OPNAORTA  /  THE CUTSCENE
POSITION PAPER · REV 0

THE CUTSCENE

The most expensive three minutes in games is a recording with a lock on the door.

Photoreal faces, motion-captured performance, an orchestral score — and the player's controller goes dead. Then a prompt appears and says PRESS F, because the only agency the scene can absorb is agency an author pre-cut a hole for. This page is about the organ that removes the lock, and about doing it without generating a single unauthored word.

Substrate · measured on one consumer GPU Cutscene integration · specified Position paper No unauthored speech required Ships with its tape
THE MOVE: RE-TIME, DON'T GENERATE
A studio has already recorded everything the scene needs — the lines, the barks, the alternate reads, the performance capture, all cast, cleared, and certified. What it does not have is anyone home to decide when. A resident mind changes only that: it watches the scene continuously, judges at every boundary whether this instant deserves a line, a look, a held breath — or the death of a sentence already in the air — and selects from what the writers wrote. Nothing new is spoken. The scene stops being a recording and becomes a performance that is happening, because the timing is being decided now instead of in 2023.

§ 01 — The diagnosis

The lock is not a design choice.
It is an admission.

Ask why control is taken away for three minutes and the honest answer is: because nothing in the scene can respond to you. The camera is on rails because a camera that could be surprised would need something watching. The prompt appears because an authored branch is the only kind of agency a recording can hold.

Every technique the industry has invented to soften this is a compensation for the same vacancy — and each is real engineering aimed one layer above the cause:

The techniqueWhat it is actually compensating for
Skip buttonThe scene cannot tell whether you are engaged, so it offers to stop existing.
Quick-time eventAgency, pre-cut into a hole by an author months ago.
Branching dialogue treeChoice, discretized into what could be recorded in advance.
Barks & ambient VOPresence, faked with a timer and a random draw.
In-engine "seamless" cutsceneThe seam removed from the rendering. The seam in the mind is untouched.
Turn-based LLM NPCA mind, admitted to the room — that exists only while it speaks.
A resident mind"Does this instant deserve anything — a line, a glance, a held breath? And does the world need the floor?"

Notice the shape. Faster inference does not remove the lock; it makes the locked interval shorter. A better voice model does not remove it; it makes the recording lovelier. Even an LLM bolted into the engine does not remove it, because a turn-based mind is absent while you talk — which is exactly the interval a cutscene consists of.

A cutscene is the industry's most expensive apology for the fact that nobody is home.

§ 02 — The inversion

Interactive already exists.
The scene is where it is taken away.

This is the strange part. Games solved presence everywhere except the story. In gameplay, the world is continuous, everything reacts, and interruption is the medium. Then the cutscene begins and the character who dodged your grenade three seconds ago cannot notice you have walked out of the room.

Every seam in a modern game is smooth except this one, and the reason is architectural rather than artistic: gameplay runs on a continuous loop and story runs on playback. The engine ticks sixty times a second; the scene ticks once, at the author's desk, years ago. A resident mind is the first thing that can put story on the same clock as the world — because it is the first mind that has a clock of its own.

And a game world is the perfect body for this organ. The hardest input for any resident mind is knowing what is happening; in a game, every happening is already an engine event — entity-keyed, timestamped, pre-registered, deterministic. Perfect instrumentation is native here and nowhere else. The percept stream a mind needs is a byproduct the engine already emits and currently throws away.

Which sets up the whole opportunity in one line: the studio has the performances, the engine has the percepts, the audience already knows how the scene should feel. The one missing piece is a mind that is present for the three minutes — and presence, not intelligence, is what nobody has been shipping.


§ 03 — The organ

What a resident mind does
with three minutes

The substrate is FUSOR: a resident, multi-lane, text-backbone mind that never takes turns. It ingests the scene as fused token streams — dialogue, engine events, player input, and now a sighted lane — holds its memory in VRAM instead of rebuilding it, and decides at every boundary whether this instant deserves anything. Four of its properties matter for a cutscene, and each is a dated bench number rather than a promise.

PropertyWhat it buys the sceneStatus
Boundary detectionKnows where a thought ends — 0.97 at a finished clause vs 0.02 mid-phrase, free on the ingest pass. A character can come in on your clause instead of after your silence.MEASURED
The un-sayA line already in the air dies mid-word when the world changes — measured at 1 ms to silence, with the exact words that reached the air on the tape.MEASURED
Judgment at boundariesHold or speak, decided from inside the weights ~30 ms after the event, not by a timer and not by a designer's threshold.MEASURED
Fork & abort economicsA second perspective on the same attention state costs 0 MiB; killing a forming branch costs 13 µs. Three minds on one trunk run at 1.208× the cost of one.MEASURED
The scene, heldSomething seen once is answerable later from the trunk with no re-look — 6/6 at distance against a 1/6 floor on the bench.MEASURED
The listening facePrediction error and boundary mass, already computed at ingest, drive brow, lids and gaze while the player speaks — the 95% of a scene where turn-based characters are structurally blank.SPECIFIED
The world's floorScene events are seated as a third speaker and may outrank both the player and the character mid-utterance.SPECIFIED
Receipt · the measured climax, another scene, same substrate A sentence was forming — "That light was the mountain ridge breaking through the cloud layer — it's why we keep the—" — when a hand entered a door-handle zone. Verdict at +30 ms, the line killed at +52 ms, a second seat on the floor at +80 ms, the corrected line committed at +430 ms. The identical model, run turn-based on the identical inputs, missed the same moment by 65.7 seconds. Wall-clock, no acceleration, on one consumer GPU.

Read that as a cutscene spec and it says: the character can be interrupted, and can interrupt herself, at the granularity of a syllable. Not at the granularity of a dialogue node.


§ 04 — The ladder

Five modes, and the first one
changes nothing on screen

A studio cannot bet a flagship on a character who improvises. It does not have to. Presence arrives in modes, each one shippable alone, each gated, and the floor mode is byte-identical to the cutscene you already cut.

0The floor

INERT — it watches, judges, and never speaks

The resident ingests the scene and the player's actions, judges at every boundary, and writes every verdict, margin, and hold to a hash-chained tape. Its output is wired to nothing. The scene plays exactly as authored, frame for frame. Ship value on day one: a replayable record of where a character would have acted — a QA instrument that tells a director which three seconds of a two-minute scene the audience is actually being ignored in.

Gate · F-INERT: byte-identical playback, or it does not ship
1The face

THE LISTENING FACE — reaction without speech

The ingest signals already being computed — prediction error, boundary mass, a valence probe — drive brow, lids, gaze and micro-orientation while the player is the one talking. No line is spoken, no branch is taken, no legal surface is opened. This alone kills the deadest tell in the medium: the character who is a mannequin until it is her turn.

Gate · F-LISTENING-FACE: in-speech reaction < 400 ms
2The stop

THE UN-SAY — a recorded line can die mid-word

The character is mid-sentence and the player draws a weapon, walks out, or the world detonates. The line stops where it stands — not at the end of the audio file — and the mind knows exactly which words reached the air, because that is on the tape. This is the single most uncanny-valley-crossing effect available, and it generates nothing: it is a kill switch on playback, decided by judgment.

Gate · F-CUT: audible mid-word death, ≤ 10 ms from the kill
3The choice

SELECTION — which recorded line, and when

The writers' bank is the vocabulary; the resident chooses within it, continuously, on the world's clock. Alternate reads that today are wasted on a random draw become ranked responses to what actually just happened. Authorship is untouched: every word was written, cast, recorded, and approved. Only the timing and the selection are live.

Gate · F-BEATS: the story still lands under player agency
4The room

THE WORLD TAKES THE FLOOR — the third speaker

A conversation in a turn-based system has exactly two parties. A scene has more: the door, the fire, the second character, the thing arriving. Seated on the same trunk, the world can outrank both the player and the character mid-utterance — which is how a scene becomes a place rather than an exchange.

Gate · F-ORDERING: the reaction precedes the player's report of it, timestamped
5Optional

GENERATIVE — unauthored speech

Last, optional, and per-studio. Everything above ships without it. This is deliberate: the entire felt effect of presence — reaction, restraint, interruption, the mid-word stop — is available before a single word leaves the writers' room, and with it the entire liability surface of a character who can say anything to anyone.

Gate · studio policy, not architecture
The conservation move. A racing team once found a second a lap not by making more power, but by catching the braking energy their rivals were dumping into the brake discs as heat. Modes 0–4 are the same trick pointed at drama: they create no performance, they re-time performance the studio already recorded — and the energy currently being thrown away is the player's own attention during the three minutes their controller is dead.

§ 05 — The hard part

The primitive is not the problem.
The policy is.

Here is the failure mode this page refuses to hide, because it is measured: a resident mind with its dial at zero fires 921 times an hour — unlivable. Tuned to a fixed threshold instead, it catches 36 of 39 planted moments and is deaf exactly where it matters. Raw capability makes a scene worse, in a specific and predictable way: a character who reacts to everything is not present, she is twitchy.

No single runtime knob separates those jaws. That is the whole reason this cannot be shipped as a plugin with a sensitivity slider, and it is where an honest program spends its year. The resolution has two halves:

1 · The disposition lives in weights, not in a slider

When to speak is not a threshold; it is taste, and taste is trained. The judgment has to be compiled into the model itself — which is unglamorous, corpus-shaped work, and is the actual reason the industry has not already done this.

2 · The scene tunes itself in a sandbox

A cutscene arrives with something no other domain has: a reference lap. The authored cut is the optimum the scene was designed to hit. So the objective can be written down honestly — maximize the player agency the scene absorbs, subject to every authored beat still landing — and scored automatically, offline, against the engine's own event log, across tens of thousands of synthetic players who interrupt at four seconds, walk out at thirty, or draw a weapon at sixty-one.

THE SANDBOX   one authored scene · its beat list · its line bank
                 the engine is the physics model — free, deterministic, entity-keyed

   synthetic players ──▶ candidate disposition ──▶ tape ──▶ score
        (thousands)          (dial, ranking,          │        agency absorbed
                              refractory, floor)      │        subject to beats landing
                                   ▲                  │
                                   └── optimizer ◀────┘
                                        output: weights, not a rulebook

   PROFILES  stalk · firefight · negotiation · confession · ambient
             same weights, opposite dispositions — tuned per moment-genre, swapped per scene

Different moments want opposite characters. A stalk sequence wants forty seconds of silence and a face doing all the work; a firefight wants a callout every three seconds; a negotiation wants interruption precisely at the clause boundary. These are not different models — they are profiles, tuned in the sandbox, swapped per scene, amortized across a genre of moments rather than paid for per title.

The trap in the objective, named before anyone falls in

Optimize a raw engagement proxy — dwell, input rate, "did they keep watching" — and the sandbox will breed a character who fidgets to hold attention. That is novelty over relevance, and it is the exact failure this estate has already measured on another instrument. So the loss ships with its own executioner: proxy-ranked policies must agree with blind human preference on held-out scenes above a pre-registered threshold, or the loss is theater and the program stops until it is repaired.

The corpus nobody has read Twenty years of quick-time events are hand-labeled moment-worthiness annotations: a designer asserting, with a timestamp and a required response window, "this instant deserves an act." Every shipped title carrying a PRESS-F prompt contains supervision for exactly the quantity that has to be trained into the weights. It has been sitting in the medium the whole time, read as UI instead of as data.

§ 06 — The contract

Falsifiers, pre-registered

This page is a position, not a product announcement. Positions that cannot lose are advertising, so each claim below is built to be killed — and the substrate rows already carry dated tapes.

NameThe claimIt loses if
F-INERTMode 0 playback is byte-identical to the shipped cut.Any divergence — the safe floor is not safe.
F-CUTA recorded line dies audibly mid-word, ≤ 10 ms from the kill decision. [substrate: measured at 1 ms to silence]The kill is inaudible or late — the anti-turn demo is inaudible.
F-LISTENING-FACEIn-speech facial reaction to semantic events < 400 ms; the turn-based arm structurally scores zero while the player talks.Misses the window — the every-frame claim dies.
F-BEATSEvery authored beat still lands, in a legal order, across the synthetic player population.The story stops landing under agency — interactivity ate the scene.
F-LOSS-PROXYSandbox-ranked policies agree with blind human preference on held-out scenes.They don't — the objective is theater and the program halts.
F-FELTBlind raters prefer the resident cut over the authored cut at parity of beats.They don't — presence is not felt, and this page says so in public.
F-QA-DETThe tape replays byte-identical with wall clock stripped; a nondeterministic-seeming character is certifiable.Any divergence — determinism is theater and QA cannot sign it.
F-ZOMBIEThe turn-based arm speaks through the interrupting event, every run. [substrate: measured, +65.7 s]Any baseline run self-interrupts — the contrast claim dies and publishes anyway.
What is measured, and what is not

The substrate numbers on this page — boundary detection, the mid-word abort and its clocks, fork and abort economics, the held scene, the turn-based contrast — are MEASURED on one consumer GPU with dated tapes in the reference build's runs ledger. Every cutscene integration described here is SPECIFIED: designed, gated, and unbuilt. No shipped game has run this. The sandbox in § 05 does not exist yet. What exists is an organ with receipts, a ladder whose first rung changes nothing on screen, and a contract that prints its own kill conditions.

MEASURED — dated receipt on the reference bench  ·  SPECIFIED — designed, gated, unbuilt  ·  BET — a claim with a printed kill condition

§ 07 — Where it goes

A cutscene is the cheapest place
to prove the most expensive thing

Nothing here is really about cutscenes. The cutscene is a bounded, instrumented, adversarial test of one claim — that a mind can be present in a scene rather than summoned into it — with an authored optimum to score against and a legal surface small enough to actually ship.

The same organ, pointed at a longer scene, is an NPC who is still there while you talk. Pointed at a room with real stakes, it is a rehearsal instrument for the hardest conversations people have to have — where the counterpart's face changes while you are still speaking, and where the craft's hardest skill is not filling a silence. The cutscene is where that reflex can be demonstrated for the price of one scene, in a domain that already has perfect instrumentation and a recorded ground truth.

The industry crossed the visual valley years ago. What is left is temporal — and the fix is not a faster turn. It is not taking turns.
The scene this argument was born in

THE CABIN — one scene, two seats, five minutes at altitude, and the first NPC who is still there while you talk. The climax has fired live on the bench: a sentence killed mid-word by a hand entering a door zone, with the tape to prove it.

THE CABIN — the scene →