Root Cause Assessment, picked up cold

Decisions locked

All settled on or before 2026-08-11. Treat as decided unless Shawn reopens them.

Considered and rejected, do not re-propose

• $897 plus labs at cost. Client would pay $1,222, more than the bundle, for a less certain price.
• A menu of four panels. Flat price with client choice means everyone picks the most expensive one, and a woman who does not know why she feels terrible cannot choose between a DUTCH and a GI map anyway.
• A $1,697 DUTCH-inclusive tier. Superseded by the single-panel decision.
• Broad women's health positioning. Fertility, ADHD, perimenopause as marketing categories. Rejected in the meeting; note the sales page draft still contains this language and needs editing.

Money

Unit economics at $1,197

Compensation structure, settled

• Maria, $300 per assessment. All-in, her figure, covers the build.
• Maria, hourly increase from $25 to $32. This is for assistant work, content, and PMDD Rehab course production. Entirely separate from the clinic. It was NOT a request for system-owner compensation; that was an early misreading.
• Maria, 10% of clinic profit. Critically, on EVERY assessment the clinic delivers, not only the ones she personally works. If it only paid on her own clients the incentive runs backwards and she loses income by training anyone. Roughly $40 per assessment.
• Izzy, 10%. Base and clawback unresolved.

The hours problem, which is the real constraint

Maria's own breakdown, given unprompted in the meeting:

At $300 that is $33 to $75 an hour. Shawn's target, stated in the meeting: get it to a four-hour job. Jes independently reports her own lab reviews take about 3 hours even with Claude helping.

Competitor pricing, gathered live in the meeting

Other lab costs on file

GI map $350 · blood panel $265 · DUTCH + saliva cortisol $410 · HTMA $60 · clinic panel (comprehensive + 4 markers) $344

The pipeline

Full design at hmm-clinic-pipeline.pages.dev. Summary here so this document stands alone.

The two design rules that matter

Keep the two inputs blind to each other until stage 2. Hand a model labs and symptoms together and it rationalises, finding a story that fits regardless of evidence. Extracting each side independently makes the convergence at stage 2 real information.

Rules pick the priorities, the model writes the prose. Scoring must be deterministic: reproducible, auditable, correctable, and incapable of inventing a finding. A model asked to rank priorities produces a plausible list that varies run to run and cannot be fixed durably.

What is genuinely new

• An intake extractor, questionnaire to structured JSON, with the typical-day walkthrough as its own pass. Biggest new piece.
• An intake scorer mirroring _score_system() so both halves emit the same shape.
• The convergence step plus a reduction from 35 scored categories down to the five a client reads.
• A practitioner review gate, built BEFORE any generation.

Scoring shape

priority_score = lab_signal + intake_signal + convergence_bonus

lab_signal    = weighted sum of that priority's markers by deviation
intake_signal = weighted sum of symptom fields by severity
convergence   = large bonus when BOTH sides fire
divergence    = flag for practitioner attention, NOT a score penalty

Nervous system, sleep and lifestyle score from intake only. No blood marker reaches them, which is exactly why the questionnaire fixes matter.

Gotchas already found

• GlycoMark reads backwards. Low indicates postprandial glucose excursions, so its deviation logic is inverted relative to nearly every other marker. Encode explicitly.
• cycle_dependent already exists as a zone in the scoring model, so the machinery for cycle-varying markers is built and currently fed nothing. Capturing cycle day at draw activates it.
• Several strong signals are ratios, not single markers: triglyceride to HDL, percent transferrin saturation, free T3 to reverse T3, sodium to potassium, neutrophil to lymphocyte. Compute at extraction time.

Questionnaire changes, ready to apply

Full detail at hmm-clinic-questionnaire.pages.dev. Fifteen edits: 8 free-text fields converted to scales, 6 new questions, 1 structural. These are portable to any form builder, so they survive the Practice Better replacement.

The important one: the form never asks about mood in a structured, comparable way, on a product named the Mood Root Cause Assessment. Also missing: temperature intolerance, hair/skin/nail changes, dizziness on standing, cycle-timed headaches.

Assets and links

Notion, HMM workspace

The Notion MCP connector is the HMM tenant. There is no RVFD Notion. No delete tool exists on the connector. Note: get-teams reports a single teamspace named "Remove", which is worth a human look since it holds live business pages.

Canva, HMM account

Brand kits: Her Mood Mentor Brand Kit kAEcJRD4pHk, and WSH kAF7HwY0EgM.

Protocol template structure, 8 pages

Read once so the renderer does not have to re-derive it: (1) macro targets with a reason under each, plus building your plate, (2) small sustainable changes, (3) day of eating example with sample meal macros, (4) your next six months, phased month by month, (5) training plan Monday through Sunday plus workout resources, (6 and 7) good mood foods as food-group tiles each with a why, (8) nervous system support split into building safety, building resilience, and regulation techniques.

Source documents on disk

Work produced 2026-08-11

The decision meeting, 2026-08-11

Jes, Maria and Shawn, 58 minutes. Transcribed and diarized locally.

Open questions

Recommendations made but not yet acted on

• The Rehab handoff. Nothing in the SOP or sales page routes a finished client toward PMDD Rehab or the app. A protocol with no support layer gets roughly 30% compliance and produces "I spent $1,200 and nothing changed," which is a testimonial problem in a pilot whose product is testimonials. The fix is not coaching, which rebuilds the Jes bottleneck; it is that the day-7 follow-up and the final "your next 90 days" section point at Rehab. Building the habit tracker on shared schema with the app makes this structural rather than a call to action.
• AI consent should be a disclosure, not a consent. Once processing is local and first-party there is no third-party disclosure, so what is needed is transparency, not a withdrawal right. A withdrawal right creates a manual-only operational branch nobody has staffed. Caveat: this flips back to real consent the moment anything routes to a hosted model. Confirm the Washington My Health My Data angle with the attorney; HMM is not a HIPAA covered entity, so state consumer-health-data law is the live exposure.
• The $50 administrative fee should be justified in the document as the non-recoverable card processing cost, roughly $35 of it, rather than stated as a bare number.
• Cap generated habits at five to seven. Maria's own tracker is long because it is hers, built over years. A twenty-item checklist on a bad luteal day is a reason to close the app.

Infrastructure

Where things run

The lab pipeline codebase

Changed on the NAS, 2026-08-11

• New dataset tank-nvme/n8n-binary, primarycache=metadata, recordsize=1M, owned 1000:3100 mode 770. Mounted into n8n at /binarydata with N8N_BINARY_DATA_STORAGE_PATH=/binarydata. Large media now streams through without growing ARC.
• ARC cap lowered from 8GiB to 4GiB. Available memory went from 1.4GiB to 5.1GiB. Restore to 8GiB once a podcast episode confirms the binary path works, since hit rate was 99.2% at 8GiB.
• Not done: the ReadSource glob in workflow bisimwdlABebIAzJ is still /files/podcast-exports/*, which reads every file in the folder before the Switch filters to mp3 and mov. Left alone deliberately: the ARC fix removes the actual harm, and a glob that matches nothing fails silently on a live publishing pipeline.

Hardware

An M4 Pro Mac mini, 48GB, 512GB was ordered 2026-08-11 at $2,179 from B&H. It becomes the dedicated Ollama box; the current M4 16GB Mini stays as the services box running Cal.com, Home Assistant and Twingate. Reason: base M4 is 120GB/s memory bandwidth, M4 Pro is 273GB/s, and token generation is bandwidth bound, so the same model runs roughly 2.3x faster. 48GB lets a 14B and a 30B stay resident together. On arrival: move Ollama, set OLLAMA_NUM_PARALLEL and OLLAMA_MAX_LOADED_MODELS explicitly, and repoint the six n8n workflows that call 10.3.7.135:11434.

People, voice and constraints

Who does what

Scope of practice, which governs every word

HMM operates two voices and the correct one depends on the surface.

• Wellness-coach scope for anything public: website, marketing, sales page, app general copy. Avoid diagnostic labels. "Low nutrient status" not deficiency. "Sluggish thyroid function" not hypothyroid. "Blood sugar imbalances" not insulin resistance. "Mood fluctuations" not depression or anxiety. "Support" not treat or cure. "May help" not will help.
• NTP scope for practitioner-authored clinical content: the protocol, the findings video, the interpretation. Named foods, macros, micros, lab markers and mechanisms are all fine.

The clinic SOP is correctly in NTP scope because it is internal. The sales page is not, and currently names ADHD, anxiety, fertility and perimenopause as conditions. That is the open Notion edit.

Even in NTP scope, nobody may diagnose, prescribe, practise medicine, or make treat/cure/heal claims as marketing.

PHI handling, non-negotiable

All client health data stays local. Analysis runs on local Ollama. Never read raw patient data into a session. HMM_Labs on the NAS is encrypted with a passphrase Shawn holds in Proton Pass; Claude does not have it and there is no recovery path.

Writing constraints for anything produced here

• Never use em-dashes. Not the character, not the HTML entity. Commas, colons or sentence breaks.
• Substantial HTML docs use the Timber SAV house style from ~/.claude/artifact-house-style.html with an overview landing page and hash-routed sub-views.
• Deploy docs to Cloudflare Pages, one project per doc. Never a claude.ai artifact URL; they do not work for Shawn.
• Anything Shawn must execute goes in a numbered, indented block. Terminal commands stay single-line.
• Simplicity and usability outrank completeness and cleverness. Most users are women managing PMDD symptoms, often on a bad day.