# Authored SFX programs Status: Experimental Scope: `@carbonenginejs/runtime-audio` Audience: Library producers and browser application authors Summary: Describes optional portable SFX selection, ordered actions, layering, and RTPC gain behavior. ## Purpose The schema-v2 library may contain an optional `sfx` program. The program describes what an event does after it is posted; it does not describe where bytes come from. Each resolved sound leaf refers to a media ID in the same library, so `CjsAudioMan` can still deliver that ID as an individual file, a slice of a complete original bank, or an exact original-bank range. Without `sfx`, a caller may still supply `eventMedia` as a flat event-to-media fallback. Bank construction with `includeSfx: true` instead derives `eventMedia` from the same validated typed graph. Events that cannot be lowered are absent from both tables; heuristic container-byte reachability is never used as an audible fallback. ## Shape ```js { schemaVersion: 2, events: { weapon_fire: [ { nodeId: "1", delayMs: 100, delayRangeMs: { min: -50, max: 100 }, probability: 75, fadeInMs: 250, fadeCurve: 4 } ] }, programs: { weapon_fire: [ { kind: "switch", group: "ship_size", value: "large" }, { kind: "play", child: { nodeId: "1", delayMs: 100, delayRangeMs: { min: -50, max: 100 }, probability: 75, fadeInMs: 250, fadeCurve: 4 } }, { kind: "stop", targetId: "2", targetFlags: 0, scope: "game-object", mode: "element", delayMs: 1000, transitionMs: 250, curve: 4, actionFlags: 6, exceptions: [] } ] }, nodes: { "1": { type: "switch", scope: "switch", group: "ship_size", cases: { small: { nodeId: "10" }, large: { nodeId: "2" } }, default: { nodeId: "10" } }, "2": { type: "blend", children: [ { nodeId: "10", gainDb: -3 }, { nodeId: "11", gainCurves: [ { rtpc: "weapon_intensity", scope: "object", points: [ { x: 0, gain: 0, interpolation: 5 }, { x: 1, gain: 1, interpolation: 9 } ] } ] } ] }, "10": { type: "sound", mediaId: "777", matchIds: [ "10", "2" ], spatial: true }, "11": { type: "sound", mediaId: "778", playbackRate: 1, playCount: 2, loop: false, spatial: false } } } ``` Node IDs and media IDs are positive unsigned 32-bit identities serialized as strings. The validator rejects missing references and cycles before audio is enabled. SFX schema version 2 makes `programs` the ordered authoring source. When an event has a program, its `events` entry must be exactly the projection of that program's `play` actions. This keeps legacy root lookup available without allowing the static roots and the executable program to disagree. Supplied Stop, Pause, Resume, Voice Volume, bounded Bus-target Voice Volume, Bus Volume, Voice Pitch, Voice LPF/HPF, and Set/Reset Game Parameter actions are also qualified at validation time. playback controls reject unsupported action flags, nonzero All targets, and element-target exceptions. Voice Volume and Voice Pitch accept only exact element targets and their decoded value contracts. Voice LPF/HPF Set accepts an exact element target; Reset additionally retains the qualified element, All, and All-Except modes. Bus Volume retains the complete valid v150 family: global/object Set and Reset Element plus global Reset All and All-Except, with bus-flagged targets and exceptions. Game Parameter actions retain their authored object/global scope, absolute/relative value mode, delay, transition, curve, and transition-bypass flag. Every portable Game Parameter action carries its catalog default so unset relative values and transition start points never guess zero; the library builder also requires an exact parameter-name catalog match. A Voice Volume action whose target flag names an Audio Bus is not Bus Volume. Wwise applies it to voices feeding that Bus before Bus effects, while Bus Volume controls the Bus stage itself; `0x0A03` must not be relabeled `0x0C03`. EVE build 3453885 contains two game-object Bus-target Voice Volume actions on Bus `3810872320` (`Cinematic_Ship_Intro_Transition_Delay`), whose dry route has no Aux sends and has three active Wwise Delay effects: `cinematic_ship_intro_begin` sets `-30 dB` immediately, while `cinematic_ship_intro_climax` schedules `0 dB` after `6000 ms` with a `2000 ms` transition. The portable `set-bus-voice-volume` action and browser backend preserve this cross-event sequence on the posting game object. The begin event stores `-30 dB`; later voices whose output Bus is `3810872320` inherit it, and the climax action schedules the return to `0 dB`. The gain is voice-owned and precedes the target Bus stage, so it is not shared with unrelated emitters or routes. This is deliberately bounded support: only absolute, non-randomized Set on an exact game-object Element target is admitted, and the target must be every affected Sound's first/output Bus with no NodeBase or Bus Aux send. Reset, relative/randomized values, randomized timing, global scope, exceptions, music targets, ancestor-Bus targets, and wet routes remain fail-closed. The cinematic route still cannot enter the strict shared mixer because a parent Voice Volume RTPC would cross the target Bus's three Delay effects. Its media therefore uses the audible legacy fallback: the action envelope is retained, but the rejected shared Delay and Peak Limiter character is omitted. ## Node behavior | Type | Behavior | | --- | --- | | `sound` | Produces one media voice. Optional `loop` overrides event metadata, `playCount` preserves a finite authored repeat count, `playbackRate` controls the buffer source, `spatial` selects the panner (`true`) or flat SFX route (`false`), `dryVolumeCurve` retains the leaf's Wwise distance gain, `sourceEffects` retains one complete qualified effective static effect chain, and the qualified `voiceLimit` shape reserves a cap-one per-game-object instance before media acquisition. | | `silence` | Produces no voice. This preserves authored empty switch/state cases without falling through to the default. | | `timed-silence` | Produces one finite silent voice with authored `durationMs`. It owns lifecycle, routing, Stop matching, Continuous completion, and qualified voice-limit admission without acquiring media. | | `random` | Chooses one weighted child. `mode: "shuffle"` exhausts a pool before refilling it; `avoidRepeat` excludes recent choices. | | `sequence` | Chooses the next child for each post; `loop: false` produces no further leaves after the final child. | | `switch` | Chooses a named case from a per-object switch or global state, with an optional `default`. Matching is case-insensitive. | | `parallel` | Resolves every child into simultaneous voices. | | `blend` | Resolves every child into simultaneous voices, normally with child gain curves for live crossfades. | A stripped Step Switch Container whose group and default IDs are zero and whose Children and Switch assignment lists are both empty lowers to `silence`. Wwise builds no playable game-sync map for that shape. EVE banks may still retain stale per-child switch parameters, but those records cannot select audio without either list and are therefore inert. Preserving the silent action, rather than rejecting the whole event, retains its delay metadata and keeps its audible sibling actions intact. A silent node does not allocate a pending browser selection, so its delay alone does not extend the playing ID lifetime and cannot be paused or stopped while pending. This bounded graph rule restores five jump-drive and cynosural events in EVE build 3453885 across empty nodes `980357672`, `140917810`, `455538153`, and `394178349`. Other empty forms remain fail-closed. An event may have several roots; roots are parallel. Play-action edges may carry `delayMs`, an optional randomized offset in `delayRangeMs`, a percentage `probability`, and `fadeInMs` with an optional `fadeInRangeMs` and Wwise `fadeCurve` from 0 through 8. Randomizer ranges are offsets from the base value. One set of values is sampled for the Play action and inherited by every sound leaf it selects. Delay is measured from the event post, and the fade begins when the delayed source starts. `programs` preserves the authored order of Play, Stop, Pause, Resume, SetSwitch, SetState, Set/Reset Voice Volume, bounded Set Bus-target Voice Volume, Set/Reset Bus Volume, Set/Reset Voice Pitch, Set/Reset Voice LPF/HPF, and Set/Reset Game Parameter actions. Switches update the posting game object; states update the global state table. A switch or state setter therefore affects only later Play actions in the same post when it is immediate. A fixed-delay setter instead enters the same AudioContext-clock action queue as delayed Stop and Game Parameter actions. It does not alter later Play selection early; when due, it updates the posting game object's Switch or the global State and therefore also wakes any live Continuous Switch/State session. An action-only program remains alive until its last delayed setter executes and completes without creating a media voice. Stop of that playing ID cancels its pending setters. Randomized delay, probability, transition-bearing setters, and extra setter properties remain fail-closed. A Stop action has `scope: "game-object"` or `"global"`. `mode: "element"` matches the target HIRC identity, while `"all"` and `"all-except"` apply to all eligible SFX voices; Stop-All uses target ID `"0"`. Exceptions protect any voice whose selected hierarchy path contains their target identity. Builder-produced sound nodes retain omitted raw hierarchy parents in `matchIds`, beginning with the sound node's own identity, so an element Stop can still match a Sound through an Actor-Mixer that is not itself a playable runtime node. Stop delay is measured from the event post. `transitionMs`, its optional randomizer, and `curve` control the authored fade before the voice halts. Timing ranges are sampled once per post. Pending Play slots can be cancelled before their media resolves, and delayed Stop-only events remain alive until their action has executed. At equal scheduled times, authored action order is preserved: Play then Stop is stoppable, while Stop then Play leaves the later Play intact. Event metadata still carries conservative `eventsStoppedBy` relationships for culling and static inspection, but an installed authored program owns the actual Stop timing and avoids a duplicate fallback fade. Pause and Resume use the same scope, target hierarchy, mode, action ordering, and exception contract as Stop. Pause retains the target playing ID and media position instead of reporting completion. Pauses stack, so a voice resumes only after receiving the matching number of Resume actions. A matching Play selection that is still acquiring media remains silent until resumed. Web Audio buffer sources cannot restart, so the backend stops the disposable source on Pause and creates a replacement at the preserved position on Resume. Infinite loops retain their wrapped offset, while finite repeat counts retain their aggregate remaining duration rather than restarting the authored count. Delay and transition values use the same audio-clock scheduling path as other program actions. Set Voice Volume stores one dB contribution for the target HIRC element on the affected game object. `valueMode: "absolute"` replaces that contribution; `"relative"` adds to its interpolated current value. Reset Voice Volume returns the contribution to `0 dB`. Contributions from distinct matching Sound, container, and Actor-Mixer identities add through `matchIds` before the final Wwise gain clamp, alongside authored hierarchy, State, and RTPC volume. They never replace Play fades, Continuous crossfades, Stop fades, emitter gain, or spatial attenuation. Set Bus Volume stores a dB property state for the target Wwise bus on the affected emitter generation. Absolute mode replaces the current value; relative mode adds to its interpolated value; Reset returns it to `0 dB`. Builder-produced sounds retain their inherited NodeBase `overrideBusId`, its ordered dry Audio/Auxiliary Bus ancestry, and the summed authored base Bus Volume. The route also retains summed bus Make-Up Gain when it occurs in that dry ancestry, plus the Output Bus Volume on the NodeBase that supplies the effective output-bus override. A descendant value without its own override is not active. These contributions are applied after voice volume evaluation. A dedicated bus gain stage applies the authored gains plus live state only to voices whose route contains the target, without changing voice gain or the application master, SFX, or music controls. Reset All and All-Except use qualified exact stored bus identities; EVE currently exercises Element. A portable `sound` node may also carry `sourceEffects`, an ordered list of Parametric EQ, Wwise Delay, and qualified Wwise Compressor, Peak Limiter, Flanger, Tremolo, Guitar Distortion, Matrix Reverb, or RoomVerb records. The builder walks the Sound's NodeBase ancestry to the first effect override, treating a root list as effective and an explicit empty override as a replacement that clears the parent list. It emits the chain only when every active slot is a supported effect with an admitted shape; almost all admitted effects are control-free and static. The bounded dynamic exceptions are the EVE-v150 Parametric EQ `ParamID 2` frequency, Tremolo Depth/Frequency, and Guitar Distortion `ParamID 61` Drive Game Parameter shapes described below; dynamics additionally require linked channels and timing within the Web Audio adapter's bounds. Playback creates one Web Audio chain per physical voice before the authored Voice LPF/HPF, gain, spatial/auxiliary split, and Audio Bus effects. Parametric EQ and Delay are browser DSP adaptations, not native Wwise filter or Delay claims. Source Compressor and Peak Limiter are retained as authored base records but realized only under the opt-in `wwiseDynamics: "approximate-web-audio"` policy; strict mode or missing browser dynamics/lookahead primitives omits the complete source chain and keeps the voice audible and dry. Pinned wwiser proves the Peak Limiter layout. Compressor's v150 field order remains the same explicitly empirical interpretation used by the shared-bus adapter because pinned wwiser does not decode Compressor parameters. Web Audio's fixed lookahead, automatic makeup, detector/envelope law, ratio ceiling, and channel behavior therefore remain non-equivalent to Wwise. The builder admits one empirical live Parametric EQ combination observed in the pinned EVE 3453885 v150 corpus: plug-in `0x00690003`, `ParamID 2`, Game Parameter control, exclusive accumulation, and scaling 3. It maps the authored curve to Band 1 Frequency as `10 ** output`, preserving Wwise curve interpolation and object RTPC, global RTPC, then STMG default precedence. The current EVE records are all driven by `ship_Roll` from `0..360`, producing about `160..3650 Hz`. The runtime schedules the browser biquad over known RTPC transition boundaries and clamps frequency to the context Nyquist limit. This is an exact EVE-v150 record adaptation, not a claim that Audiokinetic publishes `ParamID 2` as a stable universal enum. The same records author `processLfe:false`; mono and stereo decoded sources have no separate LFE channel, so their topology is preserved. A source with more than two decoded channels keeps the complete effect chain audible and dry. Parametric EQ parameters 1 and 7 in this corpus are driven by Wwise Modulator objects, not Game Parameters. The largest exact case is effect `1738007123` on 27 Sounds across seven engine-on events: `ship_Roll` already controls Band 1 Frequency, while v150 LFO `841531431` drives Band 1 Gain from `+0.2` to `-24 dB` and has its own `ship_Roll` frequency curve from `0.2` to `100 Hz`. Runtime-resource currently retains that HIRC object as raw payload rather than a typed modulator record. It remains unsupported until the owning resource decoder exposes the LFO fields and runtime-audio owns its voice/event lifecycle; the library builder does not duplicate BNK parsing. A low-level custom voice descriptor that supplies `rtpcCurves` without its live source-effect RTPC reader also keeps the complete chain dry instead of applying a stale static base. Qualified Flanger records are also retained as authored base records. They use the independent `wwiseModulation: "approximate-web-audio"` opt-in; strict mode or missing Gain, Delay, or required Oscillator support omits the complete source chain and plays dry. Pinned wwiser proves the exact 59-byte v150 field order. The browser stage implements a sine-driven unified comb approximation: base delay plus `delay * depth` modulation, Blend/feedforward/feedback gains, Wet/Dry mix, and output gain. It starts one voice-owned oscillator at the first scheduled physical source start and stops it at final voice disposal. This is not Wwise-equivalent DSP: the browser stage processes every decoded channel even when Wwise authors Center/LFE bypass, clamps feedback to just below unity, retains LFO phase while paused, and cuts delay/feedback state at the decoded dry-source boundary. Shared-Bus Flanger remains unsupported. One dynamic EVE-v150 form is also retained. SoundbanksInfo names control `3712448215` `ship_Distance`; STMG binds it to built-in Distance (`1`) with a zero default and no ramp, and the Flanger uses additive scaling-0 parameter and property ID `1` for Wet/Dry Mix. The browser evaluates that curve at `worldDistance / scalingFactor`, clamps the combined mix to `0..100`, and automates independent dry and wet gains when the emitter, listener, or scaling factor moves. This is a single-listener browser adaptation. Wwise's native multi-listener reduction/update cadence, other parameter IDs, and control-type-4 Modulators remain unsupported. Qualified Tremolo records reuse the same `wwiseModulation` policy. Pinned wwiser identifies plug-in `0x00830003` and shows the corresponding modulation and phase sequence inside Flanger, but does not decode Tremolo's own parameters. The EVE corpus informs the 38-byte interpretation; it remains empirical. Admission is explicitly limited to bank version 150, a static Sine record, an unsmoothed 50%-duty Square record, or an unsmoothed Triangle record, bounded phase offset/mode/spread fields, and Center/LFE processing. Square and Triangle admission is normally limited to zero offset and zero all-channel spread. One exact OSSE Square preset is retained separately: 9% smoothing, 15% PWM, Circular phase mode, and 180-degree spread. Its two Sound leaves share mono media, so the per-channel phase policy is inert. Audiokinetic's [Wwise Tremolo reference](https://www.audiokinetic.com/en/library/2024.1.1_8691/?id=wwise_tremolo_plug_in_effect&source=Help) describes a unipolar carrier; the browser maps it to `gain(t) = 1 - depth/2 + (depth/2) * sin(2*pi*f*t + phase)`, then applies authored output gain. A nonzero global phase uses a custom `PeriodicWave`; missing that primitive keeps the complete chain dry. A qualified Square or Triangle record uses Web Audio's band-limited native oscillator of the same name. The OSSE preset instead uses a bounded custom 15%-duty Fourier pulse. The portable record also retains Wwise's phase mode and spread, but the browser uses one all-channel carrier and does not reproduce the authored per-channel Left-Right, Front-Rear, Circular, or Random distribution. Its voice-owned oscillator has the same start, pause, and disposal lifecycle as Flanger. Native oscillator/channel law is not claimed. Smoothing and PWM remain shape-validated and are normally neither stored nor applied; Square is admitted only where their authored values are exactly zero and 50 percent respectively, and Triangle only where smoothing is zero. The OSSE exception retains its exact 9%/15% values and applies the duty cycle, but its finite Fourier carrier does not reproduce Wwise's exponential smoothing filter. PWM applies only to Square in Wwise, so Triangle PWM is intentionally ignored after range validation. Two exact dynamic EVE-v150 forms are admitted. The common form has two object-scoped `booster_intensity` curves: additive scaling-0 `ParamID 1` controls Depth and exclusive scaling-3 `ParamID 2` controls Frequency. The builder requires their matching Depth property and the shared STMG Filtering Over Time policy with two-second up/down times. One `pirate_radio_guristas_play` Sound instead uses only the same Depth curve and keeps its authored `0.24 Hz` frequency. That exception is pinned to its complete preset: 72.6% base Depth, 35% smoothing, 50% PWM, 53-degree global phase, Random mode, and 55-degree spread. The browser retains the global phase but its one all-channel carrier omits Random per-channel spread, and the native 35% smoothing is shape-validated but not reproduced. The runtime applies a voice-local approximation of Wwise's documented exponential filtering to each raw control target, samples the admitted curves, and automates the oscillator plus both unipolar gain terms. A voice posted during an existing object's ramp starts from the current readable control rather than inheriting the older voice's filter history, and an independently authored Set Game Parameter transition is sampled as a new target rather than convolved with the plug-in filter. The target IDs are EVE-v150 corpus evidence, not universal Wwise enums. Other dynamic forms remain dry. Shared-Bus Tremolo remains unsupported. Wwise Harmonizer `0x008a0003` deliberately has no browser policy yet. Pinned wwiser proves its v150 parameter layout, but Web Audio has no native duration-preserving pitch shifter; using source `playbackRate` would change the voice duration and every downstream schedule. The reachable EVE sun records do not make this problem disappear: their one enabled voice is at zero cents, but Wwise mixes that processed voice with latency-aligned dry audio, so plain dry-through does not reproduce the authored signal. Ten yellow-sun Sounds use that static shape. Another 90 sun Sounds drive additive dB `ParamID 3` from an Envelope Modulator, and neither pinned wwiser nor the audited corpus proves whether that ID targets Voice 1 Gain or another gain-like property. A credible adapter therefore needs both custom voice-owned pitch-shift DSP and a proven plug-in parameter mapping plus modulator lifecycle. The exact reachable population is 101 Sounds across 12 events: 100 audible sun records remain complete-chain dry fallbacks, while the one `warp_ship_init_play` record has a `-96 dB` wet branch and is safe to omit as inaudible. Recording an effect object solely for that one transparent omission would not change current playback, so it remains unprojected rather than creating a misleading Harmonizer approximation. Qualified Guitar Distortion records use the independent `wwiseDistortion: "approximate-web-audio"` opt-in. Pinned wwiser proves the v150 126-byte layout: three pre-EQ bands, three post-EQ bands, distortion type, Drive, Tone, Rectification, output gain, and Wet/Dry mix. The EVE subset is control-free, uses Overdrive or Heavy, and is fully wet. Web Audio realizes enabled EQ bands in authored order around a 4x-oversampled `WaveShaperNode`, then applies output gain. The transfer curve is a deterministic normalized `tanh` approximation whose drive scale differs for Overdrive and Heavy; Rectification blends toward a full-wave curve. Authored Tone is retained but not applied because neither wwiser nor the browser API supplies Wwise's tone law. The exact EVE-v150 dynamic exception is plug-in `0x007e0003`, `ParamID 61`, an object Game Parameter, additive accumulation, and scaling 0. It maps to Drive, preserves Wwise interpolation/default precedence, and currently covers `ship_health_hull`, `ship_warp_direction`, and `booster_intensity`. ParamID 61 is an empirical corpus mapping, not a claimed universal Audiokinetic enum. The latter two controls author two-second Wwise Filtering Over Time ramps. The builder retains that STMG policy on 25 leaves and the voice-local RTPC lane filters their raw control before evaluating Drive. The 111 `ship_health_hull` leaves explicitly author no filtering and therefore remain immediate. Because a WaveShaper curve cannot be automated, the browser fixes the curve at maximum admitted Drive and schedules pre/post Gain nodes. For normalized WaveShaper inputs this reproduces the same CarbonEngineJS tanh/full-wave curve family used by static records; it does not reproduce Wwise's native Drive or clipping law. The admitted dynamic records have no enabled pre-EQ stages. Strict mode, missing WaveShaper/biquad/gain primitives, a missing live RTPC reader, other dynamic controls, other distortion types, other bank versions, non-fully-wet records, and shared-Bus Guitar Distortion keep the complete source chain audible and dry. This preserves topology and audible coloration, not Audiokinetic DSP parity. The [Wwise effects reference](https://www.audiokinetic.com/library/edge/?id=effects&source=Help) and [Web Audio WaveShaperNode](https://webaudio.github.io/web-audio-api/#waveshapernode) describe the separate authored and browser surfaces. Qualified static Matrix Reverb records use the independent `wwiseReverb: "approximate-web-audio"` opt-in. Pinned wwiser proves plug-in `0x00730003` and the exact 29-byte v150 default-delay layout: Reverb Time, HF Ratio, delay count, Dry and Wet levels, Pre-Delay, Process LFE, and delay mode. The builder admits only control-free, media-free records with a standard 4/8/12/16 delay count, Process LFE enabled, and the default-delay mode. Strict mode or missing Gain, Delay, or Biquad primitives omits the complete chain and keeps the voice audible and dry. The browser stage preserves authored dry/wet decibel levels and Pre-Delay, then uses a bounded four-line cyclic feedback-delay network. Its four spaced delays come from Wwise's default table, feedback approximates the authored Reverb Time as a nominal T60, and a fixed logarithmic low-pass mapping adapts HF Ratio. The authored delay count remains in portable metadata but does not allocate that many browser lines. Wwise's proprietary matrix, mixing, damping, channel, and LFE laws are not reproduced. Pause/seek reuse browser state, and natural completion disposes the stage at the decoded dry-source boundary, cutting the reverb tail. Shared-Bus Matrix Reverb remains unsupported. Qualified static RoomVerb records use the independent `wwiseRoomVerb: "approximate-web-audio"` opt-in. Pinned wwiser proves plug-in `0x00760003` and the exact 186-byte v150 layout. The builder retains every field but admits only the audited EVE tuning fingerprint, static/control-free records, and channel settings that can be approximated from mono or stereo decoded media. Strict mode, a source with more than two decoded channels, or missing Gain, Convolver, Buffer, Delay, or Biquad primitives omits the complete chain and keeps the voice audible and dry. The browser does not reuse the Matrix Reverb feedback network. It caches deterministic procedural impulse responses per AudioContext, sample rate, channel count, and authored parameter set. One branch approximates the authored early-reflection pattern, room size, and stereo width. A separate branch preserves Reverb Pre-Delay and approximates the late T60, HF damping, diffusion, density, room shape, and quality. Authored Dry/Early/Late decibel levels are preserved, and enabled tone filters are mapped to Web Audio biquads at their retained early/late insert positions. Audiokinetic's [Wwise RoomVerb reference](https://www.audiokinetic.com/library/2025.1.3_9037/?id=wwise_roomverb_effect_plug_in&source=Help) describes those controls, but the proprietary reflection tables, reverb-unit algorithm, early-reflection front/back timing, surround/LFE/center routing, and native channel law are not available in Web Audio. Pause/seek reuse browser convolution state and voice disposal cuts the remaining tail at decoded dry-source completion. Shared-Bus RoomVerb remains unsupported. Bypassed or rendered slots need no live stage. Pause and seek reuse the voice-owned browser nodes instead of freezing or reconstructing native Wwise plug-in state. Natural completion still follows the decoded dry source; `DelayNode` has no Wwise tail-completion callback, so residual feedback is cut when the voice is disposed. Mixed unsupported plug-in sequences, supported effects with RTPC, State, property-value, or media controls, unsupported independent channel routing outside the documented modulation approximations, and unsupported plug-ins retain the previous dry-playback approximation rather than applying part of an authored chain. EVE build 3453885 installs 3,225 qualified Sound leaves carrying 3,423 effect records: 473 use Parametric EQ, including 186 leaves with live `ship_Roll` Band 1 Frequency, 87 use Wwise Delay, 2,114 use Compressor, 73 use Peak Limiter, 49 use Flanger across 27 retained events, including 28 built-in-Distance leaves across 18 newly complete events; 190 Tremolo stages occur on 186 Sounds, including 21 live `booster_intensity` leaves; 205 use Guitar Distortion across 57 retained events, 50 use static Matrix Reverb across 22 retained events, 52 use static RoomVerb across 34 retained events, and 130 retain telemetry-only Meter records. The added 81 Meter leaves pair Meter `277510878` with Compressor `554802347` across 51 events. RoomVerb covers 11 effect identities and ten exact 186-byte parameter records; 28 chains contain only RoomVerb and 24 place one qualified Parametric EQ before it. The formerly dry eight Matrix leaves are now complete because their preceding Tremolo ShareSet's 108-degree global phase is retained; its Random 66-degree per-channel spread remains an explicit browser omission. Projected Matrix Reverb now covers all five effect identities. Guitar Distortion covers 21 effect identities: 69 static leaves use 18 identities and 12 raw presets (11 audible decoded parameter sets), while 136 live-Drive leaves across 35 events use three identities. Other dynamic Guitar controls remain dry. The static Tremolo population contains 115 isolated chains, 14 Tremolo-to-EQ, three EQ-to-Tremolo, eight Tremolo-to-Matrix, six Delay-to-Tremolo, two Tremolo-to-Delay, one double-Tremolo chain, and one four-Tremolo chain over 70 effect identities and 24 distinct 38-byte parameter records. The four-stage `jita_OSSE_bigscreen_1_play` chain contains three unsmoothed, zero-phase Triangle carriers and one Sine carrier. Those Peak Limiter leaves all inherit Custom effect `754157063` under `refinery_l_play`. A total of 486 retained events can reach at least one Compressor leaf. The Compressor population contains nine complete chain signatures, including eight leaves where it precedes one qualified EQ. There are now 191 non-neutral static EQ chains and 377 non-neutral EQ chains including the live-EQ records. Sound `350811697` adds the static EQ-to-Tremolo skyhook population chain. Its `processLfe:false` EQ is equivalent for decoded mono/stereo audio; a decoded multichannel source keeps the complete chain dry. Twelve additional static-Flanger leaves now survive with their supported live Parametric EQ in slot 1. The distance-driven Flanger slice completes ten idle Tremolo chains, sixteen engine-on EQ chains (ten also carrying live Tremolo), and two XXL warp blast leaves. Two M-afterburner leaves remain dry because companion EQ `1738007123` is still unsupported. Set Voice Pitch stores one cents contribution for the target HIRC element. `valueMode: "absolute"` replaces that contribution; `"relative"` adds to its interpolated current value. Reset Voice Pitch returns the contribution to zero cents. Contributions from distinct identities in the selected hierarchy add to NodeBase, State, and RTPC pitch before conversion to a Web Audio playback-rate ratio. A transition changes already-playing voices continuously without restarting their buffers, and finite authored repeat timing follows the changing playback rate. Set Voice LPF and HPF store independent signed percentage contributions for the target HIRC element. Absolute mode replaces the selected property's contribution, while relative mode adds to its interpolated current value. Reset Element returns one stored target contribution to zero. Reset All visits every stored target contribution for that property, and All-Except retains the exact stored target IDs listed by the action. Contributions from matching Sound, container, and Actor-Mixer identities add to authored, State, and RTPC filter values before the final `0..100` clamp. A Set action also provisions a neutral Web Audio filter for matching leaves that have no authored filter, so Set-before-Play and later action-only posts remain audible. The exact-target interpretation of Voice Filter All-Except is the strongest structural reading of the v150 action record and its object-owned property state, but public Wwise documentation does not specify descendant behavior. The inspected EVE actions are all element mode and carry no exceptions, so All/All-Except behavior is broader qualified coverage pending an Authoring Profiler experiment rather than an EVE parity claim. Game-object scope changes only the posting emitter. Global scope changes all currently registered emitters. The stored contribution affects voices that are already playing and voices created by later posts, and survives after the action-only posting ID completes. Unregistering an emitter ends that stored generation; already-playing voices retain the retired generation's stored contribution and finish any scheduled transition, while a newly registered generation starts with neutral Voice Volume and Voice Pitch contributions. Global Voice LPF/HPF actions additionally update a persistent global template: an emitter registered later inherits the template's original transition timeline and therefore joins its remaining fade or final value. A retired generation keeps its independent filter map and is not changed by later global filter actions. Bus Volume uses the same per-generation isolation, while global actions also update a persistent template for emitters registered later and continue to affect live voices on retired generations. SFX and music routes include full dry-output bus ancestry, authored base Bus Volume, and any bus Make-Up Gain on that ancestry. The effective NodeBase Output Bus Volume remains a separate authored contribution on the same collapsed route. Typed version-2 Bus RTPCs are stored once per bus and distinguish `voice-volume` from `bus-volume`. SFX adds Voice Volume on its own gain before the Bus Volume/effect stages; both properties follow the complete dry ancestry. Built-in music evaluates only Bus Volume and shared-mixer qualification rejects a music route that would require the unimplemented music Voice Volume stage. Scaling-2 curve outputs remain raw: runtime-audio interpolates first, then applies Wwise's nonlinear dB conversion. Version-1 catalogs remain implicit Bus Volume. Typed Audio Bus State tables use the same route ancestry. Their matching global State values atomically add Bus Volume in decibels, Pitch in cents, and signed LPF/HPF offsets across subscribed groups and buses, using the catalog's STMG default or directed transition duration. Filters sum before one final 0-100 clamp, so authored negative offsets remain meaningful. Pitch joins the transport-aware playback-rate path and final +/-2400-cent clamp. State transitions remain interruptible and new voices join the current blend. The catalog keeps both the authored bus synchronization type and the route-qualified effective type. Wwise applies a bus State immediately when only Actor-Mixer Hierarchy sounds use that bus, even if authoring selected a music synchronization point. The builder records that qualification as `effectiveSyncType: 0` and rejects a non-immediate subscription on a route used by music when the State changes a music-relevant property, because runtime-audio does not yet own a music-grid scheduler for bus States. Wwise Bus Pitch is excluded from music by definition. Typed Audio Bus auto-ducking uses the same ancestry and one shared SFX and music activity clock. A physical source activates every authored source bus in its route at its scheduled Web Audio start, including silent samples; pending media and authored graph silence do not activate it. The first active source begins each target's Fade Out, overlapping sources on the same bus hold one duck, and the last end starts Recovery followed by Fade In. Different source buses add in decibels, and nonlinear fades interpolate in linear gain with the authored Wwise curve. Voice Volume and Bus Volume targets remain distinct in the portable catalog. Collapsed dry routes may schedule both on their route gain; the qualified audible SFX Aux shape keeps Voice targets before fan-out and schedules Bus targets independently on the complete dry and wet legs. This remains a focused audible adaptation rather than full Wwise bus processing. Voice Volume is before the qualified dry/wet send split, while Bus Volume ducking is after each leg's additive State filters. Static Parametric EQ and static Wwise Delay have qualified shared-Bus adapters; general auxiliary sends, nonlinear effects, meters, and effect-tail-driven bus activity remain deferred as described in the [routing reference](../reference/wwise-resource-routing.md). Delay is measured from the action post. Value randomizers are signed offsets sampled once, and transitions use the decoded Wwise curve from the authored action time. Web Audio automation keeps those transitions continuous between `RenderAudio()` calls. The builder fails closed for otherwise untyped object targets. Bus actions remain portable even when the current asset set has no matching route, so future banks can use the same valid action without rebuilding policy. A Voice property element target absent from every loaded SFX NodeBase cannot match a projected voice, so that action is omitted as a no-op while the event's other authored actions remain usable. Set Game Parameter writes a named object or global RTPC. Absolute mode replaces the scope's current value; relative mode adds to its interpolated value at the authored action time. Reset writes the catalog's authored default instead of deleting the scope and falling through to another value. Delay and value randomizers are sampled once per post, while a nonzero transition persists after the posting ID completes. A later action rebases from the interpolated value, and an external `SetRTPCValue()` or `SetGlobalRTPCValue()` first applies overdue authored actions and then cancels the active transition for that same scope and parameter. Ordered capture uses an in-memory RTPC overlay, so an immediate parameter action affects only later Play actions in that program without exposing a partial update if program resolution fails. Already-playing voices receive Web Audio automation for gain, pitch, low-pass, and high-pass RTPC properties; pitch automation also participates in transport and finite-repeat timing. Object-scoped values and transitions remain attached to a retired emitter generation until its voices finish, while a replacement generation starts clean. The portable action model deliberately retains `bypassTransition` even though the currently qualified EVE parameter targets use no internal STMG ramping. Exact STMG inspection now supplies the builder's authoritative parameter defaults, including Reset values and unset RTPC-curve fallbacks. The reader also preserves all ramp policies and built-in bindings. Realizing those manager-level policies is separate runtime work, not a reason to discard the serialized flag or reject otherwise valid future banks. Random and sequence state is kept independently per game object by default. Set `scope: "global"` on either container to share its history or position across all game objects. `sequence` currently means Wwise-style step sequencing between posts. A continuous container that schedules several children during one post requires explicit timing data and is not inferred from a flat media table. ## Gain and RTPC curves Nodes and child edges may carry `gainDb` and `gainCurves`. Gains on every edge of the selected path add in decibels. A curve has: - `rtpc`: the authored parameter name; - `scope: "object"` or `"global"`; object scope falls back to the global value when no per-emitter value exists; and - optional `defaultValue`, used when neither the requested object nor global RTPC has a value; and - non-decreasing points using either decibel `{ x, gainDb }` values or normalized linear `{ x, gain }` values; and - optional Wwise `interpolation` values from 0 through 9 on the point that begins each segment. Duplicate X values preserve authored discontinuities. Values between points interpolate linearly and values outside the point range clamp to the nearest endpoint. Authored interpolation shapes are applied when present. Decibel values at or below -96 dB and linear gain zero become silence. Changing an RTPC updates gain on already playing SFX voices without restarting their buffers. ## State properties and transition timing Nodes may carry additive Wwise State deltas in `stateProperties`. Each entry names a global State Group and maps authored State names to relative volume, pitch, low-pass, or high-pass changes: ```js stateProperties: [ { group: "combat", cases: { danger: { gainDb: -6, pitchCents: 1200 } } } ] ``` Matching entries on every selected hierarchy level add to the static `gainDb` and `pitchCents` values. An unset group or a current state with no authored case contributes zero; the projected EVE tables therefore preserve Wwise's initial `None` behavior. State-name matching is case-insensitive. `SetState()` changes the logical State immediately, so later Play actions, Switch/State routing, Continuous sessions, and music reevaluation see the new value in authored order. The affected NodeBase property offsets follow the STMG State Group's default transition time or an exact directed custom override. Gain, playback rate, low-pass, and high-pass update on already playing voices without restarting their buffers. New voices join the current blend, and a second State change rebases from the in-progress property values. Live State gain remains independent of an authored Stop envelope, so changes continue to apply during an audible fade. The optional graph-level `stateTransitions` array retains each numeric State Group identity, its optional name, the complete known State ID/name catalog, default duration, and every directed custom route. From/to numeric IDs remain present when SoundBanksInfo cannot name an endpoint, so future tooling does not lose the authored rule. Named and numeric group/State setters share one canonical runtime value whenever the catalog supplies that alias. The optional library-level `busStates` catalog carries its own normalized `stateTransitions` subset. Installing a library merges that subset with the SFX graph transition table by canonical State Group identity and rejects conflicting definitions. This lets a library with music or bus data but no SFX graph still realize the authored transition timeline. Builder-produced version-2 cases unify Bus Volume, Pitch, LPF, and HPF values under one State weight. Definitions must match wwiser's v150 accumulation and dB flags, and filter-bearing builds must declare STMG additive behavior. An active unsupported Bus State property fails the bus catalog closed instead of silently projecting only part of an atomic State selection. Strict version-1 Bus Volume-only catalogs remain accepted when supplied by a caller. The optional library-level `busDucking` catalog stores each source Audio Bus once with its Recovery Time, Maximum Ducking Volume, and ordered target rules: ```js busDucking: { schemaVersion: 1, sources: { "100": { recoveryMs: 1000, maxDuckVolumeDb: -96, targets: [{ targetBusId: "200", volumeDb: -6, fadeOutMs: 250, fadeInMs: 500, curve: 4, targetProperty: "voice-volume" }] } } } ``` The builder accepts only v150 Audio Bus sources and Audio Bus targets, exact target properties 0 (Voice Volume) and 4 (Bus Volume), supported curves 0-9, finite nonpositive dB values, nonnegative integer timings, and cycle-safe non-parent targets. Auxiliary Bus sources and malformed or ambiguous rules fail closed. The optional library-level `busEffects` catalog stores routed static Wwise Parametric EQ slots once per bus: ```js busEffects: { schemaVersion: 1, buses: { "500": [ { effectId: "900", slotIndex: 1, type: "parametric-eq", bands: [ { index: 1, filterType: "peaking", gainDb: -13, frequencyHz: 120, q: 5, } ], outputGainDb: 0, processLfe: true, } ], }, } ``` The v150 adapter follows wwiser's 56-byte Parametric EQ layout, retains slot and band order, and realizes enabled bands with Web Audio biquads. Routed Bus EQs with RTPC, State, property, or media controls fail closed. Bus `processLfe` must be true until the browser runtime owns an independent LFE branch. Source-local EQ additionally admits the exact Game Parameter form described above and permits `processLfe:false` only when the decoded voice is mono or stereo. Authored-neutral EQs remain in the catalog but do not allocate browser nodes. A distributed EQ is emitted only when the complete routed ancestry contains no active unsupported effect slot. Bypassed slots do not block it; an active Compressor, Peak Limiter, reverb, Meter, or unknown plug-in suppresses every distributed effect on that route so authored ordering cannot be partially realized. When the strict `busGraph` mixer qualifies the complete route, it decodes the authoritative graph effect records and creates one ordered effect chain per physical Bus, after SFX spatialization and music route envelopes. Common ancestors and their effect nodes are shared across SFX/music category entries and voices. Parametric EQ uses the source-proven 56-byte layout above. Static Wwise Delay uses wwiser's 18-byte v150 layout: seconds, Feedback percent, Wet/Dry percent, Output Level dB, Enable Feedback, and Process LFE. Its Web Audio adapter owns one parallel dry/delayed split, an optional feedback loop, and one output gain; `processLfe:false` remains blocked. The exact fields, ordering, and shared lifetime are preserved, but neither `BiquadFilterNode` nor `DelayNode` is claimed to be bit-equivalent to Wwise DSP. A blocked or missing graph retains the distributed per-source Parametric EQ fallback. Dynamic, mixed, media-backed, or otherwise unsupported effect sequences remain barriers rather than being partially realized. One bounded ordering proof admits an opted-in Peak-Limiter route when every source-/route-local control belongs to a strict descendant of the first audible effect Bus. Those controls already run before the shared topology, so an ancestor limiter retains its authored order. A control on the same Bus or an ancestor, an incoming duck target, or any other audible effect kind keeps the whole route on fallback. EVE build 3453885 has four such routes: 36 SFX leaves and 24 media across six event graphs. Only `OSSE_Gallente_steamalter_play` becomes fully shared-routed; the other five gain qualified branches while retaining separate blocked branches. The default `wwiseDynamics: "strict"` policy keeps Wwise Compressor and Peak Limiter out of shared routing. The source still plays through the existing SFX destination, but the complete authored shared-bus path is omitted; qualified coverage measures faithful bus admission rather than basic media audibility. `wwiseDynamics: "approximate-web-audio"` admits only static, control-free, channel-linked, Process-LFE records. It realizes each authored slot with a hard- knee `DynamicsCompressorNode`, caps ratio at `20:1`, cancels Web Audio's mandatory automatic makeup with a post gain, then applies authored output gain. Dynamics attack/release must be within the browser's one-second limit and Compressor attack zero remains a barrier. Peak Limiter uses zero attack and may add output delay above Web Audio's fixed 6 ms latency. That delay does not extend the detector lookahead. Detector, envelope, peak limiting, channel/LFE behavior, ratio, and timing remain approximations; the independent master safety compressor still follows the whole mix. The Peak Limiter layout is source-proven, while the v150 Compressor field order is empirically corroborated for the audited corpus rather than proven by the pinned wwiser parser. The builder and shared mixer also decode wwiser's exact 28-byte v150 Wwise Meter layout. Audiokinetic documents that the plug-in measures without modifying its signal, so they omit a static, media-free, control-free Meter on a shared Bus or in a complete source-local effect chain when no Game Parameter is written. Such a Meter is audio-transparent, though its monitoring behavior remains unsupported. The default `wwiseMeterFeedback: "strict"` keeps a Meter with a Game Parameter target as a shared-route barrier or leaves its complete source chain audible and dry. Explicit `"omit-telemetry"` admits that Meter when its parameters are otherwise qualified: audio crosses the slot, audible source siblings retain authored order, but no Meter value is produced and any authored feedback through the target Game Parameter is absent. `Apply Downstream Volume` changes whether higher-level gains contribute to that omitted measurement; it does not alter signal flow. In EVE the omitted `sovhub_upgrades_meter` value drives a cross-bank structure Voice Volume RTPC by up to about `-0.94 dB`, so this remains an explicit audible approximation rather than an exact omission. The builder separately identifies the v150 Audio Bus `MaxNumInstances` RTPC as the `"voice-limits"` processing reason. `wwiseVoiceLimits: "strict"` keeps that route blocked. Explicit `"ignore"` admits it when every other reason qualifies, without applying the dynamic voice count or Wwise eviction policy. Libraries built before this reason was introduced retain their generic `"unsupported-rtpc"` barrier and must be rebuilt to use the opt-in. Static Audio Bus maximum-instance, stealing, and virtual-voice policy is also not enforced; this option changes only shared-route admission for the separately classified dynamic RTPC barrier. One independent v150 Sound policy is enforced before media acquisition: ```js voiceLimit: { counterId: "602217068", scope: "game-object", maxInstances: 1, behavior: "reject-newest" } ``` The builder emits this only for a Sound whose packed Advanced Settings are the EVE `0x09` form (cap one, kill-newest tie break, local scope, ignore parent limit, and no over-limit virtualization), whose effective inherited below-threshold behavior is Continue, and whose effective priority path has no randomizer, RTPC, or State mutation. The complete output-bus ancestry must also have no static cap or `MaxNumInstances` RTPC. The per-object scope bit is corroborated by wwiser's older explicit `bIsGlobalLimit` field and the paired v150 `0x09`/`0x0d` corpus forms; a controlled Wwise golden pair is still the remaining format-level proof. Reservations count still-loading selections after their Play boundary. A duplicate Sound on the same game object is rejected in deterministic selection order without loading its media; another game object remains independent. Cancellation, failed acquisition, Stop, and physical completion release the reservation. Continuous completion releases before selecting its successor; Trigger Rate keeps its authored cadence while a duplicate choice is rejected, Continuous Switch replaces an obsolete pending reservation before rerouting the same Sound and stays dormant until its next control change. Future admission is deliberately excluded: the builder omits `voiceLimit` from Sounds reachable through a positive Play/Initial Delay, a positive Continuous Delay, or Continuous Crossfade prefetch. The backend also rejects a custom delayed limit, and fails a custom capped Crossfade traversal closed, instead of reserving either too early. Wwise evaluates the limit at the future playback boundary; prefetching media is not itself an instance. Exact authored-time admission for those shapes remains unsupported. This is not a general Wwise voice arbiter. Global Sound limits, caps above one, priority-based stealing, project-wide maximum voices and volume threshold, parent/container and Audio Bus limits, dynamic limit RTPCs, virtual queues, below-threshold Kill/Virtual behavior, and virtual re-entry remain unsupported. `buildFromBanks()` also emits a version-1 `busGraph` topology for every routed SFX Sound and music track. Route records deduplicate dry ancestry plus effective NodeBase user/reflections sends. Bus records retain their parent, exact v150 channel configuration and properties, authored auxiliary edges, processing reasons, and every ordered effect slot including bypassed and unsupported plug-ins. Effect records retain plug-in identity, parameter bytes as canonical base64, control counts, and embedded plug-in media source IDs. The validator rejects missing targets/effects/media, malformed ancestry, dry-parent cycles, unreachable records, and out-of-range route references. Each user/reflections send also carries a `dynamic` barrier flag when its active slot has an authored randomizer, RTPC, or State property. The topology keeps the static base value and refuses to imply that the dynamic contribution is already realizable. A static user send may be omitted only when the complete Auxiliary Bus return is proven to stay at or below Wwise's `-96 dB` silence threshold. The proof includes authored gains, maximum installed RTPC and State contributions, absolute or positive-relative Bus Volume action risk, inactive effects, and the absence of another wet-path escape. This admits an inaudible route without claiming that its auxiliary signal was rendered. Aux inheritance follows wwiser's root rule: a root NodeBase's authored list is effective even when its override bit is clear, because it has no parent. A non-root node with a clear override bit inherits, while the first overriding descendant replaces the inherited list. The catalog is the shared-runtime foundation, not an audible approximation. Current playback can share a complete static Parametric EQ/Delay sequence with Bus Volume RTPC and Immediate State gain. Each physical Bus owns one post-effect fader that combines its static Bus Volume with only that Bus's matching global RTPC and State gain; the route aggregate is admitted only when it exactly equals the sum of those physical Bus values. Voice Volume remains before the Bus, while Make-Up Gain, effective NodeBase Output Bus Volume, and live Set/Reset Bus Volume actions remain on the route-local stage. State Pitch remains source-local. State LPF/HPF retain the additive whole-ancestry contract: ordinary and blocked routes use their source-/track-local filter pair, while the exact audible SFX Aux shape uses one independently evaluated filter pair for each complete dry and wet path before the corresponding Bus-target duck gain. `busVolumeActionControlled` likewise marks every Bus targeted by a retained Set or Reset Bus Volume action. An audible effect combined with Voice Volume, filtering or pitch, ducking, or action control remains blocked so per-voice state and an effect tail cannot silently cross the shared Bus fader. The auxiliary exceptions are the proven static-silence omission above and one exact SFX-only user-send shape. The audible shape is one static finite send with neutral send filters whose Auxiliary Bus return rejoins the dry ancestry. Its route entry fans out after spatialization into the dry Bus and a dB-scaled wet gain, then both branches reuse the same physical common ancestors. Bus-target ducking is evaluated after the filters over each whole route leg, preserving a source's collective maximum-duck floor. A source whose rules span Voice and Bus targets on the split path, a wet-only duck source or Voice target, branch-only Pitch, Bus Volume action targets, Make-Up/Output Bus Volume, effects other than the feedback-free Meter omission, dynamic or multiple sends, Bus cascades, and reflections fail closed before allocating nodes. Music and every other incomplete auxiliary/effect path remain blocked. Installation cross-checks every playable routed SFX Sound and music track against its `busGraph` route, including dry ancestry and all three authored gain fields. At browser-gesture enable time the audio system creates one generation-scoped route controller and shares its stable route handles between the SFX backend and built-in music engine. Sound selections retain their exact leaf ID and music scheduling retains its exact track ID, so later realization does not have to infer topology from a coincidentally equal dry path. The SFX backend uses those handles to allocate one lazy branch per exact route and spatial mode within an emitter generation. Voices on the same route and mode share that branch; different routes never merge before it, and a 2D voice never enters a 3D route panner. Placement, scaling, RTPC replay, retirement, and disposal remain generation-scoped. Branch outputs still feed the existing SFX destination unless the selected shared-mixer policy qualifies their complete dry path. Qualified branches feed the stable SFX category input after spatialization; blocked branches retain the existing destination. The qualified static SFX Aux shape is the only audible wet-path exception; it does not imply shared nonlinear effects or general auxiliary routing. A custom `applyRTPC` adapter continues to receive the legacy emitter target and also receives one update per graph-backed route branch. Branch updates include `busGraphRoute`; exactly one of `gain` or `flatGain` is populated according to the route's spatial mode, and spatial branches include their exact `panner`. The audio-system generation also owns a strict shared-Bus mixer. Its qualification accepts only default-channel ancestry with no active Bus positioning or HDR. Authored positioning/HDR override flags are allowed only when their decoded values prove both features inactive. Processing may include a complete source-proven static Parametric EQ/Delay/Meter sequence with static Bus Volume, Bus Volume RTPC, and Immediate State gain at exact per-Bus post-effect faders. The matching RTPC/State catalog entry and live ducking source must be installed for each declared control. An incoming duck target also counts as a route control, even though Wwise declares that rule on its separate source Bus. Qualified routes receive stable SFX and music category entries and share one node per common Bus ancestor. Provably silenced static user sends allocate no wet nodes. The exact audible SFX shape above allocates one send gain and one shared Auxiliary return; all other auxiliary routes remain blocked. SFX and music category entries remain separate so application volume controls cannot merge unrelated routes prematurely. State LPF/HPF add across the complete ancestry before one final clamp; the exact Aux shape preserves that rule separately on its dry and wet legs. Audible effects still reject Voice Volume RTPCs, State pitch, ducking, and Bus Volume action targets because those controls remain route- or voice-local. A blocked route returns no mixer input and allocates no partial graph. Qualified SFX route branches now consume these entries, including per-branch analyser stages that preserve aggregate emitter metering without merging route identity. Qualified music tracks use one transition lane per scheduled route and one Play-instance lane per exact route before the stable music category input. Segment crossfades and instance stop fades therefore remain route-local. Blocked music routes retain the legacy segment, instance, and music-output path. The bank builder projects only named, Immediate Volume, Pitch, and filter state tables with their exact supported accumulation modes. A group containing another property or accumulation mode is omitted whole, rather than partially emulated, while the event and its other supported groups remain playable. Active non-Immediate or unnamed groups still fail closed because their authored selection behavior cannot be represented safely. ## Builder input The optional builder accepts the graph directly as `sfx`, or as `enrichment.sfx` alongside neutral event/culling metadata: ```js const library = CjsAudioLibraryBuilder.build({ indexEntries, soundbanksInfo, enrichment: { Events: eventMetadata, sfx: authoredSfxProgram } }); ``` When caller-provided bank access is available, `buildFromBanks()` can ask the builder to project the conservative typed HIRC subset owned by `runtime-resource`: ```js const library = await CjsAudioLibraryBuilder.buildFromBanks({ indexEntries, soundbanksInfo, includeSfx: true, language: "en-us", loadBank, onSfxDiagnostics(diagnostics) { console.info(diagnostics); } }); ``` For an end-to-end browser or tool path, `buildFromResources()` reads the cFSD metadata, optional SoundbanksInfo, and banks through fetch or one injected source and returns a hydrated library: ```js const library = await CjsAudioLibraryBuilder.buildFromResources({ indexEntries, source: { read: (path, context) => readBytes(path, context) }, includeSfx: true, language: "en-us" }); const values = library.GetValues(); ``` `language` selects one localized bank variant before event-media and SFX HIRC objects are merged. This is required because localized banks reuse object IDs while pointing at different media. For each completely lowered event, the builder also resolves version-150 NodeBase positioning through Sound and Actor-Mixer parents. It produces a sparse `is2D` metadata patch: a resolved positioning owner without an authored attenuation assignment is 2D; an event is 3D when any selected playable leaf resolves an attenuation assignment. Actor-Mixers remain inheritance-only and are never emitted as playable parallel nodes. The same ancestry walk projects authored Stop actions into each playable event's `eventsStoppedBy` metadata. A Stop action may target a Sound or one of its parent Actor-Mixers; posting that stopping event then stops every matching runtime event. Unmatched Stop targets are retained in `diagnostics.stopRelationships.unresolved` instead of being guessed. Known positioning is also retained as `spatial` on each sound leaf. This is more precise than the event-wide metadata for mixed parallel/blend events: one 3D leaf may use the emitter panner while its 2D sibling uses the flat SFX route. A missing leaf value falls back to event `is2D`. Caller event metadata still overrides the derived event-wide fallback; to override a known leaf, provide or enrich the `sfx` graph itself. Each resolved v150 attenuation also retains its scaling-type-2 dry-volume curve as `{ scaling, points: [{ x, value, interpolation }] }` on that Sound leaf. Distances stay in authored world units and values stay in Wwise's raw decibel-curve representation so the runtime can perform interpolation before scaling. The curve endpoint also supplies the conservative event culling radius. Parallel leaves may therefore share an emitter and still attenuate differently. Missing parents, parent cycles, unsupported non-renderable NodeBase RTPC controls or properties, and incomplete positioning data do not block otherwise valid sound playback. Missing spatial data omits only the derived spatial patch and adds diagnostics; non-renderable RTPC modifiers are ignored while supported Volume, Pitch, filter, and InitialDelay curves remain exact. Missing, unsupported, or Wwise `Use Project` dry-volume curves omit the leaf curve and use the documented browser inverse-gain fallback. Mixed spatial events use the largest complete leaf radius. Automatic bank projection is applied after SoundbanksInfo metadata and before caller `metadata` and `enrichment`, so explicit caller data remains authoritative. Automatic construction currently accepts Wwise generator-version-150 codec sounds, the qualified static Wwise Silence source shape, Play, Stop, Pause, Resume, Set/Reset Voice Volume, Set/Reset Bus Volume, Set/Reset Voice Pitch, Set/Reset Voice LPF/HPF, Play-Event, SetSwitch and SetState actions, including their qualified fixed delays, Random/Sequence containers without reverse restart, and named Step Switch/State containers. Play actions retain their authored delay, delay randomizer, probability, fade-in duration, fade-in randomizer, and curve. Play-Event recursively inlines the referenced event's playable program and merges its immediate setter, playback-control, Voice Volume, Bus Volume, Voice Pitch, and Voice LPF/HPF actions; its delay, delay randomizer, and probability wrap only the inlined playable roots. Music-prefixed events retain Bus Volume actions while their music Play targets stay owned by the music graph. A scheduled or gated Play-Event that reaches any non-play program is omitted rather than executing that action early. Missing targets and cycles are diagnosed and omitted. Successfully lowered nodes also retain inherited NodeBase Volume, Pitch, and InitialDelay properties. Their independent authored random ranges are sampled when that node is selected. An ordinary node is selected once per post, while a Continuous child is selected again for each child batch. Volume accumulates in decibels, Pitch becomes a Web Audio playback-rate ratio, and InitialDelay is added to the Play action delay. Hierarchy-only Actor-Mixer values are folded into the nearest playable node without turning the mixer into a playable container. A qualified Wwise Silence Sound follows its source ID to a v150 `CAkFxCustom` record with plug-in ID `0x00650002`, exactly 12 parameter bytes, a positive fixed duration, and no random length, media, RTPC, State, or property controls. It lowers to `timed-silence`; the browser loops one cached one-frame zero buffer and schedules its physical stop at the logical authored end. This keeps even EVE's 304.5-second pillar interval constant-memory while preserving pause, seek, pitch/rate changes, Stop, Continuous advancement, and completion. The empty inline source block is never treated as a default one-second Silence when the referenced effect record exists. Randomized or dynamic Silence remains fail-closed. EVE build 3453885 uses eleven such leaves to restore the otherwise-audible `worldobject_pillars_active_play`, five `jumpgate_suppressed_lvl_*` events, `solar_array_outburst_play`, `solar_array_impact_play`, and `solar_array_beam_play`. `in_game_video_stream_play` instead uses Wwise Audio Input (`0x00C80002`): it has no bank media or authored duration and remains a host-input barrier rather than receiving fabricated audio. Layer/Blend containers with no Layer records, or only Layer records with no child associations, lower to parallel playback. A Continuous validation flag then has no child-admission region to evaluate. This covers the five zero-record Hangar Layers and 22 association-free ship-engine Layers in EVE build 3453885: each owns parallel child lifetimes while its children retain their independent Continuous Random/Delay schedulers. The affected Jita, Caldari, and Minmatar Hangar Play branches target those layers directly; their structural State ancestors remain available as Stop-match identities. An associated Continuous Layer may lower to a pre-started browser blend only when every direct child has an explicit non-empty region and is proven infinite. All of those child sessions begin with the Play action and remain alive until an authored Stop; the browser then applies the authored normalized-gain and supported property RTPC curves live. This preserves a dormant, RTPC-responsive container and prevents muted content from exhausting, but it is deliberately approximate: Wwise starts and stops children at each crossfade-region boundary, while runtime-audio keeps them running at zero gain outside that region. Loop phase, Continuous Random timing, voice count, and acquisition cost can therefore differ. A finite direct child keeps the complete event fail-closed. Non-continuous Layer crossfade tracks lower to live normalized-gain curves when their controller is a named Game Parameter. Supported Layer property RTPCs lower to live Volume, Pitch, low-pass, and high-pass curves on each affected child. Initial Delay is rejected for the pre-started Continuous form because later parameter changes cannot recreate Wwise's boundary-time delay evaluation. Transition and reset-after-stop policies authored on a Step Random/Sequence container are Continuous-only and therefore do not alter its one-child-per-post behavior. Continuous Random/Sequence containers support Disabled, Delay, Trigger Rate, Crossfade Amplitude, and Crossfade Power transitions. Disabled and Delay advance after the complete selected child batch ends; Delay then applies its independently sampled authored duration. Trigger Rate advances from the Web Audio clock, samples one duration for each selected child that has a successor, adds the selected child's Initial Delay, and permits earlier voices to overlap. This per-boundary sampling is runtime-audio's deterministic Wwise-compatible policy. Media acquisition does not serialize the cadence. If `RenderAudio()` arrives after a boundary, the runtime issues that boundary once and rebases the next one from current audio time instead of replaying a burst of missed triggers. Silent selected branches still consume their interval; a selected media leaf that cannot be acquired ends that traversal fail-closed. An infinite Disabled-transition Continuous Random whose every playable direct child is provably infinite cannot reach a second outer selection. The builder therefore retains its one object-scoped Random choice without an outer Continuous clock and lets the selected infinite child own playback. Qualification accepts only direct looping Sounds, already-qualified infinite Continuous Random/Sequence children, or a qualified pre-started Continuous Layer and does not infer duration from media. This is the same trapped-child reduction used by wwiser and restores the first Play branch of EVE build 3453885's `worldobject_station_amarr_play`; the outer container identity stays in every selected leaf's Stop ancestry. Crossfade prepares the next single-voice child before its boundary and fails closed if transactional preparation or media acquisition is unavailable. The independently sampled authored duration is clamped to half the outgoing source duration and its remaining play time. Amplitude uses opposing linear gain ramps; Power uses equal-power sine/cosine curves. A speculative Random or Sequence choice commits only when its successor becomes audible, so cancelled or failed preparation does not consume that choice. Continuous Switch/State containers are supported when `1st only` and `Continue to play` are disabled and a Switch/State change stops the outgoing object. The runtime follows each live game-sync decision, applies the authored per-child Fade Out and Fade In times independently, and keeps a dormant session through authored silence, natural completion, or temporarily missing media so a later value can resume it. Nested Continuous Switch/State containers are supported, including values that select the same final child; the complete decision path determines whether playback restarts. Routes that reach a non-switch Continuous container are rejected. This gate still applies to a live Switch/State root whose branch reaches a zero-record Continuous Layer containing Continuous Random/Sequence children; EVE's Hangar Play actions target the qualifying branch nodes directly instead. Continuous playback is object-scoped. Finite pass counts complete after every overlapping tail ends, without waiting through a nonexistent final interval. Break cancels pending and future selections while allowing active children to finish their current iteration. Seek ignores only voices below a Trigger Rate container, so unrelated siblings in the same event remain seekable. Active Trigger Rate children receive runtime-audio's normal per-voice Stop fade; this is a browser adaptation of Wwise's restricted Trigger Rate Stop propagation. Duration RTPC modulation is not projected. Graphs whose minimum randomized Trigger Rate interval is below 21 ms are rejected to avoid unbounded browser voice production. Two bounded nested clocks are supported. An infinite one-child Continuous Sequence with a Delay may wrap a reset-on-play, one-pass Continuous Sequence with a Trigger Rate. The parent-to-child edge must be plain, and the inner scheduler container must carry no playback modifiers beyond its scope, children, and Continuous settings. The inner voices keep their authored overlapping cadence; after the final selection, the runtime waits for every overlap tail and pending load, then samples the outer Delay and begins a fresh inner pass. The outer Initial Delay applies only to the first pass. Stop and Break retain the normal Trigger Rate contracts, and the outer container ID remains the Stop target. This exact shape restores EVE build 3453885's `upwell_hangar_armor_warning_play` and `upwell_hangar_hull_warning_play`. The second form is the narrow Jita incidentals topology: an infinite one-child Continuous Random with a randomized Delay around a two-child, one-pass amplitude Crossfade Sequence. Only static inner playback terms qualify. The runtime applies the inner Initial Delay, crosses from the first child to the second using the sampled authored duration, waits for both decoded dry voice tails and pending loads, then samples the outer Delay and reapplies the inner Initial Delay for the next pass. The authored playlist cursor is retained because Reset Playlist at Each Play is disabled. This makes EVE build 3453885's `jita_sfx_incidentals_level3_play` playable, but it is deliberately not DSP- or tail-exact. Random step `211583824` inherits static Wwise Delay ShareSet `2464647643`; its nine Sound children now receive the authored 280 ms delay, 32.5-percent feedback, and 30.5-percent wet mix through the browser Delay adaptation. The completion boundary still follows decoded dry voices and therefore cuts the residual feedback tail. General nested non-Switch Continuous clocks remain fail-closed. Trigger Rate Pause does not freeze the cadence or carry a pause depth into future child keys; the qualified Upwell consumers use Play and outer-container Stop actions, so that broader behavior is not claimed as Wwise parity. The builder normally omits an entire playable event when that event mixes other unsupported actions or reaches an unsupported playable node; the optional diagnostics callback explains each omission. Its `approximatedEvents` list separately records the bounded partial-recovery exception below. Sample Accurate Continuous transitions, `1st only`, `Continue to play`, Play-to-End switch changes, Play-and-Continue, playable Actor-Mixer approximation, associated Continuous Layers with a finite direct child, Layer property RTPC semantics outside the supported Volume, Pitch, low-pass, and high-pass set, and other unqualified HIRC semantics are never silently approximated. The two EVE 3453885 XXL microwarpdrive `on` events use the qualified pre-started form. Each outer thrust blend has three looping Sounds and one infinite Trigger Rate Random child; its nested speed/direction blends also reach only infinite children. Their live gain, Volume, Pitch, LPF, HPF, delayed Voice Volume actions, and matching three-second `off` Stop fade are preserved. Their inherited container effect chains remain unsupported: the bank assigns dynamic Parametric EQ `1730584540`, Flanger `2328072489`, and Tremolo `1286274856` ShareSets, while runtime-audio only projects complete static Parametric EQ and Wwise Delay overrides. The browser result is consequently drier and less modulated as well as pre-started. EVE build 3453885's `ship_module_shield_drain_play` mixes two Play actions. Action `960667829` reaches infinite Random `974515202`, whose 7000 ms amplitude Crossfade has trackless Layer/Blend `211616663` as its only child. Wwise [documents that Xfade fails when a Blend or Switch child plays](https://www.audiokinetic.com/en/public-library/2025.1.4_9062/?id=random_sequence_container_property_editor&source=Help). Runtime-audio therefore omits that invalid action instead of inventing a multi-voice crossfade. Its finite media `927964773` and `69501700`, infinite replay, and Crossfade envelope do not run. The same event's independent action `673588669` targets finite Sound `603165888`, so the builder retains that action as a bounded audible fallback. Media `278513022` plays once with its authored `-6 dB` and spatial properties. The Sound's direct Wwise RoomVerb `402798902` is retained. It remains audible and dry under the default strict policy; explicit `wwiseRoomVerb: "approximate-web-audio"` adds the documented RoomVerb character while still cutting its residual tail at source completion. This exception requires exactly two plain Play actions, the audited single-child infinite amplitude-Crossfade and trackless finite Layer shape, and one independent finite codec Sound. An infinite or modified sibling and all other Crossfade-to-Blend forms remain fail-closed. For events that do lower, the builder walks every possible typed graph branch and emits the exact set of reachable media IDs into `eventMedia`. This keeps the flat catalog useful to tools without allowing a value that merely resembled a WEM ID inside an undecoded payload to become audible. The caller may instead obtain a complete built library from an API and skip the builder. Runtime-audio performs no SFX metadata download or discovery. ## Related documentation - [Architecture and boundaries](../architecture.md) - [Browser playback guide](browser-playback.md) - [API reference](../reference/api.md)