integer seed. Generated requests accept 5x5, 6x6, 7x7, 8x8, or 9x9 dimensions. Production default and Fixture remain limited to the validated 5x5, 7x7, and 9x9 presets. Production default retains the normal generated-with-fixture-fallback policy. Explicit Generated is strict: generation failure is reported instead of silently serving a fixture. These overrides are authorized on the backend and never change an existing Daily puzzle. Practice production-default and generated paths continue to evaluate generation and structural-quality flags with the caller identity; Daily uses the stable synthetic system identity because one persisted puzzle is shared by every player.

Timing and scoring

The primary score is active completion duration; lower is better. Paused intervals do not count. Display can be rounded to seconds, but persisted results retain milliseconds for deterministic ordering and an explicit tie policy.

The browser timer is presentation only. For a competitive Daily, the backend owns start, completion, and elapsed-time calculation. Completion requires the server to verify that submitted placements use the public bank exactly once, fill every maximal slot, and agree at crossings. The private persisted solution is a construction witness, not an exclusive answer key. Client clocks, browser visibility, and client-reported durations are not authoritative.

Possible future secondary metrics include mistakes, placement changes, hints, and completion rate. They should be stored separately rather than folded into an opaque score formula.

Player experience

The focused game screen should reuse the suite's shared identity, header, navigation, loading, error, and completion patterns without importing Letter Golf's visual metaphor or terminology. The board interaction must work with keyboard, pointer, and touch input and adapt to rectangular grids and small screens.

The player selects a slot and Across/Down direction, then selects an entry to place it immediately. An entry selected before its slot remains a cancellable preview until the player chooses a compatible slot and confirms it. A placed entry can be removed with the delete control and placed again; the player does not drag entries. Arrow keys move the single keyboard tab stop between playable cells. Invalid previews identify conflicting crossings without exposing the solution. After a successful placement, selection and keyboard focus advance to the next unfilled slot that can accept a remaining entry. Traversal continues in the current direction's natural reading order, wrapping once, then starts at the beginning of the other direction when needed. If no viable slot remains, the current selection is retained. Advancement never selects an entry automatically. During active play, Pause joins the direction and delete controls in one centered icon toolbar below the board. Each icon has an accessible name and tooltip; destructive Practice abandonment stays off this routine-move toolbar. When explicit placements fill every playable cell, any remaining entries that exactly match the fully spelled crossing slots are assigned automatically. The inferred assignments are persisted and submitted through the same authoritative completion validation as explicit placements.

Cells retain a stable 48-pixel size. Rectangular or larger boards use a focusable horizontal scroll region rather than shrinking text or adding a second zoom model; browser zoom remains the zoom control. A paused puzzle hides both the grid and word bank, so pause cannot become a free inspection state.

Daily stats, streaks, and groups

Reverse Cross uses the same durable anonymous-first identity and shared platform storage as Letter Golf, with game-specific result semantics.

Daily profile statistics include games started, games completed, completion rate, current and longest streak, exact best active duration, and a typical active duration. Typical time equals the exact result after one completion, then uses the bounded duration histogram median after additional completions. Letter Golf's word-count score and wildcard fields do not apply. Completion rate is derived as gamesCompleted / gamesStarted when at least one eligible Daily has started.

Statistics begin with the 2026-08-31 Daily. Eligibility is consistent for live updates and replay:

Play Started counter Completion metrics and streak
Valid global Daily on or after the epoch Yes Yes, when completed
Abandoned eligible Daily Yes No
Practice No No
Solution-viewed or otherwise spoiled No No
Invalid or missing Reverse Cross version identity No No
Before the statistics epoch No No
Superseded by account merge No No

Typical time uses a mergeable histogram rather than a recent-results list. For active duration d milliseconds, bucket b = min(floor(d / 15000), 240) is incremented. The estimate selects the lower median rank (n + 1) / 2 and reports the selected 15-second bucket midpoint; overflow bucket 240 reports its lower bound of 3,600,000 milliseconds. This bounds storage at 241 buckets while preserving exact best duration separately.

Start and completion profile deltas run only for newly persisted lifecycle transitions, so retries and concurrent losers do not double count. Repair, backfill, account merge, play reset, and player export replay persisted plays through the same typed projection used by live processing. Duplicate puzzle results use first completion, never best score or time. Reverse Cross duplicates with different persisted version fingerprints are incomparable and excluded from aggregate metrics.

Group standings compare players only when they played the same Daily puzzle definition: the same game, date/scope, language, difficulty preset, content version, and generator version. Lower server-calculated duration ranks first. Practice puzzles are excluded. Standings should avoid revealing the answer bank or solution to players who have not completed the puzzle.

Words and content policy

The authoritative shared-layer design, lifecycle semantics, migration from IWords, and physical ownership are documented in Lexical and Game Architecture. This section records the Reverse Cross product policy that design supports.

Reverse Cross and Letter Golf should consume one shared, immutable, versioned lexical catalog in memory. They should not each download and hold independent copies of the same million-entry data. Raw sources, normalization, moderation, and source provenance belong to the shared platform; eligibility decisions and search indexes belong to each game.

Each catalog entry will ultimately carry at least:

  • language and normalization version;
  • display form, grid form, and segmented graphemes;
  • source, source version, license/provenance, and source-native quality score when available;
  • content-form tags such as common word, inflection, proper noun, abbreviation, phrase, or idiom;
  • optional frequency/familiarity, part-of-speech, theme, topicality, and editorial-quality data;
  • suite-wide and game-specific moderation state.

Letter Golf's normal profile can continue to reject spaces, phrases, and most proper nouns while Reverse Cross can opt into carefully sourced crossword-style entries. These are policy choices over catalog metadata, not reasons to build unrelated dictionary systems. Denied content is filtered before generation and checked again at authoritative submission; matched denied terms are never sent to or echoed by the client.

The existing Crossword Nexus Collaborative Word List is a useful scored source, but its score is source-specific rather than a universal definition of quality. Its provenance and license must travel with a pinned, validated content version. Additional proper-noun, phrase, idiom, and topical sources require an explicit rights and quality review before ingestion; another publisher's daily clues or answers are not a content source.

Difficulty and puzzle quality

Difficulty is multi-dimensional. A named preset may eventually map to a versioned collection of knobs, but the underlying signals remain independently measurable and tunable.

Candidate structural signals include dimensions, slot count and length distribution, crossing density, checked-cell ratio, connectivity, symmetry, branching, placement ambiguity, and the generator's measured search/backtracking cost. Candidate lexical signals include familiarity, entry length, rare graphemes, proper-noun/abbreviation/phrase ratios, morphological repetition, and language-specific frequency.

"Fun" is not a synonym for generator difficulty and should not be collapsed into one unexplained number. A quality record can instead preserve a rubric: cleanliness, visual balance, interlock, variety, surprise, thematic cohesion, and empirical play outcomes. Automated rejection and ranking can narrow candidates; human playtesting and editorial judgment remain important, especially for themes and topical material.

Themes

A Reverse Cross theme is a deliberate relationship among entries, not merely a request parameter passed to a dictionary. Theme packs need a stable ID, language, provenance, editorial notes, eligible entries, freshness/expiry rules where relevant, and enough supporting fill to produce a fair puzzle. Because Reverse Cross supplies the entries rather than clues, the theme must improve cohesion without making the reconstruction trivial.

The Phase 3 generator is intentionally unthemed. Themes, proper nouns, abbreviations, phrases, idioms, and topical packs are later, independently controlled content capabilities. The broader LLM-assisted enrichment and theme-manifest direction is tracked in LLM-Assisted Content Enrichment and Theming.

Sharing

A Daily completion can share a spoiler-free result such as:

Reverse Cross 2026-08-26
2:34

See if you can beat my time: https://verbolus.com/games/reverse-cross

The final share format may include a compact non-spoiling grid pattern or difficulty marker. It must not expose the answer bank, placements, or theme reveal before the recipient plays.

Technical direction

Current state

  • The asynchronous, cancellable IReverseCrossPuzzleProvider boundary resolves to ReverseCrossProductionPuzzleProvider, which selects bounded ReverseCrossGenerator execution behind ReverseCrossGeneration and otherwise uses ReverseCrossFixtureProvider.
  • Daily uses ReverseCrossDailyPuzzleSourceProvider for missing dates. It serves fixtures unless ReverseCrossDailyGeneration evaluates true for system:reverse-cross-daily, then delegates to the Daily production-default provider. Generated Daily therefore requires both ReverseCrossDailyGeneration and ReverseCrossGeneration for that same synthetic identity. ReverseCrossStructuralQuality is also evaluated with the synthetic Daily identity.
  • Generated serving is absent/off by default. Bounded generation failures, catalog/index errors, feature evaluation failures, and per-process concurrency rejection return fixtures for fallback paths; caller cancellation is propagated.
  • ReverseCrossFillIndex retains catalog-entry references grouped by grapheme length and flat integer postings by position/grapheme. Its process-local provider coalesces construction and caches by catalog/profile identity, length bounds, and index-options version.
  • Typed domain puzzles use explicit grapheme-aware cells and keep solved cells and placements in a separate private solution. The old ReverseCross type is now only the deployed response DTO.
  • The backend rotates among validated 5x5, 7x7, and 9x9 presets within configured dimension limits. Fixtures and generated requests without explicit dimensions both use deterministic seed-based selection. Generation uses five seconds, 25 mask attempts, 100,000 search nodes, slot lengths 3-9, and two admitted generations per process.

Engine and configuration limits support square grids from 5x5 up through 9x9. Production fallback fixtures and the default automatic production selection remain the validated odd-only presets (5x5, 7x7, 9x9). Explicitly requested, strict generated Practice puzzles (the Generated override) support square sizes 5, 6, 7, 8, and 9; this path is intended for targeted validation and playtesting only. Production-default and fixture presets remain 5, 7, and 9, and the random production selection continues to pick from those odd presets. - The 9x9 preset preserves its deployed mask and Across entries. Its invalid copied 7x7 Down bank was replaced by the twelve Down entries proven by the complete private solution. - The frontend uses stable entry and slot IDs, grapheme arrays, explicit preview/place/delete controls, keyboard traversal, and a presentation-only timer that begins after server Start. - Daily and Practice puzzle/play persistence, revision-aware autosave, resume, pause, Practice abandon, and authoritative completion are implemented. Reverse Cross Daily profile statistics, streaks, deterministic replay, group presence, and comparable duration standings are implemented. Persisted Daily puzzle definitions are immutable cache hits; disabling generation flags affects only dates that have not yet been persisted. The group Today view uses a URL-backed game picker and renders only the selected game's result model; personal duration and streak summaries are also available on the profile and after an eligible Daily completion. The completion overlay follows the shared game-suite hierarchy and uses the player's chosen display name when present. It centers the Daily or Practice completion status ahead of result metrics. Best time, the current result, and typical time share one compact three-column row, with the current result emphasized in the center. Current and best Daily streak have their own section, followed by links to the Reverse Cross scorecard and Reverse Cross standings for the player's groups. While the overlay is open, it obscures the completed board so a shared screenshot does not reveal the solution. - Browser coverage verifies that generated 6x6 and 8x8 Practice requests preserve their requested dimensions through creation, keyboard placement, autosave, pause/resume, and completion. Responsive coverage keeps those sizes at 48-pixel cells with 44-pixel controls and no page-level overflow; 8x8 and 9x9 boards also retain internal horizontal scrolling at an effective 200% zoom viewport.

Validation evidence (strict generated Practice cohort)

  • A focused, opt-in production-catalog cohort exercised explicitly generated Practice puzzles (3 deterministic seeds per size). Both tested sizes (6x6 and 8x8) produced 3/3 successful constructions for the selected seeds.
  • 6x6 cohort summary: p50 = 1150.4 ms, p95 / max = 1166.3 ms; allocated sum ≈ 3.6 MiB; max attempts = 5; max nodes = 16,047; max backtracks = 16,035; blocked ratio observed 16.7%–25.0%; 100% of masks were fully checked; one connected component per successful puzzle; slot counts: 12; solution counts observed: 1 / 1 / 2 (one seed produced two valid boards).
  • 8x8 cohort summary: p50 = 1213.8 ms, p95 / max = 1962.0 ms; allocated sum ≈ 8.2 MiB; max attempts = 6; max nodes = 14,159; max backtracks = 14,133; blocked ratio observed 15.6%–20.3%; 100% checked; one connected component per successful puzzle; slot counts observed 24–26; all three seeds yielded at least one valid solution.

Measurement note: these runs were opt-in, used a pinned production catalog snapshot, and were intended for validation/playtesting rather than a broad SLO or SLA benchmark. They exercise the real production content and seed paths rather than synthetic or warmed pools.

Decision: enable the explicit Generated Practice override for targeted validation and playtesting, but do NOT expand the automatic production selection beyond the existing fixture-backed odd presets (5x5, 7x7, 9x9) until a larger, corroborating cohort establishes success, latency, uniqueness, and fallback-compatibility metrics.

Target boundaries

raw lexical sources
        |
normalization + provenance + moderation
        |
shared immutable catalog snapshot
        |
game eligibility profile + game-specific index
        |
Reverse Cross generator -> private solution + public puzzle
        |
persistence / play validation / stats / groups

Letter Golf keeps its count-signature/count-trie algorithms in its own game module. Reverse Cross uses its own positional fill index for fixed-length crossing constraints. A shared catalog snapshot and grapheme model sit below both; one game's algorithm must not leak into the other's.

The bounded generator produces a complete private construction witness and independently validates every successful result before serving it. Phase 3B established native MRV fill against validated fixture masks; Phase 3C added deterministic mask construction with fair global budget slicing; and Phase 3D added production composition, fixture fallback, concurrency admission, and replay-safe telemetry. Generation does not enumerate solutions to prove uniqueness. Discovery of alternate valid reconstructions is deferred to spoiler-protected God View/editorial tooling. See ADR 0001 and Phase 3 of the Reverse Cross roadmap.

Playtesting

Feature-gated God View tools expose persisted puzzle metadata, bounded fixed-mask/fixed-bank solution analysis, Practice generation settings, and caller-only Daily reset. Metadata and analysis come from privileged server routes rather than the public puzzle response. Practice analysis is available immediately; Daily analysis requires the requesting player to have completed that Daily. Analysis is capped at 2 seconds, 250,000 search nodes, and 100 alternates, and reports whether the search completed or why it was truncated. The UI presents the construction witness and one alternate at a time, including changed slot-to-entry placements.

Reset deletes only the requesting player's play for today's Daily. Feature flags hide controls but are not an authorization boundary; every privileged route independently enforces its policy and returns not found while disabled.

Record playtest notes by puzzle version, generator version, content version, language, dimensions, difficulty preset, device/input mode, and player familiarity. At minimum capture completion time, completion/abandonment, mistakes or rejected placements, ambiguity/confusion, and qualitative comments. Do not log raw sensitive entries or a player's typed content in telemetry.

Open product decisions

  • Define hint types and whether hints disqualify or annotate competitive results.
  • Calibrate which rectangular dimensions should be offered within the supported 5-9 row and column limits.
  • Decide whether duplicate entries are ever desirable.
  • Define the minimum editorial review required before a theme or topical pack can become a Daily.