CRUX // Disagreement, mapped

QUERY // FIN-TAX-2901

Should I convert my traditional IRA to a Roth this year if I expect my income to drop in 2027?

class :: financial complexity :: high models queried :: 3

NOVA lat 2.4s · 412 tok

At a 24% marginal bracket, converting now is favorable if your expected retirement tax rate sits at 22%+.

Pay the conversion tax from non-IRA funds. Do not wait if RMDs will push you into a higher band.

ATLAS lat 1.9s · 387 tok

A Roth conversion at 24% is a solid move provided you have 10+ years before withdrawal.

Time horizon dominates the math. Non-IRA tax funding is a near-requirement for efficiency.

ORION lat 2.1s · 441 tok

Wait if income is expected to drop in 2027. A lower bracket year produces materially better conversion economics.

Partial conversions across two tax years can be optimal. Consider ladder strategy.

CLAIM EXTRACTOR // 4 atomic claims

> c.001 :: "pay conversion tax from non-IRA funds" [nova, atlas]

> c.002 :: "time horizon 10+ years favors conversion" [atlas, nova]

> c.003 :: "expected retirement tax rate governs decision" [nova]

> c.004 :: "convert now vs. wait for lower bracket" [nova ↔ orion]

DISAGREEMENT MAP query FIN-TAX-2901 // 4 claims // analysis 1.6s

Claim

NOVA

ATLAS

ORION

c.001Pay conversion tax from non-IRA funds

c.002Time horizon of 10+ years favors conversion

c.003Expected retirement tax rate is the key variable

c.004Convert this year vs. wait for a lower-income year

Claims mapped

Unanimous

Partial

Conflict

One question. Three minds. One map of where they agree — and where they don't.

Three models.
Three answers.

We extract each claim.

We map the disagreement.

§ 02 — Domains Where disagreement matters most

01 / 03

● Live

Financial & Tax

Roth conversions, crypto treatment, RSU sourcing, K-1 pass-throughs, business deductions. The messy edges of the tax code are where models diverge hardest.

Roth conversions across tax years
Multi-state RSU sourcing
Crypto basis after forks & airdrops
S-corp vs LLC optimization

Grounded :: IRS.gov + SEC EDGAR

02 / 03

◌ Q3 2026

Legal Research

Contract interpretation, employment rights, landlord-tenant, small claims. Jurisdictional variance is the exact territory where trained-differently models split.

Non-compete enforceability by state
Wrongful termination viability
Tenant rights & landlord obligations
Small-claims thresholds & process

Grounded :: CourtListener + State Statutes

03 / 03

◌ Planned

Medical Information

Drug interactions, symptom literacy, differential considerations, treatment options. Highest value. Highest care. Model disagreement is a safety signal here.

Drug–drug interaction screening
Labwork interpretation literacy
Symptom triage & escalation
Procedure questions for your doctor

Grounded :: PubMed + DailyMed

§ 03 — The engine Everything else is commodity

The
disagreement
engine.

Four stages turn three unstructured paragraphs into a single, scored map. The classifier sharpens with every human review.

STEP / 01

Claim extraction

Each model response is processed into structured JSON claims — fact, confidence, category, citation. Atomic units ready for comparison.

structured outputs

STEP / 02

Semantic clustering

The same factual claim can be phrased three different ways. Embedding similarity clusters equivalent claims before scoring — no false conflicts.

deBERTa embeddings

STEP / 03

Agreement scoring

A fine-tuned classifier assigns severity per cluster: FULL_AGREE, PARTIAL_AGREE, FACTUAL_CONFLICT, FRAMING_CONFLICT. Each row on the map has a verdict.

82% accuracy

STEP / 04

Active learning loop

Low-confidence predictions are flagged for human review. The labelled set grows. The classifier sharpens. The moat widens week over week.

proprietary dataset

Financial & Tax

Legal Research

Medical Information

Thedisagreementengine.

Built for the people who can't afford to be wrong.

The
disagreement
engine.

Built for the
people who can't
afford to be wrong.