START HERE

HANDOFF — Prayer Delivery Offer Engine

prayer/HANDOFF.md

HANDOFF — Prayer Delivery Offer Engine

Box: sha@192.168.0.83 (or S: share) — Folder: /home/sha/vibing/prayer/ Date: 2026-09-14 — Phase A complete. Build docs rewritten to the FINAL ANSWERS, reviewed fresh-context by a reasoner, and the review's 6 blockers and 22 fixes applied. Status vocabulary: Phase A is DOCS-COMPLETE. No app code exists. Nothing is CODE-COMPLETE. Nothing is LIVE-VERIFIED.

What this is

A prayer-delivery funnel engine. The 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.

Four things define the shape:

  • A PrayerSong-shaped funnel. Thirteen surfaces: landing, 5-step survey, transition, sales page, checkout with one bump, four one-tap OTOs, thank-you, portal, membership, re-placement. Funnel identity is one signed httpOnly cookie, ph, minted at survey step 5.
  • Stripe under Numina House, not TagadaPay. (Corrected 2026-09-21 — this line was three weeks stale and directly contradicted the Stripe sections further down.) Merchant of record is Shaw's entity, which carries full chargeback liability. We own every checkout page; the rail tokenizes, charges and runs subscriptions. The synchronous settle path is authoritative and the webhook only reconciles. Statement descriptor NUMINA* HOLYCARDS.
  • A donor-conditioned two-pass image pipeline. Pass 1 regenerates the site from one approved Pinterest donor photograph per location. Pass 2 is an edit that puts a hand holding the buyer's card into the foreground. A vision-model read-back checks the card text before upload. There is no compositing code, no physical fulfilment, no fulfiller.
  • Delivery through a portal, not just an inbox. Every buyer gets /portal/<token> on a magic link: gallery with a 15-minute reveal tick, order timeline, intention field, membership manage, email preferences, one-tap offer surface. Email still carries the photo, so a portal outage never blocks a delivery.

One shared engine, many brand fronts. A brand is a config file plus a prompt pack, a donor image, seven example photos and a domain. Adding one requires zero app-code changes.

The files (read in this order)

  1. README.md — one-screen map and current status.
  2. HANDOFF.md — this file. The standing handoff. 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 and everything in the build docs that disagrees.
  4. PLAN-PRAYER-ENGINE-v2.md — the single build plan: the offer, what to port from Divine Rev, the part-2 lineup, and the A/B/C/D phase sequence.
  5. SPEC-PRAYER-ENGINE.md — the build spec. Product architecture and BrandConfig, the twelve-surface funnel plus supporting routes, the pricing ladder and SKU names, the image pipeline, the audio product, the monthly engine, seventeen emails, analytics and the one scale-and-kill rule, launch sequence, brand-expansion playbook, pre-first-dollar checklist.
  6. OPS-RUNBOOK.md — infra. Repo layout and the byte-identical BrandConfig, route table plus the ph handle spec, DB schema, TagadaPay wiring and the dunning ladder, the fulfilment worker and card-text QA, Resend templates, droplet deploy, the analytics event map, the /admin dashboard, the day 0–13 checklist.
  7. PROMPT-PACKS.md — the production prompt asset. Donor spec per location, six randomization axes, the Pass-1 and Pass-2 recipes (both locations per brand: the front end and the bump), the QA read-back contract, seven hero deliverables per brand, six location angle packs, card art, persona portraits, the 48-combination 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, the three OTOs, thank-you, portal microcopy, /membership, /replace, seventeen emails, five English Meta ads (the three -ES primaries were dropped 2026-09-30 — the launch is USA only), the 30-line ticker pool, six testimonial stubs, the positioning block.
  9. ECONOMICS.md — the money model: unit economics, spend curves, the nine brand fronts, portfolio cases, what caps this.
  10. REALISM-RULES.md — the rules every generated frame is judged against. Read this before touching any prompt, pack or QA gate. R2/R2a/R2b were INVERTED on 2026-09-21 for the held card; the file records why, so nobody re-introduces the invented-surface failure.
  11. research/ITEM4-SCOPE-2026-09-21.md — the decisions layer for the second-location bump: the new SKU identity, what item 4 has to reverse, and the open ends.
  12. research/OFFER-SURFACE-INVENTORY-2026-09-21.md — every file:line carrying the old bump SKU, the "one bump only" rule, rush, or the false at-rest claim, plus Appendix 1b (load-bearing vs cosmetic, with the ordering hazards that sequence the code refactor).
  13. specs/SPEC-BUMP-SECOND-LOCATION-2026-09-21.md — the code spec for the bump swap, in hazard order.
  14. PERSONA-FR-MARCO.md — the persona: face, story, name. He is named in full on every surface.
  15. PRAYERSONG-MAPPING.md — the target offer structure this funnel is mapped against.
  16. research/PHASE-A-REVIEW-2026-09-14.md — the fresh-context review whose blockers and fixes are now applied. Read it to understand why a rule exists before changing one.
  17. research/TAGADAPAY-DIGEST.md — rail facts, signed terms, and the five questions still unanswered in writing.
  18. research/DIVINE-REV-REUSE-DIGEST.md — what to port from Divine Rev rather than rebuild.
  19. research/RESEARCH-PASS-1-2026-09-14.md — the part-2 segment lineup.
  20. RESEARCH-DIRECTIVE.md — the open multi-religion research directive and its 6-axis rubric.

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

Decisions already made (do not relitigate)

These are the FINAL ANSWERS from DECISIONS-2026-09-14.md plus the settled architecture. They are closed.

  • Ladder. The Witnessed Carry $34.90 front end against a $69 anchor. The Apostle's Tomb $9.97 — the same card photographed in the Vatican Grottoes — and Speed, 24 hours $9.00, which hides OTO2. Two checkout options, both unchecked; the pre-checked ban is absolute. Second Prayer $14 at /oto/1, typed into a prefilled textarea. Same-Day Placement $19.90 at /oto/2. Voice of the Sanctuary $24.90 at /oto/3 (was $9.90; Shaw, 2026-09-29). Your Card by Post $19.90, postage included, at /oto/4, the last step before /thanks (Shaw, 2026-09-29). Anniversary Re-Placement $14 at /replace. Monthly Prayer Membership $9.90/mo or $79/yr. Western Wall front runs $29.90.
  • Prayer input caps at 220 characters, enforced client and server. There is no long-prayer path.
  • Five-step survey replaces the single lead form.
  • Delivery is 3 days standard, 24 hours on rush, both counted from orders.paid_at.
  • 30-day guarantee, one exact string, everywhere it appears: "100% Money Back Guarantee. Not satisfied? Get a full refund. No questions asked, no hassle. 30-day guarantee. Risk-free purchase."
  • No AI framing anywhere. No customer-facing asset says how the photograph is produced, in either direction.
  • No incentivised customer video and no $50 voucher. Reviews are asked for by email reply only. The portal has no upload surface.
  • One physical product: Your Card by Post (OTO4). Revised by Shaw, 2026-09-29: "I want to add a print and mail add-on since now the card is a print-ready file… a proper OTO that we can set up for print and mail." It posts a print of the same card file the photograph uses. Everything else delivered is a generated image, an MP3 or an email.
  • The card is printed, not handwritten. Shaw, 2026-09-28: "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… It's always going to be uniform in every way." @prayer-engine/card renders it — one geometry, one typeface, one printed header, 105 x 148 mm at 300 dpi, the buyer's own words typeset. It is deterministic, it is byte-identical across the front end and the bump, it costs nothing, and Pass 2 writes no text at all. CARD-SPEC.md is the authority. This is what makes a physical print-and-mail add-on possible at all, and it closes three delivered defects: two different hands on one card across the pair, three different hands across a hero set, and a mistranscribed name dead-lettering a paid order.
  • Payments are Stripe (Numina House). Card only for v1; wallets are off.
  • Fulfilment is image models only — GPT Image class and Nano Banana Pro class, conditioned on a per-location Pinterest donor reference. Card-in-hand is the default composition for every brand. The card itself is printed, not generated — see the 2026-09-28 decision below.
  • Delivery is through the portal, and every email links into it.
  • Support is email only. No phone number exists in any page, template, env var or descriptor.
  • Finite terms only on the recurring layer. No negative-option patterns. Cancelling is always one click and always easier than disputing.
  • Full infra isolation from every other brand: separate entity, payment store, Meta BM, domain, sending domain, droplet, and a fresh pixel.
  • Part 2 lineup: Wave 1 Judaism (Western Wall) plus Fatima, Wave 2 Buddhism, Wave 3 Hinduism, Wave 3b Makkah with a partner shape. Sikhism archived NO.

Build status (2026-09-15)

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. NOT DEPLOYED.

Shaw, three points in one pass: the avatar on every review 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"; the wall should be "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 second version of the imagery is what had it wrong. The first pass put a blank white card in every avatar (caught 2026-09-28); the fix composed each avatar against that person's own printed card, which made it worse: a photograph of a customer holding a card is a photograph of somebody posting a card. Nobody posts anything. They type a prayer at a kitchen table, @prayer-engine/card prints it in Rome, and Fr. Marco carries it. The card is out of every avatar for good, and avatar_instruction in apps/worker/scripts/review-set.py now forbids any object in the hands and generates with no reference image at all — the card reference was how a card kept getting into the frame.

Code: Testimonials renders one ReviewSlider holding all eighteen; components/ReviewCard.tsx and lib/reviews.ts are the single statement of what a review looks like, 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 rather than a hypothetical one: /create/ready was captioning a photograph of Denise's card as "Maria's" and /oto/1 was showing Denise's card under Maria's name, because the captions were typed by hand at each call site while the review data sat elsewhere. A caption is now derived from the key the photograph belongs to, so the name under a card cannot disagree with the card. useRotatingReviews deleted: nothing is rotated in or out, and with one row of eighteen there is nothing to rotate. The arrows exist at every width (under the row on a phone, in the gutters from 1024px), which is what "click on the next button" requires. Playwright rewritten: one slider, eighteen cards, every avatar an avatar-* file, and a phone-width test that the Next button is visible and moves the row.

Assets: 18 avatars regenerated (~$0.14-0.21 each), WebP variants rebuilt, and the deploy is a separate step — the lesson from 2026-09-28 that the site serves the origin, not the working tree.

Gates: pnpm typecheck 0 · 953 unit tests · Playwright 29/29.

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.

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.

2026-09-28 — the card is a print (Shaw's directive)

The change. The card used to be a branded A4 sheet attached as a blank reference, with the image model asked to hand-write the buyer's words on it in blue ballpoint. It is now rendered by code before Pass 2 runs and attached as a finished card at both locations.

Why. Shaw, and he was right on the substance: the handwriting neither mattered to the buyer nor survived contact with a generative model. Three separate defects had all been patched around it — the bump copying the front end's photograph to keep one card looking like one card, an A4 template attached to stop the logo being invented, and parseCardText stripping a printed header out of a vision read-back. Every one of those existed because something on the card was being generated.

What it costs. Nothing per order. A card render is local, deterministic and free.

What it unlocks. A physical card. 300 dpi at an exact trim with exact margins, so the identical file can go to a print house and arrive in the post — the print-and-mail add-on that has been an open thread since 2026-09-26 is now a fulfilment integration rather than a research problem.

Where the card lives. prayer-engine/packages/card/: layout.json (the geometry — the single source of truth), src/render.ts (the layout, the fit ladder, the rasteriser), src/png.ts (a PNG encoder over the platform's CompressionStream), src/generated.ts (header, footer and glyph atlases, built by scripts/build-assets.py). The package typecheck runs build-assets.py --check first, so a layout change that was never rebuilt fails the build.

What changed around it. brands/<slug>/assets/card/ is deleted and push-brand-assets.sh no longer uploads a template — the worker renders the card rather than reading one. BrandConfig.location.cardTemplate is gone. [CARD_TEMPLATE] and the unconsumed ## CARD VARIANTS section are gone from the pack format; [CARD_REFERENCE] now renders at every location. {FIRST_NAME}, {PRAYER_TEXT} and nameRecovery are gone from Pass 2, replaced by varyAxesForRetry, which steps the hand grip and card slot on for a retry. The copy sweep replaced "written out by hand" with the printed-card phrasing on every customer surface, in the emails and in COPY-CREATIVES.md.

Gates. pnpm typecheck 0 · 697 unit tests (21 new in packages/card, including the digest tests that prove the header and footer bands are pixel-identical across two different orders) · Playwright 27/27.

Two iterations on the same day. The first shipped the print; Shaw's feedback on it — "the name at the top is MASSIVE and various different sizes… needs to be completely uniform each time" — took the ladder out and reserved two lines of height for the name so the prayer starts at the same y on every card. The footer emblem was being dropped by the model because the clause never named it. And the delivered frame is now shrunk to 720x960 JPEG, 1.82 MB -> 88 KB. All 43 images regenerated again. See CARD-SPEC.md §Uniformity and packages/imaging/src/deliver.ts.

Verified live, same day. The printed card went through a real model: hero-set.ts re-ran all seven example deliverables on openai/gpt-5.4-image-2 and all seven passed the QA gate on attempt one (~$0.45). The 18 review panels and 18 avatars were recomposed against each person's own rendered card, and brand-card-stills.py re-shot the two brand stills. That is the whole imagery sweep, and it closes the "no live render" gap in the commit above.

Deployed the same day. prayer-worker 93ab1319 and prayer-web 8c470794, with the brand assets re-pushed to R2 and no migrations pending. Checked against the live origin by hash, not by eye. The lesson worth keeping: the first sweep regenerated the imagery in the repo and stopped there, so the deploy is not optional — regenerating an asset and shipping it are two steps, and the site serves the ORIGIN, not the working tree. apps/web/static/card-sweep-2026-09-28/LIVE-*.jpg are the deployed pages.

Two gaps remain, and both are the SAME location twice. No bumped order — the front end plus the crypt — has been through a live model on the printed card, and neither has a member delivery. Those are the two paths where the card is attached at a second location or from a subscription row, and photo-pair-check.ts is the rig that covers the first.

Open work (in order)

2026-09-30 — LIVE. Open, in order:

  • DECIDED (Shaw, 2026-09-30): no mailing address in the email footer. MAILING_ADDRESS stays empty and the footer prints just "Holy Cards Ltd." The OTO4 envelope return address is still open.
  • OTO4 vendor. Five inquiries sent from holycards-print@agentmail.to (Cardly, Prodigi, Stannp, PostGrid, Click2Mail), sample card attached; replies land in research/VENDOR-REPLIES-2026-09-30.md via a twice-daily scheduled check (bridge schedule #136/#137). Until then /admin/mail CSV, posted by hand.
  • Ad creatives, then launch per OPS §10 ($50 CAD/day, broad, USA only, women 40+).
    • DECIDED (Shaw, 2026-09-30): the candle is out. No candle in the offer, no candle in the copy. Closes the highest-severity open risk in the brand build and the last item gating the ad set. Recorded in brand-context-st-peters/README.md, 09-advertising-compliance.md §7/§9, COPY-CREATIVES.md §10.
    • DECIDED (Shaw, 2026-09-30): Spanish dropped, USA only. The three -ES primaries are deleted from COPY-CREATIVES.md §10.
    • DECIDED (Shaw, 2026-09-30): the personal-attributes openers in Ads 1 and 3 stay as written ("leave it"). Accepted risk — they are the two highest-risk assets in the kit under Meta's Personal Attributes rule.
    • Plan and 20+ concepts: AD-PIPELINE-PLAN-2026-09-30.md. The pipeline is built: ~/.claude/skills/prayer-ad-forge/ (SKILL + references/CREATIVE-INVARIANTS.md + scripts/prayer_gate.py + the skeleton library), plus the offer card at ~/.claude/skills/script-forge/offer-cards/st-peters.md and the style module at ~/.claude/skills/script-forge/styles/prayer-broll.md.
    • Wave-one copy is written and passing both gates: prayer/ads/WAVE-1/SCRIPT-OPTIONS-prayer-broll-v1.md — four silent B-roll options. qa_gates.py = ALL GATES PASS; prayer_gate.py = ALL CHECKS PASS. No frame is rendered — the still set on disk is the stale parchment set (PROMPT-PACKS §10).
    • WAVE 2 IS THE LIVE SET — 8 ads, not the wave-1 five. prayer/ads/WAVE-2/ads/ad-01..08.jpg + ADS-eight.md + RUN.md. Wave 1 is superseded: Shaw caught that its frames were invented scenery and that "AD 5" was an anonymous generated man, not Fr. Marco. Wave 2 uses only approved assets — donor-gated Pass-1 scenes, card-lock cards, and Fr. Marco off the identity lock. Built from competitors/INTEL-2026-09-30.md; the category comparison is on the board.
    • DECIDED (Shaw, 2026-09-30): every image ad frame uses openai/gpt-5.4-image-2 (gpt54). Pinned as invariant B0 with the explicit --model flag on every call, because DEFAULT_MODEL could drift. One documented exception: the card-lock blend and grade steps stay on Gemini 3 Pro, because gpt54 moves and enlarges the card ~1.2x and the card's geometry is the product (Shaw's own note in cardlock.json). gpt54 makes every scene and every hand plate.
    • FIVE IMAGE ADS RENDERED, NOT UPLOADED — awaiting the gate. prayer/ads/WAVE-1/run-20260930-image/RUN-20260930-st-peters-five-statics/ → final-ads/ad-01..05.jpg (1080x1350) plus a 154px thumbnail of each. Copy in ../ADS-five-statics.md (prayer_gate.py ALL CHECKS PASS). Three frames carry the card in hand through services/cardlock (real printed card, real hand plate, alignment held at scale 0.997); two are type-only over the place. Type is composited deterministically, never generated. Full record: that folder's RUN.md.
    • DECIDED (Shaw, 2026-09-30): Fr. Marco may appear in prayer ads, face included. Recorded as a brand-scoped inversion of the standing characters-on-screen directive — do not propagate to the Devine Rev offers.
    • DECIDED (Shaw, 2026-09-30): no reviews in paid creative. The 18 site reviews stay on-site; the gate enforces it mechanically.
    • ⚠️ The candle decision reached the ad frames. PROMPT-PACKS §10 AD7 and AD1 were lit by candles and AD1 was a votive-candle macro. All seven blocks were corrected to the basilica's own lamps on 2026-09-30; the first run of prayer_gate.py is what caught it.
    • Currency trap: the ad account bills in CAD while the scale-and-kill thresholds in SPEC §9 are USD. Convert before applying the rule; $35.10 USD ≈ $48 CAD, about one day's budget.
  • Grottoes bump backdrop varies run to run; one of two live renders read as a modern corridor rather than a crypt. Watch it; regenerate the donor if it recurs.
  • Docs drift from today's code: COPY §6 (OTO1 name field), COPY §9 email 3 (every photo, audio as a link) and the ORDER_REF subject, SPEC/OPS OTO3 voice + container /audio, PROMPT-PACKS §11 (the home-mic chain replaces the bells bed). Update them to match.
  • Refunds are recorded as psp_refund events and do not change orders.status or stop fulfilment (by design, confirmed live).

2026-09-29 — OTO4 Your Card by Post and OTO3 at $24.90: CODE-COMPLETE, not deployed. Spec: specs/SPEC-OTO4-MAIL-AND-OTO3-PRICE-2026-09-29.md. /oto/3 → /oto/4 → /thanks on both the OTO3 accept and decline; a mail_orders table, a mail_print worker job that stores the card PNG at <brand>/<order>/print/<id>.png, /admin/mail (Export CSV, Mark sent), the portal status line, the /thanks recap line, and oto4_take. Open:

  • Run migration 0006_oto4_mail_orders on the remote D1 before deploying (it rebuilds jobs and upsells to widen their CHECK lists; apply while the queue is quiet).
  • Choose the print-and-post vendor. MAIL_VENDOR=manual is the only adapter; lob / postgrid are stubs. Until then an operator exports /admin/mail as CSV, posts the cards, marks them sent.
  • The print file is the trim only (1240 × 1748 px = 105 × 148 mm at 300 dpi). The renderer draws no bleed; the print house extends the white edge by CARD_BLEED_MM (3 mm). Confirm with the vendor.

2026-09-27 (late) — two pricing decisions from the competitor ladder read. Source: research/COMPETITOR-OTO-MAP-2026-09-27.md. Their ladder is mapped; ours is longer and cheaper at every comparable step. Two lines are contradicted by the only market data that exists, and both are Shaw calls, not copy jobs:

  • DECIDED 2026-09-29: OTO3 is $24.90 (Shaw: "For 5, let's change the pricing to that"). Built in SPEC-OTO4-MAIL-AND-OTO3-PRICE-2026-09-29.md with OTO4; docs and code moved together. The original note follows. Voice of the Sanctuary was $9.90; the only comparable — Sacred Prayers' Sacred Audio Vault — is $24.90, and the audio step is the one our own ICP doc names as the channel the photograph cannot reach (ICP-AND-BUMP-EVIDENCE AV:94). The revenue curve is nearly flat ($9.90×15% = $1.49 vs $19.90×8% = $1.59 vs $24.90×6% = $1.49), so this is a one-line test — but changing it moves SPEC §3 and the SKU row the invariant table propagates into OPS §4/§9, COPY §4–§7, pricing.ts and BrandConfig.products. Test it in isolation or leave it.
  • DECIDE the membership's early read. Neither operator carries a single recurring SKU — not Sacred Prayers (funnel or store) and not Holy Land Prayer (five products, all one-time; their nearest is a finite $49 four-week). So the 6% join / 4-month life in ECONOMICS §2 has no external validation at all. Consider leading with the annual $79 so a weak monthly cannot drag the AOV path, and instrument the join rate from order one.

Not open, recorded as confirmed: the $34.90 front (live on both their surfaces under ~30 ads since 2019); second prayer at $14 (ours is $14 — no headroom); the no-pre-checked-bump rule (their WordPress form doesn't pre-check either; only their Funnelish surface does, and it is the one place they out-earn us on the front end, ≈$1.50–$2.50/buyer); OTO2 rush $19.90 (no competitor sells speed at any price); and the architecture's no-physical-fulfilment rule (their $29.97 mail-home bump is matched by our $29.80 OTO2+OTO3, both digital). One new rule to hold: never let a second surface — a Shopify store, a promo reskin — carry its own copy of the ladder. Their four-week service prices at $49, $89 and $140 across their own surfaces.

2026-09-27 — the three content sections are LIVE and the backfill is COMPLETE.

Wall prayers 14,602 across 2,456 days (2020-01-01 → 2026-09-27), comments 43,361, answered prayers 2,363 across 2,362 days, library pages 79. The wall's pager reads "Page 1 of 2,434". sitemap.xml carries 92 URLs and omits the wall, whose pages are noindex by design. Every route verified live by curl. Gates: typecheck 0 across 11 packages, 675 unit tests, D1 lane 16/16, next build green.

Read SPEC-CONTENT-SECTIONS-2026-09-26.md §11 before touching any of this. It records the eight bugs the build found, and four of them are the same failure — a thing that is correct in the lane it was tested in and wrong on the platform. Two had the fix already written elsewhere in this repo. That section is the reason a future change to this feature should be verified against D1 and against a live tick, not only against the test suite.

Open, in priority order:

  • DONE — the day repair ran. 190 days were short when the walk finished (four or five prayers instead of six) plus 6 missing entirely. The cause was diagnosed, not patched: 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. Fixed in four measured steps (a dropped round retries; a top-up names what is already written; situation pools widen to 7-8; a top-up may draw on a shared palette as a last resort), then applied: 157 rebuilt, 33 refused. Final: 14,727 prayers over 2,462 days, 2,429 days at exactly six (98.7%), zero gaps, zero duplicates.
  • 33 short days remain and nothing retries them. They came back short even with four rounds and the wider palette. Running content:backfill --repair-days over whatever the short-day query returns would likely fix a further share. If the residual matters, add a step to content_daily that repairs one short day a day.
  • The repair is not automatic. Nothing periodically checks the archive for short days. If the daily tick keeps writing, a malformed answer can still leave today short. Either run the repair on a schedule or add the check to content_daily.
  • No Playwright coverage of the fourteen new routes. Verified live by curl and by arithmetic against the row counts, but no browser test drives the pager, the thread render or the moderation queue.
  • Header mode="content" is specified in §5 and not built. The footer carries the three links and the sections link to each other; nothing so far justifies a nav in the funnel's header, which exists to offer exactly one action.
  • The daily tick has not been observed over a full day. content_daily at 06:20 writes six prayers, one answered prayer, the comment trickle and an occasional admin reply. It has run and written rows, but not yet through a full quiet day.

2026-09-26 — the three content sections are CODE-COMPLETE, NOT DEPLOYED, NOT BACKFILLED.

Landed: SPEC-CONTENT-SECTIONS-2026-09-26.md and research/DEEP-DIVE-HOLYLAND-2026-09-26.md; six tables + migration 0003_third_doorman; ContentStore; the @prayer-engine/content package (taxonomy with a 79-page library registry, writer briefs, offline fixture, model client, 13-rule moderation gate); ticks content_daily and content_backfill; the backfill CLI; fourteen web routes plus sitemap and robots; the footer Reading column; /admin/content. Gates: pnpm typecheck 0, 660 unit tests (was 591), next build green with every route registered.

Open, in priority order:

  • DECIDE the writer model. The default is deepseek/deepseek-chat-v3.1 at ≈$1.20/Mtok. The library is the only part worth escalating.
  • DECIDE D5 — the survey publish opt-in. The only funnel change this feature wants, and the only honest way the wall ever fills from real orders. Without it the wall is 100% generated.
  • DECIDE the answered-prayer naming policy. Ships as Variant A (personal testimony, the sender's own account, "cure" never ours). Variant B is one config value.
  • RUN the backfill. content:backfill --dry-run first (it prices the job and writes nothing), then the real run. Dry-run projection: ≈61,700 rows, ≈6.2M output tokens, ≈$7.41.
  • DEPLOY the migration before the cron. content_backfill is declared in wrangler.toml and will start walking the moment the worker deploys; it needs 0003_third_doorman applied to D1 first, or every tick fails against a missing table.
  • The ticks are inert until CONTENT_SECTIONS=on is set on the Worker. That is the guard that makes deploying safe: without it a deploy would start the backfill and its model spend by itself. Set the variable when Shaw wants the content to start.
  • Not built, deliberately: /api/content/report (needs rate limiting and an abuse path); Header mode="content" (only justified by traffic); Playwright coverage of the new routes.

2026-09-25 (later still) — Fr. Marco is on the page, and every shipped render shows our card. NOT DEPLOYED.

Superseded 2026-09-28 by the printed-card refactor. The card is no longer an attached blank template with the words generated onto it: @prayer-engine/card prints it before Pass 2 runs and Pass 2 writes nothing. Kept as history because both findings below are still true of any printed mark the card carries, and because the QA contract they describe is the one in force.

Landed: the four persona stills re-rendered from the identity lock; a new PersonaIntro band on the landing (data-persona) carrying his story and his refusal list; all seven example photographs regenerated through the real two-pass pipeline on the branded card; the proof band rebuilt as two ReviewSlider instances.

Two findings worth carrying, both recorded in full in CARD-SPEC.md:

  1. A branded card changes the QA contract, not just the picture. The first branded hero run failed all seven rows with name="Holy Cards OF ST. PETER'S" — the photographs were correct and the vision read-back had transcribed the printed header as the card's first line. The instruction now scopes card_text to handwriting only, and parseCardText strips the header defensively as well, so a model that ignores an instruction cannot dead-letter a paid order. Any future printed mark on the card — a date, a place line, a signature — must be added to that exclusion list in the same commit that adds it.
  2. cardTemplateKey() did not match the sync convention. It returned brands/<slug>/card/template.png while sync-brand-assets.ts keys every asset brands/<slug>/assets/.... Uploaded to one key, read from another, and the only symptom would be every card silently rendering unbranded. Fixed; the key now lives beside donorKey().

Open, in priority order:

  • DEPLOY. Nothing above is live. prayer-web and prayer-worker both need it (the worker owns the new prompt path and the QA instruction).
  • Get the card template into production R2. Closed 2026-09-28: there is no card template to upload; the worker renders the card. push-brand-assets.sh no longer pushes one and loadCardTemplate no longer exists.
  • Pass 2 is not conditioned on Fr. Marco. ## HAND VARIANTS selects a grip on an attached hand reference rather than rendering from the character sheet, so the delivered photograph's hand is the right shape of hand but is not the persona's own. This is the last place the persona is not load-bearing.
  • The card template is convention, not config. Closed 2026-09-28: the card is rendered from layout.json on every order, so a second brand cannot ship without one. What a second brand does need is a logo of its own — the renderer reads apps/web/public/assets/brand/logo/, which is currently one brand's mark, and that is the remaining brand-expansion gap.
  • Western Wall needs its own persona and card before it ships, and its portraitSet is still ari with no assets behind it.
  • Ad creative does not use the persona or the branded card yet (PROMPT-PACKS §12).
  • OTO photo captions per PERSONA-FR-MARCO.md ("This one is Maria's…") are still unwired.

2026-09-25 — the card becomes a branded asset, and Fr. Marco becomes a built persona. The card half is superseded by 2026-09-28: as a branded attached template it lasted three days. The persona half still stands. Superseded by the entry at the top of this list; both landed the same day. The card is A4 portrait printed with our maroon lockup, wordmark and rule, attached as a reference image rather than described in words — a model told to "add a logo" draws a different one every render. The front end attaches card/template.png; the bump attaches the front end's finished photograph, so it is still one card photographed twice. Fr. Marco was built on Leg 0 of the Video Forge skill from a Pinterest seed screened for candid realism. REALISM-RULES R9 and PROMPT-PACKS §7 both changed; CARD-SPEC.md records why "no decoration ever" was superseded rather than violated.

2026-09-25 — express pay, the short billing block and the design pass are DEPLOYED (see README status). Open: Shaw's review of the live tabs; a wallet payment tested on a real phone; register the domain for Apple Pay / Google Pay in Stripe LIVE mode at launch; the landing's weight (1,087 KiB) could come down; the brand kit and SPEC/OPS still describe the old footer (positioning block on every page) and the full billing address.

2026-09-25 — express pay, the short billing block and the design pass are DEPLOYED (see README status). Open: Shaw's review of the live tabs; a wallet payment tested on a real phone; register the domain for Apple Pay / Google Pay in Stripe LIVE mode at launch; the landing's weight (1,087 KiB) could come down; the brand kit and SPEC/OPS still describe the old footer (positioning block on every page) and the full billing address.

The persona. PERSONA-FR-MARCO.md had a story and no fixed face. He is now built on Leg 0 of the Video Forge skill (aimikoda single-composite, GPT Image 2 — Shaw's call that all image generation runs on GPT Image 2), from a Pinterest seed screened for candid realism: 234 candidates → 43 → 8 → one real phone selfie in an ordinary room → two mutation passes → 01_mutated_v2.png, the identity lock → sheet.png, the 11-section character sheet. Assets and the reproducible pipeline are at persona/fr-marco/. He self-registered as Fr. Marco Bellandi. Open: the landing, /offer and /create/ready still show the retired studio headshot; the persona has not yet been wired into any live surface.

The card. A4 portrait, printed with our maroon lockup, wordmark and rule (CARD-SPEC.md). Attached as a reference image, never described in words — that is the whole design, because a model told to "add a logo" draws a different one every render. Front end attaches card/template.png; the bump attaches the front end's finished photograph, so it is still one card photographed twice. Rendered and verified on real models at both locations, same seed: identical header, identical handwriting, identical line breaks. REALISM-RULES R9 and PROMPT-PACKS §7 both changed, and CARD-SPEC.md records why "no decoration ever" was superseded rather than violated.

Gates: pnpm typecheck 0 · 591 unit tests · prompt-render check (one second-image clause, zero leftover tokens, correct clause per location) · both locations rendered.

Open from this run:

  • The review slider is specified but NOT BUILT. research/REF-TEARDOWN-2026-09-25.md §1 has the full component spec; the substance is that a stubbed entry must render without stars, because a star row is a rating and there are no real ratings yet.
  • The card template is not yet uploaded to R2 — run scripts/sync-brand-assets.ts against the production bucket before the next live order, or the front end silently renders unbranded (loadCardTemplate returns null rather than throwing, by design).
  • The Western Wall brand needs its own card and persona before it ships.
  • The slug of the retired studio headshot fr-marco-master.jpg is still referenced by the live web app; the persona swap replaces it with the character sheet's derived frames.

2026-09-25 — express pay, the short billing block and the design pass are DEPLOYED (see README status). Open: Shaw's review of the live tabs; a wallet payment tested on a real phone; register the domain for Apple Pay / Google Pay in Stripe LIVE mode at launch; the landing's weight (1,087 KiB) could come down; the brand kit and SPEC/OPS still describe the old footer (positioning block on every page) and the full billing address.

2026-09-23 — copy pass and ten UI iterations DEPLOYED and live-verified (see README status). Superseded note follows: 2026-09-22 (night) — copy pass and five UI iterations, committed, NOT deployed. See the README status line. Deploying needs both prayer-web and prayer-worker, because emails 1, 2, 5, 7, 8, 9, 11, 17 and the email footer changed. Still waiting on Shaw: the countdown, the discount card, the $69 line, the testimonial stubs and the ticker are kept as they are by his call (COPY-REVIEW §Shaw's decision). Apple Pay / Google Pay and a shorter billing block (audit C6) are still unanswered.

2026-09-22 — the UI/UX pass (phone first, benchmarked against PrayerSong). The audit is research/UI-AUDIT-2026-09-22.md: 4 defects and 15 layout items, all fixed in af1b4b8 and 2c59482 on branch held-card-pivot. Typecheck exit 0; Playwright 25/25. The after-shots of all 27 surfaces at 390px and 1280px (54 shots, 0 with overflow, 0 with an unloaded image) are in review-site/static/ui-audit-2026-09-22/after/. Deployed as prayer-web version 5558551f. Live-verified on the public landing page only. Open, waiting on Shaw's call (the audit's C1–C10): billing fields + Apple Pay / Google Pay at checkout, the "Lit." line in the letter, the [NAME], [CITY, STATE] stubs on /create/ready and /offer, the /thanks example card, the OTO3 sample label, em-dashes in the landing hero.

2026-09-21 — the held-card pivot and the second-location bump. The delivered photograph changed from a card lying at rest to a card held up in the hand at the spot, and the failing $4.99 Wider View bump was replaced by a second location, the Vatican Grottoes, at $9.97 (The Apostle's Tomb). Read research/ITEM4-SCOPE-2026-09-21.md for the decisions and research/OFFER-SURFACE-INVENTORY-2026-09-21.md for every surface that still carries the old text.

Done and committed (/home/sha/vibing/prayer-engine, branch held-card-pivot): 9aec3b8 safety checkpoint · 9e3fcd0 two-location prompt packs, one shared realism rubric, the production gate taught composition checks for the first time · 7436db7 the front end is always the Confessio, so the location no longer rotates. pnpm typecheck exit 0; 577 tests across 48 files; the gate live-verified against two frames Shaw approved and one he rejected.

Done in the docs, uncommitted and NOT versioned — /home/sha/vibing/prayer is not a git repo, so every edit below exists only on disk: SPEC, OPS-RUNBOOK, COPY-CREATIVES, ECONOMICS, CLAUDE.md, PROMPT-PACKS and PRAYERSONG-MAPPING now carry the new bump and the new prices. The whole derived chain moved: AOV 42.83 → 44.07, max stacked 83.69 → 88.67 → 97.67 (the speed option moved it again), net 37.94 → 39.06, contribution 40.13 → 41.25, break-even ≈$40 → ≈$41.

Closed on 2026-09-22 — all three blocks in the previous handoff are done:

  1. The bump's code swap is DONE and green: SKU bump_apostles_tomb ("The Apostle's Tomb", $9.97), migration 0002_petite_ultron (column bump_taken, enum gains bump_apostles_tomb + bump_speed, photo_wide and wider_view gone, rows rewritten in place), and the worker's photo job is a four-step machine that renders the bump from its own Pass-1 frame at the Grottoes and attaches the front end's finished photograph as the second reference image, so the card is copied rather than written twice. pnpm typecheck exit 0, 587 tests green, pnpm --filter db test:d1 green.
  2. The copy sweep is DONE in the app and in COPY-CREATIVES: the at-rest claim is gone from apps/web, packages/email and COPY-CREATIVES, and the email↔COPY parity guard in packages/email/test/templates.test.ts passes.
  3. The checkout speed option is DONE: bump_speed ("Speed, 24 hours", $9.00), a second unchecked checkbox, orders.rush set at checkout, and /oto/2 (page and route) gated on tookCheckoutSpeed.

Still open, in order:

  1. Nothing is committed and nothing is deployed. Both repos have large uncommitted diffs; holycardsofsaintpeters.com still serves the OLD build (its checkout still offers "Votive Candle, $4.99"), so none of this is visible in a browser yet.
  2. The candle promise in live copy is unreconciled. PROMPT-PACKS §4.1a says the candle clause was removed ("Pass 2 places no light of any kind"), yet apps/web still promises "A candle lit beside it" on the landing, /offer, /replace, /thanks, /membership and the portal, and the fee line in copy.ts lists "candle" among what the payment covers. Either the promise is kept (a real candle, lit by hand at the balustrade — R1 allows it) or the copy comes out. Shaw's call; the doc already removed it, the live pages still sell it. Resolved 2026-09-22: the promise was removed (README status), and the last leftover, "Lit." in the /offer letter, came out in the copy pass of the same day.
  3. The OTO2 naming collision. The checkout sells "Speed, 24 hours" $9.00 beside OTO2's "Same-Day Placement" $19.90 — one promise at two prices, gated so nobody is charged twice. The cleaner fix is to give OTO2 something genuinely better ("Same-Day Placement"), which changes its meaning and its anchor. Shaw's call.
  4. The brand kit (brand-context-st-peters/) still carries the at-rest wording in its doctrine and hook-bank lines. It is the voice source for new copy, so it will keep seeding the old claim until it is swept or marked historical.
  5. review-site/public/ was not rebuilt, so the review site still shows the pre-sweep text.
  6. PROMPT-PACKS §2 and §6 still document one location's donor spec and angle pack. The keys were renamed to st_peters_confessio and the vatican_grottoes ambience row was added, but the crypt has no §2 donor spec and no §6 angle pack in the doc, even though pass1.md ships BUMP SCENE / BUMP ANGLES and the donor is approved in APPROVED.txt (donor 30, frame 30 of the 2026-09-21 pull). Write them from the packs, not from scratch.
  7. Items a–e and 1–6 below, unchanged; the per-item notes that are now closed are struck at the end of this file.

Older open work

2026-09-20 — design pass shipped (commit fd54dd8, both Workers deployed, live landing verified). Items 1 below (the sample MP3) is DONE; the seven example photos are live. Hosting rule from Shaw: nothing on the box, everything on Cloudflare. Review site is now https://prayer-review.pages.dev (deploy steps in review-site/README-DEPLOY.md). Remaining, in order:

a. Hero name mismatch: COPY §1 caption says "Carmen's card"; the H-1 photo reads "Maria" (PROMPT-PACKS §5.1). Fix one side: regenerate H-1 with {FIRST_NAME}=Carmen and update the PROMPT-PACKS row (recommended, copy stays untouched), or change the caption in COPY + three pages. b. e2e:cf DONE 2026-09-20: 24/24 (commit 3193796). Register any future Stripe webhook on the BRAND DOMAIN, never on workers.dev: brand resolution is by Host and workers.dev answers 404. c. OTO3 player click-test DONE 2026-09-20: plays on the live site; test order portal /portal/v1guQ2NA26idromMBe2P8iyFdDBNy5ne. Two test orders now sit in production D1; delete both before launch. Add a favicon (404 today). d. Vercel cleanup (Shaw's yes): delete projects prayer-review and prayer-funnel-preview. e. Items 3–6 below, unchanged.

2026-09-17 (night) — the engine is LIVE on Cloudflare in Stripe test mode and one real order ran end to end (see README Status and PLAN-CLOUDFLARE-PORT.md). What remains, in order:

  1. Voice of the Sanctuary sample. BrandConfig.audio.sampleUrl points at assets/audio/st_peters_sample_15s.mp3, which does not exist in apps/web/public, so the OTO3 player renders disabled. Produce a 15-second sample (OpenRouter TTS in the persona voice, no ambience bed in v1) and ship it under apps/web/public/assets/audio/. Same for Western Wall when its front is built.
  2. Playwright e2e:cf (OpenNext preview on local D1) has not had a clean run; every attempt so far failed on box disk pressure. Run it alone in a quiet window; fix whatever is real.
  3. Western Wall front: stonesofjerusalem.com is owned by a third party (registered 2024-11). Shaw picks a new name (free on 2026-09-17: stonesofjerusalemprayers.com, prayersatthewesternwall.com, jerusalemstoneprayers.com); then config domain + descriptor + docs, prompt pack from PROMPT-PACKS §6/§2.5, donor via the board flow, Resend domain, zone + Worker domain.
  4. Go-live switches (Shaw): Stripe live keys (pass companies/9566-0866/stripe-api-live) into the Worker secrets and a live-mode webhook registration; Meta pixel id + CAPI token; decide the free Stripe "Dispute deflection" toggle (recommended on) — research/STRIPE-RADAR-2026-09-17.md.
  5. Housekeeping: stop the obsolete prayer-pg Docker container; scope ~/.claude/hooks/prompt-gaze-guard.py to video prompt files; the box's disk (90% full, Cursor server writing continuously) is what breaks Playwright — clear it before relying on local e2e again.
  6. Then Phase D launch per SPEC §10.

Infra map (srana Cloudflare account): zone holycardsofsaintpeters.com (20f8fc48…), D1 prayer-engine (c8601c43…), R2 prayer-engine, Queue prayer-jobs, Workers prayer-web (custom domains apex + www) and prayer-worker. Secrets set with wrangler secret put from pass (generated salts + webhook secret in pass companies/9566-0866/prayer-engine-secrets; Resend key is services/resend-devine-rev). Worker logs: Cloudflare observability query API (workers/observability/telemetry/query).

Earlier open-work notes (kept for history)

Direction change 2026-09-16 (the Stripe move is now resolved — see item 1 below): Stripe under a Numina House registration replaces TagadaPay; Shaw opens the account, Fable writes the Stripe adapter behind PspAdapter. Hosting is Cloudflare for everything (Workers via OpenNext, D1, R2, cron triggers); the D1 port is the next engineering phase after the donor and the QA rule are settled. Models run through OpenRouter on the key in pass apis/openrouter.

Phase B — Shaw-owned, and it blocks the build.

  1. Stripe adapter CODE-COMPLETE and live-verified in test mode (17/17 smoke, 2026-09-16: customer, vault, charge, idempotent re-charge, decline, 3DS, off-session charge, refund, subscription lifecycle, lookupCharge); webhook registration waits for the public URL from the Cloudflare deploy, then scripts/stripe-register-webhook.ts <public-url> <brandSlug> captures the signing secret.
  2. Descriptor DONE: NUMINA* HOLYCARDS (prefix fixed by the Stripe account, suffix set in BrandConfig.site.statementDescriptor). Confirmed 2026-09-16, fits Stripe's 22-character cap, no phone number, never the word "prayer" (OPS-RUNBOOK §4).
  3. Payout, reserve, acquirer and fees — all published, nothing pending. Stripe's standard Canada payout schedule is 7-day rolling at first, then 2-day. Reserve is none by default. The acquirer is Stripe itself. Fees are in ECONOMICS.md and SPEC-PRAYER-ENGINE.md §3.
  4. Domain, Resend sending domain, Meta BM and page, fresh pixel and CAPI token, droplet, support mailbox, model API keys (image models plus the vision model for card-text QA).
  5. Donor reference image picked on Pinterest and approved for the launch location, with its APPROVED line recorded.

Phase C — build, per OPS-RUNBOOK §10 day 0–13 and PLAN §5. Six steps, each with its own gate. The fulfilment gate is the one that decides the product: 10 test orders, denominator stated 10 of 10, every photo legible, every name spelled right, every bumped order showing two photographs, one at each location. Run it before any copy is locked, because the 220-character cap and the card-size rule appear on six customer-facing surfaces and are cheap to change now and expensive later.

Phase D — launch, per SPEC §10: three creatives, $50 CAD/day broad, USA only, women 40+, no more than four new ads a day, creative rotation at day 6, CPA read at day 13 against the single rule in SPEC §9.

Also open (found 2026-09-16): brands/western-wall/prompts/ does not exist, so the Western Wall front cannot fulfil until its pass-1/pass-2/themes pack is written from PROMPT-PACKS §6 and §2.5 and its donor is picked the same way as St. Peter's (board flow). The order-confirmation email job (transactional_email) and the guest-membership Postgres insert are in a fix lane as of this note.

Also open, not blocking: the multi-religion research pass in RESEARCH-DIRECTIVE.md, which locks the part-2 lineup beyond Wave 1.

Rules of engagement

  • The build docs are complete directives. Strip nothing, soften nothing, add no compliance language, no disclaimers, no "consult X" footnotes, no legality checks.
  • Copy positioning stays categorical, never comparative: "service fee, not a donation", "we do not sell outcomes", "not the church".
  • Testimonials are presented as real customer reviews — 5 stars, a name, a city and a customer avatar, over a photograph of a held card. Revised 2026-09-25 by Shaw's directive, superseding the [REAL CUSTOMER — INSERT ONCE LIVE] stub rule: the marker and every rule that enforced it are removed from the code and from every doc. Quotes stay service-framed and never outcome-framed. No aggregate customer-count claim appears anywhere. Revised again 2026-09-29. The avatar is the customer and holds nothing — a plain candid photo of a person at home, never a card. An avatar holding a card read as the client photographing their card and sending it to us, which inverts the offer: the buyer types the prayer, we print it, we carry it. And the wall is one section, not three bands: one row of every review, walked with a Next button, one card treatment everywhere (landing wall and the proof under an example photograph on /offer and /create/ready read the same lib/reviews.ts).
  • The statement descriptor never contains the word "prayer."
  • No email says charges have stopped while the payment rail can still charge. That sentence exists in exactly one template.
  • Cross-doc invariants must move together. BrandConfig is byte-identical in SPEC §1 and OPS §1. SKU names, prices, route names, event names, the email template list and the location keys are all mapped in CLAUDE.md; change one, change all of them.
  • One persona per brand, one ad account per brand, zero shared identity surface across brands.
  • When a doc's status or the open-work list changes, update both README.md (Status line) and this file (Open work).