The formula
Four sub-scores. One composite. One threshold that changes what medicine can see.
Every prior system that detects discrepancies in health data uses a single signal: a value is outside a reference range, or two records disagree on a structured field. CORE C2D scores contradictions on four simultaneous dimensions, weighted and combined into a single C-Score that no prior art replicates.
// CORE C2D Contradiction Scoring Function
// Patent claim language, simplified for presentation
C_score = (w1 × E_sub) + (w2 × R_sub) + (w3 × T_sub) + (w4 × S_sub)
// where:
E_sub = NLI entailment score // do the two statements logically conflict?
R_sub = rule-match sub-score // deterministic domain rules (pain 2 vs pain 6)
T_sub = temporal conflict score // how close in time? same event, different accounts?
S_sub = source-role precedence // physician > caregiver > patient for clinical weight
// Default weights: w1=0.40, w2=0.25, w3=0.20, w4=0.15
// Detection threshold: TH_C = 0.6 (configurable)
if C_score > TH_C: contradiction_edge.seal(sub_scores, both_entries)
// Patent claim language, simplified for presentation
C_score = (w1 × E_sub) + (w2 × R_sub) + (w3 × T_sub) + (w4 × S_sub)
// where:
E_sub = NLI entailment score // do the two statements logically conflict?
R_sub = rule-match sub-score // deterministic domain rules (pain 2 vs pain 6)
T_sub = temporal conflict score // how close in time? same event, different accounts?
S_sub = source-role precedence // physician > caregiver > patient for clinical weight
// Default weights: w1=0.40, w2=0.25, w3=0.20, w4=0.15
// Detection threshold: TH_C = 0.6 (configurable)
if C_score > TH_C: contradiction_edge.seal(sub_scores, both_entries)
Worked example
Margaret's medication contradiction.
C-Score: 0.81
C-Score: 0.81
E_sub · NLI Entailment
0.87
Natural language inference detects semantic contradiction between "I remembered everything" and "needed prompting for medications." High entailment conflict.
R_sub · Rule Match
0.90
Deterministic rule: if patient reports medication independence and caregiver documents prompting required, R_sub = 0.9. Hard clinical contradiction.
T_sub · Temporal Conflict
0.95
Both entries sealed within 6 hours of each other on the same day. Maximum temporal weight — same event, irreconcilable accounts.
S_sub · Source Precedence
0.60
Caregiver vs. patient reporter pair. Neither source is clinically authoritative over the other — both accounts preserved. Moderate weight.
Composite C-Score
0.81
Threshold: TH_C = 0.60 · Status: CONTRADICTION SEALED
Both original entries preserved exactly as written. The contradiction edge exists alongside them — not replacing them.
Investor objection
"Anybody can compare two data points and flag a discrepancy. My spreadsheet does that."
Why that misses the point
A spreadsheet compares values. C2D compares accounts. A spreadsheet can tell you that a pain score of 2 and a pain score of 6 are different numbers. It cannot tell you that the 2 was self-reported by a patient who underreports pain chronically, that the 6 was observed by a caregiver on a cold morning four hours later, that a domain-specific rule flags this pair with 90% certainty, and that the temporal proximity makes this the same event described two incompatible ways. The composite score — four-dimensional, weighted, threshold-gated — is what the patent protects. It cannot be replicated by a simpler system without becoming a simpler system.
Investor objection
"AI can already detect contradictions in text. This isn't new."
Why that misses the point
NLI (natural language inference) for contradiction detection is well-established in academic literature. What is not established — and what the prior art search confirmed — is combining NLI entailment with deterministic rule scoring, temporal conflict weighting, and source-role precedence into a single composite score stored as a first-class graph edge with sub-score breakdown and resolution tracking. The novelty is not the NLI. The novelty is the four-component architecture applied to longitudinal, multi-reporter clinical evidence. Our patent attorney rated this cluster HIGH novelty after full prior art analysis.
What competitors must give up
| Approach | What they detect | What they miss |
|---|---|---|
| Threshold alert (Epic, most EHRs) | Value outside reference range | Reporter conflict, temporal proximity, source weight — invisible |
| Medication reconciliation AI (DrFirst) | Structured data field mismatch | NLI semantic conflict, no composite score, no graph persistence |
| NLI contradiction detection (academic) | Semantic contradiction in text pairs | No rule scoring, no temporal weighting, no source precedence, no clinical context |
| CORE C2D | All four dimensions simultaneously | Nothing — the composite is the patent |
40%
False positive rate in single-signal threshold alerting systems — eliminated by the four-component composite score
"Couldn't someone just use three sub-scores and avoid the patent?"
They can try. The moment they omit source-role precedence, they lose the ability to weight physician observations above patient self-report — a clinically essential distinction. Omit temporal conflict, they cannot distinguish same-event accounts from longitudinal change. Each sub-score is load-bearing.
"What happens to the original entries when a contradiction is flagged?"
Both are preserved exactly as sealed. The contradiction edge exists alongside them — neither account is overwritten, suppressed, or marked as wrong. The vault holds the contradiction as evidence, not as an error to be resolved. That is a legally and clinically distinct design.
The C2D engine finds what
nine years of clinical data hides.
nine years of clinical data hides.
Claim Cluster 3 shows how we detect changes that never trigger a reference range alert — because we compare the patient to herself.
Read Claim Cluster 3 →