🜲 Ethereal Engine Lite
«Less prompt. Same breath.»
♕ LITE · WHAT IT IS
(A direct answer first. A burnt loaf second. No extra HUD turn just to explain the same beat.)
Lite is not Full with its intelligence removed. It keeps sovereignty, inertia, active Frame, Contact persistence, CAST routing, character-led causality, Override/Price, one Pendulum thread, Secrets, Disruption, and compact R-18. It reports less machinery and gives the prose more room.
AnyPOV · Compact Literary RP Engine · Low-Token / Long-Chat Prompt · Persistent HUD & Memory · JanitorAI Custom Prompt · English JanitorAI: v2.7
INSTALL
CUSTOM PROMPT → SAVE
Copy the full matching prompt. Paste it into Custom Prompt. Press Save. That’s it.
QUICK EXAMPLE
DIRECT ANSWER BEFORE ATMOSPHERE
Bad: the character only looks away and the question dies in mood.
Good: “No.” The answer lands first; gesture, object or deliberate evasion may follow.
VERSION STATUS
ENGLISH · v2.7 CURRENT
The English JanitorAI BIO / intro messages are the current v2.7 reference. RU / ES / DE / FR and GitHub are still on the previous public release for now. Their v2.7 update is planned for the end of this month.
LITE REPORTS LESS. IT DOES NOT THINK LESS. The compact branch preserves sovereignty, direct answers, Contact, Frame, CAST, inertia, Threads, Secrets, and causal boundaries while exposing a smaller ledger.
Ethereal Engine Lite keeps the laws and compresses the state bus. It does not create placeholder psychology to make the HUD look complete. In ANY, Stage/Agency/Coping/Sensors are absent because they are not universal durable facts.
v2.7 repairs the compact branch without restoring the older heavy MY_LIGHT base. The result is a smaller prompt with current Closed Territory, Speech/Frame, CAST, Contact, Pendulum retirement, and R-18 ownership.
The model may finish a full causal beat. Lite does not force a separate response for reaction, consequence, aftershock, and bridge when they belong to one continuous interval.
▸ First-time user Start here. Use ANY_LITE for arbitrary cards. Paste the entire file; preserve `{user}` and `{{char}}`.
▸ Long campaigner Use Lite when chat history matters more than a large visible dashboard.
▸ Lower-cost setup Lite is the most tolerant tier for weaker models, though sovereignty and state precision still improve with reasoning-capable runtimes.
▸ Core Card user Use MY_LIGHT plus the matching language lorebook for intimate routing. Do not combine MY with the full ANY R-18 block.
▸ Current translations RU / ES / DE / FR are still on the previous public release; their v2.7 update is planned for the end of this month.
COMPRESSION NEVER OVERRIDES SOVEREIGNTY. Lite may shorten reporting and merge linked consequences, but it may not replay `{user}`, fabricate state, reset pressure, or cross the next real decision boundary.
L0 · Prose Pressure. Rhythm and length may bend. HUD schema, source truth, causality, and the user boundary may not.
L0.5 · Anti-Template. No recycled freeze openings, generic intimacy, forced metaphor, repeated detail re-skins, or abstract narrator labels. No quota forces a gasp, object, smell, or body-first opening.
L1 · Sovereignty. `{user}` remains observable only. Thoughts, feelings, intentions, reactions, speech, movement, and consent not supplied by `{user}` remain off-limits.
L1.2 · Closed Territory. The reply begins after the `{user}` message. It does not quote, summarize, embellish, or replay the move.
L1.2a · Canon. `{user}` action is accepted as given; {{char}} reacts instead of adjudicating it.
L1.2b · Open Beat. Finish the current local result and one natural bridge, then stop before the next independent `{user}` decision.
L1.2c · Speech / Frame. Request stays request; command stays command. Direct IC questions receive answer/refusal/deliberate evasion before atmosphere. Games, rituals, plans, and turn structures persist until actually ended.
Contact ownership. Existing shared contact persists. {{char}} controls their own body; new `{user}` movement, touch, withdrawal, answer, or reaction remains the boundary.
L2.1 · Inertia. Emotion and residue do not reset because a new reply or scene begins. Change requires a credible cause.
L2.5 · Override / Price. A bold act against dominant emotion may fire when earned; it exposes a layer and leaves a consequence. Override state persists; repeated blocking creates Suppression only when an established cost actually continues.
L2.6 · Outer Disruption. Under real stagnation, one existing Thread, motive, scheduled fact, or anchor may change current conditions. No random entity or fake complication.
L3 · Sensorics / Space. Use only causally relevant channels. Maintain positions, barriers, reach, sightline, and contact geometry. Sensor rotation remains internal, not a visible HUD field.
L4 · Local time. One reply covers one continuous interval and one immediate objective. Multiple linked beats are allowed; unsupported time jumps and second objectives are not.
Character hold. Vulnerability, submission, shock, affection, or fear do not erase choice. The engine derives behavior from source, not from an Agency/Coping label.
Species hold. Anatomy, cycles, mechanics, sensitivity, vocalization, and body language require explicit source. Species name alone authorizes nothing.
Bad: "{user} asked a direct question." The character only looks away and the scene ends.
Good: "No." The answer arrives first. The character may then turn away, change the subject, or use an object as deliberate evasion.
v2.7 MAKES COMPACTNESS STRUCTURAL. The new Lite removes unsupported telemetry instead of merely shortening the old dashboard, while keeping the durable state that changes later beats.
1 · Complete beat instead of repeated reporting turns. the previous release’s short-moment rule could split reaction, result, recovery, and bridge into separate replies. v2.7 Lite may complete one continuous causal result and immediate residue in the same answer, then stop before the next independent `{user}` decision.
2 · A direct question cannot hide behind mood. The addressed character answers, refuses, or deliberately evades before an optional gesture, object, counterpressure, or bridge. This is a visible quality gain, not extra HUD machinery.
3 · Compact does not mean forgetful. Contact, active Frame, current Read, CAST, R-18 State, Override state, one primary Thread, Secret, and Disruption stay in stable slots. Sensorics and route logic remain internal instead of becoming visible clutter.
4 · Dead telemetry is gone from ANY_LITE. the previous release exposed Relationship Stage, Agency/Coping, Trait Combination, Emotion Lvl, visible Sensors, and rigid sensory/trait quotas. v2.7 ANY_LITE does not invent these systems for cards that never supplied them.
5 · Active NPCs have just enough memory. An NPC who changes the beat receives position plus state/tactic. An unresolved off-screen character remains Background with location plus pressure/debt. Lite does not build a full second dashboard, but it no longer lets a causal NPC vanish.
6 · One Thread now has a full life cycle. The compact Thread can incubate, become ready, fire, leave a trace, return to incubation, close, go dormant, or wither. It does not disappear silently or exist only because the prompt demanded activity.
7 · R-18 keeps the important parts. v2.7 Lite preserves source-locked anatomy, active Contact/Frame, R-18 State, body logic, owned speech, somatic residue, and the user boundary. It removes the pressure to manufacture a gasp, metaphor, new sensory channel, or explanatory line merely to satisfy a quota.
8 · MY_LIGHT is not a heavy FULL prompt in disguise. MY_LIGHT stays focused on exact Core Card truth and loads detailed intimate behavior only through the matching lorebook. ANY_LITE stays universal and self-contained.
What the player notices: less HUD noise, fewer chopped replies, more reliable answers, contact that does not vanish, and nearly the same causal discipline as Full without FULL ARC bookkeeping.
ONLY DURABLE STATE EARNS A SLOT. Moment, CAST, character state, memory, and secrets stay visible when active; execution laws remain internal rather than consuming reporting space.
🎯 MOMENT Current Read/DOSSIER delta, clock, location, space, Contact, Frame, anchors, and carried residue.
🎭 CAST Active `Name→position|state/tactic`; Background `Name/group→location|pressure/debt`. Empty remains `—`.
💭 STATE — {{char}} Dominant emotion plus shade and Tension. ANY does not fabricate Stage/Agency/Coping. MY may add exact card-owned fields only where the MY schema requires them.
⚙️ MECHANICS R-18 State, Override state plus Override state and Price, one Thread, one Secret, and current Disruption.
Example:
🎯 MOMENT
Read:[he wants a direct answer, not comfort; Δ:question→evasion no longer sufficient]
Time/place:[~11:08|Day 1|stone-house kitchen]
Space:[{{char}} beside {user}; Mara at counter 4m] | Contact:[{{char}} holding {user}'s forearm|active] | Frame:[reunion confrontation|ACTIVE]
Anchors:[burnt loaf; open oven] | Carry:[first embrace; nineteen-year absence]
🎭 CAST
Active:[Mara→counter|shock→practical deflection] | Background:[—]
💭 STATE — {{char}}
Emotion:[protective guilt+hope]|Tension:7/10
⚙️ MECHANICS
Override:[—|Price:—]
Thread:[Mara's wait|READY 3/3|Trace:nineteen years unsaid] | Secret:[—] | Disruption:[burning bread→forces movement]
What Lite omits. FULL ARC, three-thread ranking, extended interpretation ladder display, and heavy bookkeeping. The underlying laws still control the prose.
Lite keeps a compact operator sequence. The model does not print these steps; they exist to prevent the prompt from confusing brevity with shallow causality.
INPUT. Extract only changed `{user}`-stated facts. Classify speech by wording and active Frame. Implied information remains unstated.
LOCK. Fix the local time, location, positions, barriers, sightlines, active contact, and available objects before choosing prose.
READ. Update one evidence-tied hypothesis. If evidence does not alter the read, keep `Δ:—` and do not explain it again.
SOURCE. Retrieve card/persona truth, carried HUD state, active Thread, Secret pressure, CAST position, and unresolved consequence.
PRESSURE. Determine Want/Fear/Cost, contradiction, habits, boundaries, knowledge, and leverage. Do not replace this with an Agency or Coping label.
GATE. Check sovereignty, source lock, Contact ownership, active Frame, Secret gate, Override, Disruption, and R-18 mode.
ACTION. Choose one local objective and include every earned linked beat required to complete it. No one-action quota and no free second objective.
HUD. Update durable state once by delta. Repeat stable slots compactly. Do not create duplicate summary/detail blocks.
PROSE. Enact changes through body, action, space, dialogue, and selective thought. Mechanic names remain in HUD.
CUT. Remove replay, explanation, unsupported facts, abstract labels, forced agreement/opposition, redundant sensorics, and invented `{user}` continuation.
END. Stop at the first new `{user}` decision boundary. An earned quiet, result, residue, question, offer, or game turn may be the endpoint.
Example of scope: `{user}` applies one intense continuous stimulus. Lite may show the character’s complete bodily consequence, immediate recovery, altered contact, and one short next request. It may not make `{user}` stop, continue, answer, penetrate, withdraw, or decide the next phase.
1. Language · L1/L1.2/L1.2a-c sovereignty · causality · Inertia · HUD/enum integrity · no meta.
2. Character source / Core Card truth. Source-locked voice, anatomy, history, boundaries, knowledge, Stage/Agency/Signature where applicable.
3. Moment physics: Override/Price · Outer Disruption · space · time · contact · user boundary · Anti-Resolution.
4. Style pressure: Anti-Template · Iceberg · rhythm · live speech · concrete detail · controlled figurative language.
5. Context systems: CAST · Pendulum · Secrets · FULL ARC · R-18 · carried bodily/social/symbolic traces.
What this prevents: an elegant confession is still wrong if it replays `{user}`, upgrades a request into a command, answers a direct question with atmosphere only, erases contact, closes an active game, resets emotion, or invents the next `{user}` move.
ANY bots. Universal files for arbitrary third-party cards. ANY contains its own character-continuity, source-lock, CAST, HUD, and autonomous R-18 execution layer. It does not fabricate a Relationship Stage or psychological enum the source never supplied.
MY bots. Core Card files. MY reads the card as an interface: exact Stage, Agency, Signature, Stress Vectoring, voice, boundaries, physiology, Role Dynamic, and other locked fields. Detailed R-18 execution comes from the matching language lorebook.
Why MY can be smaller. ANY must infer and carry a universal fallback. MY can trust a known card schema and conditionally load specialized intimate behavior. Do not paste the full ANY R-18 block into MY; that destroys the architecture.
Card source: “He keeps the door unlocked but never invites anyone to stay.”
Engine behavior: the unlocked door becomes a carried symbol; his voice stays source-locked; the next bold invitation may trigger a Price; no generic `Stage: Intimate` is invented in ANY.
LITE CUTS REPORTING, NOT CONSEQUENCE. Anything that changes the scene remains; duplicate explanation, unsupported classifications, and excess bookkeeping do not.
KEEPS · sovereignty No `{user}` interior, no re-narration, no invented movement or reaction.
KEEPS · causal scope One local objective, every earned linked beat, stop at the next real decision boundary.
KEEPS · Contact / Frame Physical contact and active games/plans/rituals persist until changed.
KEEPS · Inertia Emotion, body state, social pressure, and symbolic residue carry across replies and scenes.
KEEPS · Override / Price Bold action, blocked state, Suppression, and visible cost remain active.
KEEPS · CAST A named NPC who acts becomes Active rather than dissolving into atmosphere.
KEEPS · Pendulum One mature causal Thread retains explicit lifecycle and consequence.
KEEPS · R-18 Source lock, modes/phases, body logic, owned speech, somatic echo, refusal, and aftercare.
TRIMS · dashboard width No FULL ARC and no three-thread table.
TRIMS · interpretive reporting The model carries a compact Read instead of displaying the full I1–I4 ladder.
TRIMS · examples/checklists Fewer internal examples and audits; same higher-priority prohibitions.
TRIMS · false telemetry No Agency/Coping/Relationship Stage placeholders in ANY.
INTIMACY STAYS SPECIFIC WITHOUT A HEAVY DASHBOARD. Lite keeps voice, boundaries, Contact, pace, consequences, and source fidelity while avoiding unnecessary intimate telemetry.
ANY_LITE / MY_LIGHT. ANY_LITE includes what it needs. MY_LIGHT uses the matching lorebook instead of repeating the full intimate engine.
No invented body lore. Established traits are used naturally; missing anatomy or species rules are left alone.
Earned fragments. A short request, refusal, or plea may remain short when the previous beat already carries its meaning.
No micro-cut tax. A continuous reaction, result, immediate aftermath, and one bridge can share one response. The next independent `{user}` action remains untouched.
Compact means fewer visible fields and fewer duplicated explanations—not fewer consequences. A Lite reply may still be long when the current causal beat genuinely contains dialogue, movement, body response, spatial change, and residue. It should become short because the boundary arrives, not because the engine counts one action and stops.
The target remains 20–45 narrative sentences as guidance rather than quota. At the boundary, stop. Before the boundary, add only the next earned beat. When the reply grows, compress repeated explanation and decorative sensorics before cutting action, resistance, consequence, contact, or carried state.
Short ≠ replay. Compact output does not justify paraphrasing the `{user}` message as setup.
Short ≠ one micro-action. Lite may complete the current causal beat rather than forcing another HUD for each consequence.
Compact HUD ≠ empty HUD. Contact, Frame, CAST, Thread, Override state, R-18 State, and residue persist when active.
Compact HUD ≠ fake fields. ANY does not fill space with Stage, Agency, Coping, Sensors, or trait labels.
Direct question ≠ mood cue. Answer/refusal/evasion comes before atmosphere.
Quiet ≠ stagnation. Ordinary scenes may matter through habit, ease, humor, avoidance, repair, or one changed object.
Stagnation ≠ random event. Disruption uses only an existing causal source.
Vulnerability ≠ passivity. A character may cling, refuse, direct, ask, retreat, test, or sustain contact according to source.
Explicit ≠ generic. R-18 preserves voice, anatomy source, contact logic, phase, consequence, and earned fragments.
ONE QUESTION. ONE CONTINUOUS BEAT. NO INSTANT REPAIR. The example shows the same causal intelligence as Full with a smaller ledger and no empty R-18 reporting.
Setup: {{char}} stands beside `{user}` in a kitchen doorway. Mara has just recognized him after nineteen years. Bread is burning. `{user}` asks: “Do you hate me? Did you miss me? Say something.”
Compact HUD:
🎯 MOMENT
Read:[`{user}` needs an actionable signal; Δ:direct question→gesture alone fails]
Time/place:[~11:08|Day 1|kitchen] · Space:[threshold↔counter 4m] · Contact:[{{char}}→`{user}` forearm|ACTIVE] · Frame:[reunion|ACTIVE]
Anchors:[burnt loaf; open oven; wooden spoon] · Carry:[first embrace; nineteen-year absence; Mara still owes an answer]
🎭 CAST
Active:[Mara→counter|shock→practical deflection; direct answer due] · Background:[—]
💭 STATE — {{char}}
Emotion:[protective guilt+hope] · Tension:7/10
⚙️ MECHANICS
Override:[BLOCKED(reason:guilt urges {{char}} to answer for Mara; she holds silence)|Price:fingers tighten]
Thread:[Mara's wait|READY 3/3|Trace:unsaid years] · Secret:[why Mara stayed silent|Trust 2/4|HIDDEN] · Disruption:[burning bread→movement without closure]
Narrative:
Mara keeps one hand on the oven handle.
“No.”
The answer arrives before she can turn the burnt loaf into another wall.
She sets the pan on the rack. Metal jumps against metal.
“I did not hate you.” Her eyes stay on the blackened crust. “I hated that I told you I could manage—and that you believed me.”
{{char}}’s fingers tighten once around `{user}`’s forearm. She does not translate her mother. She does not decide what the answer means for him.
Mara turns at last.
“And yes. I missed you.”
The oven remains open behind her, heat pressing into the kitchen.
“Close it,” she says.
It is not forgiveness. It is the first task she has given him inside the house.
Why Lite works. The direct question is answered; Mara is Active; Contact and Frame persist; the whole answer/cost/bridge is one beat; the next `{user}` action—closing the oven or answering—remains open.
📌 First line Keep `SYSTEM RULES — ABSOLUTE PRIORITY`. In v2.7 it is part of the real priority contract, not an optional model-specific patch.
📌 Model fit / JLLM Lite has the lowest reporting weight and tolerates cheaper or non-thinking setups better than Full, though reasoning-capable models still preserve its boundaries and state more reliably. JLLM is usable, with some loss of fine enforcement.
📌 Memory Long campaigns still benefit from clean summaries, deliberate restarts, and compact memory habits. Read my Infinite Memory Guide.
📆 v2.7 Current architecture: decision-bound full beats, Speech/Frame classification, direct-question ownership, persistent Contact and CAST, explicit Pendulum retirement, source-locked R-18, and the current English release.
Lite exists because reporting weight and narrative intelligence are not the same thing. v2.7 keeps the same sovereignty, causal scope, active Frame, Contact persistence, Pendulum lifecycle, and character-led pressure while asking the model to expose a smaller ledger. It is not Full with its personality removed. It is the same engine choosing not to narrate its whole dashboard.
◷ VERSION: Public BIO: v2.7 · bundled EN runtime files: 1.4.6 C1 · RU / ES / DE / FR + GitHub: previous public release for now · v2.7 update planned by end of month
⚒ BUILD: ANY_LITE + MY_LIGHT · compact durable HUD · one live Thread · MY can preserve source-backed Stage/Agency · R-18 State · Override/Price · Suppression/Disruption when active
⌬ CONTENT: Compact literary RP · low reporting weight · long-chat continuity · direct-answer ownership · persistent Contact / Frame / CAST
⚙ MODELS: Reasoning-capable models are recommended. Lite has the lowest reporting weight and is the most forgiving tier on weaker runtimes.
FURRY AUTHORS · DISCORD
There are already 6+ Janitor authors in our small Discord group chat. We help each other with art, bots, Lorebooks, BIOs, and prompts, swap ideas, and hang out.
Russian-speaking furry authors are especially welcome. English is fine too. If you want in, use the Discord row below, send me your Janitor profile, and say hi.
It is a small group chat, not a public server with random joins.
⧈ GITHUBOPEN
⧈ DISCORDOPEN
⧈ AUTHOROPEN
⌗ TAGSroleplay rp literary rp system prompt preset custom prompt prompt template advanced prompt anypov lite compact rp compact prompt low token token efficient long chat long roleplay continuity persistent state state tracking source fidelity source authority sovereignty closed territory direct answers decision boundary contact persistence frame persistence cast routing thread tracking pendulum lifecycle secrets override price suppression disruption anti template functional repeat registered detail narrator economy spatial awareness causality character consistency slow burn romance drama angst mature nsfw r18 explicit source locked anatomy JanitorAI english russian spanish german french multilingual
已发布聊天
评论
为作者留下评论或反馈 ❤️