START HERE

Prayer delivery offer engine

prayer/README.md

Prayer delivery offer engine

A buyer writes a prayer in a five-step survey, pays, and receives a photograph of their own prayer card held up in the hand at a famous holy site, with their name and their words legible on it. One engine, many brand fronts. v1 is St. Peter's Basilica, and the checkout bump is a second location — the same card in the crypt beneath it.

Twelve funnel surfaces plus eight supporting routes. The payment rail is Stripe under Numina House, with Shaw's entity as merchant of record. Fulfilment is a donor-conditioned two-pass image pipeline over a card we print ourselves: the buyer's words are typeset by @prayer-engine/card before either pass runs, and Pass 2 only holds the finished card up. A six-check vision gate over the card text, plus the realism rubric in REALISM-RULES.md. Delivery is a tokenised portal, and every email links into it.

Status

2026-09-30 (wave 2 — eight ads, built from the recon) — 8 statics, gated, NOT uploaded. prayer/ads/WAVE-2/ads/ad-01..08.jpg (+ a 154px thumbnail each), copy in ads/WAVE-2/ADS-eight.md (prayer_gate.py = ALL CHECKS PASS), run record in ads/WAVE-2/RUN.md. Every frame now uses an approved asset — the wave-1 defect was invented scenery and a stranger as the carrier. Pass-1 scenes come from the repo's own generator, which runs the donor approval gate; cards from services/cardlock (real print stamped in, alignment 0.9966–0.9998); Fr. Marco from the identity lock in persona/fr-marco/sheet/. A real defect was caught mid-run: the first Fr. Marco pass had the model invent the card text (one card read the Hail Mary, another a generic petition) — all four card frames were remade through card-lock. Built against competitors/INTEL-2026-09-30.md; the side-by-side with the category is on the competitor board. Primary text 615–890 chars, deliberately longer than wave 1.

2026-09-30 (five image ads rendered, wave 1 — superseded) — Five 1080x1350 statics, built and gated, NOT uploaded. prayer/ads/WAVE-1/run-20260930-image/RUN-20260930-st-peters-five-statics/final-ads/ (ad-01..ad-05.jpg, plus a 154px thumbnail of each for the feed-legibility read); copy in ../ADS-five-statics.md, prayer_gate.py = ALL CHECKS PASS. Three frames carry the card in hand through services/cardlock — the real printed card on a real hand plate, Gemini blending only light and colour, alignment verified at scale 0.997 / shift <2.6px, $0.276 a frame — which is what makes the buyer's name crisp and the card identical across frames. Two are type-only over the place. Type is composited deterministically (prayer-ad-forge/scripts/composite_prayer_static.py, Cormorant Garamond + the brand's contrast floor), never generated; it took four passes to get the scrim and the type placement right and each failure is now a comment in the file. Every frame generated with openai/gpt-5.4-image-2 per Shaw's directive, pinned as invariant B0 with one documented exception (the card-lock blend/grade stays on Gemini 3 Pro because gpt54 moves the card ~1.2x and the geometry is the product). Not yet run before upload: img_gates.py, REALISM-RULES, the §7 pre-flight. Full record: that folder's RUN.md.

2026-09-30 (ad pipeline built) — the ad-creative pipeline exists. prayer-ad-forge (~/.claude/skills/prayer-ad-forge/): a router that owns the prayer binding and calls the four forges for engines, carrying references/CREATIVE-INVARIANTS.md (the laws, including the five forge laws that are inverted or superseded for prayer), scripts/prayer_gate.py (the prayer floor: candle, card-on-a-surface, production-method claims, outcome claims, the verbatim boundary line, aggregate counts, reviews in paid, price in a frame, ordination claims, Devine Rev characters) and intel/SKELETONS/README.md (the direct lane mapped from adlib-sweep, the parallel lane seeded). Supporting: the offer card ~/.claude/skills/script-forge/offer-cards/st-peters.md, the style module styles/prayer-broll.md, the results-ledger seed, and st-peters wired into qa_gates.py. Wave-one copy passes both gates: prayer/ads/WAVE-1/SCRIPT-OPTIONS-prayer-broll-v1.md, four silent B-roll options. Fr. Marco may appear in prayer ads, face included (Shaw) — a brand-scoped inversion, not to be propagated. No reviews in paid creative (Shaw). prayer_gate.py earned its keep on day one: it caught four live candle references in the frame briefs (AD7 "candlelit", AD1 a votive macro) and PROMPT-PACKS §10 was corrected to the basilica's own lamps the same session. Nothing is rendered. Plan and 20+ concepts: AD-PIPELINE-PLAN-2026-09-30.md.

2026-09-30 (ad-launch decisions) — Shaw's four ad-launch calls, applied to the docs the same day. The candle is out — "we are not doing a candle anymore remove" — which closes the highest-severity open risk in the whole brand build and the only one that gated the ad set (it could have turned the brand's own honesty positioning into the weapon used against it; the one first-hand St. Peter's account in the evidence pack reports no candles available). The copy was already candle-free; the claim is now recorded as closed in brand-context-st-peters/README.md, 09-advertising-compliance.md §7/§9, and COPY-CREATIVES.md §10. Spanish dropped, USA only — the three -ES primaries are deleted; the segment was orphaned across all nine stages per Stage 8 §7. Budget $50 CAD/day. The personal-attributes openers in Ads 1 and 3 are left as they are by decision ("leave it") and remain the highest-risk assets in the kit. Ad pipeline plan and idea board: AD-PIPELINE-PLAN-2026-09-30.md. Note the currency trap: the account bills CAD and the scale-and-kill thresholds ($33.40 / $35.10) are USD.

2026-09-30 — LIVE on Stripe live keys (Numina House), live gate passed. One real purchase ($44.87, front + Tomb bump, card ending 6687) and one full refund on production: webhook 200 for payment_intent.succeeded and charge.refunded, psp_refund recorded, CAPI Purchase sent live (event id = order id). Before that, three Stripe test-mode orders through the live pipeline (bump, Speed, OTO1, OTO2, OTO3, membership, /replace): every photo through card-lock passed QA, the member's first photo arrived ~15 min after signup carrying the September default prayer.

Fixed and deployed today: Meta CAPI never sent its access token (every server event was a 400 since the code was written); browser/server Purchase dedupe (eventID as the 4th fbq argument); CAPI kept alive with waitUntil and given _fbp/_fbc; the delivery email (it carried one photo of four, a raw audio URL, a blank date, a dead membership button and no portal link — now every photo, a "Listen to it here" link, the date, the membership and portal links); follow-up delivery emails for photos made after the first email (OTO1, late OTO3, /replace); OTO1 now asks "Who is this one for?" and prints that name; the checkout date follows the Speed box; confirmation subject shows ORDER_REF (8 chars), not a UUID; the member email's SEASONAL_NOTE and portal link. OTO3 is now Shaw's chosen voice (ElevenLabs s2wvuS7SwITYg8dqsJdn, multilingual v2, one take) mixed to a "home mic" sound by a new POST /audio on the card-lock container; the /oto/3 sample is the approved "D" take.

Meta: pixel/dataset 1555593462990623 "Holy Cards of St. Peter's" in the Numina House BM, connected to ad account 1092968693202859; domain verified (DNS TXT); CAPI token in pass companies/9566-0866/meta-capi-st-peters. Production D1 test data cleared (backup prayer-engine/.backups/d1-before-golive-20260930-1221.sql). Versions: worker 48c1100c, web ed16a13e; engine commit 8cd9f32 (held-card-pivot = master). No mailing address in the email footer (Shaw).

2026-09-29 (latest) — OTO4 Your Card by Post ($19.90, postage included) and OTO3 Voice of the Sanctuary at $24.90: CODE-COMPLETE, not deployed. Shaw: "a proper OTO that we can set up for print and mail" and "For 5, let's change the pricing to that". /oto/3 → /oto/4 → /thanks; the posted card is a print of the same card file the photograph uses, rendered by a mail_print job to R2 and fulfilled through a MailVendor adapter (manual: /admin/mail CSV export and Mark sent). New AOV $47.56, net before ads $41.67, contribution $43.86 (ECONOMICS §2). Spec: specs/SPEC-OTO4-MAIL-AND-OTO3-PRICE-2026-09-29.md. Before deploy: remote D1 migration 0006_oto4_mail_orders (HANDOFF open work).

Update, same day: LIVE. Deployed and live-verified on holycardsofsaintpeters.com: web cc336df5, worker edd3cdd9 with CARDLOCK_MODE = container and the Cloudflare Container application prayer-worker-cardlockcontainer (4 x basic). Two Stripe test-mode orders ran end to end (survey, checkout, OTOs declined, worker, container, R2). The first exposed an old bug: the worker printed the buyer's name, not the person prayed for; fixed in 0dbc4be (cardName()), and order two delivered "Elizabeth" correctly. Per photo: scene $0.19 + container $0.28 + QA $0.02, about $0.49; about 2 minutes of scene and 47 s of container. Deploy note: run wrangler with a clean XDG_CONFIG_HOME (a stale OAuth login on disk overrides the token's containers:write), and the deploy box needs Docker.

2026-09-29 (later) — The card is stamped, not drawn, and it takes 320 characters. Shaw, after seven review rounds on the board prayer-card-uniformity: "the card should look standard on every image", then "Fantastic i think we have a winner". Audit and every decision: AUDIT-CARD-UNIFORMITY-2026-09-29.md.

Card-lock (prayer-engine/services/cardlock, config brands/st-peters/cardlock.json): the printed card is stamped into a fixed rectangle on one of five green-screen grip plates (tilted with the card), Gemini 3 Pro blends light and colour, the print is stamped back on, Gemini grades, the print is stamped back on again; 1080×1440 JPEG q70. No model ever draws the card or the thumb on it. Cap 220 → 320 characters at a 54px body (copy-only; no price or SKU change). All 25 site card photos rebuilt (7 examples, 18 review panels; apps/worker/scripts/cardlock-site.py), published into the working tree with WebPs. Live orders: a Cloudflare Container runs the same Python (CARDLOCK_MODE, default legacy), live-verified in local Docker only.

Gates: pnpm typecheck 0 · 968 unit tests · 9 cardlock tests · survey 320/321 e2e green. Not deployed, nothing committed (engine branch held-card-pivot). Deploy needs Docker on the box and Workers Paid; container switch-over is a second deploy.

2026-09-29 — The review wall is one section, and its avatars are people, not cards. Shaw, three points in one pass: "the avatar is someone holding a card, which makes no sense… it looks like the clients are sending us the card, and then we're going to St. Peter's"; "let's just have one section for reviews… they can click on the next button to scroll to the next image"; and "the reviews need to be standardized at all times."

He is right about the structure, and the imagery had it backwards. The 2026-09-28 sweep had put each person's own printed card into their avatar, which made the wall read as customers photographing a card and posting it to us. Nobody posts anything: the prayer is typed at home, @prayer-engine/card prints it in Rome, Fr. Marco carries it. The card is now out of every avatar for good. avatar_instruction in apps/worker/scripts/review-set.py forbids any object in either hand and generates with no reference image at all — the card reference was the mechanism putting a card in the frame — and all 18 avatars regenerated as candid phone photos of the person at home.

One section, standardized. Three bands became one row of eighteen, walked with Next, which is the affordance Shaw asked for; the arrows now render at every width (under the row on a phone, in the gutters from 1024px) instead of only on a desktop. components/ReviewCard.tsx and lib/reviews.ts are the single statement of what a review is, and /offer, /create/ready, /membership, /thanks and /oto/1 take the quote, the stars, the face and the caption under their example photographs from the same source by key. That closed a live mismatch, not a hypothetical one: /create/ready captioned a photograph of Denise's card as "Maria's", and /oto/1 showed Denise's card under Maria's name. A caption is now derived from the key the photograph belongs to, so the name under a card cannot disagree with the card. useRotatingReviews is deleted: nothing rotates, and there is nothing to rotate.

Gates: pnpm typecheck 0 · 953 unit tests · Playwright 29/29 (one slider, eighteen cards, every avatar an avatar-* file and none a panel, and the Next button visible and moving the row at 390px). Not deployed — the assets and the code are in the working tree, and regenerating an asset and shipping it are two steps.

2026-09-28 — The card is printed, not handwritten — and it is a print we can mail. Shaw's directive: "I don't think handwritten matters… It's going to be a printed card… we can generate the card first, then we can superimpose the card through an LLM process with the hands and the background whatever it is. It's always going to be uniform in every way."

New package prayer-engine/packages/card: a deterministic renderer that typesets the buyer's first name and prayer onto our card — 105 x 148 mm (A6), 1240 x 1748 px at 300 dpi, one geometry in layout.json, one typeface (EB Garamond), one printed maroon header and rule, a footer emblem, a fit ladder so the whole 220-character cap is legible at one size. No model touches the text. The worker renders it before Pass 2 and attaches the same bytes at both locations, so the front end and the bump cannot disagree about the card and nothing is handed between them.

Three defects it closes, all of them delivered: the bump used to write the prayer a second time in a different hand ("it looks like the person wrote two different cards", 2026-09-21); three renders of the hero set drew three different hands ("some of the images look like its different hands", 2026-09-25); and a vision read-back that transcribed the printed header as the card's first line failed all seven rows of a paid hero run (2026-09-25). All three existed because something on the card was being generated.

What it unlocks: the physical card. Exact trim, exact margins, 300 dpi — the identical file can go to a print house, which turns the print-and-mail thread open since 2026-09-26 into a fulfilment integration rather than a research problem.

A card a buyer can be handed. The intake caps the first name at 24 characters, the name wraps to two lines before it shrinks, and a character the name's glyph tier lacks falls back to the full-repertoire body tier — so no first name can dead-letter a paid order, and none is ever silently shortened. The prayer is stricter: an unprintable character in the prayer throws, because a card carrying a different prayer from the one she wrote is worse than a failed job.

How it is built, and the two traps. The Worker has no font rasteriser and a WASM one costs about 1 MB gzipped on the critical path of every paid order, so every glyph the card can print is rasterised once, at build time, into a deflated coverage-mask atlas; the runtime only measures advance widths and blits. Two things bit. Fractional glyph advances silently discard a typed-array write — no throw, no warning — so an unrounded pen draws the first glyph of every line and nothing after it, which looks exactly like a broken atlas. And a character the typeface lacks must fail loudly: dropping it would print a different prayer from the one the buyer wrote and then fail a correct card on the QA word-match, so the renderer folds to NFC, substitutes ? to keep the word's place, and throws UnsupportedCardCharacterError.

Swept with it: BrandConfig.location.cardTemplate, the [CARD_TEMPLATE] pack block and the unconsumed ## CARD VARIANTS section are gone; [CARD_REFERENCE] now renders at every location; {FIRST_NAME}, {PRAYER_TEXT} and the name-recovery spelling trick are gone from Pass 2, replaced by varyAxesForRetry; brands/<slug>/assets/card/ is deleted; the copy reads "printed" on every customer surface, in all seventeen emails and in COPY-CREATIVES.md; and the survey's card preview now draws the real printed card instead of a plain index card.

Every card in the app was then re-rendered on the printed card. Changing the pipeline does not change the assets, and the assets are what a visitor sees, so the imagery was swept separately: the 7 example deliverables re-run through the real two-pass pipeline (all seven passed QA on attempt one, ~$0.45), the 36 review assets (18 panels + 18 avatars) recomposed against each person's own rendered card — every avatar had been holding a blank white card, which is what Shaw caught — and the last two brand stills behind a new rig, brand-card-stills.py. journey/mothers-cards.jpg is deliberately untouched: his mother's prayers on the backs of receipts are narrative, not product. WebP variants regenerated for all of it. Screenshots of the running funnel under apps/web/static/card-sweep-2026-09-28/.

2026-09-28 (later) — The card is uniform, the footer is back, and the delivered file is a fifth of the size. Shaw, after the first pass: "the name at the top is MASSIVE and various different sizes… needs to be completely uniform each time", and "reduce the quality of the picture and make it much smaller so people can't open it up in a new tab and review/scrutinize them."

Uniform. The name was on a ladder, so Maria rendered at 168px and a two-part name at 140. There is now one name size (120px) and one prayer size (66px), held by three rules: the name block reserves two lines of height whatever it wrapped to, so the prayer starts at the same y on every card; the name is never set in another size; and the glyph repertoire is one set, so an odd character cannot change the size. Measured and recorded in layout.json — the longest prayer the 220-character cap admits in ordinary English is six lines at 66px in a field that holds seven, and every realistic name fits two lines at 120px. Asserted by digest: the header and footer bands are pixel-identical across six cards with different names and prayers.

The footer emblem. Cards were losing it. The render always drew it; the prompt clause never mentioned it. It does now, and the emblem is 53% larger.

Smaller. New packages/imaging/src/deliver.ts: every frame is decoded, area-averaged to 960px and re-encoded as JPEG 74 before storage. Measured through the live Worker on real workerd — a real model render — 1.82 MB in, 88 KB out, printed name crisp at 1:1. QA runs on the original, before the shrink, so it can never fail a check; a codec failure stores the original rather than withholding a paid order. The site's own example imagery is published at that same size, so the funnel shows the file a buyer actually receives.

DEPLOYED 2026-09-28 — prayer-worker 93ab1319 and prayer-web 8c470794, plus the brand assets re-pushed to R2. Verified against the live origin by hash: hero_landing, ready_01, offer_02, panel-maria, avatar-maria, avatar-herbert and three WebP variants all serve the new files, and the live HTML reads "Your card is printed". apps/web/static/card-sweep-2026-09-28/LIVE-*.jpg are the deployed pages.

Gates: pnpm typecheck 0 · 697 unit tests · Playwright 28/28. Live-verified: the printed card renders correctly through a real model — the 2026-09-28 hero run passed all seven rows on the first attempt, and the review panels and avatars were generated against the real render. Still open: no bumped order (front end + crypt pair) has been through a live model with the printed card, and neither has a member delivery.

2026-09-27 (late) — The competitors' OTO ladder is mapped, and two of our lines are the only ones the market contradicts. research/COMPETITOR-OTO-MAP-2026-09-27.md: Holy Land Prayer fully read — bumps $14 / $29.97 / $9.97 on the live $34.90 form, offer pages at $49 (4-week) / $40 / $30, a gift line at $34.89–$44.89, a $20 gift card, video sold only on the legacy $19.95 form at $10/$15/$20, and their own service-to-price sheet leaked in a hidden dropdown on /promo/ ($89 / $44 / $25 / $24.99). Their .org Shopify surface sells the same front at $34.99 and the same four weeks at $140. Sacred Prayers' post-checkout steps are not reachable — Funnelish custom paths, 490-path sweep clean, API 401, no Wayback capture, and their email validator rejects every address, so no session can be minted; their SKU list came back from shop.yoursacredprayers.com instead, including a $24.90 audio product.

Two open calls, both Shaw's: OTO3 Voice of the Sanctuary at $9.90 is 2.5x under the only comparable ($24.90) and the curve is nearly flat — a one-line price test, but it moves a SPEC §3 price across the invariant rows. The membership has zero category validation — neither operator carries a single recurring SKU, so the 6%/4-month assumption in ECONOMICS §2 has no external read and wants testing early. Confirmed: the $34.90 front, second prayer at $14 (dead level), the no-pre-checked-bump rule (their own WordPress form doesn't pre-check either), and OTO2 rush as the one line nobody else sells.

2026-09-27 — The three content sections are LIVE and the backfill is complete. The Prayer Library, the wall and the answered prayers are serving on holycardsofsaintpeters.com, backfilled continuously from 2020-01-01: 14,602 wall prayers across 2,456 days, 43,361 comments, 2,363 answered prayers and 79 library pages. sitemap.xml carries 92 URLs and deliberately omits the wall, whose 2,434 pages are noindex — the reference site draws the same line, and DEEP-DIVE §0 is why.

The read that shaped it is research/DEEP-DIVE-HOLYLAND-2026-09-26.md; the scope is SPEC-CONTENT-SECTIONS-2026-09-26.md, whose §11 is the build record.

Eight bugs, none of which a green test suite could see. Four are the same failure — correct in the lane it was tested in, wrong on the platform: D1 rejects a multi-row insert over ~100 bound parameters and libsql does not (so CI was green while production wrote zero comments); a Workers-native fetch refuses to be called as a method of anything but the global; two overlapping cron expressions are not both delivered, so the walk had never started; and a chunk's days inherited the time of day the cursor was created at, putting a decimal point in public URLs. Two of the eight had the fix already written elsewhere in this repo — rows.ts documents the parameter cap and the imaging client carries the .bind(globalThis) with a comment explaining why. Two more made the run ledger lie, reporting created=0, refused=0, errors=[] for jobs that had spent money and failed.

Built: a new @prayer-engine/content package (taxonomy with a 79-page library registry, writer briefs, a 13-rule moderation gate, the distinctness guard, the schedule); six tables and two migrations; two worker ticks; a resumable walk with a write lease, a day-skip guard and a wall-clock budget; a backfill CLI with a real dry-run price, an emit-SQL fast path for the library, and a targeted day-repair path; fourteen web routes plus sitemap and robots; a footer Reading column; and a moderation queue at /admin/content. The D5 survey publish opt-in is built — unchecked, honoured at payment, landing as a draft a person passes.

The comment section is new, not copied — the reference has comment_status: closed on every prayer and zero comments site-wide. Ours is moderated and intercessory only: a comment never reports an outcome, the gate enforces it, and three distinct reader reports return one to the queue.

Gates: typecheck 0 across 11 packages · 675 unit tests · D1 lane 16/16 in real workerd · next build green · every new route verified live by curl.

The short days were diagnosed, not patched. 190 of 2,462 days came back with four or five prayers instead of six. The cause was not the model and not the gate: every category carried exactly six situations, so the sixth request produced a rephrasing the distinctness guard correctly refused, and the top-up round re-sent the identical request and got the identical answer. Fixed in four measured steps — a dropped round retries, a top-up names what is already written, the situation pools widen from 6 to 7–8, and a top-up may draw on a shared palette as a last resort. Six previously-stubborn days measured 6/6/6/6/6/6.

The repair pass rebuilt 157 of the 190 short days. Final distribution over 2,462 days: 2,429 at six prayers (98.7%), 21 at five, 12 at four — with zero gaps, zero duplicate slugs, zero duplicate bodies and zero over-full days.

Open: the 33 days the repair refused are not scheduled for another attempt; no Playwright coverage of the new routes; Header mode="content" is specified and not built.

2026-09-26 — The three content sections are built: the Prayer Library, the wall, and the answered prayers. Modelled on holylandprayer.com, which was read page by page and through its WordPress REST API for research/DEEP-DIVE-HOLYLAND-2026-09-26.md. The scope is SPEC-CONTENT-SECTIONS-2026-09-26.md.

The finding that shaped it: their three sections are three different products. The Library (/prayer_for/, 78 pages) is index and is the real SEO engine. The wall (/prayer-wall/, ~71,000 prayers) is noindex — a liveness surface that buys nothing in search. The Answered hub and its five category pages are index; the individual entries are not. Their answered archive also stops dead for 2021 and 2022, which is the discontinuity our backfill exists to avoid.

Built (engine, 660 unit tests green, typecheck clean, next build green): six tables and a migration; a ContentStore over D1/libsql and in memory; a new @prayer-engine/content package with the taxonomy, a 79-page library registry, the writer briefs, an offline fixture and a model client, and a 13-rule moderation gate with table-driven tests; two worker ticks — content_daily at 06:20 and a resumable content_backfill every five minutes that walks backwards to 2020-01-01 and then costs nothing; a backfill CLI whose --dry-run prices the job; fourteen web routes plus sitemap.xml and robots.txt; a footer "Reading" column; and a moderation queue at /admin/content.

Verified: a real backfill against a throwaway libsql database wrote 79 library pages, 192 wall prayers, 576 comments and 60 answered prayers with cursors advancing exactly one whole day at a time and zero refusals.

The wall's comments are new, not copied — the reference has comment_status: closed on every prayer and zero comments site-wide. Ours is a moderated, admin-owned thread: generated comments are intercessory only and pass the gate before they are stored, and nothing is outcome-framed. That rule is what keeps a comment section from being an exhibit.

Open: the real backfill has not been run (dry-run cost: ≈61,700 rows, ≈6.2M output tokens, ≈$7.41 — it needs Shaw's go); the survey publish opt-in (D5) waits on Shaw and is the only funnel change; /api/content/report and Header mode="content" are specified and not built; no Playwright coverage of the new routes.

2026-09-25 (later) — Fr. Marco is a built persona, and the card is a branded asset. Two changes, both Shaw's directive.

The persona. Fr. Marco had a story (PERSONA-FR-MARCO.md) and no fixed face — fr-marco-master.jpg was a studio headshot that contradicted DESIGN-BRIEF §0. He is now built on Leg 0 of the Video Forge skill (the aimikoda single-composite method, GPT Image 2), the same way every other persona in the stack is built: 234 Pinterest candidates harvested from six realism-qualified queries → 43 through a vision screen → 8 finalists → one seed, a real phone selfie in an ordinary room (not a studio, not retouched) → a two-pass light identity mutation → 01_mutated_v2.png, the identity lock → sheet.png, the 11-section composite character sheet. It self-registered his full name, Fr. Marco Bellandi. The resemblance check is clear on a candid dim render; the two checks that failed did so on the other render of the other face, which is the documented pattern — a FAIL is a signal to re-light before it is a signal to recast. One asset, many processes: landing portrait, /offer letter, ads, and the pass-2 hand.

The card. The card was a plain white index card with nothing on it. It is now A4 portrait, printed with our own maroon lockup, wordmark and rule (CARD-SPEC.md), so it looks like it came from us. The design decision that matters: the card is attached as a reference image, never described in words — a model told to "add a logo" draws a different one every render. Front end gets card/template.png as its second image and reproduces the header; the bump gets the front end's finished photograph and copies the whole card, so it is still one card photographed twice. RENDERED on real models: front end and bump, same seed — identical printed header, identical handwriting, identical line breaks, in the basilica and in the crypt. Both prompts carry exactly one second-image clause and zero leftover tokens. Gates: pnpm typecheck 0, 591 unit tests, prompt-render check pass. The card changed REALISM-RULES R9 and PROMPT-PACKS §7; CARD-SPEC.md records why "no decoration ever" was superseded rather than violated.

Open: the Western Wall card and persona; the review slider (research/REF-TEARDOWN-2026-09-25.md §1) is specified but not built; the two remaining branded orders are still rendering.

2026-09-25 — DEPLOYED (prayer-web 579989ca): express pay, the short billing block, and the design pass Shaw asked for. Engine commits cf61381…36ad24b. Checkout: Apple Pay / Google Pay (Stripe Express Checkout Element, Link off; domain registered in test mode, NOT yet in live mode) and every card form cut to name, card, postal code and country; the card path is live-verified in Stripe test mode, the wallet path is NOT LIVE-VERIFIED (no wallet in the lane's Chrome; test on a phone). Design: a dome-in-arch logo (apps/web/public/assets/brand/logo/), EB Garamond headings with Inter for body and numbers, 14 licensed photographs of the basilica (apps/web/PHOTO-CREDITS.md), a compact footer with the Meta disclaimer (the positioning block now lives on /support and the legal pages), checkout trust marks and a 30-day seal, a photograph under every survey step, her own card on /create/ready, photographs and a four-stage timeline in the portal. A Fable cross-model review (verdict FIX) was applied. Gates: typecheck 0, 591 unit tests, Playwright 26/26. Live Lighthouse mobile on /: performance 89, accessibility 100, best practices 100, LCP 3.7 s, 1,087 KiB (heavier: photographs and a second face). 2026-09-23 — DEPLOYED and LIVE-VERIFIED: the copy pass and ten mobile-first UI iterations (prayer-web 5920244c, prayer-worker 1ddc7f2e; engine commits 5d32eb0…6e50896). Live mobile Lighthouse on /: performance 92 (was 76), accessibility 100, best practices 100, LCP 3.2 s (was 6.7 s), 377 KiB (was 1,697), zero console errors. Live phone walk in Stripe test mode: survey → checkout (Stripe card field with no second ZIP) → a declined card → address kept → 4242 on retry paid → OTO1 with its sticky Yes and decline → thanks → portal ("Prayer for Rosemary", "order #8ED23810"). A real defect was found and fixed on the way: a retry after a declined card re-charged the declined card, so a buyer whose first card failed could never pay (8674e69, pinned by a new Playwright test). Gates: typecheck 0, 587 unit tests, Playwright 26/26. Screens: review-site/static/ui-iter-2026-09-22/final-10/. Open: Shaw's copy calls (countdown, discount card, testimonial stubs, ticker), Apple Pay / Google Pay, the shorter billing block. 2026-09-22 (night) — Copy pass + five mobile-first UI iterations, CODE-COMPLETE — NOT LIVE-VERIFIED (not deployed). Commits 5d32eb0, d326528 (copy: no candle, no bells, "handwritten" replaces "printed" on every surface and email, COPY-CREATIVES in step, research/COPY-REVIEW-2026-09-22.md) and 1b25e84…7959a44 (UI: sticky phone CTA on / and /offer; OTO buttons before the photo; shorter checkout; phone-sized WebP stills and no background video on phones; AA contrast on grey text; autofill and keyboard hints; a country picker; faster survey steps; phone crops; post-purchase portal-first; desktop hero fits a 1366x768 laptop). Landing on a local production build, Lighthouse mobile: performance 76 → 95, accessibility 100, LCP 6.7 s → 2.9 s, 1,697 → 369 KiB. Gates: typecheck 0, 587 unit tests, Playwright 25/25. Screens: review-site/static/ui-iter-2026-09-22/final/. Open: deploy (prayer-web and prayer-worker, the emails changed). 2026-09-22 (evening) — The UI/UX pass is DEPLOYED (prayer-web version 5558551f). Live-verified on the public pages only: the phone landing page shows the new layout, with the CTA inside the first screen, 0 overflow and 0 console errors. The funnel steps behind the survey were not walked live. Every surface was captured at 390px and 1280px and judged against PrayerSong (research/UI-AUDIT-2026-09-22.md). All 4 defects and 15 layout items are fixed in af1b4b8 + 2c59482: on a phone the first action now sits above the fold on the landing page, the survey, checkout and the OTOs; the three OTOs share one template; the portal names the recipient, shows a short order number and offers a Download button. Typecheck exit 0, Playwright 25/25. Open: Shaw's review of the live pages, plus the ten copy/offer calls (C1–C10) in the audit. 2026-09-22 — The bump swap, the copy sweep and the checkout speed option all landed in code. The checkout no longer sells a wide crop of the front end's own frame: it sells The Apostle's Tomb, $9.97 — one more photograph, at a second location, the Vatican Grottoes — plus a second unchecked option, Speed, 24 hours, $9.00, which hides OTO2. The worker renders the pair from two Pass-1 frames and attaches the front end's finished photograph as a second reference image, so the card is copied, not written twice. Migration 0002_petite_ultron renames the column to bump_taken and rewrites the enums in place. The at-rest claim is gone from apps/web, packages/email and COPY-CREATIVES.md; the canon is the card is held up over the tomb of St. Peter, and photographed where it is held. The candle promise was removed from customer copy — the pack had already dropped the clause, because a votive cannot be lit at the papal altar and the real ones there are electric (R1); candles remain scenery in the photographs, never a promised deliverable. OTO2 was renamed Same-Day Placement ($19.90) to end its collision with the $9.00 checkout speed option. Verified: pnpm typecheck exit 0, 587 unit tests green, pnpm --filter db test:d1 green, email↔COPY parity green, Playwright 25/25 including the new OTO2-hiding gate test. Also fixed: a pre-existing red in packages/imaging/test/prompt-builder.test.ts that still asserted the old 80/150 card-distance bands. Open: the brand kit's unswept at-rest wording, the review-site rebuild, the Cloudflare R2 key rotation, and the UI pass (page-cro benchmarked against PrayerSong). 2026-09-21 — The delivered photograph changed shape, and so did the bump. The card is now photographed held up in the hand at the spot, not lying at rest, and the failing $4.99 "Wider View" bump — a second photograph of the same placement, created this morning and dead by evening — is replaced by a second location: the same card in the Vatican Grottoes, sold as The Apostle's Tomb, $9.97. Commits 9aec3b8, 9e3fcd0, 7436db7 on branch held-card-pivot in /home/sha/vibing/prayer-engine. Verified this run: pnpm typecheck exit 0, 577 unit tests across 48 files, and the new realism gate live-checked against three frames — the two Shaw approved pass all nine checks, the one he rejected fails on no_surface_invented, light_is_real and nothing_invented, naming the shipped defect ("numerous invented wax candles with drips on an invented foreground surface"). Two findings worth carrying: the production QA gate had no composition check at all — its six checks were text and identity — which is how a frame with an invented candle field shipped green; and packages/imaging/src/realism.ts is now the one home for the rubric, shared by the rigs and the gate. The docs are updated but NOT versioned — /home/sha/vibing/prayer is not a git repo. Open: the 64 at-rest lines in COPY-CREATIVES and 33 live apps/web surfaces, the bump's 111-reference code swap, six mock orders rendering for review, and git init for this directory.

2026-09-20 — Design pass shipped and LIVE. Commit fd54dd8 in /home/sha/vibing/prayer-engine: every surface and the seventeen emails follow DESIGN-BRIEF-2026-09-18.md, with the components/ui kit, seven example photos from the real pipeline, persona stills, three muted video loops and the Voice of the Sanctuary sample MP3. Verified this run by Fable: typecheck 10/10, 554 unit tests, Playwright memory lane 24/24. Deployed prayer-web (version bf1f76b1) and prayer-worker (version b6d859db). LIVE-VERIFIED: https://holycardsofsaintpeters.com renders the new landing with zero console errors (screenshots review-site/static/design-2026-09-18/live-landing-*.png); hero photo, sample MP3 and video loop return 200. Later the same day (commit 3193796, prayer-web version 126bf6b2): one Stripe test order placed on the live redesigned funnel at 390px ($34.90 + OTO3 $9.90, both PaymentIntents succeeded); the OTO3 player LIVE-VERIFIED playing (currentTime 1.77 s → 9.28 s of 15 s). e2e:cf is green for the first time, 24/24, and the memory lane is 24/24. DEFECT FIXED: the Stripe test webhook was registered on prayer-web.srana.workers.dev, which answers 404 unknown_brand, so the 2026-09-17 order's payment_intent.succeeded was never delivered; the endpoint now points at https://holycardsofsaintpeters.com/api/psp/webhook (same signing secret) and both new charges show pending_webhooks=0. Known small: /favicon.ico is 404. Hosting rule (Shaw, 2026-09-20): nothing on the box, everything on Cloudflare. The review site moved to Cloudflare Pages: https://prayer-review.pages.dev. Two Vercel projects (prayer-review, prayer-funnel-preview) still exist and wait for Shaw's delete call. Open: hero caption says "Carmen" (COPY §1) while the H-1 card reads "Maria" (PROMPT-PACKS §5.1); favicon; Western Wall front; go-live switches; test-order cleanup in production D1.

2026-09-17 (night) — LIVE on Cloudflare, test mode. https://holycardsofsaintpeters.com serves the funnel from a Worker (OpenNext) with D1, R2 and a Queue; prayer-worker runs the five crons and the queue consumer. LIVE-VERIFIED end to end by Fable: one Stripe test-mode order placed in Chrome ($34.90, PaymentIntent succeeded, exactly one charge intent settled), confirmation email delivered from care@holycardsofsaintpeters.com (Resend domain verified), the photo job ran Pass 1 → Pass 2 on the Worker with real models and passed QA on the first attempt (deliverable in R2, prayer photo_ready, image at review-site/static/live-2026-09-17/), delivery email queued for +3 days. Commits 52305ec (D1), 992cfb3 (R2 + Worker), e642910 (OpenNext), then three live fixes (form read after await, charges missing the Stripe customer, checkout event name) and the portal link labels. Stripe test webhook registered on the web URL. Unit 554 tests, typecheck 10/10. Playwright: 24/24 memory, 23/24 SQLite (one navigation abort that does not reproduce alone), e2e:cf not yet green (box disk pressure). Open: Voice of the Sanctuary sample MP3 missing at /assets/audio (OTO3 player disabled), Western Wall domain (stonesofjerusalem.com is taken) + prompt pack, e2e:cf pass, Stripe live keys (Shaw flips), Meta pixel/CAPI values, dispute deflection toggle (Shaw's yes/no).

2026-09-16 (evening) — Second commit d5398a2 in /home/sha/vibing/prayer-engine: verify-pass defects N1–N10 fixed, confirmation-email job added, Pass 2 moved to GPT 5.4 image 2 with the plain-card handwriting pack, QA checks 3–5 made frame-relative. LIVE-VERIFIED: the full pipeline passed the QA gate on the first attempt on donor #11 (gallery run 8). Playwright 24/24 on both stores, 312 unit tests, typecheck clean. Open: donor choice #11 vs #14 (board d3), Stripe adapter, Cloudflare port, Western Wall prompt pack. Stripe adapter (Numina House account) is now CODE-COMPLETE and live-verified in test mode, 17/17; webhook registration waits for the public URL. The fee model in ECONOMICS.md and SPEC §3 is now recomputed on Stripe Canada pricing (5.7% + $0.22 effective, no fixed platform fee, no reserve), replacing the retired TagadaPay figures.

2026-09-16 — Phase C build is CODE-COMPLETE and committed (initial commit in /home/sha/vibing/prayer-engine, branch master). LIVE-VERIFIED this run by Fable: typecheck clean across 9 packages; unit tests engine 57 / db 38 / payments 58 / brands 11 / email 42 / imaging 31 / audio 10 / web 5 / worker 24; Playwright 22/22 on the in-memory store and 22/22 on live Postgres 16; worker drain runs all five crons. Code-review blockers B1–B11 and fixes F1–F15 applied. Image, vision and TTS run through OpenRouter (live-verified: one demo order produced a legible card photo and a voice track). Direction change 2026-09-16 (DECISIONS addendum): payments move to Stripe under Numina House (Shaw sets up the account; adapter boundary kept), hosting moves to Cloudflare (Workers, D1, R2; the local Postgres is a test harness only). Open: the St. Peter's donor photograph (board prayer-donor-st-peters), the QA flame-count rule (delta cannot pass with a candle rack in frame; proposed: one new votive beside the card, foreground only), the Stripe adapter, the Cloudflare port. Preview (mock rail): https://prayer-funnel-preview.vercel.app

2026-09-15 — Phase C build is CODE-COMPLETE; core paths LIVE-VERIFIED locally. Repo /home/sha/vibing/prayer-engine/ (no commits yet). Verified by Fable this run: typecheck clean; tests engine 57 / db 37 / payments 43 / brands 11 / email 42 / imaging 27 / audio 7 / worker 20; Playwright 17/17 on the mock store and 17/17 on a live Postgres 16 with the migration applied (11 tables); worker drained 108 real job rows on fake models to 110 deliverables, 0 dead letters. Not live: TagadaPay (adapter typed, untested against the real API), real image/vision/TTS models, Resend sends, deploy. Reasoner code review in research/CODE-REVIEW-2026-09-15.md once it lands. Phase B (Shaw) still gates the live rail, models, and launch. 2026-09-14 (late) — Phase C started. Foundation at /home/sha/vibing/prayer-engine/ (monorepo, schema, engine, payments mock + Tagada adapter, 2 brands; 146 tests, typecheck clean, brand resolution live-verified on a local server). In progress: funnel pages, fulfilment worker, email package. Then portal + dashboard, then Postgres live check, then code review.

2026-09-14 — Phase A complete and reviewed. The build docs are rewritten to Shaw's FINAL ANSWERS, then twice reviewed fresh-context: 6 blockers and 22 fixes applied, then 8 more findings from a second verify pass. Phase A is DOCS-COMPLETE. No app code exists yet, so nothing is CODE-COMPLETE and nothing is LIVE-VERIFIED.

The three bullets that stood here were the 2026-09-14 status and are all superseded — kept only so the change is legible. Phase B is done (Stripe under Numina House, descriptor NUMINA* HOLYCARDS, the rest tracked in HANDOFF.md). Phase C is built and live. The review site moved to https://prayer-review.pages.dev — the two Vercel URLs below are dead.

The package, in read order

  1. README.md — this file. The one-screen map and the status line.
  2. HANDOFF.md — the standing handoff: decisions already made, open work, rules of engagement. Read before changing any doc.
  3. DECISIONS-2026-09-14.md — Shaw's dictated decisions. The FINAL ANSWERS block at the top overrides everything below it.
  4. PLAN-PRAYER-ENGINE-v2.md — the single build plan: the offer, what to port from Divine Rev, the part-2 lineup, the A/B/C/D phase sequence.
  5. SPEC-PRAYER-ENGINE.md — the build spec. BrandConfig, the thirteen surfaces plus nine supporting routes, the pricing ladder and SKU names, the image pipeline, the audio product, the monthly engine, seventeen emails, the one scale-and-kill rule, launch sequence, expansion playbook.
  6. OPS-RUNBOOK.md — infra. Repo layout and the byte-identical BrandConfig, the route table with the ph and resume-token specs, DB schema, Stripe wiring and the dunning ladder, the fulfilment worker and its QA gate, Resend templates, droplet deploy, the event map, /admin, the day 0–13 checklist.
  7. PROMPT-PACKS.md — the production prompt asset. Donor spec per location, the randomization axes, the Pass-1 and Pass-2 recipes for both locations a brand ships (the front end and the bump), the QA contract, seven hero deliverables per brand, six location packs, card art, persona portraits, the monthly rotation, ad stills, ambience presets.
  8. COPY-CREATIVES.md — finished launch copy for every surface: landing with the live ticker, survey, transition, sales page, checkout, three OTOs, thank-you, portal, /membership, /replace, seventeen emails, five English Meta ads (the three -ES primaries were dropped 2026-09-30 — the launch is USA only), a 30-line ticker pool, six testimonial stubs, the positioning block.
  9. ECONOMICS.md — the money model: unit economics, spend curves, nine brand fronts, portfolio cases, what caps this.
  10. REALISM-RULES.md — the rules every generated frame is judged against. Read before touching any prompt, pack or QA gate. R2/R2a/R2b were inverted on 2026-09-21 for the held card, and R9 was revised on 2026-09-25 for the branded card; the file records why in both cases.
  11. CARD-SPEC.md — the branded prayer card: A4, the printed header, the layout invariants, how the pipeline attaches it, and what is deliberately not printed on it.
  12. research/ITEM4-SCOPE-2026-09-21.md — the decisions layer for the second-location bump.
  13. research/OFFER-SURFACE-INVENTORY-2026-09-21.md — every file:line still carrying the old bump or the false at-rest claim, with the ordering hazards that sequence the code refactor.
  14. specs/SPEC-BUMP-SECOND-LOCATION-2026-09-21.md — the code spec for the bump swap.
  15. PERSONA-FR-MARCO.md — the persona: face, story, name, and the identity lock and character sheet built 2026-09-25. He is named in full on every surface.
  16. PRAYERSONG-MAPPING.md — the target offer structure this funnel is mapped against.
  17. research/REF-TEARDOWN-2026-09-25.md — the three competitor pages taken apart, including the review-slider component spec. Every item marked ADOPT / ADAPT / DO NOT ADOPT.
  18. persona/fr-marco/ — the persona's build: the donor harvest and its manifest, the numbered contact sheets the screen was run on, the approved identity lock, the 11-section character sheet, and render-persona-stills.py.
  19. research/PHASE-A-REVIEW-2026-09-14.md — the fresh-context reviews whose findings are applied. Read it to learn why a rule exists before changing one.
  20. research/TAGADAPAY-DIGEST.md — rail facts, signed terms, and the five questions still unanswered in writing.
  21. research/DIVINE-REV-REUSE-DIGEST.md — what to port from Divine Rev rather than rebuild.
  22. research/RESEARCH-PASS-1-2026-09-14.md — the part-2 segment lineup.
  23. research/COMPETITOR-OTO-MAP-2026-09-27.md — the competitors' upsell ladder and every price behind it: Holy Land Prayer's bumps, offer pages and hidden price dropdown, their Shopify SKU set, and the Sacred Prayers SKU list recovered from the store on their own root domain. Pairs with 17 — that one takes apart their pages, this one takes apart their ladder.
  24. RESEARCH-DIRECTIVE.md — the open multi-religion research directive and its 6-axis rubric. It locks the part-2 lineup beyond Wave 1 and does not block the build.

Background only

MASTER.md and REVIEW.md (the 2026-09-12 deep dive), plus research/adlib-sweep.md, research/funnel-walk.md, research/rogovy-case.md, research/sacred-prayers-verbatim-copy.md and research/web-research.md. Read them for context. Do not carry MASTER.md's compliance framing forward into the build docs.

Also on disk: brand-context-st-peters/ (the brand context build) and review-site/ (the source of the review site above).