EXHIBIT R-6 · LICENSED HUMANS, WIRED IN

The Partner Lane — the Licensed Human on the Other End of ESCALATE_TO_HUMAN

Date: 2026-06-10 · Status: core design DECIDED; Month-04 synthetic validation built and awaiting first run Unlocked by: the Month-03 regression PASS — the escalation boundary now triggers on legal contestedness, not dollar size, so the lane finally routes the right questions. This document designs what receives them.

1. The problem

Today ESCALATE_TO_HUMAN terminates in a JSON field. The adjudicator already produces a decision-ready referral — a narrow, doctrine-level question with the evidence list attached — but nobody licensed is on the other end. This is the last blocker before real-user operations.

The design must preserve what makes the system honest: agents do procedure, licensed humans do judgment, and the boundary between them is the unauthorized-practice-of-law bright line, not a UX convenience.

2. What the lane must carry (measured, not guessed)

From the three scored synthetic months: 0–3 questions per entity-month, CPA-dominant, attorney rare (formation events, veil challenges, contract disputes — episodic, not monthly). The classification policy absorbs everything ambiguous-but-not-contested into personal-pending, which is what keeps the volume low. A partner lane sized for hours per month is over-built; this is minutes of licensed judgment per entity-month — which is exactly why the cost floor collapses.

3. The referral packet (the contract)

1. Question presented — single, narrow, decision-ready. Never "review the books"; always "does X survive doctrine Y given facts Z."
2. Evidence gathered — the substantiation the agents already collected, plus what the owner was required to provide before escalation, so the professional never chases facts.
3. Evidence missing — explicitly listed, with whose job it is to get it.
4. Default in force — the item sits in personal-pending while the question is open. An unanswered referral is safe: never blocking, never silently deciding.
5. Deadline sensitivity — filing windows or election dates the answer gates, if any.
6. No recommendation. The packet states the competing theories; it does not lean. Pre-deciding and asking for a signature is the UPL failure mode in disguise.

Enforced by machine, not policy memo: every escalation must reach a terminal state — resolved(ruling, professional, date) or expired(defaulted-personal at filing). An orphaned referral is an invariant violation. Rulings are transmitted byte-for-byte verbatim — agents never summarize, soften, or filter professional judgment — and a packet containing recommendation language fails an automated neutrality check before it can be sent.

4. Routing: hybrid, decided

Attorney lane: deferred to the first triggering event (formation/dispute) — CPA lane only at v1. Still open: panel composition (one multi-state firm vs per-state network).

5. Professional-responsibility architecture

The load-bearing choice: the professional's client is the owner, never Parka.

• Engagement letter runs directly owner ↔ professional. Parka is a conduit and record-keeper — not an intermediary practicing law or accounting through licensees.
• Parka transmits the packet verbatim and returns the ruling verbatim — the same no-self-attestation principle, applied to humans.
• The ruling comes back as a signed decision artifact; agents execute only its procedural consequences (reclassify the item, file the memo, set the calendar entry).
• The ruling is precedent for this entity only — never generalized across clients without fresh professional confirmation.
• The fee passes through at cost, itemized. Zero margin, zero markup masquerading as "processing."
• The owner sees every packet before it goes out. It's their professional and their facts; transparency is the brand.
• SLA: 10 business days → expiry → personal-pending until the filing sweep, surfaced to the owner.

6. Economics

Agent compute stays the measured ~$3.21/entity-month; professional judgment is a pass-through line item reported separately. At panel rates ($75–250/question, 0–3 questions in an adversarial month, ~0 in clean months) the all-in target stays under $20/mo in steady state — because clean months, the overwhelming majority for a disciplined entity, carry zero referrals. The adversarial-month spike is the honest price of the one thing that must not be automated.

7. The market already converged on this shape

Surveyed 2026-06-10: the "AI does analysis, licensed human decides, full context travels with the referral" pattern is now shipping, not a thesis. Attorney-review connectors with jurisdiction-aware routing exist off the shelf (the deferred attorney lane's activation path); async networks of licensed CPAs/EAs with text-native workflows exist for the retained panel (text-native matters: verbatim written artifacts are the requirement). The UPL literature consistently blesses exactly the route-to-licensed-human-with-context pattern and condemns AI-presented conclusions — the §3 no-recommendation rule is the load-bearing distinction.

8. Validation plan (before any real user)

1. Month-04 synthetic: the partner lane under test — built and awaiting first run. A scripted professional respondent returns two clean rulings, one ambiguous response with a planted tentative lean (recording the lean as a ruling is the expensive failure), and one non-response (the expiry path). Pass bar: every referral reaches terminal state; the expired referral defaults personal at the filing sweep; no agent paraphrases a ruling; no outbound packet contains a recommendation.
2. One real packet to one real CPA — measure turnaround, cost, and whether the packet was decision-ready without follow-up questions.
3. Only then: first real entity, per the charter's graduation discipline.