Stealth

What it is. Guards with legible vision cones and noise radii patrol a space. You are weak in the open and safe in the seams. The game is a live routing puzzle: read perception, find the gap, move through it, don’t be seen.

Player fantasy / why it’s fun. “I threaded the needle.” The held breath of slipping behind a guard whose cone you watched, timing your dash to the exact beat his back turns. Power fantasy inverted: you win by being unseen, and the world is more dangerous than you.

Pillars

Perception is legible. Cones, hearing radii, and alert states are drawn or inferable. The player plans against visible danger — a threat you can’t read is a cheap death, not tension (FUN.md §5 demands both-ways affordance proofs).
Concealment actually conceals. Shadows hide, cover blocks sight, a distraction pulls a gaze — reliably, through a full patrol loop. If hiding is a coin flip, planning is pointless (pillar 1’s dual).
Exposure is punished fast. Being seen escalates quickly — an alert, a hunt, a fail-or-flee. The knife-edge between hidden and caught is where the genre lives.

The loop stack

Scale	The loop
Moment	Watch a cone sweep; time a move through the gap it leaves.
Encounter	Cross one guarded room: read the patrol, find the safe pocket, chain single-guard windows.
Session	A level/heist: string encounters into a route to an objective and back out.
Meta	Optional ghost (0-alert) runs, alternate routes, tools that reshape perception — mastery of the space.

Essential systems

System	Why it’s needed
`system-enemy-ai`	Patrols, gaze, alert-state escalation, aggro/threat — readable, beatable minds; the danger is the AI.
`system-telegraphs`	The alert ramp (suspicion → search → chase) must be telegraphed so the player can react before it’s terminal.
`system-encounter-design`	Composing cones and pockets; the safe-pocket-at-midpoint and chained-single-guard rules live here.
`system-save-and-checkpoint`	Getting caught should rewind to a fair, recent checkpoint — long re-stealth is punishment tax, not tension.
`system-onboarding`	Teach cone-reading and noise in a no-fail room before the first real guard.
`system-difficulty-and-dda`	Difficulty = cone density and window tightness; assist options widen windows or slow patrols.

Content & difficulty model

Balance on cone-shadow duration, not guard distance. The real constraint is how long the safe gap in a guard’s sweep stays open along your path — the chord of shadow, not raw spacing (FUN.md §5). Compute it against the actual route.
A safe pocket at every long traversal’s midpoint. No open crossing longer than a player can hold under one window; give a place to breathe (FUN.md §5).
Chain single-guard windows; never require joint phases. Waiting for two patrols to align by chance explodes combinatorially into unfair wait times — design so each step depends on one guard’s timing (FUN.md §5).
Prove affordances both ways. Exposure punished fast AND hiding conceals through a full loop. “The bot got through” proves neither on its own — you need the punished-exposure and concealment-holds assertions.

Reference wiring: examples/veilstep — the canonical stealth sim: cones drawn from world.state, the heist-bot 0-alarm run plus the punished-exposure and concealment-holds proofs. Grep docs/API.md for the raycast/FOV and audio primitives.

Signature-mechanic seeds

“X but Y” (process-the-twist) — bend what perceives you or what you control.

Stealth but you control the guards’ distractions, not the thief — you route an AI thief by pulling gazes. (perspective)
Stealth but light is your resource — you carry the only lamp; every step lit is a step seen, so you ration your own visibility. (constraint — pairs with survival-horror light)
Stealth but the guards remember and adapt — patrol a room twice the same way and they pre-empt it. (mechanic-swap — see anchor-shadow-of-mordor)
Stealth but sound is the only sense — no cones; guards are blind and you hide in silence, routing by noise alone. (constraint)
Stealth but you’re the guard hunting an invisible intruder — read their noise, cut off their route. (perspective)

Common pitfalls

Illegible perception. Cones you can’t see or infer make stealth a save-scum lottery. Draw the danger.
Concealment that doesn’t. If shadow only sometimes hides, the player stops trusting the map and starts fleeing — the loop collapses. Prove it holds a full loop.
Joint-phase gates. Requiring two patrols to align is an unfair wait. Chain single-guard windows.
Pocketless marathons. A long open crossing with no midpoint refuge is a memorisation gauntlet, not a plan.
Combat as the out. If fighting your way through is easier than sneaking, it’s a shooter with cones. Make exposure genuinely costly.

Anchors

anchor-shadow-of-mordor — systemic guard memory; enemies that adapt to your routes turn stealth into an evolving rivalry.

Verify

Prove it with FUN.md §5 · Stealth — heist bot 0-alarm run + a punished-exposure assertion + a concealment-holds assertion, balanced on cone-shadow duration with a midpoint pocket per long traversal. Design the plannable danger here; prove both affordances there.

Composes with

system-enemy-ai — readable, beatable patrol minds are the danger.
pattern-pacing-and-tension — the hidden/caught knife-edge is a tension curve.
system-encounter-design — cones and pockets composed into a route.