Mathematics matters in crossword construction because a grid is not merely a collection of clues. It is a network of cells, slots and shared letters. A decision in one answer restricts several others; a block can force a rotated partner; and a plausible local fill may create an impossible region many steps away.
For students asking why mathematics is important in design and computing, crosswords reveal geometry, coordinate mappings, graph connectivity, combinatorics, constraint satisfaction and optimisation. They also teach restraint. A program can satisfy formal constraints while producing an awkward or unfair puzzle. Mathematical validity and editorial quality overlap, but they are not identical.
This article treats invented grids and word lists. Published crossword conventions vary by outlet, language and puzzle type. Rotational symmetry is common in some blocked-grid traditions but is not a universal rule. A recent research example, Generating News-Centric Crossword Puzzles as a Constraint Satisfaction and Optimization Problem, explicitly formulates automatic construction as constraints plus an objective; it supports the modelling approach without declaring one style mandatory.
Quick Reading Route
- Map rotational symmetry with coordinates rather than visual guessing.
- Extract slots and turn crossings into equations.
- Prune word domains before search expands.
- Trace backtracking instead of hiding failed branches.
- Measure several qualities without reducing editorial judgement to one number.
- Use the learning plan to build a small reproducible solver.
Why Mathematics Is Important in Crossword Grids
A crossword grid has at least three layers. The geometric layer decides which cells are open or blocked. The combinatorial layer identifies across and down slots and the equality constraints where they cross. The language layer decides which words and clues are acceptable.
Mathematics connects the layers. Coordinate transformations describe symmetry. Graph theory tests whether open cells form one connected region. Constraint satisfaction represents each slot as a variable whose domain is a set of candidate words. Search algorithms explore choices when propagation alone does not decide the fill.
The layers can disagree. A perfectly symmetric pattern may isolate open cells. A connected pattern may create many two-letter slots disallowed by a chosen house style. A fill may satisfy every crossing but rely on obscure entries. There is no single equation that replaces design judgement.
That makes crosswords an unusually honest mathematics lesson. Formal reasoning is powerful, but a good model begins by stating what it covers and what remains qualitative.
Rotational Symmetry Is a Coordinate Rule
Consider an n-row by m-column grid with zero-based coordinates. A half-turn, or 180-degree rotation, maps cell (r, c) to (n − 1 − r, m − 1 − c). If the pattern has rotational block symmetry, a block at one coordinate requires a block at its image.
For a 7×9 grid, cell (1, 2) maps to (5, 6): 7 − 1 − 1 = 5 and 9 − 1 − 2 = 6. Applying the mapping again returns (1, 2). This “do it twice and return” property is a useful verification check.
In an odd-by-odd grid, the centre cell maps to itself. In a 7×9 grid, the centre is (3, 4). Every other cell belongs to a pair under half-turn symmetry. Therefore a single centre block changes the total block count by one, while adding a noncentral block orbit changes it by two.
Orbits count independent decisions
An orbit is the set of positions reached by repeatedly applying a transformation. Under a half-turn, most orbits have two cells and the centre orbit may have one. A constructor choosing symmetric blocks is really choosing orbits, not individual cells.
If a 15×15 odd grid has one centre cell plus 112 noncentral pairs, there are 113 block-orbit decisions before other constraints. The raw number of symmetric binary patterns is at most 2¹¹³, not 2²²⁵. Slot rules, connectivity and block-count targets reduce it dramatically.
This is symmetry as compression: a geometric rule reduces independent variables.
Symmetry does not guarantee quality
A pattern can satisfy the coordinate mapping and still contain disconnected open areas, awkward narrow corridors or undesirable slot lengths. Symmetry is one constraint among many. It should not be presented as proof that a grid is elegant.
Some crossword styles use mirror symmetry, no symmetry, barred cells or other structures. The correct transformation belongs to the stated format, not to crossword mathematics as a whole.
A Grid Becomes Variables When Slots Are Extracted
Once blocks are fixed, scan each row for maximal runs of open cells. Every run meeting the chosen minimum length becomes an across slot. Scan columns similarly for down slots. Record start coordinate, direction and length.
For example, an across slot A1 might start at (2, 1) with length five. A down slot D3 might start at (0, 3) with length four. If A1’s third cell and D3’s third cell occupy the same coordinate, their indexed letters must be equal.
Using zero-based word indices, the crossing becomes A1[2] = D3[2]. Other crossings create more equality equations. The entire fill problem can be represented as variables, candidate domains and constraints.
Off-by-one errors are real mathematical errors
If a slot begins at column 1, its third character is at column 3 under zero-based indexing: 1 + 2. Confusing third with index 3 shifts the crossing by one cell. A solver may then reject valid fills or accept impossible ones.
Draw the coordinate table and test extraction on tiny grids. A correct algorithm should find the expected number of slots, lengths and crossings. Visual confidence is not enough.
Crossings Are Equality Constraints
Suppose A1 has domain {CATER, LATER, WATER} and D3 crosses its first position with domain {CLOUD, LEMON, WIDEN}. If the relevant D3 letter is also its first, candidate pairs must begin with the same character.
The possible starting letters for A1 are C, L and W; the same are available for D3, so no branch disappears yet. If another constraint reduces D3 to {LEMON}, then A1 must become {LATER}. Information propagates through the shared letter.
This is the heart of a constraint satisfaction problem. A variable has a domain. A constraint removes combinations that cannot coexist. Repeated filtering may force values without guessing.
Checking percentage
An open cell belonging to both an across and a down answer is checked by a crossing. A cell belonging to only one slot is unchecked. One simple checking percentage is checked open cells divided by all answer cells, under a declared counting convention.
The number is useful but not absolute quality. Two grids with equal checking percentages can distribute unchecked cells very differently. A long answer with several consecutive unchecked letters may feel harder to verify than isolated unchecked cells.
Constraint Propagation Removes Impossible Candidates
Arc consistency is one useful idea. For every candidate word in one slot, ask whether at least one candidate in each crossing slot supplies the required letter. If not, remove that word. Continue until no domain changes.
Suppose an across slot’s candidates are {RATE, RITE, ROTE} and its second letter crosses a down slot whose compatible crossing letters are only {A, O}. RITE is removed. If a later constraint leaves the down crossing letter only O, ROTE becomes forced.
Propagation can trigger a cascade. A forced word fixes letters, which shrink neighbouring domains, which may force more words. The order of revisiting constraints affects efficiency but should not affect the final stable set for the same propagation rule.
Empty domains are valuable evidence
If a domain becomes empty, the current partial choices cannot lead to a solution under the model. The solver should report which constraint removed the final candidate. An empty domain is not a software embarrassment; it is a precise contradiction signal.
In construction, the contradiction may mean an earlier word choice must be undone, the block pattern needs revision or the candidate dictionary is too restrictive. The mathematics locates the conflict but does not decide the editorial remedy.
Backtracking Is Disciplined Reversibility
When propagation stops with several candidates remaining, choose a variable, try one candidate, propagate its consequences and continue. If a contradiction occurs, restore the earlier state and try the next candidate. This is backtracking search.
The word “guess” understates the process. A good solver records a decision stack and reverses every dependent domain change accurately. If restoration is incomplete, later branches inherit false restrictions. If restoration is too broad, useful earlier deductions are lost.
A search tree visualises the alternatives. Nodes are partial assignments; edges are chosen candidates; failed leaves contain contradictions. The first solution is not necessarily the only solution or the best-scoring solution.
Minimum remaining values
A common heuristic chooses the unfilled slot with the smallest domain. If one slot has two candidates and another has 300, branching on the two-candidate slot often exposes failure sooner. This is called the minimum-remaining-values idea.
It is a heuristic, not a theorem that it always explores the fewest nodes. A slightly larger domain may connect to more neighbours and constrain the grid more strongly. Solvers often combine domain size, degree and value-ordering measures.
Least-constraining value
Within a selected slot, try the candidate that removes the fewest options from neighbours. This may preserve flexibility. In crossword construction, however, a less flexible thematic answer may be required. Search ordering serves the objective; it does not define it.
Worked Example: A Tiny Crossing Network
Imagine three length-three slots X, Y and Z. X[1] = Y[0], and Y[2] = Z[1]. Let X = {CAT, DOG}, Y = {ARE, ODD}, Z = {RED, EEL}.
If X = CAT, then X[1] is A, forcing Y = ARE. Y[2] is E, so Z needs E at index 1. EEL has E at index 1, while RED has E at index 1 too; both remain. This branch has two completions under the stated constraints.
If X = DOG, X[1] is O, forcing Y = ODD. Y[2] is D, but neither RED nor EEL has D at index 1. The domain of Z becomes empty, so the branch fails.
Propagation proves X must be CAT, but it does not make Z unique. If puzzle uniqueness matters, another constraint or clue distinction is required.
Completion is not uniqueness
Finding one valid fill establishes existence. It does not establish that no second fill exists. To test uniqueness, continue search after the first solution or add a constraint excluding it and ask whether another solution remains.
For a human crossword, clue answers and semantic interpretation add constraints beyond letter patterns. A grid fill may be structurally non-unique while well-written clues distinguish the intended entries.
Connectivity Is a Graph Problem
Treat each open cell as a vertex. Connect orthogonally adjacent open cells with edges. Choose one open cell, run breadth-first or depth-first search, and count the vertices reached. If the count equals the total number of open cells, the open region is connected.
This graph test is more reliable than visual inspection, especially around narrow corners. Rotational symmetry does not imply connectivity. Two disconnected regions can be perfect rotated copies.
Slot connectivity can also be modelled. Let every answer slot be a vertex and connect two slots if they cross. A connected cell graph usually supports a connected crossing network under common slot rules, but the precise relationship depends on minimum lengths and grid conventions.
Articulation points and fragile grids
An articulation point is a vertex whose removal disconnects a graph. In a crossword grid, a single open cell or narrow bridge may be structurally fragile. Blocking it would separate regions, and fill pressure may concentrate there.
This does not make every articulation point unacceptable. It gives a diagnostic. Graph measures help a constructor compare patterns before investing effort in clues and theme entries.
Combinatorics Explains the Search Explosion
If ten independent slots each have 100 candidates, the naive combination count is 100¹⁰ = 10²⁰. Crossings make the variables dependent and eliminate almost all combinations, but brute-force enumeration remains hopeless.
Propagation uses constraints early, while heuristics seek contradictions before the tree becomes wide. This is why problem representation matters more than raw processor speed. A solver that waits until the grid is full to check crossings wastes enormous effort.
The candidate dictionary also changes complexity. Larger domains offer more possible fills but expand search. Stricter language filters improve acceptability but may make a pattern unsatisfiable. Construction becomes a balance among vocabulary, geometry and computational effort.
Symmetry reduces pattern search too
When designing blocks, selecting one noncentral cell forces its rotated partner. Connectivity and slot-length checks can be applied after each orbit choice. Early rejection prevents the program from generating every symmetric pattern and testing only at the end.
This mirrors fill search: encode consequences as soon as possible.
Optimisation Needs More Than One Score
A construction system may reward theme entries, common vocabulary, crossing density and smooth slot lengths while penalising obscure fill, repeated words or awkward clusters. A weighted objective might be score = 5T + 2C − 4O, where T is theme coverage, C a checking measure and O an obscurity penalty.
The coefficients express editorial priorities, not mathematical truth. Changing them may select a different grid. A transparent system reports component scores and sensitivity rather than hiding everything behind one total.
The 2023 news-centric crossword research frames generation as a constraint satisfaction and optimisation problem that seeks inclusion of news-derived words. That is a useful demonstration of objectives layered on feasibility. It does not prove that topical inclusion alone makes a good puzzle.
Pareto trade-offs
Sometimes one grid has more theme content while another uses more familiar fill. Neither dominates the other on every criterion. They lie on a trade-off frontier. Showing both to an editor may be better than forcing a single arbitrary weight.
Mathematics can organise choices without pretending preferences disappear.
Clues Add a Different Constraint Layer
Grid algorithms operate mainly on letter patterns and word lists. Clues involve meaning, grammar, register, cultural knowledge and fairness. A candidate word fitting every crossing is not automatically clueable for a target audience.
Ambiguous clueing can be intentional, but the answer should be supported by the clue and crossings. Proper nouns, abbreviations and variant spellings may require explicit policy. These are editorial constraints that can be recorded as tags and filters but still need human review.
For students, this is a chance to connect mathematics and language. Formal structure narrows possibilities; precise wording selects among meanings. Neither subject is a decorative extra.
Alternative Mathematical Formulations
The same crossword fill can be expressed in several mathematical languages. A constraint satisfaction formulation uses slot variables and word domains. A Boolean satisfiability formulation creates true/false variables for choices and clauses that enforce exactly one word per slot, consistent crossings and no forbidden repetitions.
An exact-cover formulation creates rows representing candidate placements and columns representing requirements. Choosing rows must cover each requirement exactly once. The fit is natural when the requirements are discrete and exact, although letter-consistency details may need careful encoding.
Integer programming can use binary variables x_{s,w}, equal to one when slot s receives word w. Constraints require Σ_w x_{s,w} = 1 for every slot. Crossing compatibility can forbid incompatible pairs, and an objective can reward preferred entries.
No formulation is automatically best. A CSP solver handles domain propagation directly; SAT solvers exploit highly optimised Boolean reasoning; integer programming supports weighted objectives. Representation changes which structure an algorithm sees easily.
One-hot encoding and its cost
If a slot has 5,000 candidates, one-hot encoding creates 5,000 binary variables for that slot. Exactly-one constraints then need an efficient encoding. A compact character-based model may use fewer variables but require more complex word-validity constraints.
Students should compare model size as well as mathematical elegance. A correct formulation that creates millions of unnecessary clauses may be impractical.
Designing the Block Pattern Is a Second Search Problem
Filling a fixed grid assumes the blocks are known. Constructing the grid itself requires choosing block orbits while satisfying a target size, acceptable slot lengths, connectivity and theme placement.
Suppose an odd grid allows the centre to be open or blocked and all other decisions occur in pairs. If the target block count is odd, the centre must be blocked; if even, it must be open, under exact half-turn symmetry. Parity gives a deduction before search.
Adding one block orbit may split one long slot into two short slots. It can also create new down starts or isolate a region. Therefore pattern construction should update slot and graph consequences after every choice.
Theme entries create geometric anchors
A long theme answer may require a specific row, central position or crossing pattern. Fixing it reduces the available block orbits. Two theme entries may demand incompatible cells even before candidate words are considered.
Model theme placements as reserved open cells and length constraints. Search can then reject block patterns that damage those slots rather than discovering the conflict after full generation.
Counting independent binary choices
If k noncentral block orbits remain undecided, the raw pattern count is 2ᵏ. A block target changes this to combinations. If exactly b of the k pairs must be selected, the count is C(k,b), possibly multiplied by a centre choice fixed by parity.
For k = 20 and b = 6, C(20,6) = 38,760, much smaller than 2²⁰ = 1,048,576. Connectivity and slot-length pruning reduce it further.
Entropy Can Guide Search but Needs Care
Domain size is a simple measure of uncertainty. If all candidates are treated equally, a slot with d candidates has log₂d bits of choice. A two-word slot has 1 bit; a 256-word slot has 8 bits.
Real candidate probabilities are unequal. A probability-weighted entropy is H = −Σpᵢlog₂pᵢ. A slot with one highly likely candidate and nine unlikely ones may have lower entropy than a uniform ten-candidate slot.
Choosing the lowest-entropy slot may expose a strong decision early. But probabilities based on word frequency or a language model can be biased toward common prose rather than crossword conventions. They should guide, not dictate.
Information gain at crossings
A candidate that fixes several rare crossing letters may reduce neighbouring entropy sharply. Expected information gain estimates the average reduction before making the choice.
This idea resembles active learning: choose the decision expected to reveal the most. Calculating it exactly can be expensive, so solvers use approximations such as domain size times crossing degree.
Measuring Grid Structure Fairly
Block count alone misses slot shape. Record the distribution of slot lengths: minimum, median, maximum and counts at each length. A pattern dominated by minimum-length entries behaves differently from one with varied longer slots.
Crossing degree counts how many other slots a slot intersects. In ordinary American-style geometry, it often equals the number of checked cells in the slot. Low-degree slots can be loosely constrained; high-degree slots transmit information widely.
Graph diameter measures the largest shortest-path distance between slots in the crossing graph. A smaller diameter can help constraints propagate across the grid quickly, but it does not guarantee easier human solving.
Averages can hide weak regions
A grid may have a good average checking percentage but a corner with several consecutive unchecked letters. Map measures spatially. Report the minimum checking fraction across long slots and identify clusters of low constraint.
Similarly, average domain size can look manageable while one theme slot has zero candidates. Feasibility depends on every required variable, not the mean.
Worked Example: Pattern Parity and Symmetry
Imagine an 11×11 grid with half-turn block symmetry and a target of 25 blocks. The centre is the only fixed cell. Every noncentral block choice contributes two blocks, so an odd total requires the centre to be blocked.
After placing the centre block, 24 blocks remain, corresponding to 12 selected pairs. If 30 candidate block orbits remain after reserving theme cells, the raw count of ways to select 12 pairs is C(30,12) = 86,493,225.
That count is already much smaller than choosing 25 arbitrary cells from 121, but still large. Now reject any partial pattern that creates a one- or two-cell slot under a minimum-length-three rule. Test connectivity when a newly placed pair could seal a corridor. Respect reserved theme spans. Early pruning turns a huge combination count into a tractable search for many teaching cases.
The exact count does not predict construction time because pruning strength and implementation matter. It does show how parity and group orbits reduce the problem before computation.
Debugging a Crossword Solver
Create tiny cases with known answers. A 3×3 open grid has three across and three down slots of length three. Every open cell is checked. The crossing graph has six slot vertices and nine crossing edges.
Test a grid with one disconnected open cell and confirm connectivity fails. Test a slot whose domain becomes empty after one fixed letter. Test restoration by exploring a failing branch and verifying that every removed candidate returns before the next branch.
Log reasons, not only results
When removing a word, store the reason: fixed-letter mismatch, unsupported crossing, duplicate-use rule or policy filter. When a branch fails, report the first empty domain and the decisions leading to it.
Reason logs make solver behaviour explainable and help distinguish bad data from bad algorithms. If every candidate for a slot is removed by the same crossing, inspect that index and neighbour domain first.
Property-based tests
Some properties hold across many generated grids. Applying a half-turn twice returns every coordinate. Extracting slots and rebuilding their cells should cover exactly the intended open runs. A reported solution should assign one word per slot and satisfy every crossing.
Generate random small cases and assert these invariants. This catches edge cases that hand-picked examples miss.
Human Solving and Computer Filling Are Different Tasks
A constructor knows the word list and seeks a complete consistent assignment. A human solver reads clues, makes uncertain hypotheses and uses crossings to revise them. The mathematical structures overlap, but the information sources differ.
Human pencil marks resemble candidate domains. Writing a tentative answer is a branch; checking letters propagate; erasing after contradiction is backtracking. Expert solvers also use semantic probability and theme insight that a basic fill algorithm lacks.
This analogy can make algorithms intuitive without reducing human reasoning to brute force. People compress cultural and linguistic knowledge into powerful heuristics.
Fairness is audience-relative
A clue considered easy by one community may be inaccessible to another. Word-frequency data, school level, local knowledge and language variety affect difficulty. Report the target audience and avoid universal claims.
Mathematics can analyse clue lengths, crossing support and response data, but fairness remains partly a content and inclusion question.
Misconceptions Worth Correcting
“A symmetrical grid is automatically good”
Symmetry controls paired positions. It does not guarantee connectivity, acceptable slot lengths, a clean fill or fair clues.
“The longest answer should be filled first”
Length can matter, especially for themes, but domain size and crossing degree may make another slot more constraining. Variable choice is contextual.
“Finding one fill proves uniqueness”
One solution proves existence only. A second search is needed to investigate alternatives under the same model.
“A bigger dictionary always helps”
It may improve feasibility while adding obscure, duplicated or inappropriate entries and increasing search. Dictionary quality and tagging matter.
“The highest numerical score is the best puzzle”
A score reflects chosen features and weights. Editorial quality includes aspects the model may omit.
Did You Know?
A half-turn on an odd grid has exactly one fixed cell: the centre. That single coordinate explains why the parity of a symmetric block count can depend on the centre state.
Crossword crossings are equations between indexed characters. A familiar pastime becomes a network of tiny equality constraints, and one fixed letter can propagate across the grid.
Failed search is useful data. A contradiction reveals that a partial design cannot satisfy the declared constraints. Recording failures can improve the pattern, dictionary and heuristics.
What Students Should Practise
- Map coordinates under a 180-degree rotation and verify the inverse.
- Extract maximal across and down slots from a binary grid.
- Build a crossing table with exact character indices.
- Filter domains until arc consistency reaches a fixed point.
- Draw a backtracking tree and restore state accurately.
- Test open-cell connectivity with breadth-first search.
- Compare existence, uniqueness and best-score questions.
- Report which puzzle qualities remain editorial judgements.
A Four-Week Learning Plan
Week 1: geometry
Represent a small grid as zeros and ones, implement the half-turn mapping and generate block orbits. Compare odd and even dimensions and identify fixed points.
Week 2: graph and slot structure
Extract slots, crossings and connected components. Write tests for edge cells, one-cell runs and off-by-one indices. Preserve a coordinate diagram beside the data.
Week 3: constraint solving
Create small candidate domains, propagate crossing letters and add backtracking. Count search nodes under random, longest-slot and minimum-domain ordering.
Week 4: evaluation
Construct two feasible grids with different trade-offs. Report block count, slot lengths, connectivity, checking pattern, theme coverage and vocabulary concerns. Ask a human reviewer to explain preferences the score missed.
Guidance for Parents and Teachers
Begin with graph paper and coloured pencils. Let students rotate coordinates by hand, circle crossing cells and cut candidate words into strips. Physical manipulation reduces the cognitive load before code is introduced.
Celebrate contradictions. When a domain becomes empty, ask which earlier choice caused it and what evidence would have warned sooner. This builds debugging habits without treating failure as wasted work.
Require precise conventions: grid size, coordinate origin, minimum slot length, symmetry type and permitted vocabulary. Many disagreements vanish when definitions are written first.
Balance computing with language. After a valid fill is found, ask whether every word is appropriate and clueable for the audience. Mathematics should improve editorial reasoning, not replace it.
Frequently Asked Questions
Do all crosswords use 180-degree symmetry?
No. It is common in some blocked-grid traditions, while other formats use different symmetry or none. State the selected convention.
What is a crossword constraint satisfaction problem?
Slots are variables, candidate words are domains and crossings impose equality between specific letters. Additional constraints can cover repetition, themes and grid rules.
Why not try every combination?
The combination count grows exponentially. Propagation and good search order eliminate incompatible choices early.
Is backtracking the same as guessing?
It explores alternatives systematically, propagates consequences, records decisions and restores state after contradiction. Random guessing lacks those controls.
How can I check connectivity?
Model open cells as vertices with orthogonal neighbour edges, traverse from one open cell and compare the reached count with the total.
Can a crossword have two valid fills?
Yes. Grid constraints alone may permit multiple fills. Clues or extra constraints can distinguish an intended solution.
What makes a candidate domain good?
Correct length and fixed-letter compatibility are necessary. Editorial tags for familiarity, spelling, register and theme relevance also matter.
Does an AI-generated grid need human review?
Yes. Formal feasibility does not guarantee clue accuracy, fairness, cultural appropriateness or enjoyable language.
Useful Next Reading
Compare crossword symmetry with textile patterns, symmetry transformations and repeats. Compare its search process with Sudoku, constraints and backtracking search. Then use the Mathematics Learning Hub for graph theory, combinatorics and coding routes.
For a research example, read Majima and Ishihara’s 2023 constraint-satisfaction and optimisation formulation. Treat its specific objective and experimental setting as one documented approach, not a universal definition of crossword quality.
Conclusion
Crosswords answer “why mathematics?” by turning a page of squares into a network of consequences. Geometry pairs cells. Graphs test connection. Indexed equations link letters. Constraint propagation removes impossibilities, and backtracking makes uncertainty reversible.
The mathematics does not write every good clue or settle every aesthetic choice. It does something equally valuable: it makes the structural problem explicit, testable and reproducible, leaving human judgement to work with better evidence.
A Practical Investigation Studio
These investigations use paper pieces or synthetic data. Complete two or three for a short project or the sequence as a portfolio. The aim is to expose assumptions, constraints and error signals, not to imitate regulated operations or claim that one classroom model captures every real case.
Investigation 1: Symmetry map
Number rows and columns from zero and apply the half-turn mapping to every cell in an odd and even grid. Mark fixed points and paired cells, then verify that applying the mapping twice returns the start.
Investigation 2: Block orbit
Choose a few seed blocks and generate all cells forced by rotational symmetry. Count independent block decisions rather than counting every displayed block as a separate choice.
Investigation 3: Slot extractor
Scan a small grid left-to-right and top-to-bottom to list every across and down slot. Record start coordinate, direction and length so crossings can be generated mechanically.
Investigation 4: Crossing equations
Give each slot a variable and write equality constraints for intersecting letter positions. Check indices carefully; an off-by-one error can make a valid fill appear impossible.
Investigation 5: Domain pruning
Start with short word lists for several slots, apply fixed letters and crossing constraints, and record how each domain shrinks. Distinguish a forced choice from a merely convenient one.
Investigation 6: Variable ordering
Solve the same toy fill by choosing the longest slot first, the smallest domain first and a random slot first. Compare search nodes rather than claiming one heuristic always wins.
Investigation 7: Backtracking tree
Draw the first three levels of a search tree, including failed branches and restored domains. The record should show why undoing consequences is part of the algorithm.
Investigation 8: Uniqueness test
Construct a grid with two valid fills, then add one clue or fixed letter to remove the ambiguity. Explain the difference between a completed grid and a uniquely determined puzzle.
Investigation 9: Connectivity check
Model open cells as a graph with orthogonal adjacency and run a simple reachability test. Compare visual symmetry with actual connectivity; one does not guarantee the other.
Investigation 10: Theme trade-off
Give bonus scores to a few theme entries and penalties to obscure fillers in a synthetic objective. Show how a higher numerical score can still need editorial review.
Investigation 11: Pattern comparison
Compare two grids with equal block counts but different slot-length distributions and checking percentages. Use several measures instead of declaring quality from one number.
Investigation 12: Puzzle report
Publish the grid coordinates, symmetry rule, slot list, domains, search choices and unresolved editorial judgements. A peer should be able to reproduce the fill or locate the first discrepancy.