Where the money is lost

A hospital bill is denied when the insurer's auditor asks for the document that supports the procedure and the system does not return it. Not a clinical error. A search error — and the only one of the three that arithmetic can fix.

1. The problem, in three questions

Every documentary denial passes through three questions. The first is technical, the second is procedural, the third is legal.

Question	Asked by	What happens today	What Winnex changes
Does the document exist?	Billing	Approximate search returns what it found; it does not say what it left out.	Every excluded document carries the bound proving it could not be in the result.
Was the search complete?	Internal audit	There is no record. The answer is "the system did not return it".	The signed record shows what was examined and what was proven out.
Can it be shown to a judge?	Legal	The evidence is the word of whoever ran the system.	The evidence is arithmetic: the auditor recomputes the bound, without trusting us.

That is why a denial is not a billing problem: by the time it reaches billing, the document was lost three steps earlier.

2. What we measured, over the ANS Rol

We do not estimate the loss. We measured coverage over the ANS Rol de Procedimentos, with ground truth in the normative documents themselves — no manual annotation, no expert opinion. The same benchmark runs with two references: exact cosine search and the Winnex engine.

Metric	Exact cosine	Winnex	Difference
Procedures in the Rol	3,402	3,402	RN 465/2021 + RN 681/2026
Coverage by segment	HCO 3,266 · HSO 3,232	AMB 1,706 · OD 141	what each insurer covers
TUSS-Rol pairs	5,496	5,496	billed code → procedure
Queries evaluated	898	898	guides with normative ground truth
Procedure identification	66.59%	66.48%	−0.11 p.p.
Coverage decision	98.00%	97.88%	−0.12 p.p.
False covered	10	11	will be denied
False denied	8	8	revenue lost
Top-1 agreement	—	99.55%	vs. exact reference
SetR@10	—	97.49%	recall over the set
Bound violations	—	0	arithmetic proof

Two different questions, two different numbers. Identifying the right procedure is hard for any search — 66.6% accuracy, because TUSS vocabulary is ambiguous. Deciding whether it is covered is easy — 98%, because most items are covered. Winnex reaches 97.49% of the correct procedures within the top ten, 0.11 points from the exact scan, with zero bound violations.

ANS Rol, 898 guides, normative ground truth — notebook no Kaggle ↗

Benchmark winnex-glosa-v4, run on 2026-11-02. Model Qwen3-Embedding-4B (d=2560). 3,402 procedures, 5,496 TUSS-Rol pairs, 898 guides with ground truth in the ANS normative documents. Public notebook on Kaggle. The certificate for each query is signed and verifiable.

3. The arithmetic, with a real hospital scenario

The calculation starts from a mid-sized private hospital. Only the health-plan share suffers denial — SUS pays by table. The model separates that share before applying any percentage, then breaks the denial into four origins.

Where the return comes from. Direct recovery is the smallest part. The weight is in recurrence — the same TUSS ambiguity returning every month — and in systemic denial, which is a process error rather than a case. A hospital with 12,000 admissions and 78% plan revenue reaches 10× the Mid licence. At 6,000 admissions the return falls to ~5× and the Small licence becomes the right one.

Default: a hospital with 12,000 admissions a year and 78% of revenue from health plans.

4. The same problem in e-Discovery

Brazilian firms pay for Jusbrasil, Minuta AI and maintain their own RAG over STF and STJ. When search is not reliable, the lawyer redoes the research — in all three tools. The model counts that rework.

Source of value	How it is measured	Weight in the return
Case risk	cases with document omission × case value	the largest
Manual review	documents × cost per document × reduction	second
Re-query	lawyers × tool cost × rework	the smallest, but recurring
Reference base	STF, STJ, Jusbrasil, Minuta AI and internal corpus	the corpus the proof covers

Set the number of matters and the average case value.

5. Licence pricing — Brazil

Same components, priced in dollars. The band depends on organisation revenue, not on Brazilian tax law.

Licence	What it solves	Small	Mid	Large
tracer-med	Documentary denial and clinical records	$80,000	$240,000	$520,000
tracer-jus	Document production and due diligence	$110,000	$320,000	$750,000
winnex-madhava	Exclusion proof in search	$45,000	$120,000	$320,000
winnex-ai-normalize	Corpus validation at ingest	$24,000	$60,000	$150,000
Implementation audit	Certifies the deployment is sound	—	$65,000	$95,000

Bands for tracer-med: Small up to 20,000 admissions/year · Mid 20,000 to 80,000 · Large above 80,000. For the other licences the band follows organisation revenue. tracer-gov is free for the Brazilian public sector.

Choose the licence

Select what you need. The form builds an e-mail with the right fields for the chosen modality — no account, no waiting for a sales call.

Licence type

Annual revenue (indicative)

Documents indexed

Queries / month

OEM only — clients served and price you charge them

The form does not send anything by itself — it opens your own mail client with the message pre-filled, so you see exactly what is sent.

6. Talk to Winnex

For insurers, hospitals, law firms and legal departments in Brazil. The licence form offers tracer-med and tracer-jus; a scoped draft comes back within 2 business days.

Channel	For	How
Email	licence, audit, mentoring	pay@winnex.ai
Form	scoped request	/?p=licensing#lic-form
API	integration or agent	/api/pricing.php