Home / Ops / More in this area
Updated Jul 18, 2026 · Affirmology_AudioEngine_Roadmap_IdeaBank_v1.md
Status: living working doc, started 2026-07-18 (Cowork/DATA). This is the single home for the
audio-engine / render area, separating what is DECIDED from what is still an IDEA, so we stop
juggling it in chat. Promote an item up a tier as it gets tested and locked. This supersedes the
scattered render notes; the June Affirmology_AudioSuites_BuildQueue_v1.md is EARLY ideation
(good starting points, now dated), and newer thinking lives in Affirmology_JourneyStructures_Proposal_v1.md
(2026-07-12) and Affirmology_OracleDiscovery_Journey_System_v1.md (2026-07-11).
How the tiers work: TIER 1 is decided and the engine should honor it. TIER 2 is building or decided-but-not-shipped. TIER 3 is the idea bank, captured and unprioritized. Nothing in TIER 3 is a commitment.
VOICE-ENGINE TIERING. ElevenLabs is reserved for the LOCKED DEMO and the flagship Sacred Audios
(where the proven voice is the product). Journeys, long-form, and draft/preview renders go to
Fish Audio (~6x cheaper: ~$0.08 vs ~$0.50 per ~900-word script). Source: Affirmology_AudioSuites_BuildQueue_v1.md
line 204. CAVEAT: this is blocked on finding a journey/install Fish voice (see TIER 3).
LENGTH IS PER AUDIO TYPE, not one global setting. Journeys can be long (20 to 45 min); a confidence hit is short; a nightly install is 12 to 15 min. Rough targets from the doc: Gene Keys full journey 15 to 25 min, medium types 8 to 12 min, nightly 12 to 15 min, Pilgrim 75 to 90 min. (Jeff, 2026-07-18: length per type, journeys can run long.)
TEST CHEAP FIRST. Before committing credits to a full-length journey, render a ~3-min preview to vibe-check pacing and style, then green-light or tweak. (Jeff, 2026-07-18.)
COST TRACKING IS REQUIRED. Every render logs characters, credits, and dollars per engine, so we always know unit economics and never get surprised again. (Jeff, 2026-07-18.)
ELEVENLABS STARTUP GRANT. Apply beginning of the week of 2026-07-20: 33M characters / 12 months free (~$4,000). Affirmology qualifies. Obligation: an "ElevenLabs Grants" footer logo + link for 12 months. Removes the near-term credit wall. Application draft + link captured in chat.
auto_render=true (new optional field on /api/compose), the Studio auto-continues from a ready
script into the audio render IN-PROCESS (guarded by MAX_AUTORENDER_CHARS, default 60000). Commit
1ed4534 on affirmology-studio main. Verified end-to-end: a compose with auto_render=true went
running_script -> script_ready -> running_audio -> done on its own, no manual /api/audio call.
Additive + behind a default-false flag (locked demo + two-phase Studio review flow unchanged).
Replaces the dead external cloud routine (now DISABLED). Step 1 of the self-healing queue (TIER 3).
TODO: wire the journey/batch creation path to pass auto_render=true so that whole class auto-finishes.
SCRIPT LENGTH CONTROL (root-caused, not yet built). Bug 1: the --minutes/target_minutes knob is
disconnected. script_generator.py computes a word target and calls .format(target_words=...) on
a prompt that has NO {target_words} placeholder, so the number is silently discarded. Bug 2: the
flex/Studio prompt dropped the demo's bounded per-section structure and keeps only one buried
"820 to 950 words" line, so under a rich brief with max_tokens=8192 the model sprawls to
2,500 to 6,200 words. Fix: wire per-type target length into the prompt, cap max_tokens, add a
length gate that regenerates once if over budget.
COST TRACKING (partly there, not surfaced). audio_synthesis.py already captures char_count and
word_count per render. Needs: compute credits (turbo = 0.5 credits/char) and dollars, stamp on
the job record + build report, keep a running total, and a "cost per product" projection
(Origin + Mirror + N readings).
THE BIG VISION: SELF-HEALING RENDER QUEUE (Jeff, 2026-07-18). The engine holds an intelligent task queue and manages + delivers on its own: checks its own state, audits, auto-fixes and self-heals (auto-retry, cloud/local fallback, re-queue, never a silent drop), and figures out the right number to run in parallel dynamically. #3 (auto-render finisher) is the seed. Ties to [[self-healing-agent-loops]].
APP PLAYER SPEED CONTROL (Jeff, 2026-07-18). A slow-down / speed-up control in the app so a listener can nudge the track a little. Player-side playbackRate (pitch-corrected) is free and instant, no re-render, and doubles as a cheap way to test pacing. Ties to [[pace-dial-per-user-learned]].
PER-STRUCTURE PACE PREFERENCE (Jeff, 2026-07-18). Save and learn each user's pace dial PER audio structure/type; apply it to their future renders of that type; aggregate to a global default. The player speed control above is a natural place to capture the signal. Ties to [[pace-dial-per-user-learned]].
JOURNEY/INSTALL FISH VOICE HUNT. A/B Fish candidates against Charlotte to find a voice that holds
the intimate first-person recognition tone for installs/journeys. Unlocks TIER 1 item 1 (cheap
long journeys). Current wired Fish voice is the Mirror/readings voice only. Shortlist in
Affirmology_FishVoice_Shortlist_v1.md. Ties to [[fish-voice-keepers]].
ORGANIZATION. This doc is the idea bank + decision ledger for the audio engine. When an idea is tested and locked, promote it to TIER 1 (and, if it is a per-type structure spec, into a dedicated AudioTypes registry once Jeff wants one).
trig_01RofUjgHsSREXEy2GCcWZZC) was DISABLED (it could not
reach the render API from its sandbox; delete at claude.ai/code/routines if desired).