pivot: tabletop pharma / Ayurveda-first + LOOP_SPEC v2

LOOP_SPEC.md

@@ -0,0 +1,194 @@
+# Shoploop — LOOP_SPEC v2
+
+_Last updated: 2026-07-24 · Status: drafted, not scheduled_
+
+## The Product = The Reverse-Feedback Content Loop
+
+```
+Customer connects (Nango) → PostHog reads impact → Creative Mutation generates variants
+→ 3D Render Batch produces MP4/PNG/USDZ → Ads Auto-Publish (Postiz) & Website Update PR
+→ PostHog reads impact → posterior compounds nightly
+```
+
+Product is the loop. Reference sites (shoploop.studio, ashwagandha.shoploop.studio, studio.shoploop.studio) are customer #0-3.
+
+## The 10 Loops
+
+Each loop is defined against the 9-primitive framework: Target · Fitness · Search Space · Stopping · Mutation · Memory · Ground Truth · Boundary · Decay.
+
+---
+
+### Loop 1 — Connector Ingest
+- **Runs:** every 15 min
+- **Job:** Nango pulls fresh data from customer's Google Ads + Shopify + PostHog + Meta + Klaviyo + GitHub into their own Postgres partition
+- **Target:** 100% connector uptime, <5min data freshness lag by 2026-09-30
+- **Fitness:** rows-synced-per-15min / expected-rows; alert if <95%
+- **Search Space:** connectors enabled at customer onboarding
+- **Stopping:** none (continuous)
+- **Mutation:** on API error, backoff + retry per Nango retry policy
+- **Memory:** last-cursor per connector in Postgres
+- **Ground Truth:** connector's own API response
+- **Boundary:** OAuth expiration triggers customer email, does not auto-refresh
+- **Decay:** cursor state resets on connector reconnect
+
+### Loop 2 — PostHog Reverse-Feedback Analyst
+- **Runs:** every 4 h
+- **Job:** join ad clicks × sessions × ATC × checkout × PDP dwell. Surface top 3 opportunities: low-CVR-high-CTR, high-search-no-match, high-ATC-abandoned-checkout
+- **Target:** 1 actionable opportunity per cycle for every customer
+- **Fitness:** opportunities-acted-on / opportunities-surfaced ≥ 40% (measured weekly)
+- **Search Space:** last 7 days of PostHog + ad platform data
+- **Stopping:** none
+- **Mutation:** SQL query templates iterate based on which surfaced opportunities converted to Loop 3 wins
+- **Memory:** rolling 30-day opportunity log per customer
+- **Ground Truth:** PostHog events + ad platform CVR
+- **Boundary:** opportunities require Loop 3 to accept; not auto-executed
+- **Decay:** opportunities expire in 72h if not picked up
+
+### Loop 3 — Creative Mutation Kernel (Thompson Sampling)
+- **Runs:** every 6 h
+- **Job:** pick 6 candidate variants per active opportunity across (3D angle × molecule × material × copy × language). Thompson Sampling on Beta posterior over CVR
+- **Target:** variance in top-5 candidates < 0.05 for 3 consecutive cycles = converged
+- **Fitness:** posterior mean CVR uplift vs baseline; weighted by trilingual coverage
+- **Search Space:** all molecules in customer's PIM × all Studio templates × hi/en/mr
+- **Stopping:** convergence (3 cycles < 0.05 variance) OR 30 iterations
+- **Mutation:** top 20% arms survive, bottom 80% replaced with mutations along dimensions correlated with win in surviving arms
+- **Memory:** Beta(α, β) per arm in Postgres, versioned in git
+- **Ground Truth:** Loop 2 measured CVR after publish
+- **Boundary:** candidates must pass Language QA (Loop 7) before Loop 4 renders
+- **Decay:** posteriors halved every 30 days (freshness bias)
+
+### Loop 4 — 3D Render Batch
+- **Runs:** every 6 h (after Loop 3 releases winning candidates)
+- **Job:** Blender + MolecularNodes + NVIDIA Omniverse + usd-content-agents produces MP4/PNG/USDZ per candidate. Validation Agent quality-gates
+- **Target:** ≥95% render success, ≤$0.02/frame, PBR fidelity score ≥0.85
+- **Fitness:** (success_rate × 0.4) + (1/cost_per_frame × 0.3) + (pbr_score × 0.3)
+- **Search Space:** Blender cycles + Omniverse Kit + USD asset library
+- **Stopping:** batch complete OR AWS spot pool preempted
+- **Mutation:** on Validation Agent rejection, retry with adjusted material/lighting
+- **Memory:** render cache keyed on (asset_hash + material_hash + camera_hash)
+- **Ground Truth:** Validation Agent + spot-check by human weekly
+- **Boundary:** cost cap $50/day per customer; hard-stop
+- **Decay:** cached renders expire in 90 days
+
+### Loop 5 — Ads Auto-Publish (Postiz)
+- **Runs:** every 6 h (after Loop 4)
+- **Job:** winning ad variants push to Meta + Google Ads + YouTube Shorts via Postiz + Nango
+- **Target:** zero manual ad uploads for enrolled customers by 2026-08-31
+- **Fitness:** ads-published / ads-approved-by-Loop-3 = 100%
+- **Search Space:** customer's active ad accounts via Nango
+- **Stopping:** weekly $ cap set by customer reached
+- **Mutation:** budget allocation shifts toward highest-CVR variants (Loop 3 posterior)
+- **Memory:** publish log per customer per platform
+- **Ground Truth:** platform confirmations (Ads API responses)
+- **Boundary:** hard $ cap per week per customer; customer sets in onboarding
+- **Decay:** paused ads deleted after 30 days idle
+
+### Loop 6 — Website Update PR / API Push
+- **Runs:** daily 3am PT
+- **Job:** for website changes (new hero, updated PDP 3D asset, copy tweaks, molecule swaps), open PR against customer's GitHub repo OR push to Shopify admin API. Adapter chosen at onboarding
+- **Target:** website changes go live within 24h of Loop 3 approval
+- **Fitness:** median-time-from-approval-to-live < 24h
+- **Search Space:** all pages tagged for auto-update in customer config
+- **Stopping:** customer approval required on GitHub PR path; auto on Shopify API path with rollback watch
+- **Mutation:** on PR rejection, learn from customer's commit diff (feed back to Loop 3)
+- **Memory:** PR history + Shopify version log per customer
+- **Ground Truth:** post-deploy PostHog session recording sample
+- **Boundary:** irreversible destructive changes (delete page, remove SKU) always require PR + human approve, never auto-API
+- **Decay:** PR auto-closes after 7 days no activity, re-opens with fresh context
+
+### Loop 7 — Language QA (hi/en/mr)
+- **Runs:** daily 9am PT
+- **Job:** VLM samples all new hi/en/mr copy across all customer sites + ad creative. Not just translation accuracy — cultural correctness (Ayurvedic terminology, dosha references, respectful register)
+- **Target:** zero cultural mistranslations by 2026-08-31
+- **Fitness:** flagged-strings / total-strings-sampled ≤ 1%
+- **Search Space:** all copy shipped in last 24h across Studio + Postiz + Website + Ads
+- **Stopping:** batch complete
+- **Mutation:** on flag, propose alternative + surface in Loop 10 digest
+- **Memory:** flag log per customer, feeds Loop 3 with rejected copy patterns
+- **Ground Truth:** monthly native-speaker review (Mumbai + Pune Marathi + Delhi Hindi)
+- **Boundary:** blocking gate before Loop 5 + Loop 6 publish
+- **Decay:** approved strings re-validated every 90 days
+
+### Loop 8 — Platform Health (Nango + Embed + Renders)
+- **Runs:** hourly
+- **Job:** pings every customer's Nango OAuth + embed instance + render pipeline. Alerts on breakage. Weekly regression tests against top 20 Shopify + top 10 WooCommerce themes for embed compatibility
+- **Target:** 99.9% embed uptime, MTTR < 30min, zero silent connector failures
+- **Fitness:** (uptime × 0.5) + (1/MTTR × 0.3) + (regression_pass_rate × 0.2)
+- **Search Space:** all live customer instances + top 30 host themes
+- **Stopping:** none (continuous)
+- **Mutation:** on regression failure, auto-file fix PR to packages/embed
+- **Memory:** incident log per customer, theme compatibility matrix
+- **Ground Truth:** synthetic monitoring + customer support tickets
+- **Boundary:** customer-facing status page auto-updates; no auto-disable of customer instances
+- **Decay:** incident log archived after 90 days
+
+### Loop 9 — Competitive Radar
+- **Runs:** weekly Mon 8am PT
+- **Job:** watches Vagon.io + Innoactive + grip.tools + Threekit + Emperia releases + LinkedIn + pricing. Also upstream: MolecularNodes + Postiz + Nango + NVIDIA usd-content-agents commits + Pimcore
+- **Target:** 1 actionable insight/wk in founder digest
+- **Fitness:** insights-acted-on / insights-surfaced ≥ 25%
+- **Search Space:** competitor sites + GitHub RSS + LinkedIn
+- **Stopping:** weekly cycle
+- **Mutation:** query patterns refined based on which insights led to founder action
+- **Memory:** 90-day competitive timeline
+- **Ground Truth:** competitor's own announcements
+- **Boundary:** insights land in digest, no auto-action
+- **Decay:** insights expire from active-set after 30 days
+
+### Loop 10 — Founder Digest
+- **Runs:** daily 5pm PT
+- **Job:** roll up loops 1-9 into one WhatsApp/email. <60s read. ≤3 actions requiring founder judgment
+- **Target:** founder acts on ≥2 of 3 daily actions
+- **Fitness:** actions-completed / actions-suggested ≥ 66%
+- **Search Space:** last 24h across all loops for all customers
+- **Stopping:** daily cycle
+- **Mutation:** digest template iterates based on which sections founder engages with (measured via email opens + reply patterns)
+- **Memory:** 30-day digest archive
+- **Ground Truth:** founder reply / action taken
+- **Boundary:** digest never auto-executes anything
+- **Decay:** unactioned items escalate after 3 days
+
+---
+
+## PostHog reverse-feedback wiring
+
+Every UTM includes `shoploop_creative_id={loop3_arm_id}` and `shoploop_lot_id={ledger_lot_id}`.
+Every configurator + trace event: `posthog.capture('shoploop.engage', {creative_id, lot_id, action})`.
+Loop 2 joins creative_id → engage → purchase for real CVR (not vanity CTR).
+Loop 3 uses PostHog session recordings (VLM samples 5/cycle for genuine vs rage-click).
+
+## Scheduling policy
+
+- Loops 1, 8 always-on (platform reliability)
+- Loops 2, 3, 4, 5 gated by "customer has enrolled + Nango connected"
+- Loop 6 gated by "customer chose GitHub or Shopify adapter"
+- Loop 7 gated by "customer has trilingual enabled" (default: yes)
+- Loops 9, 10 always-on (founder-facing)
+
+## Cost envelope
+
+At 5 enrolled customers:
+- Loops 1, 8: ~$30/mo (always-on)
+- Loops 2-7 (5 customers × cycles/day × credits): ~$180/mo
+- Loops 9, 10: ~$20/mo
+- **Total loop credits: ~$230/mo**
+
+At 50 customers: ~$1500/mo. Priced into $299/mo × 50 = $14,950 gross.
+
+## Boundary conditions (human gates)
+
+- Loop 3 candidates → Loop 4 render (auto after Loop 7 pass)
+- Loop 4 renders → Loop 5 publish (auto within $ cap)
+- Loop 4 renders → Loop 6 website (PR path: human approve; API path: auto within safe-change whitelist)
+- Loop 9 insights → founder judgment (never auto)
+- Loop 10 digest actions → founder judgment (never auto)
+
+## Not in this spec (deferred)
+
+- Supply-chain ledger sync (Loop-11 candidate; needs first customer with real supply data)
+- ICP discovery / lead sourcing (Loop-12 candidate; not urgent pre-first-customer)
+- AR/mobile capture ingestion (Loop-13 candidate; needs mobile app v1)
+
+## Review cadence
+
+This spec re-reviewed on the 1st of each month via git blame + Loop 10 digest tag.

POSITIONING.md

@@ -0,0 +1,110 @@
+# Shoploop — Positioning
+
+_Last updated: 2026-07-24 · Branch: jared/tabletop-pharma-pivot_
+
+## One line
+
+**Full-funnel positionless AI marketing for tabletop pharma products — Ayurveda-first.**
+
+## What "tabletop pharma" means
+
+Countertop/shelf-visible SKUs a consumer can grip, rotate, and inspect in 3D:
+
+- **Ayurvedic supplements** — Chyawanprash, ashwagandha, brahmi, triphala, tulsi, chyawanprash-adjacent jars/tubes
+- **Nutraceuticals** — vitamins, minerals, protein powders, adaptogens
+- **OTC wellness** — cough syrups, digestive aids, immunity boosters, topical balms
+- **Traditional systems** — Ayurveda, TCM, homeopathy, Unani
+- **Not**: injectables, medical devices, prescription-only, cold-chain biologics
+
+## What "positionless full-funnel" means
+
+One agent, one loop, one graph — from ad impression to purchase to reorder — with no human "position" (media buyer, designer, copywriter, dev) required in the middle.
+
+The customer connects their stack once. The agent then:
+
+1. Reads their ad performance (Google Ads, Meta, Klaviyo)
+2. Reads their site engagement (PostHog sessions, session recordings)
+3. Reads their commerce (Shopify orders, ATC, checkout drop-off)
+4. Mutates creatives (3D angle × molecule × material × copy × language)
+5. Renders (Blender + MolecularNodes + NVIDIA Omniverse + USD)
+6. Publishes ads (Postiz to Meta/Google/YouTube)
+7. Updates the website (GitHub PR or Shopify admin API)
+8. Measures impact
+9. Compounds the posterior nightly
+
+## Why Ayurveda first
+
+- **Molecular-native storytelling.** Every Ayurvedic herb has characterized active compounds with published PDB structures (curcumin 3IPU, withaferin A 4EY7, piperine 1WCQ, bacoside A, gymnema...). No competitor shows the actual molecule in the ad. Shoploop does. MolecularNodes → OpenUSD → Omniverse composited onto the product bottle. Category-defining visual.
+- **Trilingual native.** Indian pharma requires hi/en/mr day one. Every UI, every ad, every PDP, every trace page. Postiz LLM translation + our language-QA loop guarantees cultural correctness, not just literal translation.
+- **Regulatory forcing function.** AYUSH + FSSAI + upcoming EU DPP (Digital Product Passport 2027) all require provenance, batch traceability, and disclosure. Shoploop's supply-chain layer (Merkle-anchored via OpenTimestamps) becomes the default rails.
+- **Market gap.** $18B Indian Ayurveda market, growing 15% CAGR, near-zero 3D commerce adoption. Kapiva, Baidyanath, Himalaya, Dabur, Sandu — none ship 3D configurators or molecular ad creative today.
+- **Grip.tools serves LVMH.** No one serves this vertical. That's the wedge.
+
+## Positioning versus alternatives
+
+| Alternative | What it does | Where it stops |
+|---|---|---|
+| **grip.tools** | 3D commerce for luxury (LVMH, L'Oréal) | $500k/yr enterprise, agency-installed, not self-serve, no pharma vertical, no molecular layer |
+| **Vagon.io / Innoactive** | Remote 3D streaming | Renders on their servers, no domain expertise, no ad-ops integration |
+| **Threekit / Emperia** | 3D product configurators | Configurator only, no ad-ops loop, no trilingual, no molecular |
+| **Postiz (upstream)** | Social scheduler | Scheduler only, no rendering, no PostHog feedback, no rendering pipeline |
+| **Uplane** | Automated ad buying | Ad-ops only, no creative rendering, no website updates |
+| **Optimove** | Post-acquisition CRM | Lifecycle only, no top-of-funnel creative |
+| **Shoploop** | The full loop, agent-run, AGPL-3.0, pharma-first, trilingual | — |
+
+## Pricing
+
+- **OSS** — $0 forever, self-host, AGPL-3.0
+- **Trial** — $0 for 30 days, managed hosting, no card
+- **Managed** — $299/mo, hosted embed CDN + Nango + Studio + Postiz + 10 loops running on customer data
+
+No enterprise tier at v0. Enterprise = OSS + implementation consulting engagement.
+
+## Languages day-one
+
+- **Hindi** (हिन्दी) — Devanagari, Noto Serif Devanagari
+- **English** — IBM Plex Serif
+- **Marathi** (मराठी) — Devanagari, Noto Serif Devanagari
+
+All UI, all copy, all embed prompts, all Studio, all Postiz.
+
+## The moat (one sentence)
+
+**Vagon.io + grip.tools + Postiz + PostHog + Google Ads feedback + NVIDIA Omniverse + Blender with materials — Hindi/Marathi/English support — targeting Ayurvedic tabletop category.**
+
+The composition IS the moat. The vertical IS the wedge. No competitor sits at the intersection of even 3 of these 9 components. We compose all 9 and target ONE vertical.
+
+## Nine components, one composition
+
+| Component | What it contributes | Who else has it in this stack |
+|---|---|---|
+| **Vagon.io-class remote 3D streaming** | Browser-native, no-install 3D interaction. Runs Blender/Omniverse in the cloud, streams frames | Vagon has streaming, no pharma vertical, no ad-ops. Fillable via Vagon SDK or self-host on AWS g5.xlarge spot pool |
+| **grip.tools-class 3D commerce** | Product configurator, PBR fidelity, embed-on-any-site | grip.tools closed, $500k/yr. We ship OSS embed under AGPL-3.0 |
+| **Postiz** | Multi-channel scheduling (Meta/Google/YouTube/LinkedIn), 14 langs already, AGPL fork-able | Buffer/Hootsuite closed; Mixpost no Indic. We fork + add hi/mr locales |
+| **PostHog** | Session recordings, funnel, feature flags, self-hostable | Google Analytics is drift-only; Mixpanel closed. PostHog is our truth layer |
+| **Google Ads feedback loop** | Real CVR data closes the mutation loop | Every ad-tech tool reads it; nobody feeds it back into 3D rendering. We do |
+| **NVIDIA Omniverse + Kit + usd-content-agents** | USD composition, Material/Physics/Validation Agents, Kit apps | Enterprise-only awareness. We wrap it in a $299/mo product |
+| **Blender + PBR materials + MolecularNodes** | Open-source render farm, material accuracy, molecular hero renders | Standard 3D tool. Nobody in pharma uses it for ads |
+| **hi/en/mr Devanagari day-one** | Regulatory + market fit for Indian pharma | International tools translate literally; none ship cultural QA. Loop 7 does |
+| **Ayurvedic tabletop vertical** | Molecular storytelling + AYUSH-compatible + EU DPP-ready | No 3D commerce platform is pharma-vertical. We are |
+
+**No competitor assembles even 3 of these 9. We assemble all 9 for one vertical.**
+
+## Three moats (supporting the composition)
+
+### 1. Regulatory + Molecular Provenance
+Every rendered ad, configurator asset, and SKU is provenance-signed to a PDB ID, a batch ID, a farm GPS coordinate, an HPLC lab report URI, and an OpenTimestamps hash on Bitcoin. No pharma competitor ships this. When EU DPP hits 2027, we're the default rails.
+
+### 2. Bidirectional 3D authoring
+Every competitor authors on desktop and consumes on mobile. Shoploop authors on mobile and desktop, consumes on mobile and desktop. Native iOS/Android app with ARKit/ARCore + MediaPipe finger tracking captures the harvest on-site, publishes to trace page in minutes.
+
+### 3. The Agent Layer
+The 10 loops running on the customer's own data via Nango. We don't sell software. We sell a 24/7 marketing employee whose weights compound on their data. Weights live in the customer's own repo (AGPL). No closed-source competitor can match this.
+
+## Design partner boundary
+
+An Ayurvedic pharma operator (introduced through personal relationship, floated $10k/$20k/$100k investment) is currently evaluating Shoploop. **Their brand name never appears in code, docs, reference sites, or any customer-facing surface.** The brand-neutral Ayurveda demo (ashwagandha.shoploop.studio) is Shoploop's own reference implementation, built on generic molecules and a fictional SKU. This boundary is enforced by system prompt.
+
+## Anti-scope-drift rule
+
+One new stack component per session, max. If a session generates more than one new framework/service/product surface, stop, list 3 goals, converge. Offense is one thing landed, not ten things planned.