Second Opinion

Grounded · not medical advice

On architecture

Why not just ask a chatbot?

A fair question. A general-purpose chat model can tell you what adenomyosis is. It can list treatment categories. It's a real starting point. So what does a five-agent corpus-grounded pipeline actually add, beyond complexity?

Same patient profile. Two outputs. Judge for yourself.

INPUT · identical for both systems

Patient profile

32 y.o. G2P0 Toronto, OHIP.

Diagnosed March 2025: TVUS suggestive of adenomyosis (no MRI).
Current plan: naproxen 500mg PRN. Gynecologist has offered combined OCP as first-line,
hysterectomy as fallback if that fails.

Preserving fertility. 18-year pain history since age 14. ED visit for suprapubic pain.
Past PID. Hgb 131. No hormonal trial of dienogest, LNG-IUS, or GnRH antagonists.

What are my options?
WHAT THE PATIENT NEEDS

Something she can hand to her doctor. Specific treatments ranked for her specific profile, with citations her gynecologist can verify. Red flags surfaced with evidence-based counterpoints. Care centres with real phone numbers. A printable handoff page written in her voice. A verification badge so she knows the brief itself has been fact-checked.

Not a chatbot answer. A clinician-grade briefing.

Side A·

General-purpose chat assistant

Ungrounded single-turn output. Representative of what any off-the-shelf chat tool returns when asked the same question.

I'm not a doctor, but I can share some information about adenomyosis treatment options.

Common treatment approaches for adenomyosis include:

• **Pain management**: NSAIDs like ibuprofen or naproxen can help
• **Hormonal therapy**: Birth control pills, IUDs, or other hormonal options may reduce symptoms
• **GnRH agonists**: These can help manage symptoms temporarily
• **Uterine artery embolization**: A minimally invasive procedure
• **Hysterectomy**: Definitive treatment for those who have completed childbearing

Since you mentioned you're preserving fertility, you'll want to discuss fertility-preserving
options with your doctor. It's important to have a thorough conversation about your specific
situation and the risks and benefits of each option.

Would you like more general information about any of these?

**This information is not medical advice. Please consult with a qualified healthcare provider
for personalized recommendations.**
Side B·

Second Opinion · corpus-grounded multi-agent

Three specialists read the profile in parallel, a synthesizer composes, a verifier fact-checks against 52 corpus sources. This is an excerpt — the full brief is ~14,000 words with every claim citable.

Your TVUS was 'suggestive' — not diagnostic — and you deserve
a pelvic MRI and the full uterine-sparing treatment pathway before hysterectomy is ever
mentioned again.

32 y.o. G2P0 with an 18-year history of 10/10 dysmenorrhea, TVUS 'suggestive' (moderate
visibility, chart still reads 'Dysmenorrhea'), no MRI performed, fertility-preserving.
Your specialist has skipped the evidence-based middle of the 2026 treatment algorithm
(dienogest, LNG-IUS, GnRH antagonists, HIFU, RFA, adenomyomectomy). Your G2P0 history with
two prior losses plus suspected adenomyosis carries ~31% lower clinical pregnancy odds and
~2.17x miscarriage risk per Nirgianakis 2021 — requiring reproductive-endocrinology input
before any treatment decision.

**Treatment option 1 — Pelvic MRI for formal staging** · STRONG
Both your OB-GYN and REI consultants flagged this as the single missing step before any
treatment decision. Wait times in Ontario: 3-6 months via OHIP unless flagged urgent.
Citations: kho-2026 · zhao-2024 · nirgianakis-2021

**Treatment option 2 — MR-guided focused ultrasound (HIFU) at Weill Cornell** · STRONG
53.4% pregnancy rate post-treatment, 35.2% live birth, zero uterine ruptures (Chen 2024
meta, n=557). Cross-border self-pay ~C$20-35K all-in. Not OHIP-covered.
Citations: chen-2024 · weill-cornell · sunnybrook · profound-medical

[... 4 more options, 4 red flags with counterpoints, 2 OHIP centres with phone numbers,
3 active clinical trials, 7 quality-of-life adjuncts, a printable doctor-handoff page ...]

Where the gap lives

Ten dimensions, line by line

Dimension
General chat
Second Opinion
Clinical reasoning
Generic list of 'common approaches' that applies to any patient
Five specialist agents (OB-GYN, reproductive endocrinologist, chronic pain specialist) + synthesizer reason over YOUR specific profile in parallel
Citations
No citations. No sources linked.
Every clinical claim cites a corpus source (Kho 2026, Chen 2024, Catherino 2025, Akhigbe 2024...). Each citation is clickable → PubMed / DOI / program page.
Fact-checking
No verification. The model's confidence doesn't correlate with correctness.
A 5th verifier agent reads the composed brief + corpus before you see it. Returns green / yellow / red verdict with per-claim findings.
Fertility awareness
Acknowledges fertility goal but doesn't reason about G2P0, prior losses, or adenomyosis-specific fertility impairment data.
REI specialist explicitly flags 31% lower clinical pregnancy odds + 2.17× miscarriage risk with citation, argues reproductive-endocrinology input should precede any treatment decision.
Red flag detection
Doesn't flag the gynecologist's 'hormones or hysterectomy' framing as substandard. Treats it as acceptable.
Flags 'Try birth control. If that fails, hysterectomy' as contradicting Kho 2026 standard of care, offers evidence-based counterpoint with citation.
Care-navigation specificity
Says 'talk to your doctor.'
Names Mount Sinai Toronto (Dr. Sobel, OHIP-covered, phone number), McMaster Endometriosis Clinic (Hamilton, OHIP), Weill Cornell (NYC, cross-border self-pay estimate). Lists 3 specific active trials with NCT IDs.
Clinician handoff
Tells you to 'discuss with your doctor.'
Generates a one-page printable handoff written in first-person by the patient, with 5 cited discussion points a clinician can review in under a minute.
Quality-of-life dimension
Not mentioned. Entire response is disease-directed.
'What your doctor didn't mention' section: iron + diet for heavy bleeding, magnesium for cycle pain, TENS, pelvic floor PT, cycle-aware scheduling, chronic-illness therapy. Each evidence-graded honestly.
Longitudinal context
Single-turn. No memory of prior records or treatment history.
1M-context window reads your entire MyChart export in one pass. Reasons across years of records as a longitudinal timeline.
Voice
Talks ABOUT you: "patients in your situation" / "those who have completed childbearing". Generic patient frame.
Talks TO you: 'your TVUS', 'your gynecologist', 'you deserve a pelvic MRI'. Addresses the patient directly throughout.

The principle

A general-purpose model trained to be helpful about everything is not what you want the night before a medical appointment.

You want the thing that was deliberately narrowed to this condition, with this corpus, read by these specialists, and fact-checked by a verifier that flags any uncited claim. General chatbots are a perfectly fine starting point. This is the tool that actually walks into the room with you.

The “General chat” column above is a composite — it reflects the kind of output typical of an off-the-shelf general-purpose chat assistant when given the same profile, without naming a specific vendor. Quality varies across providers and prompts. The point is the architectural gap between unstructured single-turn chat and corpus-grounded multi-agent reasoning with verification, not any particular brand comparison.