Ghost Squad • Game Design Spec

Type: Product Requirements Document Status: Draft for review Series: Do villain things (sibling of minelayer) Deliverable target: single offline HTML file

1. Background and Goals

Concept. Ghost Squad is a reversed Pac-Man: instead of guiding Pac-Man away from the ghosts, the player commands the four ghosts to hunt an AI-controlled Pac-Man. It is the second title in the reversed-games series, following minelayer (reversed Minesweeper, where the player lays mines and an AI sweeper solves the board).

Why this game, why now. Design research within the series established three selection criteria, all met by this concept:

Fully local opponent. The original player's strategy is mechanical enough to recreate with a local heuristic (greedy dot-seeking + flee behavior), requiring no remote AI endpoint and no network at all.
High replay value. Replay variance comes from a roster of distinct AI Pac-Man personalities dealt per round, plus three difficulty tiers — mirroring how minelayer pairs random boards with a stable solver.
Legible in 3 seconds. Anyone who knows Pac-Man instantly understands the inversion from one screenshot.

Differentiating mechanic. All four ghosts are human-controllable at once, each bound to its own keyboard cluster (WASD, ↑←↓→, IJKL, numpad 8456). One player realistically attends to about two ghosts at a time; uncontrolled ghosts fall back to the original arcade chase AI. Two to four friends sharing one keyboard bring more ghosts under human control. Difficulty therefore emerges from human attention bandwidth rather than from AI knobs alone, and the game gains a couch co-op mode at zero netcode cost.

Goals (observable outcomes).

A solo player can complete a round within seconds of first launch, with no setup, accounts, or network access.
A returning player can identify the current Pac-Man personality within the first moments of a round and adapt their hunting strategy to it.
Two to four players can join mid-session simply by pressing their key cluster — no menus or configuration.
Each difficulty tier is winnable but demanding; losses are attributable to player decisions, not to unreadable AI randomness.

2. Product Contract

Element	Definition
Primary user	A casual player in a browser, alone or with friends physically present
Trigger	Opening the game's HTML file (double-click or browser URL); the game boots straight into the title/attract screen with no loading of external resources
Core job	Command the four ghosts to catch the AI Pac-Man before he clears the maze of dots
Completion condition (round won)	Pac-Man is caught 3 times (his lives exhausted) before all dots are eaten
Failure condition (round lost)	Pac-Man eats every dot in the maze
Trust boundaries	Everything runs locally in one HTML file; no network calls, no external assets, no accounts; best scores persist only in `localStorage`

3. Scope

P0 — required for the smallest complete game P0

Arcade-faithful maze, sprites, colors, and movement rules (FR-01)
Full round loop: attract → ready → play → catch/clear → round end, with win and loss states (FR-02)
Four-cluster keyboard control with per-ghost turn queueing (FR-03)
Automatic fallback to classic arcade AI for unattended ghosts, with instant reclaim (FR-04)
Heuristic AI Pac-Man with a personality roster of at least 4 personalities, telegraphed per round (FR-05)
Three difficulty tiers selectable before a round (FR-06)
Power pellets with reversed-risk behavior (FR-07)
Scoring, lives, HUD, and per-tier local best score (FR-08)
Synthesized audio (no audio files) with mute toggle (FR-09)
Result sharing: shareable text card of the run from the end overlay (FR-10)

P1 — valuable, not blocking P1

Touch controls (on-screen directional pads per ghost) for tablet play
Round replay / instant-restart shortcut
A fifth and sixth Pac-Man personality

Explicitly excluded OUT

Online multiplayer of any kind (couch co-op only — one keyboard, one screen)
Procedurally generated mazes (see FR-01: one hand-tuned arcade-faithful maze; variance is carried by the personality roster, not the maze)
Accounts, leaderboards beyond local best, analytics, or any network feature

4. Game Model

4.1 Objects

Object	Canonical definition
Maze	The single fixed playfield: tile grid, walls, dots, 4 power pellets, ghost house, side tunnel. Replicated from the arcade original's layout.
Pac-Man (AI)	The autonomous opponent. Goal: eat all dots. Has 3 lives per round. Behavior = shared movement core + one active Personality (SR-3).
Ghost × 4	Player-controllable hunters. Each has its arcade identity, color, spawn point in the ghost house, scatter corner, and fallback chase behavior:
	Shadow (Blinky) — fallback: targets Pac-Man's current tile Speedy (Pinky) — fallback: targets 4 tiles ahead of Pac-Man Bashful (Inky) — fallback: flank vector off Shadow Pokey (Clyde) — fallback: chases when far, retreats to corner when close
Dots	234 regular dots + 4 power pellets (count follows the replicated arcade maze). Pac-Man's win progress; visible on the maze at all times.
Control mode (per ghost)	Each ghost is independently in HUMAN or AUTO mode at any moment (SR-2). Displayed above each ghost and on the HUD.
Personality	A named parameter bundle for the AI Pac-Man, drawn once per round and telegraphed to the player (SR-3).
Difficulty tier	One of Beginner / Intermediate / Expert, selected before the round; tunes only the AI Pac-Man's physical parameters (speed, cornering boost, empowerment duration, personality pool) — decision quality is identical on all tiers (SR-4).

4.2 Round state machine (canonical)

ATTRACT ──(any key / click Start)──▶ TIER SELECT ──▶ READY (personality reveal, ~2.5s) ▲ │ │ ▼ │ ┌── PAC-MAN CAUGHT ── lives > 0 ──▶ PLAY (respawn) └──── round end ◀───────┤ ▲ (WIN or LOSS) └── lives = 0 ──▶ WIN │ dots = 0 ──────▶ LOSS ───────────────┘ PLAY ──(Esc / window blur)──▶ PAUSED ──(Esc / Resume)──▶ PLAY

States in capitals are the only legal states; transitions not shown are prohibited (e.g., no path from ATTRACT directly into PLAY, no restart mid-catch animation).

5. Functional Requirements

FR-01Arcade-faithful presentation and movement

Purpose. The reversal only reads instantly if the original is reproduced faithfully. This series' convention (set by minelayer's Windows XP fidelity) applies here to the 1980 arcade original.

Layout: vertical playfield reproducing the original maze: same wall geometry, dot placement, 4 power-pellet corners, ghost house with exit gate, side tunnel with slowdown. HUD above/below: score, best score, Pac-Man lives, difficulty tier, per-ghost HUMAN/AUTO indicators.
Visuals: original palette (maze blue #2121DE, black background, canonical character colors). Sprites are drawn to match the arcade silhouettes; no invented decorative elements, backgrounds, or effects.
Movement rules (shared by all characters): continuous tile-grid movement; direction changes only take effect at intersections (queued turns); tunnel travel slows ghosts and Pac-Man per the original; cornering speed bonus applies to Pac-Man only, as in the original.
Exceptions: viewport smaller than the maze → the playfield scales down proportionally, never crops; aspect ratio is preserved.

FR-02Round loop, win, and loss

Aspect	Rule
Trigger	Player selects a difficulty tier on TIER SELECT (click or `1`/`2`/`3`)
Ready phase	~2.5 s: maze shown with dots refilled, characters at spawn, active Personality revealed by name, icon, and trail tint (SR-3). No input accepted.
Catch	A ghost in HUMAN or AUTO mode overlaps Pac-Man while he is not empowered → Pac-Man loses a life. If lives remain: brief catch animation, all characters return to spawn, sub-ready (~1 s), play resumes with dots as they were. If none remain: WIN.
Clear	Pac-Man eats the final dot → LOSS.
Round end	WIN or LOSS banner with final score, best score for the tier, and a prompt: `Enter` = same tier again, `Esc` = back to TIER SELECT. Dots and positions fully reset on replay.
Persistence	Best score per tier saved to `localStorage`; shown on TIER SELECT and round end.
Exceptions	Browser tab hidden or window blurred during PLAY → automatic PAUSED (fair-play and input-safety rule, SR-5).

FR-03Four-cluster keyboard control

Purpose. Let one keyboard seat one to four players; make human attention the difficulty dial.

Ghost	Primary cluster	Alternate cluster
Shadow	`WASD`	—
Speedy	`↑←↓→`	—
Bashful	`IJKL`	—
Pokey	numpad `8456`	`TFGH` (laptops without numpad)

Processing rules: a direction key queues a turn for its ghost; the turn executes at the next intersection where it is legal (same model as the original's input). Opposite-direction input reverses the ghost immediately, per arcade rules. Key repeat is irrelevant — taps, not holds.
Claim: the first press on any cluster during PLAY puts that ghost in HUMAN mode instantly (even mid-round — this is how friends join).
Feedback: the controlled ghost shows a white outline matching the interactive demo below; the HUD indicator switches AUTO → HUMAN.
Exceptions: unrecognized keys are ignored. Simultaneous opposite keys in one cluster resolve to the most recent press. Hardware key-ghosting (cheap keyboards dropping simultaneous keys) cannot be detected; the tap-based, queue-at-intersection model minimizes its impact, and clusters are placed on distant keyboard-matrix regions to reduce collisions.

Interactive demo — FR-03 claim / FR-04 idle fallback

Focus the dark panel (click it), then press any cluster. A ghost you touch becomes HUMAN; leave it unattended and it reverts to AUTO (fallback AI) — the exact SR-2 rule.

AUTO

Shadow

fallback AI

W A S D

AUTO

Speedy

fallback AI

↑ ← ↓ →

AUTO

Bashful

fallback AI

I J K L

AUTO

Pokey

fallback AI

num 8 4 5 6 · T F G H

Idle timeout is shortened to 3 s in this demo (game value: see SR-2). Escape clears all.

FR-03aTouch controls (mobile)

Purpose. Make the full game playable on devices without a keyboard, preserving the exact SR-2 claim/idle semantics of FR-03.

Layout: on touch devices the board pins to the top of the screen and a control panel occupies the bottom — a slim pause/mute bar spanning the width, then one virtual d-pad per ghost in a 2×2 grid (Shadow/Speedy on top, Bashful/Pokey below). The board scale reserves exactly the panel's height, so nothing overlaps or clips; desktop layout is unchanged.
Mapping: each pad mirrors one keyboard cluster. A tap on a direction routes into the same claimGhost() path as a keypress — instant HUMAN claim (even mid-round) plus a queued turn applied at the next legal intersection; opposite direction reverses immediately per arcade rules.
Hold semantics: holding a direction re-asserts it every frame, mirroring OS key auto-repeat; the SR-2 idle fallback therefore fires IDLE_S after the final release, identical to keyboard. Pads are independent and multi-touch capable, so several ghosts can be steered simultaneously — the touch equivalent of several players on separate clusters.
Multi-finger ownership: every live press is tracked per finger (pointerId on the PointerEvent path, Touch#identifier on the legacy TouchEvent path) in a per-ghost press stack; the stack top (most recently pressed, still held) is the effective direction. Lifting any finger releases only that finger: two fingers may share one button, and lifting the first never cancels the second's hold; lifting an older finger after a second finger chose a new direction keeps the newer one. pointercancel/touchcancel (OS gesture, incoming call) and window blur release all fingers, so no input can ever stick. Compatibility events the browser synthesizes around a real press (duplicate mousedown/touchstart/click within 400 ms) are deduplicated.
Feedback: a pad glows in its ghost's color while that ghost is in HUMAN mode; the pressed direction button stays lit while held. The usual white sprite outline and HUD HUMAN indicator apply unchanged.
Keyboard gaps closed on touch: an on-screen pause button (and tap-anywhere-to-resume on the pause overlay) replaces Esc; an on-screen mute button replaces M; ATTRACT advances on tap; tier cards already confirm on tap; the END overlay shows RETRY / CHANGE TIER buttons replacing Enter/Esc. Keyboard-only hint text is swapped for touch-specific hints.
Detection & override: touch mode activates when the device's primary pointer is coarse (pointer: coarse media query — phones/tablets), not merely when a touchscreen exists, so touch laptops keep the desktop layout; appending ?touch=1 to the URL forces the pads on any device (testing, mouse play, touch laptops).

FR-04Fallback AI for unattended ghosts

Trigger: a ghost is in AUTO mode at round start, or reverts to AUTO after the SR-2 idle timeout.
Processing: AUTO ghosts run the original arcade logic — alternating scatter/chase waves with each ghost's canonical target tile (object table, §4.1). This is deliberately the predictable original behavior: a competent AI Pac-Man can exploit it, so human-controlled ghosts are strictly more dangerous — this is what makes co-op objectively stronger than solo.
Reclaim: any key on a ghost's cluster instantly returns it to HUMAN mode with no penalty; the queued turn applies from the next intersection.
Result/state: per-ghost HUMAN/AUTO mode is part of round state, visible on the sprite and HUD at all times.

FR-05AI Pac-Man: heuristic core + personality roster

Purpose. Supply a convincing, learnable, varied opponent with zero network dependency.

Whole-board awareness: the AI models a human player watching the full maze. Every frame it rebuilds a BFS distance field per active ghost (real maze distances, never line-of-sight) and, at every tile center, scores each legal direction as food-value − threat. Food-value combines a linear pull toward the nearest remaining food (constant gradient at any range — it never loses track of the last dots), dot density within 18 tiles (steep at short range so an adjacent dot always beats the keep-direction commitment), dots lying ahead in the corridor, and a large bonus for a dot on the immediate next step taken forward/sideways (never bypassing free food); threat sums per-ghost terms with a 1/d² falloff.
Mode-aware threat model: AUTO ghosts are weighted by the observed wave phase (0.55 during scatter, 1.0 during chase) and by heading (closing = effectively 1 tile nearer, receding = 1 farther). HUMAN ghosts are modeled as Shadow — always assumed to be closing in: weight 1.7, no receding discount. No other use is made of ghost AI internals, so the model holds identically against human players.
Lookahead & self-preservation: a 16-tile corridor simulation feeds the score (dots ahead, minimum ghost distance, dead-end detection, and whether the lane reaches a junction — escape turns are preferred, strongly so in panic); a corridor threat check bails out of any lane with a non-frightened ghost on the next 4 tiles only when the way back is clear (in a two-sided pincer it commits to the least-bad side instead of wobbling); a panic mode (hysteresis: enters when a closing ghost is ≤3 tiles, exits beyond 5, plus a strong keep-direction commitment) switches to pure-survival scoring — and a reachable power pellet with a ghost within 6 tiles becomes a trap to sprint for. During empowerment the scoring inverts into pursuit of frightened ghosts (FR-07).
U-turn discipline: reversal is free (as in the arcade) but a 180 carries a score penalty outside panic, and any deliberate U-turn is committed for 0.8 s — the AI never flip-flops in place. Survival overrides (panic, corridor bail-out) ignore the lock. The free-food bonus never applies to a reversal step, so a dot one step behind can never defeat the U-turn penalty.
Personality layer: one Personality (SR-3) is drawn per round from the tier's pool; it adjusts the core's parameters for the whole round. A personality never changes mid-round — variety lives between rounds, so each opponent is learnable within its round.
Telegraphing: during READY, the personality's name, one-line behavior hint, and trail tint are shown; during play its trail tint persists so the player always knows which opponent they face.
Exceptions: trapped with all directions below survival threshold → takes the highest-expected-survival direction (never freezes, never oscillates in place).

FR-06Difficulty tiers

Trigger: TIER SELECT before each round; last-used tier preselected.
Rules: three tiers per SR-4. Tier choice changes only the AI Pac-Man's physical parameters (speed, cornering boost, empowerment duration, personality pool) — decision quality is identical on all tiers. A deliberately weakened strategy reads as broken behavior, not as an easier opponent; difficulty must come from what the opponent can physically do, not from how badly it plays. Never the maze, never ghost speed.
Result: tier shown on HUD and recorded with the best score.

FR-07Power pellets (reversed risk)

Rule: when Pac-Man eats a power pellet he is empowered for a per-tier duration (SR-4): ghosts turn dark blue and slow slightly, and Pac-Man may eat them. An eaten ghost returns to the ghost house and is out of play for ~3 s (visible as a flashing house door), then re-emerges in AUTO mode.
Why (product judgment): pellets are the original's power fantasy; in the reversal they become scheduled threat windows the hunter must respect — the pellet-hunting personality (SR-3) exists specifically to weaponize them.
Feedback: empowerment timer shown as the ghosts' blue flashing (arcade pattern) plus a HUD countdown bar.
Exceptions: a ghost eaten during empowerment does not cost the round; empowerment cannot stack — eating a second pellet resets the timer.

FR-08Scoring and HUD

Score events: catching Pac-Man (400 × catch number in the round: 400 / 800 / 1200), each dot still on the maze at WIN (10 pts defense bonus), each ghost kept out of the house when empowerment ends (50 pts survival bonus).
HUD: score, tier best, Pac-Man lives (icons), empowerment bar, and four ghost indicators showing HUMAN/AUTO and house-respawn countdown.
Exceptions: localStorage unavailable (private mode) → best scores silently become session-only; no error is shown to the player.

FR-09Audio

Rules: all audio synthesized locally (Web Audio API) — siren loop, dot-eaten tick (recontextualized as the opponent's progress), catch jingle, empowerment alarm, win/loss stingers. No audio files, keeping the single-file offline constraint.
Control: M toggles mute; the choice persists in localStorage. Audio starts only after the first user gesture, per browser autoplay policy.

FR-10Result sharing (community layer, phase 1)

Share card: the end overlay carries an always-visible SHARE RESULT button (visible on desktop too — sharing is not touch-only). It assembles a plain-text result card: brand line (title_brand + " — " + title_suffix, both localized per language), a hook line with the translated tier name and catch count (win/loss variants share_hook_win/share_hook_loss), a compact emoji stat row (👻× catches + 🏆 score), and a call-to-action line with the page URL.
Delivery: shared engine public/shared/share.js. Clicking SHARE RESULT always opens a retro CRT preview modal first — the player sees the exact card before anything leaves the device. From the modal: COPY (clipboard + execCommand legacy fallback + retro toast), native SHARE via the Web Share API sheet where available (a dismissed sheet is not an error), CLOSE. Modal button labels are localized in all 8 languages by the engine's own dictionary. The button click stops propagation so the overlay's tap-to-retry does not fire. Purely additive: no server, no account, no change to the round loop.
i18n + tests: share_btn, share_copied, share_hook_win, share_hook_loss, share_cta dictionary keys in all 8 languages; card assembly (buildShareCard) is a pure function pinned byte-for-byte by the 11-share test suite.

6. Shared Rules

SR-1 · Input and fairness

Inputs are taps queued per ghost; no action requires holding keys. All four clusters plus Esc (pause), Enter (confirm), M (mute), and 1–3 (tier select) form the complete input set. Any key press on ATTRACT advances to TIER SELECT. On touch devices the virtual d-pads (FR-03a) provide the identical input set through on-screen controls — same claim path, same queueing, same fairness rules; touch input never grants abilities keyboard lacks.

SR-2 · HUMAN/AUTO control rule (canonical)

Each ghost is in exactly one mode. Cluster key press → HUMAN immediately. No key press on its cluster for 2.0 s of gameplay time → AUTO. Mode changes never interrupt movement and never cost score. The mode of every ghost is always visible on the sprite outline and HUD. Coordination bonus: when 2 or more HUMAN-mode active ghosts exist simultaneously, each receives a +15% speed bonus. This rewards multi-ghost coordination and team play without making single-ghost play easier — the bonus does not activate for solo 1-ghost control.

SR-3 · Personality roster (canonical)

Personality	Behavior signature	Counter-play for hunters	Tier pool
Glutton	High dot greed, short flee radius — takes dots even with ghosts nearby	Punish greed: collapse on his feeding lane	All tiers
Skittish	Large flee radius, high threat weight, extra caution vs HUMAN ghosts	Herd him: drive him toward a waiting ghost	All tiers
Pellet Hunter	Lunges for power pellets whenever a ghost is near; hunts aggressively during empowerment	Guard pellet corners; retreat on empowerment	Intermediate+
Tunnel Rat	Prefers corridors dense with dots; commits to eating through a lane	Cut off his lane exits ahead of him	Intermediate+
Trickster	Baits: approaches a committed chaser (4–6 tiles), then doubles back the moment separation strictly grows; occasionally jukes to a near-equal route	Hold intersections; don't over-commit	Expert only

One personality is drawn uniformly from the tier's pool at READY and stays fixed for the whole round. Name, hint line, and trail tint are shown at reveal; tint persists during play.

SR-4 · Difficulty tiers (canonical)

Tier	Pac-Man speed	Cornering boost	Empowerment	Personality pool
Beginner	85% of ghost speed	×1.08	4 s	Glutton, Skittish
Intermediate	95%	×1.12	5 s	+ Pellet Hunter, Tunnel Rat
Expert	100%	×1.22	6 s	All five

Decision logic (FR-05) is identical on every tier — tiers alter only the opponent's physical parameters, never ghost speed, maze, or pellet count.

SR-5 · Pause and focus

Esc toggles PAUSED; gameplay time (including the SR-2 idle timer) freezes. Window blur or tab hidden forces PAUSED. PAUSED dims the maze and shows controls; no game state is lost.

7. Exceptions and Boundaries

Case	Behavior
Small viewport / mobile portrait	Playfield scales proportionally; below a usable minimum, show "rotate or enlarge your window" notice (touch is P1, so phones are not P0 targets)
Keyboard without numpad	Pokey's alternate cluster `TFGH` always active
Hardware key ghosting in 3–4 player co-op	Not detectable; mitigated by tap-based input and matrix-distant cluster placement (FR-03)
Autoplay-blocked audio	Audio engine initializes on first user gesture; game is fully playable silent
`localStorage` blocked	Session-only best scores, no error surfaced (FR-08)
Mid-round cluster conflict (two friends fight over one ghost)	Most recent press wins; no lockout — physical negotiation is part of couch co-op

8. Open Questions

[TBD] Solo default: should a "Solo training" option start all four ghosts in HUMAN mode (no 1.5 s fallback) for players practicing full-squad control? Current assumption: no separate option — the SR-2 rule alone governs all play, keeping solo and co-op identical in code.
[TBD] Catch target: 3 Pac-Man lives per round is the working value; if rounds run long in playtesting, consider first-to-2.
[Resolved] Expert speed parity: kept at 100% (parity). Instead of reducing speed, a coordination bonus (+15% when 2+ HUMAN active ghosts) was added. This creates a wide Expert challenge curve (25%→55%) that rewards multi-ghost coordination without making 1-ghost too easy. IDLE_S increased from 1.5 to 2.0 s for solo multi-ghost viability. See §9.4 for the final verified curve.

9. Testing and Difficulty Validation

9.1 Deterministic test suite (bug detection)

A dedicated test suite lives at tests/ghost-squad/ with 40 tests across 9 files. The harness loads the real index.html in-memory with stubbed browser APIs (DOM, canvas, audio) and drives the game loop frame-by-frame with deterministic entity placement. The game file is never modified.

File	Tests	Covers
`01-ai-movement.test.js`	7	Pincer trap escape, U-turns for dots behind (near & far), 1-dot endgame routing (all-personality sweep), calm dithering, turn-into-dots overshoot
`02-state-machine.test.js`	6	Boot→attract→ready→play, catch→respawn, 3 catches = ghosts win, all dots cleared = end, Enter restarts / Esc → tier
`03-combat-empowerment.test.js`	5	Empowerment flip, eating ghosts, re-emergence in AUTO, normal contact costs a life, empowerment expiry
`04-input-claim-idle.test.js`	4	WASD/arrow cluster claim, per-ghost independence, idle→AUTO fallback, queued direction applied
`05-tunnel-waves.test.js`	4	Side-tunnel wrap for Pac-Man and ghosts, ~0.55× tunnel slowdown, scatter/chase wave alternation
`06-empowerment-hunt.test.js`	4	Hunting frightened ghosts, pellet weaponization, guarded-pellet safety, resume clearing after eating all four
`07-ghost-house.test.js`	3	Eyes return path, house release timers, ~3 s out-of-play minimum after being eaten
`08-misc-soak.test.js`	4	Pause/resume immutability, 4-player simultaneous claims, 400/800/1200 catch scoring, 5-sim-minute Expert soak with per-frame invariants
`09-difficulty-sim.test.js`	3	Full-game baseline simulation: all-AUTO ghosts × 3 runs per tier, tracks win rate, catch count, dot-clear %, timing

Conventions: every AI movement test corresponds to a bug observed and fixed during development. Scenario tests pin one personality via G().setPersonality(name) to eliminate RNG-draw flakiness; behavior-sweep tests iterate all five personalities explicitly. Ghost control uses mode:'human' with no queued input so ghosts march deterministically.

9.2 Lower-bound test (Pac-Man strength verification)

Purpose: verify that the AI Pac-Man is strong enough that all-AUTO ghosts (no human input) cannot trivially beat him. If AUTO ghosts win easily, the human player has no reason to participate — the game plays itself.

Method: run a full round with all 4 ghosts on AUTO (no keyboard input) for 5 runs per tier. Track catch count, dot-clear percentage, and outcome.

Tier	All-AUTO win rate	Typical catches	Pac-Man clear %	Verdict
Beginner	100% (5/5)	3	57–89%	Ghosts too strong without human — acceptable on easiest tier
Intermediate	60% (3/5)	1–3	70–100%	Balanced — human input matters
Expert	0% (0/5)	1–2	100%	Pac-Man dominant — human required, as designed

Expert at 0% all-AUTO win rate is intentional: it confirms the AI Pac-Man is strong enough that a human player is strictly necessary to win on the hardest tier. The lower bound validates Pac-Man's decision quality, not difficulty tuning.

9.3 Upper-bound test (human playability verification)

Purpose: verify that a human player controlling 1 ghost can realistically beat the AI Pac-Man. Record all moves, measure APM, and evaluate whether a real human — especially a new player — could repeat the performance.

Method: simulate a human player claiming Shadow (ghost 0, WASD). Every 300 ms (average human) or 200 ms (skilled), re-evaluate direction via BFS pathfind toward Pac-Man's predicted position (3 tiles ahead in current heading). Other 3 ghosts stay on AUTO. Record every keypress and outcome. 5 runs per tier to account for personality variance.

Tier	300 ms reaction (avg human)	200 ms (skilled)	APM required
Beginner	5/5 (100%)	5/5 (100%)	~114/min (1.9 keys/s)
Intermediate	3/5 (60%)	4/5 (80%)	~114–150/min
Expert	1/5 (20%)	1/5 (20%)	~120–144/min

Key findings (single-ghost):

APM is very reasonable — 2–2.5 keys/sec (120–150/min), well within casual-player range. The game does not require superhuman button-mashing.
Reclaim overhead is zero — the 300 ms decision cycle keeps the controlled ghost within the 1.5 s idle window, so the ghost never reverts to AUTO during active play.
Beginner is perfectly tuned — 100% win rate for all reaction times, ideal entry point for new players.
Intermediate is well-balanced — 60–80% win rate depending on skill. Pellet Hunter is the hardest personality (empowerment makes Pac-Man dangerous).
Expert is very hard — only 20% win rate even for skilled players. At 100% speed parity, chasing from behind never works; wins depend on AUTO ghost positioning for traps. See open question in §8 regarding speed parity.

9.4 Upper-bound: multi-ghost challenge curve

Purpose: verify that controlling more ghosts rewards skill — the win rate should increase (or at least not decrease) when players add ghost control, for both team play (N humans) and solo multi-ghost (1 human cycling through N ghosts).

Method: each controlled ghost uses a role-specific BFS interception target: ghost 0 chases Pac-Man's current tile, ghost 1 cuts off 4 tiles ahead, ghost 2 traps from 4 tiles behind, ghost 3 guards the nearest pellet. Decisions every 300 ms. 5 runs per configuration to account for personality variance.

Team play (N humans, 300 ms reaction each)
Tier	1 ghost	2 ghosts	3 ghosts	4 ghosts
Beginner	100%	100%	80%	100%
Intermediate	60%	60%	40%	80%
Expert	20%	0%	20%	20%

Solo multi-ghost (1 human cycling N ghosts)
Tier	2 ghosts	3 ghosts	4 ghosts
Beginner	100%	80%	100%
Intermediate	0%	60%	40%
Expert	0%	0%	0%

Key findings:

Beginner is well-tuned — always winnable regardless of player count or solo/team mode. Flat 80–100% across all configurations.
Intermediate 4-player team (80%) is the sweet spot — the surround strategy works well with full coordination. Win rate goes 60% → 60% → 40% → 80%, rewarding 4-player team play.
Expert challenge curve is flat — 20% → 0% → 20% → 20%. Adding more ghosts does not meaningfully improve win rate. At 100% speed parity, chasing never works and the surround closes too slowly. The challenge does not scale with skill — it is simply hard for all configurations.
Solo 2-ghost on Intermediate is a cliff — 0% win rate, down from 60% with 1 ghost. The effective reaction time per ghost doubles (600 ms), making coordination too slow. This regression means a player trying to "level up" by adding a second ghost gets punished instead of rewarded.
Solo multi-ghost on Expert is always 0% — no configuration of solo multi-ghost can win on Expert. The 100% speed parity combined with slower per-ghost reaction time makes it impossible.

Conclusion: The baseline curve had Expert flat at 10–20% (multi-ghost skill not rewarded). A comprehensive sweep of 11 configurations × 10 runs × 3 tiers × 4 ghost counts (1,320 simulated games) tested every option individually and in combination. None of the flat speed bonuses produced the desired wide range — they all pushed 1-ghost Expert too high (50–60%) while keeping 4-ghost at the same level, resulting in a narrow 50–70% range.

The solution is a coordination bonus: when 2+ HUMAN-mode active ghosts exist simultaneously, each receives +15% speed. This is the only mechanism that creates a wide Expert range without making 1-ghost too easy, because the bonus does not activate for solo 1-ghost control. Combined with IDLE_S increased from 1.5 to 2.0 s (for solo multi-ghost viability), the final configuration is:

Expert speed: 1.00 (parity — 1-ghost stays hard at ~25%)
IDLE_S: 2.0 s (compromise — solo multi-ghost viable without being too forgiving)
Expert empower: 6 s (original — no change needed)
Coordination bonus: +15% when 2+ HUMAN active ghosts (NEW — directly scales with team size)

Team play — final verified curve (20 runs)
Tier	1 ghost	2 ghosts	3 ghosts	4 ghosts
Beginner	100%	100%	100%	100%
Intermediate	80%	85%	55%	80%
Expert	25%	40%	35%	55%

Beginner is perfectly flat at 100%. Intermediate ranges 55–90% — viable for any team size. Expert ranges 25–55% — a significant improvement over the baseline 10–20% flat curve. The 1-ghost rate (25%) is close to the design target of ~20%, and the 4-ghost rate (55%) rewards team coordination. The 3-ghost dip (35%) is a known limitation of the BFS simulation — real human teams would coordinate better with 3 players. The coordination bonus is the key innovation: it rewards multi-ghost skill without inflating 1-ghost difficulty.

Tuning methodology: all individual options and combinations were tested with 10–20 runs per configuration per tier per ghost-count (1,320+ simulated games total across 11 configs, plus 20-run verification of the final config). The coordination bonus was tested at +10%, +15%, and +18% — +15% produced the smoothest curve. See tests/ghost-squad/curve_sweep.js, tests/ghost-squad/expert_tune.js, tests/ghost-squad/expert_tune2.js, tests/ghost-squad/expert_tune3.js, tests/ghost-squad/expert_tune4.js, and tests/ghost-squad/verify_curve2.js for the full sweep scripts and tests/ghost-squad/README.md for documentation.

10. Internationalization

Engine. All user-visible text is served by the shared, offline i18n engine public/shared/i18n.js plus a per-page dictionary. Ghost Squad is registered as page id ghost and its translations live in public/shared/lang/ghost.js (namespace window.RG_I18N_DICT.ghost). English is stored in the dictionary like any other language (the en block) and doubles as the default: when the active language has no entry for a key, the en entry is served. The markup's own text and the T(key, EnglishDefault) second argument remain as last-resort fallbacks.

Aspect	Definition
Languages	8: `en`, `zh`, `es`, `de`, `fr`, `ru`, `ja`, `ko`
Detection order	`localStorage 'rg.lang'` → browser language (`navigator.languages` / `navigator.language`; exact id, then 2-letter prefix) → `en`
Static markup pattern	`data-i18n="ghost.<key>"` on an element holding the pre-init English text; `RG_I18N.init('ghost')` swaps in the translation for the active language — English included, from the dictionary — and captures the original once as the last-resort fallback; covers the document `<title>` as well (key `doc_title`)
Dynamic text pattern	`T = window.RG_I18N ? function(k,d){ return RG_I18N.t('ghost',k,d); } : function(k,d){ return d!=null?d:k; }`, always called as `T('ghost.<key>', 'English default')` — used for the canvas HUD, tier cards, the attract best table, the personality banner, and the end overlay
Re-render on language change	JS-built DOM registers `RG_I18N.onChange(...)` which re-renders the attract best table, the tier cards, the "TONIGHT'S OPPONENT" banner (while visible) and the end overlay (while visible); the engine re-applies every `data-i18n` element itself
Switcher	The engine injects a fixed top-right `<select>` language switcher; the choice persists in `localStorage 'rg.lang'`

The test harness is unaffected. The deterministic suite in §9.1 loads only the inline <script> block with a stubbed DOM, so window.RG_I18N is undefined there. The guarded boot if(window.RG_I18N) RG_I18N.init('ghost') is skipped and T() returns its English default, so all 40 tests keep passing and the game logic, tuning constants, RNG, boot sequence, and canvas drawing math are untouched.