reference
Phase 1-2 Friction Points
A companion one-pager surfacing seven decisions where the May architectural commitments conflict with the v1 phased plan — each with a recommendation and named decider — designed to drive a 30–45 minute sign-off session before Rev 2 is approved.
Drafted: May 5, 2026
Companion to: phased-plan-v2.md
Audience: Blair, Seth, Cindy
Purpose: Surface the seven decisions where the architectural commitments don’t quite fit the v1 phased plan, with recommendations. Designed to drive a 30–45 minute working session before signing off on Rev 2.
How to read this
Each friction point follows the same shape:
- Situation — what’s happening now in the v1 plan
- The friction — why the architecture and the plan disagree
- Recommendation — what Rev 2 proposes
- Decision needed — the specific call to make in this meeting
- Decider — who has the call
The friction points are ordered by consequence to launch, not by where they sit in the timeline.
1. The brain content lead question
Situation. V1 of the phased plan has Sarah Olaleye joining in Phase 3 (July) and being responsible for marketing-site copy + launch content + AEO + customer education + email campaigns + brain authoring + cross-pollination review.
The friction. That’s not one job — it’s three. The brain authoring and cross-pollination review require editorial judgment, MP domain knowledge, and privacy literacy (catching re-identification risk in cross-pollinated patterns). The marketing/launch content role is closer to Sarah’s existing skill set. Collapsing them into one role means brain content gets ~10 weeks of authoring time before GA — likely landing the SPEC’s #1 risk (differentiation from generic AI) on the wrong side of credible.
Recommendation. Hire a dedicated brain content lead in Phase 1 (May), separate from Sarah. The brain lead’s only job is brain authoring + cross-pollination review + eval scenario authoring + thumbs-down review. Sarah joins on the original Phase 3 schedule for her existing domain. By GA there are two content roles; by V1.5 the SPEC v2 already calls for the second hire — this just pulls it forward by ~6 months.
Decision needed. Approve the brain content lead hire for Phase 1, or accept that V1’s differentiation lands on the SPEC’s optimistic case rather than the realistic one.
Decider. Blair (with Cindy + Seth input). Budget call. This is the highest-leverage hire decision in the project.
2. Phase 2 milestone language is optimistic
Situation. V1’s Phase 2 milestone reads: “Initial chat advisor working with Haiku model (basic Free experience).”
The friction. Phase 2 is four weeks with Cindy, Seth, Ronald, and a brand-new Senior AI Engineer still ramping. The realistic Phase 2 deliverable is auth + sync + entitlement + chat skeleton (a route returns a Haiku response). Citations, eval coverage, per-customer brain integration, tier-routing polish — those are Phase 3 work. The v1 milestone language reads as if all of that ships in June, which it doesn’t.
Recommendation. Rev 2 calibrates the language: “Initial chat skeleton: a route returns a Haiku response (citations, eval, per-customer brain → Phase 3).” Same trajectory, honest description.
Decision needed. Acknowledge the calibration. No scope change; just stop calling it “advisor” in Phase 2.
Decider. Blair, Cindy, Seth jointly. This is a labeling fix more than a planning change.
3. The differentiation eval needs a Phase 3 baseline, not a Phase 5 discovery
Situation. V1 of the plan doesn’t call out the differentiation eval (MemberIntel scored vs vanilla Sonnet on MP-operator scenarios) as a milestone item. It’s implicit in “AI eval suite operational” at Phase 4.
The friction. The SPEC’s #1 risk is differentiation from generic AI. The eval suite is the primary instrument for measuring whether the brain + retrieval + per-customer context are pulling their weight. If the gap-vs-baseline isn’t widening by month 3 (end of Phase 3), the team has time to course-correct: more playbooks, retrieval tuning, prompt iteration. If the discovery happens at Phase 5, the only correction is “slip GA.”
Recommendation. Make the differentiation eval subset (30–50 scenarios scored against baseline LLM, with a baseline measurement) a Phase 3 gate item. Reviewed monthly with Blair starting Phase 4.
Decision needed. Adopt the differentiation eval as a Phase 3 milestone and a monthly executive metric.
Decider. Blair. The monthly review is on his calendar.
4. Privacy counsel needs a substantive Phase 1 architecture review, not a Phase 3 deep-dive
Situation. V1 has counsel engaged by June 1 with compliance work intensifying in Phase 3 (PRDs for cross-pollination opt-out, ToS, etc.).
The friction. Counsel’s input shapes the cross-pollination security boundary, the audit dataset structure, the secrets management approach, the data deletion pathway, and the consent flow. These are Phase 1–2 architectural decisions. If counsel sees them for the first time in Phase 3, the rework cost is real and the GA timeline tightens.
Recommendation. Schedule a substantive architecture review with counsel + Seth + Cindy in late May (before May 31 milestone). Not a get-acquainted meeting — a working review of the per-tenant isolation, audit, and cross-pollination architecture. This is a half-day investment that pays back across Phases 2–4.
Decision needed. Approve the late-May counsel review and budget Seth’s time for prep (one day).
Decider. Cindy (counsel engagement is her deliverable per JD).
5. The infrastructure-leaning hire belongs on the V1.5 roadmap, not deferred to V2
Situation. V1 has steady-state team at 4 dedicated engineers post-launch (Cindy + Seth + Ronald + Senior AI Engineer). V2 (BuddyBoss) considers further hiring.
The friction. Once V1.5 ships the agent + greenfield wizard, the system has materially more failure surface (write actions on customer sites, multi-step plans, undo orchestration, expanded eval scope). Two engineers (Seth and Senior AI Engineer) absorbing ops on top of feature work is workable for V1; for V1.5 it’s a burnout risk. The architecture we’ve built is good but not magic — kill switches and runbooks reduce ops load, they don’t eliminate it.
Recommendation. Plan for an infrastructure-leaning engineer as the next hire after Senior AI Engineer, slotted into V1.5 ramp (Q1 2027) rather than V2 (mid-2027). Not a Phase 1 decision — just a budget signal so it’s on the roadmap.
Decision needed. Acknowledge the hire on the V1.5 roadmap. Specific timing and scope can land later.
Decider. Blair. Budget signal.
6. The model abstraction layer is a Phase 1 architectural decision, not “we’ll add it later”
Situation. V1 of the plan and the SPEC both say “Anthropic API directly, no LangChain or other orchestration framework in V1.”
The friction. That’s the right call for orchestration frameworks. It’s the wrong call for the SDK boundary. Calling Anthropic SDKs directly throughout the codebase means a future strategic question (“should we have a backup model provider for resilience or cost”) becomes a multi-week refactor across every LLM call site. A thin abstraction layer that wraps the SDK — one day of work in Phase 1 — makes the future swap a config change at the wrapper level.
This isn’t about implementing a second provider. It’s about not foreclosing the option.
Recommendation. Add the model abstraction layer to Phase 1 architectural commitments. Implementation is one day. Refactor cost if deferred is a month-plus.
Decision needed. Approve the abstraction layer as a Phase 1 deliverable.
Decider. Seth (technical architecture is his domain per the decision-rights matrix).
7. Per-customer brain entries need versioning from day one
Situation. V1 of the plan and the SPEC describe per-customer brain entries as updateable markdown files, with update_customer_brain as a tool call. No mention of versioning.
The friction. The per-customer brain is the most sensitive data store in the system — it contains the customer’s specific thoughts about their business in their own words. RLS protects against unauthorized access; it doesn’t protect against accidental corruption (a buggy update_customer_brain call) or against the “the AI used to know X about me and now it doesn’t” customer experience. Versioning is the answer: every update is a versioned write, prior versions retained, customer can see history and revert.
Storage cost at this scale is negligible. Adding versioning to the schema in Phase 1 is a column and an index. Adding it after launch is migration work plus a customer comms moment (“we’re improving how your brain entries are stored”).
Recommendation. Add per-customer brain versioning to the Phase 1 schema design. Same data, plus a version column, plus a superseded_at timestamp. Done.
Decision needed. Approve the schema addition. Cost: ~30 minutes of schema work in Phase 1.
Decider. Seth (schema design is his domain).
Decisions summary — for the meeting
| # | Decision | Decider | Cost / Risk |
|---|---|---|---|
| 1 | Approve brain content lead hire in Phase 1 | Blair | One additional hire vs v1; mitigates SPEC’s #1 risk |
| 2 | Calibrate Phase 2 milestone language | Joint | Labeling fix; no scope change |
| 3 | Differentiation eval as Phase 3 milestone + monthly review | Blair | Adds eval scope to Phase 3; surfaces #1 risk early |
| 4 | Substantive privacy counsel architecture review in late May | Cindy | Half-day Seth prep; saves rework across Phases 2–4 |
| 5 | Infrastructure-leaning hire on V1.5 roadmap | Blair | Budget signal only; specifics land later |
| 6 | Model abstraction layer in Phase 1 | Seth | One day of work; preserves provider optionality |
| 7 | Per-customer brain versioning in Phase 1 schema | Seth | 30 min of schema work; prevents future migration |
What this conversation is not
- Not a re-litigation of the SPEC. SPEC v1 stands.
- Not a re-litigation of the decision-rights matrix. The matrix stands.
- Not a re-litigation of GA timing (mid-October 2026). The architectural commitments fit the timeline if these seven decisions land cleanly in Phase 1.
- Not a budget conversation about V1 total spend. Items 1 and 5 have budget implications worth acknowledging; the rest are architectural or labeling.
What this conversation is
A 30–45 minute working session with Blair, Cindy, and Seth to decide the seven items above. After that, Rev 2 of the phased plan goes live and the team executes against it.
If the conversation lands on items 2, 3, 4, 6, and 7 cleanly and defers items 1 and 5 to a follow-up, that’s still a useful outcome — five of seven friction points resolved, two with budget implications kept on Blair’s desk.
Document version: Draft v1 — to be discussed and resolved before the phased plan v2 is signed off.