Two agents, not one magic model — Brand Onboarding + Screening

TL;DR. A fit-test programme does not need one magic model that promises ≥99% from day one. It needs a recipe: first onboard a mask brand with paired measurements and an honest gate; then screen new wearers only among brands that made it through. Two agents. Clear jobs. Evidence you can defend.

Full Distribute narrative (methods figure, step tables, anti-claims, EN + 繁體): see the marketing pages under marketing/two-agents-n95-pipeline.md (#763). This post is the short web twin.

The wrong slide

Buyers have seen this pitch:

"Our AI picks the right respirator. Accuracy ≥99%."

That sentence compresses three different claims into one badge:

1. The model can suggest among options.
2. Those options are safe to suggest (brands are actually characterised).
3. The measured pass rate on your programme matches a marketing number.

Job (1) without (2) and (3) is a demo. Job (3) without enough paired cases is a hope. Fit-test work deserves a calmer story.

The recipe: two agents

Think of a kitchen that learns a new dish, then serves it safely:

Agent	Facing	Job (plain language)
Brand Onboarding Agent	Nurse / fit-tester introducing a new mask brand	Baseline the brand → pair AI screening with a PortaCount (+/−) form → refine brand-specific weights → provisional promote when the gate is met
Screening Agent	Day-to-day fit-test programme	For a new wearer, suggest among onboarded brands only → confirm with PortaCount

Same platform. Different verbs. The Screening Agent must not invent a brand that never finished onboarding — that is the whole point of separating the jobs.

Naming note: some trial documents call the second agent a "Production Agent." In the product catalogue we use Screening Agent — same job, clearer name.

Honest evidence gates (say the number you have)

Founder-agreed gates for this recipe (order-of-magnitude, not a lab certificate):

Gate	Paired cases (ballpark)	What you may say
Provisional promote	≈ 30	Ops sweet spot — brand may enter the Screening Agent
≥99% claim	~100–300	Long-term production / marketing bar — not provable at n≈30

So: a brand can be useful in the programme long before anyone should print ≥99% on a slide. Publish the gate you are at. Hide nothing that would surprise an auditor later.

Why this shape is the product

Three buyer benefits fall out of the two-agent recipe:

1. Catalogue honesty — only onboarded brands appear as screening options.
2. Reusable method — the recipe (baseline → pair → refine → promote; screening loop) can travel to the next fit-test programme; brand data and trained heads stay customer-specific.
3. Governance-friendly — roles, audit, and confirmation steps are part of the agent cards, not an afterthought bolted onto a chat window.

This is agent packaging, not "one LLM with a respirator prompt."

How this connects to what we build

OrchAI Agents is where these named agents live as product cards — one login, the right people see the right tools, admins grow the catalogue from a registry. Library holds the governed knowledge the agents may need; Eval is the honest quality gate when you want regression-safe claims; Portal wraps the full building when the programme outgrows a single-team deploy.

If your team is past "one magic accuracy number" and ready to talk about onboarding gates + screening among known brands, talk to us or try the demo.

Frequently asked questions

Why two agents instead of one model? Onboarding a brand and screening a wearer are different jobs. Separating them keeps the catalogue honest: screen only among brands that passed onboarding.

When can we say ≥99%? Not at the provisional-promote sweet spot (~tens of paired cases). A high bar needs far more paired evidence — roughly hundreds. Say the gate you actually have.

What is OrchAI Agents here? The governed front door and catalogue for named agents — roles, registry, control — not an unbound chatbot inventing mask brands on the fly.