The gate, in your language.
On the gated text channels — web chat and Telegram — a zenrep's reply is constructed and checked before it reaches the HCP: grounded in your MLR-approved corpus, verified claim by claim, fair-balanced, and written to an inspectable record. This page is the mechanism an MLR reviewer needs to sign off.
The path every answer takes
It can only speak approved material.
There's no general medical model improvising underneath. A zenrep cannot assert anything that isn't in your medically and legally approved source material — that corpus is the ceiling on what it can say.
Every claim is checked before it's delivered.
Each turn passes the pipeline: in scope, fair-balanced, grounded in an approved passage, and citable. Then, and only then, is it spoken.
It refuses rather than guesses.
When a claim can't be grounded in approved material, the gate is fail-closed: the zenrep declines, redirects, or routes the question to a human. Refusing correctly is treated as a success, not a failure.
Every turn is audit-traceable.
Each exchange records what was asked, what was answered, what the gate decided, and the exact approved passage it was anchored to — reconstructable for your compliance team any time, and exportable as CSV.
What the gate enforces, control by control
Corpus-grounding
A zenrep answers only from the MLR-approved materials you load — product monograph, ISI, approved claims. There is no general-model answer underneath.
Per-claim verification
Every factual claim in a drafted reply is checked against your approved corpus. If any claim can't be supported, the whole reply is refused rather than sent — there is no partially-approved answer.
Fail-closed
When a question can't be answered from the corpus, a zenrep refuses and routes to a human rather than guessing. Silence over speculation, by design.
Off-label refusal, fail-closed
Structural refusals (pediatric-on-adult, cross-tenant) fire before any model runs. Off-label and out-of-indication claims are refused at the fail-closed pre-send verifier, with deterministic per-class refusal templates — a reply that cannot be grounded on-label is never improvised.
Fair balance & ISI
Efficacy statements are automatically paired with the required safety information; a reply cannot send benefit without the balancing risk.
Adverse-event detection & PV routing
A two-layer detector — a deterministic pattern floor plus an LLM backstop — flags adverse-event reports, records the regulated intake, and pages pharmacovigilance. The deterministic layer is live-verified; the LLM backstop has shipped and is validated offline, with live verification in progress.
Immutable audit trail
Every turn is written to an append-only record with database-level tamper protection; reconstruct any conversation and export CSV for an inspector.
MLR-approval corpus workflow
Your MLR team reviews and approves the corpus at onboarding — nothing a zenrep can say has not already been cleared, and we walk your reviewers through the refusal templates the gate uses. Onboarding is "here is our approved corpus," not a code project.
Jurisdiction & class adaptation
The gate adapts its rules to the market and the product class: Canada (PAAB / Health Canada) and the United States (FDA OPDP) today, for both drug and device. Additional jurisdictions are on the roadmap.
What a zenrep will not do
- Make a claim that isn't in your approved corpus
- Send efficacy without the required safety information
- Guess when a question goes off-corpus — it refuses
- Speak off-label to sound helpful
- Talk to patients — HCP-facing only, by design
The proof is the receipt
The honest proof is the mechanism itself — talk to a live gated rep and open the audit view: every claim tagged to the exact approved passage it came from. Quantified verification metrics will appear here as we accumulate them under real corpora; we won't show a number we haven't earned.
Put the gate in front of your own corpus.
A demo runs against a corpus that looks like yours — bring the off-label probe, the comparison your data doesn't support, the safety edge case, and read the receipt.