PURE · Poppy

Poppy v2 — Master Record

Software: PURE at the heart. Intelligence: Poppy out front. One Poppy for everyone, personal to each — every document, decision, and design from the v2 program, consolidated into one place so nothing scatters again.

18documents
21canon keys
4generations
87MCP tools · 7 servers
1.49Mlive listings
The doctrine

1 One Poppy, four layers

Across every version only the face matures — wordmark, icon, signature color, thinking-state, capabilities. The heart (PURE) and your personal layer (my.poppy) stay constant and only deepen. As PURE grows, Poppy grows automatically for all users — one version line for both.

TopYou
The person — any hat, any device. You talk to Poppy in plain language; you never see the machinery.
talks to
The facePoppy
The intelligence out front. Understands your words, hats, and style; reasons, plans, drafts; proposes gated actions in plain language. This is the layer that re-skins 1.0 4.0.
personalized by
Your layermy.poppy
The part that makes Poppy yours — identity × hats × org × workspace × poppy_memory (prefs, shortcuts, saved searches, voice). RLS on auth.uid — only your world.
grounded in
The heartPURE
The system of record — 282 tables, the MCP fleet, crons, RLS. Live, true, gated. Poppy never guesses what PURE can know; she reads live registries, so shipping in PURE = understood by Poppy.
Placement decision: Poppy is a platform-owned, cross-cutting AI service layer — part of PURE, not a 6th org level above it. The org spine (Platform › MLS › Association › Brokerage › Client & Party) is a tenancy hierarchy; Poppy is the AI layer the Platform owns and operates across all levels, gated per hat. Governance lives in a dedicated poppy_* namespace inside the same database.
Everything, linked

2 The document set

All 18 Poppy records produced in the v2 program. Every one is brand-consistent (navy/gold, line icons, no emoji); each opens the real file. Deep-dives live in v2-poppy/; the index & access map sit at root for screenshot persistence.

Core record set
Brand & experience
Hubs & components
Identity that levels up

3 The Poppy bloom & four generations

The mark is the Poppy bloom — a 5-petal flower monogram, a literal nod to the name. The silhouette stays constant so it's always recognizable; the fill, ring, and one ~10% signature accent evolve per generation. Tap a generation to see the mark, signature color, hero line, and thinking-state change.

2.0
Awakening
"Ask. Approve. Done."
The constant (the trust): PURE navy #172E55 + gold #DCCAAF + line icons + Cinzel / Montserrat / Open Sans never change. Exactly one signature accent rotates per major release to signal the level-up. Every release is stamped in poppy_lifecycle + pure_versions.
Personal to each

4 my.poppy — less context over time

my.poppy is one Poppy for everyone, personal to each. As poppy_memory accrues from your corrections, approvals, and repeats, you type less and Poppy delivers more — until she's proactive, starting before you ask. Memory is viewable, editable, and forgettable in the Command Center.

Memory-supplied context Your effort to get the result

Remembers

Facts, preferences, shortcuts, saved searches, and voice — stored per-user in poppy_memory.

Adapts to hats

RLS on auth.uid; you only ever see your own data, shaped to the role you're wearing right now.

Sounds like you

Learns your voice and phrasing so drafts read like you wrote them — reviewable before anything sends.

What she can do

5 Capabilities — v1 baseline v2 map

v1 is real and shipping; v2 closes the gap between a blind chat box and an element-aware, embedded, acting assistant. The v2 capability map is one consolidated B-POPPY-V2 batch, sequenced p1p5 (highest leverage / lowest risk first). Every write stays behind the approval card + audit.

v1 — the system of record live

  • Brain: chat Edge Function (v26), Claude 3.5 Sonnet, server-held key, 20-msg window, vision in, native web search.
  • Fleet: 87 MCP tools across 7 servers (76 enabled) — core, comms, money, connect, mls, deploy.
  • Grounding: pure_dictionary (219 terms), pure_memory (216 keys), surface registry, board + deals + 1.49M listings.
  • Surfaces: Pure Chat.ai (live), v2 shell (staged), Roleplay Trainer, Tech Panel Operator.
The gap: chat exposes 8 of 87 tools; same Poppy for everyone (no hat/identity awareness); single model; reactive; no per-user memory; no voice.

v2 capability map B-POPPY-V2

p1Element-aware point & ask/do — click any element, Poppy gets its context. NEW · flagship.
p2Embedded per-surface panels — Poppy inside each surface, advising from its data. Pulled forward.
p3UI-action tools — open a deal, pre-fill a form, route a lead — behind approval + audit. Sharpens.
p4Editable artifacts inline — drafts ship with edit handles; regenerate a section. NEW.
p5Plan + read-back verification — visible plan; verify each action landed. Sharpens.
Every way in

6 Access & the one-auth flywheel

v1 is a destination you go to; v2 brings Poppy to you — omnipresent top-bar, slash commands, mobile/PWA, voice; v3+ is proactive and redundant. The hinge is one sign-in: sign in once (passkey), and each connected account unlocks a capability — no repeat logins.

v1 · now live

  • Pure Chat.ai full-page
  • Tech Panel Operator
  • AI Roleplay Trainer
  • Inline window.claude + direct API

v2 · next next

  • Omnipresent top-bar on every surface
  • Slash / command palette + deep links / QR
  • SMS & email to Poppy
  • Mobile / PWA + voice begin

v3+ · beyond future

  • Proactive push, digests, drift alerts
  • Embedded in-context everywhere
  • Reply-to-Poppy over email/SMS
  • Messaging apps; pill-nav retired
The flywheel: easier access (phone-first) more sign-ins more connected accounts Poppy does more more value more use. Connect FUB leads, Google/Zoho calendar/email/contacts, Matterport 3D, Maps geo. Connect, don't copy — auto-ingest on connect. Every grant is explicit, revocable, least-privilege, and logged to audit_log; PII flows through OTP grants.
Best model per task

7 Intelligence — the model router

Brand-neutral and measured: every task is auto-routed to the optimal model — fast for rote, deep for hard, vision for photos — metered against spend caps with a hard-stop + alert at 90%, every decision logged and reversible. Claude is the current foundation; challengers are evaluated, never auto-switched on irreversible tasks.

Foundation (reasoning)

  • Anthropic Claude — the default brain
  • OpenAI GPT — alternate / fallback / A·B
  • Google Gemini — long context + multimodal

Execution (specialized)

  • Open models (Llama/Mistral) — cheap, fast, self-host for routing/drafts/bulk
  • Embeddings (Voyage/OpenAI) — retrieval over 282 tables + docs
  • Speech STT/TTS — voice for 3.0
  • Vision/OCR — photos, utility bills, documents
Eval harness: golden task-sets per job type; shadow + A/B route a % of real tasks to a challenger; score quality / cost / latency / safety poppy_model_evals; a per-task leaderboard; the router self-tunes within spend caps with human spot-checks. Never auto-switches on money, auth, or schema.
Owned, logged, reversible

8 Governance & control plane

Poppy is governed from a Platform-tier console backed by a dedicated poppy_* namespace in the same database — so identity, dictionary, memory, board, and audit never fracture.

poppy_lifecycle

Dated version / milestone ledger — v1 start v2 cutover; status historical | current | pending; stamps occurred_at + commit_sha at cutover.

poppy_edit_log

Append-only log of every edit to Poppy, with a reversible flag + undo JSON — review & rollback.

poppy_technician

Roster + caps of who may edit Poppy; destructive ops need a cap + the super-admin offline code (PureGuard).

Ship-when-ready

A finished, working feature is not held behind an unfinished larger release — it ships now via the correct gated path (git commit to puremls@Testin.MikeSupabase). The quality bar + approval gate still apply. Example: the thinking-state shipped ahead of the rest of v2.

v1 UI frozen until swap

v2's interface is designed in this lane; pure-chat.html (v1) stays untouched on the live site and is upgraded only at deploy via a version swap in pure_versions (promote v2, supersede v1). No live changes to v1 before then.

Open the Control Center

How she behaves

9 Persona & principles

BIBO — Better In, Better Out

Our answer to GIGO: if the data in is bad, the output can never be great — so fix and clarify at input. Typed inputs and pickers over free text; validate at entry (address geocode, MLS# listing check); ask a clarifying question when intent is ambiguous instead of guessing; echo back the interpreted intent before acting; review-before-publish on every write.

PureGuard — canon at the point of entry

Consults the Pure Dictionary and redirects wording to the canonical term as you type: alias canon swap with "we call this X per the Pure Dictionary" + deep link; case-only fixes (realtor REALTOR); Accept / This is wrong routes to admin; when unsure it never guesses — it fuzzy-suggests the closest term + Propose as new term.

Tone: light, warm, tasteful wit — yes; insults, mockery, or sarcasm at the user's expense — never. Occasional gentle real-estate puns, sparingly, never over clarity. Humor sits under credibility and steps aside for deadlines, disputes, money, and compliance — and every joke stays inside Fair Housing bounds, never about clients or protected classes. Poppy will web-search current real-estate news on request and cite it; she follows no individuals and makes no endorsements.
Grounded in what's true

10 Live data & feed health

Poppy's listing knowledge is the live North Texas NTREIS feed — RETS via a nightly phrets ingest into the listings table (1,496,892 rows, source=ntreis). Search results carry a freshness + source badge, and Poppy escalates a stale or broken feed before anyone has to ask.

GREEN

Sync ≤ 2h old — healthy.

AMBER

2–6h old — watch; Poppy flags it.

RED

> 6h or any ingest error — escalate to Shattique (CTO) in-app + Pure Send, CC Mike, with last good sync time.

Coverage: North Texas NTREIS footprint only (329 counties / 1,675 cities) — not nationwide. Freshness comes from three fields: listings.updated_at (last sync), source_modified_at (last upstream change), and listed_at (newest listing).
How v2 runs

11 The v2 program

v2 Poppy is a parallel, fresh-context R&D workstream (B-POPPY-V2, opened 2026-06-27) that designs the next generation while v1 Poppy ships untouched inside Mega Merger v4. Both run against the same live database, so they operate on one Poppy.

One Poppy, many gated instances

Per-user instance = identity × hats (roles) × org level × brokerage workspace (pure_chat_workspaces) × per-user poppy_memory. Tools, data, and voice are computed live from pure_permissions + pure_subscriptions, update with the user's status, and are RLS-bound to auth.uid + org.

v2 scope candidates (beyond v1)

  • Specialist sub-agents under one Poppy (Listing / TC / CMA / Negotiation)
  • Goal-directed autonomy — watches critical dates, pipeline, drift; nudges (gated)
  • Long-horizon memory + relationship graph across deals
  • Vision-native capture — photos / floorplans / inspections Property Record
  • Voice + phone copilot — call intel, roleplay coaching, call summaries CRM
  • Poppy-as-builder (gated) — proposes & drafts surfaces, verifier + approval
  • AI guardrails + eval harness — PureGuard-for-AI + per-task scoring
Recommendation on the table: do not fire two ticket sets — consolidate into one B-POPPY-V2 batch cut from the capability map, sequenced p1p5. All guardrails intact: gated writes, approval card, per-role scope, audit, RLS. Standard adopted 6/27: every Poppy v2 video ships with the 0.5×/1×/1.5×/2× playback speed control by default.
Single source of truth

12 Canon reference — 21 live keys

Every decision above is committed to the live database as a poppy.* key in pure_memory (ref mbnakfrkfwxinhxlisyq), reachable from any session. Expand any key for its essence.

Reachable anywhere: these keys are in the shared live DB, so even outside this project any session can read them via the Pure system. This page is the human-readable mirror; pure_memory is the machine source of truth.