Home / Ops / More in this area

Affirmology Lost-Chat Recovery: Name Pronunciation, QC Audit, Identity Intake

Updated Jul 18, 2026 · Affirmology_LostChat_Recovery_NamePronunciationQC_v1.md

Affirmology Lost-Chat Recovery: Name Pronunciation, QC Audit, Identity Intake

Recovered from a Cowork/DATA chat lost to a desktop update. Source transcript on disk: ~/.claude/projects/-Users-jeffreyparker-CLAUDE-AFFIRMOLOGY/66d1fae1-6d73-4e02-885a-123887e66f6e.jsonl (dated 7/12, ~2876 messages, opened "i'm back" during a Mercury-retrograde day of wifi/router drops). This file reconstructs it so a NEW chat can resume with zero context loss. Everything below was VERIFIED against the transcript and against files on disk.

1) CONTEXT

This was a long marathon DATA/backend session (the engine + affirmology-studio cloud side), running in parallel with a separate CODE/apps chat that was slowly rendering 33 "Mirror" readings all day (~25 min each). Jeff's laptop wifi kept dropping. Coordination between chats ran through _CHAT_SYNC.md. The session switched the model to Fable partway through (at Jeff's request, around the name-pronunciation build). It ended cleanly with Jeff saying "ok awesome. closeing this and restarting" and the assistant confirming everything was committed/deployed/documented and reboot-safe.

Four big threads ran in this chat, roughly in order: 1. Self-heal + demo-reserve spend cap (early). 2. The QC Audit and Plan (structural-drift quality control) - the "scared AF about quality" thread. 3. Naming session (audio/star names). 4. The 7/12 identity handoff: voice-note transcription, name pronunciation system, intake/identity fields, time-invariance filter, sung-name deduction, name review queue, and the chant runner. 5. Wheel art wiring + star-scrubbing the 12 zodiac figures (end).

All referenced docs exist on disk in the AFFIRMOLOGY root (verified): Affirmology_QC_Audit_and_Plan_v1.md, Affirmology_NamePronunciation_Spec_v1.md, Affirmology_IntakeForm_and_Identity_Spec_v1.md, Affirmology_AudioNaming_Registry_v1.md, Affirmology_AppExperience_MasterPlan_v1.md, Affirmology_FirstReading_Structure_v1.md. UI ref: CLAUDE OUTPUTS/Affirmology/Affirmology_NamePronunciation_UI_v1.html. Wheel art: CLAUDE OUTPUTS/Affirmology/WheelConcepts/.

2) WHERE IT WAS LEFT (end state at close)

The assistant's final message declared everything reboot-safe. Concrete standing state:

• Studio (cloud, Render) - DEPLOYED and verified live on commit b99efbd (later health=ok): the full 2026-07-12 identity handoff (items A-E), plus the chant runner (v2 "build-and-converge" template). The cloud does not care about a laptop reboot.
• App (affirmology-app / "Maya") - COMMITTED, awaiting a build. Wheel alignment + pronunciation screens committed (5f762fc and earlier). They ride the next internal TestFlight build whenever Jeff says "build it." Latest referenced internal build was build 31; wheel figures hot-swap on next open (no rebuild).
• Chant: Jeff's own chant rendered live end-to-end. Take 2 promoted to canonical; an experimental take 4 (build-and-converge arc) was rendered and was WAITING ON JEFF'S EAR at close. All takes logged in the VoiceRegistry.
• QC drift-gate: built, tested (16/16), committed (fd46119). Running LOG-ONLY, blocks nothing, beat-markers OFF until Jeff signs off. (Note: Jeff DID approve the Mirror blueprint later in the session - see below - but enforce was not confirmed flipped on.)
• Self-heal + demo-reserve: built, tested, committed locally (c98818f). Push was "held" for the mirror render window; may have shipped later in the session.
• Wheel: all 12 zodiac figures star-scrubbed (baked-in tuning stars removed via Higgsfield edit, NOT repaint, to preserve alignment), live on R2, originals archived to wheel/figures/original/. Alignment wired to Affirmology_WheelAlignment_FINAL_v1.json.
• Coordination: every thread closed out in _CHAT_SYNC.md.

Nothing was left mid-edit or exposed. The only true OPEN decision at close was take-4-vs-take-2 for the chant (see Open/Next).

3) KEY DECISIONS (concrete)

QC Audit and Plan (Affirmology_QC_Audit_and_Plan_v1.md)

• The audit finding: three real QC gates exist and work - audio_qc.py (dropouts/loudness), chart_guard.py (chart-contamination guard, caught the Josh-on-Jeff's-record bug), and script_qc.py (mechanical: strips em dashes, enforces fixed close, bans HD gate numbers, flags length). The gap: nothing judges whether writing is GOOD, and nothing catches STRUCTURAL DRIFT. A structure's "shape" is stored as free-text prose (structures.py:62-63 directive/overview); the only shape-enforcement in the whole codebase was enforce_origin_close. So a render could skip a beat, reorder, drop the threaded symbol, or leave the open loop open and pass every gate. verify_send.py existed but nothing called it.
• Jeff's steer: structural drift is the #1 target. Craft-judge must score four dimensions: structural drift, generic/not-personal, wishy-washy, length/pacing/delivery. Decisions given: "1. green light, 2. propose smart and go, 3. fine with 90% and tuning."
• The plan (3 moves): (1) each structure becomes a machine-readable BLUEPRINT of required beats; (2) a CONFORMANCE GATE = deterministic checks (free: order, markers, length, chart-match) + a cheap LLM craft-judge; hard drift is blocked and regenerated with the failed beat as feedback; (3) ACCOUNTABILITY = per-person + per-structure verdict logging and a structure scoreboard; a structure cannot be marked "approved/production" until it clears a conformance pass-rate.
• Built: conformance engine + Mirror & Origin blueprints + bed-aware length logic (single source of truth word_band in pacing.py), wired into compose()'s existing gate-and-revise loop (Apollo runs gates; failures feed back to Orpheus) as an additive structure_id param (default "" = no-op, so the byte-locked demo is untouched). 16/16 tests pass. Commit fd46119.
• The bed insight = the length rule (Jeff's, load-bearing): length tolerance is PER-STRUCTURE, keyed to whether the bed can stretch. Loopable/ambient or spoken+silence = STRETCHABLE = generous band, soft flag. Composed-with-a-fixed-ending (the Origin's 6:15 "Heavenly Circuit") = FIXED = tight band, HARD block so the ending lands on the music. Encoded bands: Mirror 1373-2145 words @13min (loose); Origin 765-899 (tight).
• Mirror blueprint = APPROVED/LOCKED by Jeff. The 5 beats (INTERNAL scaffolding labels, stripped before voicing, never heard): Promise -> Being Seen -> Three Doors -> Install -> Seal. Plus: one threaded bespoke symbol, an open loop, seal must be last. Confirmed against Affirmology_FirstReading_Structure_v1.md (Rule 12: "The listener is the hero, always"; Rule 3: never five competing metaphors; master symbol drawn from the person's own metaphor bank).
• Testing decision: do NOT re-run the gate on the existing 33 mirrors (generated before markers existed, so only length is checkable). Instead dogfood: regenerate ONE fresh founder Mirror (script only, no Fish spend) with markers + enforce ON to watch the gate work. Better still, break in the gate during founder audio testing Sun/Mon.
• Craft north star (locked to memory): every audio celebrates THIS unique person and casts them as the hero; hero-cast and bespoke-not-cliche-metaphor are HARD craft-judge checks, not hopes. (Matches the "every-audio-celebrates-them" memory.)

Self-heal + demo reserve

• Self-heal auto-completes any half-built person: builds chart (free) then Understanding (~$0.20), retries failed builds across restarts, exposes one status endpoint ("37 done, 0 pending"). Commit c98818f.
• Demo-reserve spend cap (Jeff's requirement): self-heal must leave credits for the demo. Added a PROACTIVE per-pass cap SELF_HEAL_MAX_SPEND_USD (default $1.00/pass) that stops BEFORE a build would exceed it, not just after a credit error. Tested: $1.00 cap builds 5 leaves 1; $0.50 builds 2.
• Hard limitation: Anthropic has NO live "check my balance" API (confirmed across the codebase). Every reserve works by Jeff telling the balance; cap = balance minus demo reserve. Practical floor = auto-reload ON. Rule locked to memory: ask Jeff the credit situation before any BULK/expensive job (with the billing link), don't ask for the small autonomous trickle; scarcity is temporary (funding coming post-raise). Billing: https://console.anthropic.com/settings/billing

Identity handoff (5 items A-E) - all DEPLOYED live, verified against the running cloud

• A. Identity fields persisted: name_called (verbatim, never cleaned), name_birth, and tri-state birth_time_known (the "Randy fix": a ticked "I don't know" is now distinguishable from a real noon birth and from legacy rows). Round-trip tested. This UNBLOCKS the site intake form deploy.
• B. Time-invariance filter: if birth time is unverified, every model input is masked to the invariant set (things the birth hour cannot change); Origin-only gate; no PDF/Mirror until time is rectified. Live cusp-day proof: Sun flip masked, rising masked. Randy now renders safely from invariants.
• C. Sung-name deduction (api/identity.py): base = the chosen name ALWAYS (never the birth name); body/surname added only when provably safe; divergent cases are absolute. All 7 spec examples pass verbatim. Adds name_sung column. Runs inside _capture_profile.
• D. Name review queue: every name logged (passes included); hold matrix holds ONLY vulgarity/impersonation (<1% hold-rate guard); human verdicts both directions. identity.py endpoints (/api/name-review/...).
• E. Name pronunciation system (spec Affirmology_NamePronunciation_Spec_v1.md): engine matcher does WHOLE-TOKEN substitution only (never rewrites inside a name; "Ajnabee" survives; Scunthorpe rule holds); person override beats the lexicon; a person named after a lexicon word is protected. 10/10 tests. Hooked at the SINGLE choke point every voice path funnels through (ElevenLabs, Fish, both paced renderers all say it right) while any script a HUMAN reads keeps true spelling. Lexicon seeded with 17 terms (Affirmology hand-fix now permanent), founder-gated admin CRUD. Verified voiced: "rob, welcome to af-er-MOL-uh-jee. Your AHZH-nah hums, rob HART-line."
• Spec correction found live: the spec's 2-point time-invariance algorithm (00:00 vs 23:59) had a hole - the ascendant returns to nearly the same degree after ~23h56m, so a diff check passed "Rising: Gemini" while it actually swept all 12 signs between. Fixed with 5 samples across the day + regression test. The spec docs should absorb this correction.

App pronunciation UX (front-end, built to match Affirmology_NamePronunciation_UI_v1.html)

All 5 surfaces committed (typecheck clean), entry = Explore -> Profile -> "Your name, said right.": 1. Chant confirm screen with a small "Not how I say it" link (proactive catch). 2. "Did we get it right?" [Yes, that's me] / [Not quite]. 3. Three fix paths, easiest first: (a) variant picker with real TTS previews, (b) say-it-yourself recording (the recording is KEPT), (c) respelling box with IPA behind an "advanced" toggle. NO bare IPA box anywhere. 4. Confirm + lock with a "hear it sung" preview (sung != spoken) and an OFFERED (never forced) Origin re-render. 5. Admin lexicon table (term / pronunciation / sung form / category, add + edit). Fire rules: confirming a pronunciation change AUTO-QUEUES the chant re-render; Origin remake is offered not forced.

Chant runner (POST /api/chant/run {person: <email>})

• Runs CLOUD-SIDE, fire-and-forget, reads the queue table directly. Every pronunciation confirm queues a re-render + kicks the runner; a boot worker polls every ~3 min so nothing sits.
• Recipe: Eleven MUSIC API (POST https://api.elevenlabs.io/v1/music, music_length_ms: 30000), the proven 7/11 Gregorian prompt - a-cappella male monastery choir, melismatic, cathedral reverb, ABSOLUTELY NO instruments (model-added instruments bury the name; mix OUR bed under at ~0.18 in ffmpeg afterward if a bed is wanted). Sings the CHOSEN name (name_sung, never birth) with syllable hyphens + ellipses, e.g. "Jeff... Par-ker... Love.... I am Jeff Parker Love."
• 3 takes per person (generation is a lottery); take 1 initially canonical at chants/<person>.mp3 (app welcome reads this); chant_url now on every profile read.
• Held names never sing; a verdict releases them; a pronunciation fix auto-re-renders. Cost-safe: capped per pass, paced, quota error stops the whole pass, failures record reason and never blind-retry, Render kill-switch.
• v2 template upgrade (from Jeff's craft note): Jeff liked take 2 and the way the choir converges on "Love" but not the weak opening "Jeff." That became structure - take 4 uses a single solo voice opening, voices joining phrase by phrase, full choir converging on the final words. If take 4 wins, the runner's prompt TEMPLATE updates so every member's chant inherits the build-to-convergence arc.

Naming session -> Affirmology_AudioNaming_Registry_v1.md

• Principle: keep the mechanical "Subconscious OS / software update / version numbers" language in the APP UI only; the AUDIO NAMES themselves must be reverent/rooted (never "install / claim / sync / boost" software-chore verbs). Names are DISPLAY LABELS only (stable internal id underneath), so they can be renamed by user, rotated seasonally, or A/B'd without touching the engine.
• Two-register system: Register 1 (elemental/calm, everyday) vs Register 2 (heroic/cinematic, charged/deep).
• LOCKED first set:
• Morning -> First Light (calm)
• Sleep -> The Night Tide (calm)
• Meditation (category) -> The Stillness, holding The Release (EFT/tapping) + The Seeing (visualization) (calm)
• Movement/gym -> The Forge (charged; beat out "The Rising"/"The Kindling")
• Deep journeys (category) -> The Descent, The Becoming (hero's journey), The Passage (Chiron)
• The Rising reassigned to a NEW concept: a stronger, EVOLVING Origin Soul Song (the Sun/core-identity anthem that deepens and re-renders as the person evolves) - distinct from the fixed one-and-done origin_soul_song entry gift.
• Retired (do not revisit): "The Kindling," the "Daily Pair / Night Install / Morning Claim / S5/S11" mechanical labels. Unplaced palette kept for future structures: The Current, The Ascent, The Pulse, The Turning, As I Sleep / As I Wake.

First-suite / app structure (from Jeff, line-level detail)

First constellation = at least 4 stars: Morning, Movement, Meditation (a category giving an EFT audio + a visualization + a new reading), Sleep. On entry the user is prompted for a weekly CONTEXT (selection box + optional type-or-voicenote) that shapes that week's audios and reading. Specialty pieces to start: a Chiron healing journey, a long hero's-journey audio. Other pieces greyed out / rolled out on discovery. The weekly-context prompt is the real engine unlock (chart + fresh current context, not chart-static) and slots into the Understanding's "context store."

Voice notes v1 (transcribe endpoint)

POST /api/app/transcribe (app auth pattern), multipart audio (m4a/caf/webm, cap ~3min/~10MB), transcribe via ElevenLabs Speech-to-Text (scribe_v1) using the existing key, return {"text":...}. TRANSCRIPT ONLY is a hard rule: raw audio never stored, temp file deleted in a finally block, log only person/duration/char-count. Backend was Jeff's/DATA lane; the composer mic button (RN, next to send) was the CODE/apps lane. Early mic-button attempts failed ("Transcription failed") until a NEW full-access ELEVENLABS_API_KEY was put in .env (does TTS + Music + STT), after which transcription worked. There is NO OpenAI key with credits (checked) - Scribe fallback to OpenAI Whisper is not available.

Olga birth-time rectification (recalled from Affirmology_Olga_Rectification_v1.md)

Resolved range ~9:20pm-9:55pm (Feb 25, 1966, Dubievka, Ukraine); working pin 9:35pm (21:35), the midpoint. Any time in that 35-min window gives the same chart. Q1 (Moon 8th->7th ~9:20, answer B = after 9:20) and Q2 (Mars+Mercury 6th->5th ~9:55, answer A = before 9:55) clamp the window; Q3 (MC Cancer->Leo ~10:30, answer A) confirms.

Wheel (end of session)

Wired SkyWheel to Affirmology_WheelAlignment_FINAL_v1.json (in WheelConcepts): render only included:true stars at normalized anchors with the 0..1 fill component; figure transforms (rotateDeg/flipX as negative scale/scale/dx/dy) in a 400x640 space scaled to the wedge; lines from linksEffective with brightness following fill; omitted stars stay in data (flip to enable later, no art/geometry change); figure opacity token = 1.0. Parity proven against the tuner's own render math (Product/WheelPreview/alignment_parity_check.png). Aries chain in day order: 41 Ari = First Light -> Hamal = The Forge -> Sheratan = The Stillness -> Mesarthim = The Night Tide. The painted figures existed ONLY on Higgsfield's ephemeral CDN and were RESCUED to R2 wheel/figures/. All 12 had baked-in tuning stars/numbers; removed via Higgsfield EDIT (not repaint, to preserve pixel-exact alignment); live on R2, originals in wheel/figures/original/, review sheet Product/WheelPreview/all_figures_clean_12up.png.

4) OPEN / NEXT (what a new chat should pick up)

1. Chant take 4 vs take 2 (the one live open decision at close). Canonical = take 2, template unchanged, take 4 (build-and-converge) waiting on Jeff's ear. If take 4 wins: "make 4 canonical and update the template" -> also update the runner's prompt template so all future chants inherit the arc.
2. Queue Sol's chant (POST /api/chant/run {person: solballard@gmail.com}) so both founders' chants can be reviewed before any member hears one. Not yet done at close.
3. Flip the Mirror drift-gate from log-only to ENFORCE. Blueprint is approved; the one-fresh-founder-script proof (script only, no Fish) + a craft-judge sample check were the planned validation before enforce. Confirm markers strip clean (no leak into audio) first.
4. Verify the 33 mirrors' LENGTHS before the Mirror emails go. Some early ones may predate the length-fix (25-29min overshoots); the last 3 are known-good (10-13min). CODE has the render manifest. Re-render only the long ones (pennies). This gates the Sunday-morning Mirror email send.
5. Launch timeline (endorsed): Mirror emails Sunday ~10-11am (gift before ask), beta-join texts Sunday early afternoon, app teaser now framed "unlocks in a few days, likely Tuesday, sooner if ready," WhatsApp group Mon/Tue/Wed. Monday = buffer for live conversations + tightening (flip gate, hone beds, founder testing). NOTE: this timeline is from 7/12 and is almost certainly SUPERSEDED - the beta is now gated until launch (~7/26-8/1) and market launch is Aug 11 per memory. Treat these dates as historical.
6. Site intake form deploy - backend is ready (item A won't drop a field); waiting on Cowork/CODE wrangler deploy.
7. App build - pronunciation screens + wheel alignment are committed; ride the next TestFlight build on Jeff's "build it."
8. Spec doc correction - fold the ~23h56m ascendant-return fix into the pronunciation/identity time-invariance spec.
9. Beds - buy a small set of loopable/ambient beds by mood (gentle/calm/warm/charged) with first incoming money; these unlock the whole stretchable shelf cheaply. Jeff to hand a list of stretchable audios to wire.
10. Optional: a to-the-minute pin inside Olga's 9:20-9:55 band via an oracle run through the agent (not required).

5) FILES and REFERENCES

Docs (AFFIRMOLOGY root, all verified present): - Affirmology_QC_Audit_and_Plan_v1.md - the audit + drift-gate plan + build status. - Affirmology_NamePronunciation_Spec_v1.md - pronunciation spec. - Affirmology_IntakeForm_and_Identity_Spec_v1.md - identity/intake spec. - Affirmology_AudioNaming_Registry_v1.md - locked names, two-register philosophy, unplaced palette, retired names. - Affirmology_AppExperience_MasterPlan_v1.md - Explore tab / Wheel master plan (build read sections 1,2,4,5,15.4,15.8b,16.8). - Affirmology_FirstReading_Structure_v1.md - the Mirror structure + 12 rules (source for the blueprint). - Affirmology_Olga_Rectification_v1.md - Olga birth-time range. - CLAUDE OUTPUTS/Affirmology/Affirmology_NamePronunciation_UI_v1.html - the pronunciation UI/UX reference the RN screens matched. - CLAUDE OUTPUTS/Affirmology/WheelConcepts/ - the 12 zodiac + Sun wheel art and Affirmology_WheelAlignment_FINAL_v1.json. - Coordination: _CHAT_SYNC.md (minute-to-minute), _SHARED_CONTEXT.md (durable shared brain). The naming registry + two-register system were posted to both, addressed to the Cowork app-experience chat to fold into the master plan.

Code (in affirmology-studio cloud repo unless noted; commits are from this 7/12 session): - api/identity.py - sung-name deduction (C) + name review queue (D). - pacing.py - word_band single source of truth for bed-aware length; existing pacing_gate/negation_gate. - structures.py (~lines 62-63) - free-text directive/overview (the drift risk the blueprint replaces). - compose() - gate-and-revise loop (Apollo/Orpheus); conformance gate slots in via additive structure_id. - The conformance engine module + Mirror/Origin blueprints + pytest (16/16), commit fd46119. - Self-heal orchestrator + SELF_HEAL_MAX_SPEND_USD cap, commit c98818f. - Existing gates in src/affirmology/: audio_qc.py, chart_guard.py, script_qc.py; orphaned verify_send.py. - App (affirmology-app / Maya): pronunciation screens + SkyWheel, commit 5f762fc and earlier. - Cloud deploy commit b99efbd (later health=ok).

Endpoints: - POST /api/app/transcribe (voice notes, transcript-only). - POST /api/chant/run {person:<email>}; chant queue is self-draining cloud-side. - /api/name-review/verdict and related (review queue). - Studio base (headless): https://affirmology-studio-api.onrender.com, Bearer STUDIO_AUTH_TOKEN / CLOUD_STUDIO_AUTH_TOKEN. Hermes: POST /api/hermes. Health: GET /api/health.

Secrets / keys - THEY EXIST, values NOT printed here: - A NEW full-access ELEVENLABS_API_KEY in affirmology-studio/.env (does TTS + Music/chant + Speech-to-Text). Old key is superseded. - STUDIO_AUTH_TOKEN / CLOUD_STUDIO_AUTH_TOKEN, EXPO_TOKEN (permanently stored per memory), Resend key. - No OpenAI key with credits exists (checked in-session). - Anthropic has no live balance API; billing/top-up: https://console.anthropic.com/settings/billing

Media URLs (public): canonical chant https://media.affirmology.ai/chants/<person>.mp3; takes at .../chants/<person>/takeN.mp3.

NOTE ON CONTEXT NOT VISIBLE HERE

This recovery is built ONLY from this one DATA/backend transcript. It cannot see the PARALLEL CODE/apps chat that rendered the 33 mirrors and built the RN app screens, nor the Cowork app-experience chat that owns the master plan and email/text copy - their side of these handoffs lived in _CHAT_SYNC.md and _SHARED_CONTEXT.md, which a new chat MUST read for the current cross-chat truth. It also predates later sessions: the 7/12 launch timeline here is superseded by the current gated-beta / Aug 11 market-launch plan in memory, and any commits marked "push held" or "awaiting build" may have shipped afterward. Verify live state via GET /api/health, the current _CHAT_SYNC.md date block, and Affirmology_PROJECT_STATE.md before acting on anything dated.