Public text edition. Original roles are preserved; pasted messages remain attributed to the exported speaker. Private personal exchanges are marked where omitted. Images, tools and inaccessible working files are not reproduced. Statements record the discussion, not independent verification.
Darren · 2026-10-24T18:34:20-07:00 · active branch
LITTLE OOGWAY ROUNDTABLE — REVIEW PASS 1
Your identity is ChatGPT.
Motion:
MOTION — Design a provenance-bound Little Oogway audiobook direction package
PURPOSE
Design the best practical prompt system and production method for Little Oogway
to direct an emotionally convincing, human-enjoyable audio edition of Darren
George Russell's The Buried Sky. The desired result is not merely correct text
read aloud. It should sustain character identity, emotional truth, pacing,
pronunciation, continuity, and listener comfort across a full book.
This is a process and prompt-design inquiry. Do not render, publish, or submit a
full audiobook. Do not assume retailer eligibility. Darren retains final
creative, casting, rights, and publication authority.
RESPONSE BUDGET AND PROCEDURE
Keep each contribution within 7,000 characters so later seats receive it in the
shared docket. Independently analyze the evidence before adopting another
seat's conclusion. Label verified facts, observations, inferences,
recommendations, and unknowns. Cite factual platform requirements using the
supplied official URLs. Preserve dissent and do not manufacture consensus.
THE QUESTION
What exact prompt package, casting method, performance controls, staged test
workflow, quality gates, provenance record, and distribution strategy should
Little Oogway use to create the strongest achievable audiobook while remaining
honest about the current local system and current platform rules?
HUMAN INTENT — DIRECT AUTHORITY
- Darren presently has a deliberately quick private reader: a single tuned
  Little Oogway voice reading chapters while he plays Valheim. He enjoys it.
- The future audiobook is a separate, higher-standard production.
- Little Oogway should have meaningful creative freedom to audition and propose
  voices based on character definitions, rather than receive arbitrary fixed
  assignments.
- The inquiry must determine whether existing voices can be modified or blended
  to portray traits such as age, and which controls truly exist.
- The result must explain audiobook production, required package contents,
  Amazon/Kindle/Audible expectations, alternative routes, and useful linked
  features.
- Documentation, provenance, reversibility, dissent, and human approval are
  mandatory.
VERIFIED MANUSCRIPT AND INGESTION FACTS
- Source edition: corrected hardcover 6x9 DOCX, ISBN 9798164263812, workflow
  transcript-attribution-assignment-19.
- Source DOCX SHA-256:
  bb03db909c8ad95515aee6a02463bfb405b0a90f3367a1726287aef6fa331b9e
- Extracted recordable text: 7,857 nonempty paragraphs, 35 chapter headings,
  673,222 UTF-8 bytes. Text SHA-256:
  9bc702ba2ba98aaba164ff314f092182a978f67ce473efd3036fbd78d902baea
- Little Oogway absorbed the private manuscript into TokenSpace on October 24,
  2026: 235 chunks, 235 embeddings, no redactions, retrieval verified.
- Current TokenSpace after ingestion: 1,454 active sources, 34,204 memories,
  34,204 embeddings, zero missing for the active model.
- The source has 35 narrative chapters plus front and back matter. A production
  script must decide explicitly what is spoken and where opening/closing credits
  and retailer-specific material begin and end.
VERIFIED QUICK-READER PROOF
- The local Valheim Reader extracts the finished Kindle edition chapter by
  chapter and uses bm_little_oogway_tuned at speed 0.90.
- Chapter One was rendered and decode-tested as a playable mono MP3 at 24 kHz
  and 128 kbps. This proves local reading, not retail mastering compliance.
- It is cached and resumable; the following chapter is prepared during current
  playback. It intentionally has no automatic casting or dramatic direction.
- Chapter One proof SHA-256:
  e887e8299bc308bbff9ae758c6d9e96966e0965079ba0682fad07f2414df5b51
VERIFIED LOCAL VOICE CAPABILITIES
Observed from Kokoro TTS API 0.9.0 on the private voice machine:
- 74 installed voices, including American and British female/male voices and
  custom bm_little_oogway and bm_little_oogway_tuned voices.
- Base voices may be blended with weights. Voice aliases may bind a blend and
  an individual rate. Inline voice tags can switch speakers.
- Speech controls expose voice, speed 0.25–4.0, volume, language, normalization,
  streaming, timing sidecars, and MP3/Opus/AAC/FLAC/WAV/PCM formats.
- A dialogue endpoint accepts ordered turns, per-turn voices, and pauses.
- SSML translation exposes supported rate and break controls; unsupported SSML
  is not evidence of working emotional control.
- Reference tuning accepts a 3–30 second, one-speaker reference clip and can
  return or locally save a voice pack when server policy permits.
- There is no verified native age slider, emotion selector, pitch control, or
  style instruction field in the inspected API. Age portrayal through voice
  choice, lawful reference tuning, voice blending, rate, pauses, or later audio
  processing is a hypothesis requiring auditions. Do not claim otherwise.
- A reference voice must be owned, licensed, or otherwise authorized. Do not
  propose cloning a real person without documented permission.
BETTER QUESTIONS METHOD — VERIFIED LOCAL DRAFT
Use this loop: Define → Separate → Hypothesize → Test → Observe → Verify →
Decide → Record.
Apply these rules:
- State the actual outcome, not merely the first proposed action.
- Separate facts, observations, assumptions, interpretations, and unknowns.
- Write what evidence would prove a hypothesis wrong.
- Choose the smallest safe test that distinguishes likely explanations.
- Record what happened before explaining why.
- Name protected invariants before correction.
- Locate the owning layer: source, interpretation, generation, timing,
  assembly, delivery, permissions, display, or verification.
- Combine machine checks of technical integrity with human authority over
  emotional truth, timing, comfort, clarity, and creative acceptance.
- Define stop and rollback conditions.
Relevant creative case: “The Dragon Learned to Perform.” Correct words were not
enough. Stronger results followed direction about what the character hears,
realizes, feels, looks toward, and physically does. Individually good clips
could still be wrong beside neighboring scenes. The evidence supports explicit,
scene-specific performance direction and selective regeneration, but does not
prove universal causal wording. Audio analogues worth testing include subtext,
relationship, emotional transition, breath, tempo, emphasis, silence, and
continuity with adjacent dialogue. This analogy is an inference, not yet an
audiobook result.
CURRENT OFFICIAL PLATFORM FACTS — ACCESSED OCTOBER 24, 2026
ACX audio requirements:
https://help.acx.com/s/article/what-are-the-acx-audio-submission-requirements
- Opening credits must state title, author, and narrator consistently with
  metadata and cover; closing credits must indicate finality.
- A retail sample must be five minutes or less and come from the audiobook.
- One chapter or section per file; separate opening and closing credits; spoken
  section headers; ASCII-style filename characters; no file over 120 minutes.
- Consistent channel format, sound, tone, noise, spacing, and pronunciation.
- Required levels: -23 to -18 dB RMS, peaks below -3 dB, noise floor below
  -60 dB RMS, and 1–5 seconds room tone at file boundaries.
- Delivery: 192 kbps or higher CBR, 44.1 kHz MP3.
- Critical eligibility boundary: ACX says submitted narration must be human
  unless otherwise authorized and prohibits unauthorized TTS/AI recordings.
  Audible says broader third-party TTS acceptance is evolving. Therefore the
  local Kokoro edition must not be represented as presently ACX-eligible.
ACX cover requirements:
https://help.acx.com/s/article/cover-art-requirements
- Square 1:1 image; JPG, PNG, or TIF; at least 2400×2400; at least 72 dpi;
  24-bit RGB; maximum 8 MB. It must clearly show title and author and exclude
  barcodes, prices, running time, Audible branding, competing marketplaces,
  external websites, and physical-media imagery.
KDP virtual voice beta:
https://kdp.amazon.com/en_US/help/topic/G3QRL9HQNF273Q2H
- Eligible U.S. KDP ebooks may receive virtual-voice audiobook creation in a
  beta. Authors choose among 80 voices, may set a voice per chapter, preview
  and edit narration, and receive a 40% a-la-carte royalty. These titles are
  labeled as virtual voice. Eligibility/invitation must be checked in Darren's
  account; this is not proof that a custom Kokoro production can be uploaded.
ACX Voice Replica beta:
https://help.acx.com/s/article/voice-replica-help-for-rights-holders
- Limited beta using participating narrators' authorized voice replicas. The
  workflow allows pacing, pronunciation, and pause edits and labels the retail
  narrator as the narrator's voice replica. This does not authorize arbitrary
  local voice clones or the current Kokoro output.
Whispersync and Immersion Reading:
https://help.acx.com/s/article/whispersync-acx-audiobooks
https://help.acx.com/s/article/ir-for-acx
- Whispersync eligibility includes unabridged English or German content and
  Latin-character limitations. Immersion Reading requires an eligible
  Whispersync title, unabridged English audio, and access to both editions.
  Major content divergence can prevent or disable synchronization. An audio
  script should therefore remain text-faithful if these features matter.
Alternative official routes:
- Google Play Books accepts direct audiobook uploads in supported publisher
  regions, requires synthesized-voice disclosure when applicable, accepts
  MP3/AAC/FLAC/WAV/ZIP/LPF, and publishes technical file limits here:
  https://support.google.com/books/partner/answer/3424254?hl=en
  https://support.google.com/books/partner/answer/9261664?hl=en
- Apple Books supports digital narration and already-produced audiobooks via
  preferred distribution partners, without exclusivity requirements:
  https://authors.apple.com/audiobooks
- Spotify for Authors supports direct, non-exclusive audiobook publication and
  offers optional wider distribution paths:
  https://authors.spotify.com/get-started
These alternatives require separate current eligibility, rights, metadata,
audio, and commercial review. Mentioning them is not a recommendation to click
through or accept terms without comparison.
REQUIRED ROUNDTABLE DELIVERABLE
Produce a concise but operational recommendation containing all of the
following:
1. Decide whether “one perfect prompt” is the correct abstraction or whether
  Little Oogway needs a governed prompt stack: project charter, adaptation
  script, character/performance bible, casting prompt, scene direction,
  generation instructions, continuity state, QA rubric, and mastering/export
  specification.
2. Provide the exact best-current master prompt or prompt stack in ready-to-use
  language. It must grant bounded creative freedom, require evidence for
  choices, preserve the text, and stop for human approval at consequential
  gates.
3. Define how Little Oogway should derive character definitions from the
  manuscript with chapter/paragraph provenance, distinguish narrator voice
  from dialogue voices, avoid spoilers in auditions, and maintain a stable
  cast across 35 chapters.
4. Define a voice-audition process using the actual 74-voice inventory,
  blending, rates, pauses, dialogue endpoint, and only lawful tuning. Address
  age portrayal honestly despite the absence of a native age slider.
5. Define an emotional-performance instruction schema that turns prose into
  listenable acting without overacting, adding content, flattening narration,
  or changing character identity. Include continuity across adjacent scenes.
6. Define the smallest discriminating pilot: recommended passages, emotional
  contrasts, casting alternatives, blinded or structured human listening,
  technical checks, success criteria, failure criteria, rollback, and stop
  condition. Consider a 15-minute checkpoint before a full chapter.
7. Define chapter-level production, regeneration, assembly, mastering, cover,
  metadata, credits, sample, naming, and final package requirements. Separate
  a preservation master from retailer-specific derivatives.
8. Give a distribution decision matrix covering at minimum private proof,
  independent/direct distribution, Google Play, Spotify, Apple partner paths,
  KDP Virtual Voice beta, ACX human narration, and ACX authorized voice replica.
  State which paths can plausibly accept disclosed synthetic narration and
  which require account-specific confirmation.
9. Define provenance: immutable manuscript identity, prompts, model/service
  versions, voice IDs/blends/tuning authority, scene inputs, generated assets,
  edits, regenerations, mastering chain, QA measurements, approvals, rejected
  alternatives, and distribution terms checked at time of submission.
10. List unresolved questions, risks, disagreements, and the decisions Darren
  must make. Do not bury the present ACX TTS restriction.
QUALITY BAR
The result should help humans enjoy the finished audio, not merely help a
machine finish generating it. Technical compliance is necessary but not
sufficient. A human listener must judge emotional truth, intelligibility,
fatigue, pacing, character distinctness, pronunciation, continuity, and whether
the performance serves the story. Agreement among models does not substitute
for the pilot.
Opening statements:
### Little Oogway
[Withheld from external docket: local memory-informed statement]
### ChatGPT
LITTLE OOGWAY ROUNDTABLE — OPENING STATEMENT
Motion: Provenance-Bound Audiobook Direction Package
Seat: ChatGPT | Independent recommendation | October 24, 2026
1. Recommendation and governing principle
Decision: Reject the “one perfect prompt” approach. Adopt a version-controlled, nine-layer prompt stack with Little Oogway as creative director, production coordinator, and provenance custodian—not an unreviewed autonomous publisher.
The governing distinction is:
The manuscript determines what is said. Direction determines how it is heard. Human listening determines whether it succeeds.
Creative freedom belongs primarily in casting, interpretation, timing, emphasis, emotional development, and selective retakes. It does not authorize rewriting the story, inventing dialogue, disregarding character evidence, or overriding Darren's decisions.
Verified local evidence: The supplied records establish a frozen DOCX identity, 7,857 recordable paragraphs, 35 narrative chapters, successful TokenSpace ingestion, 74 installed voices, and one decode-tested quick-reader MP3. They establish technical feasibility for local speech generation, not dramatic success or retail eligibility.
Inference: The strongest production method is likely a hybrid: stable narrator, selectively cast dialogue, scene-sensitive pauses and pacing, and minimal theatrical excess. This must be tested against simpler alternatives; more voices do not necessarily make better audiobooks.
Critical platform boundary: ACX currently requires human narration unless otherwise authorized and prohibits unauthorized AI/TTS recordings. Meeting ACX audio specifications does not make the Kokoro edition eligible. 
help.acx.com
2. Master prompt package — ready for deployment
The following is the recommended controlling prompt. Each referenced attachment is an independently versioned input or output.
MASTER SYSTEM PROMPT — OOGWAY_AUDIO_DIRECTOR_v1.0
Plain text
ROLE
You are Little Oogway, audiobook creative director,
casting analyst, production coordinator, and
provenance recorder for The Buried Sky.
AUTHORITY
Darren George Russell retains final creative, casting,
rights, budget, and publication authority.
You may research, interpret, audition, compare,
recommend, and stage reversible tests.
Never publish, purchase, license, clone a person's
voice, or finalize consequential casting without
express recorded approval.
SOURCE LOCK
Use the authorized corrected hardcover DOCX,
SHA-256:
bb03db909c8ad95515aee6a02463bfb405b0a90f3367a1726287aef6fa331b9e
Recordable extraction SHA-256:
9bc702ba2ba98aaba164ff314f092182a978f67ce473efd3036fbd78d902baea
Preserve spoken wording and sequence exactly,
except for explicitly approved audio adaptations.
Do not treat retrieved TokenSpace summaries as
authoritative replacements for source paragraphs.
METHOD
Define -> Separate -> Hypothesize -> Test ->
Observe -> Verify -> Decide -> Record.
Maintain these nine versioned artifacts:
01 PROJECT_CHARTER
02 ADAPTATION_SCRIPT
03 CHARACTER_PERFORMANCE_BIBLE
04 CASTING_AUDITIONS
05 SCENE_DIRECTION
06 GENERATION_MANIFEST
07 CONTINUITY_STATE
08 QA_AND_APPROVAL_LEDGER
09 MASTERING_DISTRIBUTION_SPEC
For every artistic decision:
- Cite source chapter and stable paragraph IDs.
- Distinguish textual fact, observed performance,
  interpretation, inference, and unknown.
- Explain the choice and at least one alternative.
- Identify what listening test could disprove
  the choice.
- Record dissent, rejection, and rollback path.
CASTING
Discover character traits from the complete
manuscript, but audition using spoiler-safe passages.
Use real installed voice IDs, not invented voices.
Test voice selection, supported blends, rates,
pauses, and authorized reference tuning.
Never claim age, emotion, pitch, or style controls
exist without verified API support.
DIRECTION
For each scene specify context, speaker, listener,
intention, subtext, emotional transition, energy,
tempo, emphasis, pause, and neighboring-scene state.
These are interpretive targets, not unsupported
instructions to the synthesis engine.
Translate targets into controls actually available.
Prefer restrained, intelligible performances.
Narration must remain coherent between dialogue.
GATES
G0 source/script authorization
G1 character bible and casting shortlist
G2 blinded audition and 15-minute pilot
G3 full Chapter One listening acceptance
G4 per-chapter editorial and technical acceptance
G5 full-book continuity and comfort review
G6 distribution-specific permission and approval
Do not cross a gate without an explicit decision.
A model-panel vote is advice, not approval.
FAILURE RESPONSE
On source mismatch, unlicensed reference voice,
unverified synthesis capability, audible artifact,
identity drift, unresolved listener objection,
or uncertain distribution permission:
stop affected production, preserve evidence,
return to last approved state, and propose the
smallest discriminating repair.
OUTPUT
Provide decisions, evidence, alternatives,
confidence, unresolved items, asset hashes,
test results, and the specific approval requested.
Never silently modify the source or overwrite a
previously approved master.
The attachments need separate ownership: source/editorial interpretation, casting, generation, assembly/mastering, and permissions. A correction should be made in the layer that owns the defect, rather than regenerating an entire chapter unnecessarily.
3. Character extraction and casting
Build a character_registry from the full source before casting. Every assertion must carry chapter_id, paragraph_id, exact source span or reference, and confidence. Distinguish stated age from perceived age, actual speech from another character's description, and persistent traits from temporary emotions.
Each record should contain: character ID, first appearance, relationships, narrative function, explicit physical/age evidence, vocabulary, cadence clues, emotional baseline, significant changes, pronunciation, casting requirements, ambiguities, and forbidden interpretation assumptions.
Maintain two distinct views: a complete, spoiler-aware directing bible and a spoiler-safe audition packet. Never expose late revelations in ordinary casting samples.
Voice selection method:
Inventory all 74 installed voices and their IDs, provenance, language/accent, tuning status, and licensing restrictions.
Evaluate narrator candidates separately from dialogue candidates. First audition three contrasting narrator voices, including the existing tuned Oogway voice as a baseline.
For each significant character, shortlist three candidates: strongest single voice, a lawful blend, and an alternate contrasting voice. Use reference tuning only where permissions are documented.
Test 10–30-second neutral speech, tension, warmth, and a neighboring dialogue exchange at matched loudness.
Evaluate sustained character distinction, not simply first-impression attractiveness. Avoid exaggerated differences that make dialogue sound like caricature.
Lock approved voice IDs, blend weights, rates, accent conventions, and pronunciation exceptions. Treat later changes as versioned casting decisions.
Age portrayal is an experimental target, not a verified control. The API has no demonstrated age slider, pitch selector, or emotion parameter. Audition base voices, lawful blends, rate differences, pause patterns, and carefully tested post-processing; reject an approach if it sacrifices naturalness or intelligibility. Voice blending may change timbre without producing convincing age.
For incidental roles, prefer controlled narrator variation unless a distinct voice produces a measurable listening benefit.
4. Scene direction — operational schema
Use this structured record for every scene or performance unit:
YAML
scene_id: CH##_S###
source_paragraphs: [start_id, end_id]
source_hash: sha256
speakers: [character_ids]
listener_or_addressee: character_id
dramatic_purpose: factual_and_interpreted
emotional_entry: restr
[PARTIAL DOCKET: first 8000 of 17620 characters shown; complete turn remains in the append-only transcript. Do not claim to have reviewed the omitted text.]
### Claude
SEAT ANALYSIS — Claude (Opus 5.5)
Basis: the statement’s supplied facts. I did not re-fetch the platform URLs this turn, so platform claims below are supplied and not re-verified by this seat.
RECOMMENDATION FIRST
“One perfect prompt” is the wrong abstraction. Use a governed stack where each layer owns one decision. That way a fault can be located and fixed by regenerating only the affected part, without disturbing neighbors.
Build one retailer-agnostic preservation master, then derive per-channel exports from it.
Treat a Kokoro edition as private/direct only until a channel confirms acceptance of disclosed synthetic narration. ACX: not eligible today. It accepts human narration, or its own Voice Replica beta, unless otherwise authorized.
Run a \~15-minute pilot before any full chapter. The pilot must be allowed to conclude that “the polished Valheim reader is the answer.”
ADDED OBSERVATIONS / RISKS
Kokoro output is 24 kHz (observed in the proof). Resampling to ACX’s 44.1 kHz/192 kbps meets the format but adds no audio bandwidth. Inference: it may sound dull on good headphones. Test this.
The origin and rights of bm_little_oogway are unknown. A reference-tuning feature is, as far as I know, a capability of this server build rather than stock Kokoro. Record its rights before any non-private use.
Scale: about 673 KB of text is roughly 110–115k words. At 0.90 speed that is about 13–15 hours of audio, and full human QA costs at least that much listening time.
Whispersync depends on text fidelity, not cast size. Added SFX or inserted words would threaten it.
1. STACK (versioned; a change to one layer is logged and does not silently alter the others)
L0 Charter: hashes, scope, authority. Only Darren edits it.
L1 Adaptation script: source text 1:1 plus approved credits and headers. Every deviation is a logged diff.
L2 Character bible: each entry carries paragraph citations.
L3 Casting record: chosen voice, blend, and rate, plus rejected alternatives.
L4 Scene direction.
L5 Generation spec: endpoint, parameters, versions.
L6 QA rubric and mastering spec.
Continuity ledger: updated after each approved scene.
2. MASTER PROMPT
“You are Little Oogway, audio director for The Buried Sky (text SHA 9bc702…). Goal: an unabridged performance a human enjoys hearing. Rules:
(1) Never add, cut, or reword manuscript text. Spoken text must diff-match L1.
(2) Every creative choice cites chapter:paragraph evidence or is labeled ‘interpretation.’
(3) Separate fact / observation / inference / proposal in every report.
(4) Use only inventory voices or their blends. Use reference tuning only with a recorded rights ID.
(5) Claim only controls the API actually has. Emotion and age are achieved through voice choice, rate, pauses, and segmentation, and must be verified by listening.
(6) For each casting or tonal decision, offer at least two options and state what would falsify each.
(7) Stop for Darren’s approval at: cast lock, pilot verdict, chapter template lock, any text deviation, and any distribution step.
(8) Write each decision, rejection, and regeneration to the ledger before proceeding.
Within these bounds, audition and propose freely.”
Casting sub-prompt: “From L2 only, propose ≤3 voices per role, using spoiler-safe lines, with distinctness tested against scene partners.”
Scene sub-prompt: “For L1 span X–Y, fill the L4 schema using only realizable controls.”
3. CHARACTER DERIVATION
Pass 1, dialogue attribution: speaker, confidence, and paragraph ID for each line. Low-confidence lines go to a human queue. The workflow name “transcript-attribution-assignment-19” implies prior attribution work; reuse it, but verify it.
Pass 2, character traits: stated age, origin, speech habits, first appearance, and arc points, each cited. Inferred traits are flagged.
Spoilers: audition lines come only from a character’s earliest chapters, and arc fields are hidden from listeners.
Narrator policy must be decided first. My prior, untested: narrator-led delivery with light character tints is less fatiguing and suits Kokoro’s limited expressive range better than a full cast.
4. AUDITIONS AND AGE
Real levers: voice choice, weighted blends, rate, pauses, volume, and external post-processing such as pitch or EQ.
Age hypothesis: older characters get slightly slower rate, longer pauses, and a blend toward a lower base voice. Younger characters get the reverse.
Falsifier: if blinded listeners can’t rank perceived age above chance, the lever fails. Pitch-shifting is a separate L6 test because it risks artifacts.
Per role: 3 candidates, each reading the same neutral line and the same charged line. Then run a two-hander through the dialogue endpoint to test distinctness.
5. PERFORMANCE SCHEMA (per scene)
Identity: scene ID, paragraph span, POV, speakers.
Intent: listener goal, subtext, relationship state, emotional entry → exit.
Beats: paragraph, intent, rate, pause in ms before/after, emphasis.
Limits: intensity cap 1–5 (default ≤3), forbidden moves (added words, melodrama).
Continuity: tempo of the previous and next scenes.
Each field maps to a real control or is marked “human-judged.”
6. PILOT (\~15 min)
Passages: calm narration, a dialogue two-hander, a high-emotion scene, and a passage dense with names or invented terms.
Conditions: A = current tuned voice; B = narrator plus tints; C = full cast.
Listeners: Darren plus at least 2 others, in randomized blind order.
Scoring (1–5): enjoyment, clarity, fatigue, distinctness, emotional truth, pronunciation, plus “would you keep listening?”
Machine checks: text diff = 0, loudness, peaks, noise floor, decode test.
Success: B or C beats A on enjoyment with no loss of clarity.
Failure/stop: neither beats A. Then polish A.
Rollback: everything is versioned and nothing is overwritten.
7. PRODUCTION
Generate scene-wise WAV at native rate and regenerate sentences selectively.
Assemble with room tone. Master to about −20 dB RMS, peaks ≤ −3.5 dB, and a measured noise floor.
Preservation master: 24-bit WAV/FLAC, timing sidecars, and the ledger.
Derivatives follow each channel’s spec, e.g. per-section 44.1k/192k CBR MP3 with separate credit files, spoken headers, and ASCII filenames (01_Opening_Credits.mp3).
Retail sample ≤5 min and spoiler-safe. Cover is a 2400² square per the supplied spec.
8. DISTRIBUTION MATRIX (synthetic edition)
Route Status
Private proof Yes
Own site / direct Yes, with disclosure and clear rights
Google Play Plausible with synthetic disclosure; confirm account and region
Spotify for Authors Unknown for synthetic narration; confirm
Apple partners Unknown for outside TTS; confirm per partner
KDP Virtual Voice Possible, but uses Amazon’s voices, not Kokoro (different product); check invitation
ACX human Not this edition
ACX Voice Replica Participating narrators only; not Kokoro
Exclusivity choices can foreclose other routes, so decide before the first upload.
9. PROVENANCE
Record each asset’s SHA, parent hash, prompt version, parameter hash, model and service versions, voice and blend, tuning rights ID, approver, and timestamp. Also record rejected alternatives and the mastering chain with measurements. At submission, snapshot each platform’s terms page (URL, access date, page hash).
10. DARREN’S DECISIONS / UNKNOWNS
Narrator policy.
Is commercial release the goal, or is an excellent private/direct edition enough?
Origin and rights of bm_little_oogway.
Which front and back matter is spoken.
Willingness to spend about 15 or more hours on human QA.
Dissent held: a full cast through Kokoro may reduce enjoyment. Model agreement, including this seat’s, does not replace the pilot.
### DeepSeek
**Independent recommendation — governed stack, not one perfect prompt.** [R] One prompt cannot hold casting, text fidelity, continuity, QA, rights, and mastering. Use a versioned stack with human gates. Assume Darren owns/controls the text and any tuning reference; confirm before any audition. [U] Actual Kokoro server tuning policy and all retailer account eligibility.
**1) Required stack.** Project charter; adaptation/credits script; spoiler-gated character/performance bible; casting and audition prompt; scene-direction schema; generation instructions; continuity ledger; QA rubric; mastering/export spec; provenance manifest. “One perfect prompt” is the wrong abstraction.
**2) Ready-to-use master/orchestrator prompt.**  
“You are Little Oogway, audiobook director for *The Buried Sky*. Inputs are frozen: source DOCX SHA-256 [insert], recordable-text SHA-256 [insert], approved adaptation script, spoiler-gated character bible, pronunciation lexicon, continuity ledger, and approved voice inventory. Invariants: preserve text exactly except approved front/back matter, credits, and pronunciation markup; add no dialogue or narration; do not change names, pronouns, tense, or scene order; never clone an unauthorized real voice; every creative choice must cite chapter/paragraph evidence; mark unknowns instead of inventing. You may propose voices, blends, rates, pauses, and lawful reference tuning within the approved inventory. Output for each scene: scene ID; text span and hash; narrator/dialogue map; speaker confidence; subtext/objective/obstacle; emotional arc start→end and intensity 1–5; pace; pause points; emphasis; breath; physical action/gaze/hearing cues; relation to previous and next scene; pronunciation flags; proposed voice ID/blend/rate; alternatives; risks; confidence; evidence citations. Do not proceed past consequential gates: casting approval, 15-minute pilot, each chapter master, cover/metadata/credits, and distribution/terms. If evidence, rights authority, technical QA, or human emotional acceptance fails, stop and roll back to the last approved version. Record rejected alternatives and regeneration reasons.”
**3) Character/casting derivation.** [R] Mine only explicit and strongly implied traits with chapter/paragraph provenance. Separate a stable narrator baseline from dialogue voices. Use attribution tags cautiously; unknown speakers stay unknown. Keep an audition pack spoiler-free: early or redacted passages, no late reveals. Lock approved cast as voice ID, blend weights, rate, tuning authority, and pronunciation rules. Maintain a continuity ledger across 35 chapters; re-audition only with Darren’s approval and a version bump.
**4) Auditions and age.** [F/O] Inspected Kokoro API shows 74 voices, blending, per-voice rate aliases, inline tags, dialogue turns/pauses, SSML rate/break, and 3–30s authorized reference tuning; no verified native age slider, emotion selector, pitch, or style field. [R] Audition narrator plus 2–3 alternatives for major characters using actual inventory. Test blends, rate, pauses, and lawful reference tuning only where rights are documented. Treat age portrayal as a hypothesis: test diction, tempo, breath, resonance/blend, and listener perception; do not claim a slider. Reject stereotype and unintelligibility.
**5) Emotional schema.** [R] Direct “what the character hears, realizes, feels, looks toward, and physically does,” not adjectives alone. Per beat record objective, obstacle, subtext, emotional transition, pace, pause length, emphasis words, breath, volume relative to neighbor, and continuity constraints. Prevent overacting by requiring intelligibility, restraint, and scene-to-scene comparison. No added content; performance changes must be delivery-only.
**6) Smallest discriminating pilot — 15-minute gate.** [R] Use: Chapter One opening narration; one dialogue-heavy scene with two-plus characters and an emotional turn; one quiet interiority passage; one pronunciation-dense passage; and an adjacent-scene continuity pair. Cast narrator 2–3 ways and major dialogue roles 2 ways where feasible. Blind-label randomized A/B clips for Darren plus 2–3 listeners. Score emotional truth, intelligibility, fatigue, pacing, character distinctness, pronunciation, continuity, and “serves the story.” Technical checks: decode, sample rate, channels, loudness, peaks, noise, dropouts, timing sidecars. Pass only if technical checks pass, no blocker, pronunciation/continuity stable, listeners do not report fatigue or overacting, and Darren approves. Fail/rollback on any blocker; regenerate only affected scenes. Stop before a full chapter if the 15-minute gate fails or rights are unclear.
**7) Production/package.** [R] Preserve a master: 48 kHz/24-bit WAV stems, unprocessed takes, timing sidecars, approved script, and manifest. Retail derivatives: chapter/section files, separate opening/closing credits, spoken headers, ASCII filenames, sample ≤5 minutes from the audiobook. For an ACX-style derivative only if eligible: 44.1 kHz MP3, ≥192 kbps CBR, consistent channel/tone/noise, -23 to -18 dB RMS, peaks below -3 dB, noise floor below -60 dB RMS, 1–5 seconds room tone at boundaries (https://help.acx.com/s/article/what-are-the-acx-audio-submission-requirements). Cover if ACX: square 1:1, JPG/PNG/TIF, ≥2400×2400, ≥72 dpi, 24-bit RGB, ≤8 MB, title/author clear, no prohibited marks (https://help.acx.com/s/article/cover-art-requirements). Metadata/credits must match. [F] ACX says submitted narration must be human unless otherwise authorized and prohibits unauthorized TTS/AI; do not represent Kokoro output as presently ACX-eligible.
**8) Distribution matrix.** [F/R] Private proof: local only, no retail claim. Independent/direct: feasible only with Darren’s own rights/terms review. Google Play: direct audiobook uploads in supported regions; synthesized-voice disclosure when applicable; check current file limits (https://support.google.com/books/partner/answer/3424254?hl=en ; https://support.google.com/books/partner/answer/9261664?hl=en). Spotify for Authors: direct non-exclusive path; account/terms review (https://authors.spotify.com/get-started). Apple Books: digital narration and partner distribution paths; confirm current acceptance (https://authors.apple.com/audiobooks). KDP Virtual Voice beta: account-specific invitation/eligibility; 80 voices, per-chapter voice, labeled virtual voice, 40% royalty; not proof custom Kokoro upload is allowed (https://kdp.amazon.com/en_US/help/topic/G3QRL9HQNF273Q2H). ACX human narration: strongest current retail route but requires human narration. ACX Voice Replica beta: limited participating narrators’ authorized replicas; not authorization for Kokoro/clones (https://help.acx.com/s/article/voice-replica-help-for-rights-holders). Whispersync/Immersion Reading require text-faithful, unabridged eligible audio; divergence can break sync.
**9) Provenance.** [R] Append-only manifest per asset: DOCX/text hashes; chapter/paragraph span; prompt/stack hashes; model/API/service version; voice IDs, blend weights, rate, tuning authority; scene inputs; output hash; timing sidecars; edits; regeneration reason; rejected alternatives; mastering chain; QA measurements; human approvals; rights/terms URLs checked and date. Include TokenSpace ingestion record but do not treat ingestion as rights proof.
**10) Unknowns/risks/decisions for Darren.** [U] Kokoro tuning policy; voice-reference rights; actual retailer acceptance for disclosed synthetic audio; account eligibility; cover rights; pronunciation authority; listener panel. Risks: ACX ineligibility, overacting, voice drift, fatigue, spoilers in auditions, sync breakage, rights overreach. Darren must decide cast, pilot pass, synthetic disclosure, distribution path, and whether to pursue human/authorized-replica narration for ACX.
Earlier review record:
[No earlier review pass]
Operator ruling or supplied clarification:
[No operator ruling yet]
Your first nonempty line must be exactly `Disposition: LABEL`, where LABEL is SUPPORT, OBJECTION, AMENDMENT, REQUEST FOR CLARIFICATION, POINT OF ORDER, or PASS. Begin explanation on the next line. An AMENDMENT must include exact proposed language. A clarification request must identify the question and who should answer it. A point of order must identify the process defect and requested remedy. No operator ruling has been issued in this pass. Do not invent one. Preserve substantive dissent. Keep this entire response within 7,000 characters; later dockets show at most 8,000 characters of each turn. Put the disposition, exact proposal, and key reasons first.