§B2 · The consult itself, in person

The procedure, end to end

Eleven steps, in order. Who acts, what happens, what is recorded — and, marked as such, the steps where no software is involved at all.

Modality

In person, at the dispensary. No telemedicine. This is the § 5(a) design decision — “I will not do it via telemedicine” — and it is stricter than Pennsylvania law requires: § 1161a.25(a) would permit synchronous interaction. Being narrower than the statute is the point.

Actors: the cardholder (recipient — no account), the pharmacist (professional), the door’s budtender, a Rishi reviewer.

The eleven steps

offline marks a step with no software in it. Software does not reach these steps, and the honesty of where it does and does not reach is a feature of this document.

01

Walt arrives, card checked at the door by dispensary staff

System

dispensary's system only

What happens / what is recorded

Rishi is not in this step and has no visibility into it
02

Budtender routes him to the consult desk — either because the certifier did not specify form/dosage (§ 1181a.27(c)(7)(ii) makes the consult mandatory here), or because Walt asked

System

offline — no software

What happens / what is recorded

GAP-13: nothing in Rishi records why a consult was triggered, so the mandatory-vs-elective split — which is exactly what Marcus wants to audit — is not measurable
03

The door's system issues an opaque consult_ref

System

dispensary's system

What happens / what is recorded

This is the entire identity handoff. The PHI stays on their side of the line
04

Priya opens the consult. She asks what else Walt takes before discussing any product

System

offline — no software

What happens / what is recorded

The conversion moment (A4). Also the clinically load-bearing one: warfarin, clobazam, tacrolimus, CNS depressants (memo 03 §4)
05

Protocol drives it. Priya works from the approved protocol for the condition cluster

System

protocols where status='approved'

What happens / what is recorded

A draft/in_review/retired protocol may never be used — enforced downstream at step 9 with 422
06

Products are restricted to this door's shelf, fresh

System

GET /api/shelf-observations/recommend?locationId=…&conditionCode=…&maxAgeDays=14

What happens / what is recorded

see below
07

Interaction screen runs

System

protocols.interaction_flags (jsonb) → read by a human

What happens / what is recorded

GAP-14. There is no interaction-screening engine. interaction_flags is content on the protocol; consult_records.interaction_flags_raised is a jsonb array the caller supplies. Grep confirms no route computes, matches or validates a flag. The screen is Priya reading the protocol's flag list against what Walt tells her, and then typing what she raised. This is the single largest gap between the pitch and the software.
08

Priya advises: form, starting dose, titration step, onset/duration, stop rules, what to do if nothing happens

System

offline — no software

What happens / what is recorded

from protocols.titration_steps, route, onset_profile, contraindications
09

Documentation written

System

POST /api/consult-records → consult_records

What happens / what is recorded

fields: consult_ref, protocol_id, protocol_version, condition_cluster, products_recommended (jsonb), interaction_flags_raised (jsonb), route_used, duration_min, plus location_id, professional_user_id, optional shift_id. No patient column exists. Starts qa_status='pending'
10

Priya may correct the record — only while pending

System

PATCH /api/consult-records/[id]

What happens / what is recorded

rejected once reviewed. A reviewed consult is immutable
11

QA review

System

POST /api/consult-records/[id]/review → passed \

What happens / what is recorded

flagged, with reviewed_by_id + reviewed_at from the session, qa_notes

Known limitations — stated, not buried

GAP-14 · blocks a pilot

There is no interaction-screening engine.

No interaction-screening engine. protocols.interaction_flags is content; consult_records.interaction_flags_raised is caller-supplied jsonb. Nothing computes, matches or validates a flag. The screen is a human reading a list

The screen is the pharmacist reading the protocol’s flag list against what the patient reports, and then typing what she raised. This is the single largest gap between the pitch and the software. Nothing computes, matches or validates a flag.

Journeys B2, C2 · the largest gap between the pitch and the software

GAP-13 · will be hit in the first month

Nothing records why a consult was triggered.

Nothing records why a consult was triggered — the mandatory § 1181a.27(c)(7)(ii) case vs elective. Exactly the split Marcus wants to audit

Journeys B2 · one column

GAP-15

POST /api/consult-records does not require an active coverage assignment or contract. The credential gate fires; the "was this door covered then" gate does not exist
GAP-16

products_recommended is unvalidated jsonb. Nothing checks the products recorded were on that door's fresh shelf — read path and write path are structurally disconnected
GAP-17

No consult UI. Every route is an API; what Priya uses at the desk does not exist

What is missing from this journey — the source list, verbatim

GAP-14 (above): no interaction-screening engine.
GAP-15: POST /api/consult-records does not require an active coverage_assignment or an active coverage_contract. A consult can be recorded at a door outside any covered window. The credential gate fires; the "was this door actually covered at that moment" gate does not exist.
GAP-16: products_recommended is an unvalidated jsonb array. Nothing checks that the products recorded were actually on that door's fresh shelf — the recommend endpoint is a read path, structurally disconnected from the write path. A consult can record a product the recommender would have refused to surface.
GAP-17: there is no consult UI. Every route above is an API. What Priya actually uses at the desk does not exist.

What the patient IS told

What form and what starting dose, and the titration step (Busse 2018's 1–2.5 mg THC increments framing).
Onset and duration for that route, so he knows when to judge it.
The interaction finding, named. "You're on warfarin. There is a Level 1 interaction signal and a published case of an INR at 7.2. I am not telling you not to do this; I am telling you that your INR needs checking, that your cardiologist needs to know, and that you do not change the dose between checks."
The stop rules and what "not working" looks like.
What the evidence actually says — small effect, high certainty for pain (Busse 2018, BMJ) — and that 15–19% of people in comparable cohorts do not respond at all (Allan 2018).
Which of the products in front of him are on this shelf today, with measured cannabinoid values and the lot's test date and lab.
That his name is not going into the pharmacist's company's records.

What the patient is NOT told

This list is the compliance surface. It is not trimmed.

That cannabis treats his back pain, his surgery outcome, or anything else. protocols carries a first-class what_this_does_not_claim column for this exact reason.
Any diagnosis. Priya does not diagnose. She works a condition cluster the certifier already named.
Anything about certification or renewal. Barred by 28 Pa. Code § 1161a.25(e) and by the noncompete. Referral only (§B3).
That a product is safe with warfarin. The honest statement is a flag and a referral, not a clearance.
That it replaces his oxycodone, his gabapentin, or a conversation with his prescriber.
A strain name or indica/sativa framing as though it carried clinical meaning.
A causal or comparative claim about a brand. Nothing Sloane bought entitles anyone to say a word to Walt about it.
Anything from a grade-D protocol (§B3).

What step 6 actually returns

GET /api/shelf-observations/recommend (app/api/shelf-observations/recommend/route.ts):

404 NO_APPROVED_PROTOCOL if the org has no status='approved' protocol for that conditionCode. Ranking is never attempted. No approved protocol, no recommendation — full stop.
404 LOCATION_NOT_FOUND for a cross-org door.
results[] — products on this door's location_products whose last_seen_date >= cutoff (default 14 days) and whose latest shelf_observations row is not in_stock = false.
stale[] — a separate array, never mixed into results (§B4).
Ranking: ratioClassMatch (MEASURED total_thc_pct/total_cbd_pct from the latest product_lots row, classified at a 2:1 dominance threshold) → then formMatch (protocol route → product form) → then name, then id. Terpenes, "entourage", indica/sativa and strain name are never ranking variables and are returned tagged "claimed — manufacturer assertions; display-only, NEVER a ranking or matching variable" (memo 03 §5; Jikomes 2022).
A claimRung string on every response: "rung-2/3: … the product is ON THE SHELF at this door and lab-MEASURED … It does NOT say the product works for this condition."