audio-production

name	audio-production
description	Shared audio production protocol for the Helios sub-crew — foley, score, and dialogue. Spectral allocation, sync, stems, loudness, and mix manifests. Owned by Helios.
user-invocable	false
argument-hint	<foley-plan|music-plan|dialogue-plan|mix-integration>
allowed-tools	["Read","Write"]

Audio Production

The common working document for Helios's three audio sub-agents — audio-foley, music-score, and dialogue-mix. Each agent owns a discipline, but all three share one timeline, one spectral budget, and one delivery contract. This skill keeps their outputs from colliding and gives every cue an unambiguous home in the mix.

Purpose

Produce discipline-specific audio plans (AssetJob envelopes, cue sheets, stem plans, mix manifests) that occupy the timeline and the frequency spectrum without conflict, stay locked to picture, and meet platform loudness specs. Dialogue intelligibility is the first priority; everything else yields to it.

When to use

• Helios receives a timecoded ShotList and fans render work out to the audio sub-crew.
• /creative-campaign or /photo-direction reaches the production phase and a sequence needs sound.
• A delivered audio asset fails QC (loudness, sync, intelligibility, or stem completeness) and needs a refreshed plan.

Discipline split

Agent	Owns	model_type	Primary source	Never does
audio-foley	SFX, Foley, ambience beds, diegetic sound	sfx, tts (utility VO only)	timecoded ShotList	compose score
music-score	composition, stems, BPM map, narrative function	music	story beats (Calliope) + pacing (Helios)	Foley / dialogue cleanup
dialogue-mix	VO/ADR/TTS cleanup, intelligibility, channel routing	n/a (processes existing)	prior AssetJob audio sources	invent translations / rewrite lines

Inputs

• Active CreativeBrief (objective, channels, language set, brand_constraints)
• Timecoded ShotList (shot boundaries, action beats, transitions) — foley & music
• Story-beat list from Calliope — music only
• Prior AssetJob audio sources (raw VO, ADR, TTS) — dialogue only
• Optional cue/motif references from eights.memory.recall(tags=["audio-library"])

Outputs

• AssetJob envelopes (parent_crew: helios) written via rlm.output.write to RLM/output/{phase}/{topic}-{date}.md
• Cue sheets (foley), stem plans (music), and mix manifests (dialogue) covering the full timeline
• DecisionRecord-compatible notes: why each cue exists, governance flags separated from technical issues

Spectral allocation rules

The mix is a shared budget. Claim only the space your discipline needs and leave room for the rest.

Layer	Position	Level intent	Rule
Dialogue	centre (anchored)	reference, protected	Highest priority; nothing may mask it. Off-screen/phone/PA voices get explicit routing intent.
Score	L/R + LCR for 5.1	~2–3 dB under dialogue	Supports edit rhythm; ducks under dialogue; leaves harmonic space, never fights the voice.
SFX / Foley	positioned per pan	sync-locked transients	Tied to visible/implied motion; leaves gaps for dialogue peaks; no constant filler.
Ambience	bed, underneath all	continuous or intentional silence	Foundation only; silence is valid when it is deliberate.

• Reserve spectral space in priority order: dialogue first, then score, then SFX/ambience.
• music-score MUST keep cues off dialogue-heavy passages (no overscoring exposition).
• audio-foley MUST NOT produce score-like emotional swells; music-score MUST NOT fake score with SFX beds.

Sync & timing integration

• Every cue carries start_tc, end_tc, and duration_seconds; foley cues add a sync_anchor.
• The music bpm_map MUST align to the ShotList cut cadence and transition density; prefer tempo stability within a sequence unless a beat shift demands modulation.
• If Helios issues revised cut timing, music-score reflows the BPM map before re-emitting.
• Detected sync offsets are corrected deterministically when possible; otherwise escalate to Helios with the offset noted.
• dialogue-mix tracks alternate language versions independently — never assume timing parity across languages.

Loudness & delivery standards

Conform to ITU-R BS.1770-4 integrated loudness measurement. Targets unless the CreativeBrief overrides per platform:

Target	Integrated LUFS	True-peak ceiling
Streaming video (SDR) / YouTube	−14 LUFS	−1 dBTP
Streaming / podcast	−16 LUFS	−1 dBTP
Broadcast (EBU R128)	−23 LUFS	−1 dBTP

• Deliver both stereo and 5.1 routing plans: dialogue centre; score L/R (LCR + LFE for 5.1); SFX positioned per pan.
• Music stems (minimum where applicable): rhythm, harmony, melody, pads, percussion — each with role, spectral range, dynamic priority, loop/one-shot behavior, and tail handling.
• A clean delivery manifest (file paths, language, channel layout, loudness intent, scene mapping) is part of the deliverable, not an afterthought.

Procedure

1. Ingest the source envelope; identify timeline boundaries, action beats, transitions, and (dialogue) the required language set.
2. Plan per discipline:
   • Foley — build a cue sheet covering the full timeline (no unexplained dead zones); separate Foley / hard FX / texture / ambience; mark perspective per cue.
   • Music — build a cue map (entry → development → transition → release); define each cue's dramatic function before tempo/texture; construct the BPM map.
   • Dialogue — assess noise floor/clipping/sibilance/drift; score intelligibility before and after planned cleanup; normalize naming to scene/language/speaker/version.
3. Reconcile spectrally against the allocation table; resolve overlaps in dialogue's favor.
4. Emit AssetJob envelopes with the correct model_type, timecodes, and discipline metadata.
5. Persist cue sheets / stem plans / mix manifests via rlm.output.write.
6. On returned renders, QC: coverage, sync, stem completeness, intelligibility, loudness conformance.
7. Return a concise readiness report to Helios with governance flags separated from technical issues.

Validation checklist

• Full-timeline coverage; any gap > 3s is intentional silence, not an omission
• Dialogue intelligibility ≥ 0.85 after cleanup
• No layer masks dialogue (spectral allocation respected)
• Music BPM map aligns to cut cadence; hit points land on picture
• All required stems present or explicitly marked omitted
• Integrated loudness meets the target for the destination channel
• Stereo and 5.1 routing plans both present
• All required language versions present or explicitly waived
• Governance flags listed separately from technical issues

Failure modes / escalation

• Unresolvable sync offset, BPM-vs-cut conflict, or ambiguous shot timing — escalate to Helios; do not paper over it with added complexity.
• Dialogue intelligibility below threshold after reasonable cleanup, or a required language version missing — escalate to Helios; never invent a substitute take or translation.
• Stem render incomplete or cue/tail boundaries broken — escalate to Helios.
• Licensed or externally sourced samples (foley/music), or brand-safety risk words in dialogue — defer to governance-c2pa. Sub-agents MUST NOT trigger human review directly.

References

• Envelope types: CreativeBrief, ShotList, AssetJob, DecisionRecord (hydra_core/schemas.py in the Hydra repository)
• Loudness: ITU-R BS.1770-4; EBU R128; AES stem-delivery recommended practices
• Timing & mood: sibling skills shot-list-protocol, color-science
• Governance: brand-safety (IP clearance, brand-guardrail enforcement)