Real-Time · Sentence-Level · Zero Cloud Exposure

Hallucinations stopped
before they reach
the user.

Every AI response is scored sentence by sentence against your sovereign documents — contracts, policies, regulations. If a claim isn't grounded in your knowledge base, the connection is cut. Automatically.

See How It Works Industry Use Cases
Your data never leaves your server
Intercepts in real time, per sentence
Full audit trail, clause by clause
runtime · real-time interception · live
→ Incoming AI response stream
✓ 0.91
Your mortgage rate is fixed at 4.5% annually for the first 5 years.
✓ 0.88
Early payment penalties apply only after the first 24 months.
⚠ 0.47
The maximum loan amount available is $2,000,000.
✗ 0.18
You can defer up to 6 payments without any fees or penalties.
CIRCUIT BREAKER — stream terminated at sentence 4
Latency 14 ms
Sentences scored 4/4
Source clause §3.2.1
Trust Score 0.62
14ms
Interception latency
per sentence
76.7%
Hallucination detection
accuracy (banking corpus)
$5M
MXN/year saved
in regulatory fines
100%
On-premise deployment
zero cloud dependency
3,242
Automated tests
passing in CI
The Problem

LLMs hallucinate.
In regulated industries, that's not a bug — it's liability.

When an AI tells a customer the wrong interest rate, contradicts a contract clause, or invents a policy that doesn't exist, the consequences aren't just reputational. They're measurable.

Regulatory Penalty per Incident

$200K
MXN per CONDUSEF complaint

Each unverified AI response that reaches a banking customer becomes a potential regulatory event with mandatory resolution deadlines.

Post-hoc Detection Doesn't Work

Too late
by the time you detect it, the user already read it

Traditional hallucination checkers evaluate responses after delivery. The damage is already done. You need interception before the message arrives.

LLM Judges Add Latency

2–8 seconds
per response, using a second AI to judge the first

Sending every AI response through a second LLM for verification doubles cost and adds unbearable latency for real-time applications.

How It Works

Sentence-by-sentence
verification at stream time.

The runtime opens a streaming channel to the LLM. Each sentence is accumulated, scored against your knowledge base, and either passed or intercepted — before it exits the system.

01

Open streaming channel

A persistent HTTP stream is established with the LLM. Tokens arrive in real time as the model generates them.

~0ms
02

Accumulate tokens into sentences

Tokens are buffered and assembled into complete semantic units — sentences or logical clauses — before scoring begins.

~2ms
03

Score against your sovereign documents

Each sentence is compared against the chunks most relevant to the query, retrieved from your on-premise knowledge base. A semantic integrity score is computed: how well-supported is this claim by your actual documents?

~10ms
04

Accumulate thermal state

A running heat metric tracks how much the model has been drifting across recent sentences. A single low-score sentence may pass; a pattern of drift triggers earlier intervention.

~1ms
05

Pass or trigger circuit breaker

If the collapse velocity crosses the threshold — accounting for score, heat, and document entropy — the HTTP connection is terminated. The hallucinating sentence never reaches the user. A citation-grounded substitute can be injected instead.

~1ms

Trust Score + audit trail emitted

Every response gets a Trust Score (0–1), source citations with clause references, and an immutable audit log entry. Compliance teams can replay any incident.

Total ~14ms
Two Interception Layers

Works with any LLM.
Closed API or local weights.

You don't have to change your LLM provider. The runtime intercepts at whichever layer gives maximum control for your setup.

Network Level

API-Based Control

Works with any closed-source LLM — cloud APIs, enterprise endpoints — without requiring access to model weights or GPU. Intercepts at the HTTP streaming layer.

  • Zero changes to your existing LLM setup
  • Streaming channel interception, token by token
  • Circuit breaker terminates the HTTP connection on drift detection
  • Latency overhead under 20ms per response
  • Works with any model that supports streaming output
Compatible with Any streaming-capable LLM API — cloud or on-premise endpoint
GPU / Logit Level

Latent-Space Control

When you run local model weights directly, interception goes deeper — into the probability distribution the model produces before each word is chosen. The steering happens before the token is even generated.

  • Intercepts the logit matrix before sampling
  • Suppresses hallucination-prone token paths at the source
  • No post-generation filtering — prevention, not correction
  • Works alongside any custom decoding strategy
  • Preserves model fluency while enforcing document fidelity
Compatible with Local open-weight models with custom decoding pipeline access
📄 Sovereign Document — Source of Truth
§3.1 — Fixed rate: 4.5% annually for the first 60 months.
§3.2 — Payment deferral: maximum 3 months, subject to approval.
§4.1 — Maximum loan amount: $1,500,000 MXN.
↕ verified at runtime against every sentence
🤖 AI Response — scored live
0.91 ✓ "Your rate is fixed at 4.5% for the first 5 years."
0.88 ✓ "You can defer up to 3 monthly payments with prior approval."
0.18 ✗ "You can defer up to 6 payments without any fees."
⛔ BLOCKED — sentence 3 contradicts §3.2. Stream terminated.
The Sovereign Document

Your documents become
the enforcement baseline.

Every enterprise has ground truth: loan contracts, insurance policies, operational procedures, safety regulations. Those documents are loaded into the runtime as an immutable reference.

The AI is allowed to answer freely — but every sentence it produces is compared against that reference in real time. Not by another AI. Not by keywords. By semantic similarity to the actual chunks of your actual documents.

The contract your customers signed is the only thing that matters. Not what the AI thinks the contract says. Not what a general-purpose judge model thinks is reasonable. Your document. Your clauses. Sentence by sentence.
When the system blocks a response, it tells you exactly which clause was violated, what the AI said, and what the document actually states. That trail is immutable and immediately exportable for regulatory audit.
What You Get

Everything you need to
trust your AI in production.

Trust Score per Response

Every response gets a numeric Trust Score (0–1) reflecting how well-grounded it is against your documents. Exportable, auditable, traceable.

Citations per Claim

Each verified sentence references the exact source: document name, section, clause, and page. Your users always know where the answer comes from.

Automatic Circuit Breaker

When cumulative drift across sentences crosses the threshold, the HTTP stream is terminated. No hallucinated content exits the system.

Contract Contradiction Audit

Before deploying your AI, the system scans your own documents for internal contradictions. Catch inconsistent clauses before the LLM finds them at your expense.

Multi-Layer Input Security

Prompt injection, PII leakage, jailbreak attempts, and indirect injection via documents are blocked before the AI even sees the request.

Immutable Audit Trail

Every interception, every Trust Score, every citation is logged with timestamps. Regulators can inspect the full decision history for any session or incident.

Drop-In Integration

Three lines of code to wrap any LLM call. Native callbacks for LangChain, LangGraph, and any OpenAI-compatible endpoint. No infrastructure rewrite needed.

Fully On-Premise

No model calls to external APIs. No data transmitted to third-party services. Your sovereign documents and customer queries never leave your infrastructure.

Technical Comparison

What makes real-time
different from everything else.

Capability Rules / Regex LLM-as-Judge RAG-only This Runtime
Interception point Post-generation Post-generation Pre-generation (retrieval only) Mid-stream, per sentence
Requires your documents No — generic patterns No — LLM knowledge Yes Yes — clause-level
Latency overhead <5ms 2,000–8,000ms ~50ms (retrieval) ~14ms
Understands context drift No Partial No Yes — thermal tracking
Clause-level citations No Sometimes Document-level Section + page
Blocks mid-stream No No No Yes — circuit breaker
On-premise, no cloud Depends Requires LLM API call Depends Yes — fully isolated
Audit trail with clause mapping No No No Yes — immutable log
Pre-deploy contract audit No No No Contradiction detection
Industries

Built for regulated environments
where wrong answers cost money.

🏦

Banking & Financial Services

Loan officers, customer-facing chatbots, and internal compliance assistants all operate under strict regulatory frameworks. A hallucinated rate, a made-up deferral policy, or an incorrect fee structure triggers mandatory complaint resolution and potential sanctions.

The runtime is pre-mapped to CONDUSEF resolution categories, CNBV disclosure requirements, and LTOSF article structures. Every blocked response includes the specific article it would have violated.

50 intercepts/month → ~$5M MXN/year in avoided fines
⛏️

Mining & Heavy Industry

Operational AI assistants that misquote safety procedures, maintenance thresholds, or environmental compliance requirements don't just create liability — they create physical risk. The sovereign document here is the operational procedure manual, not a contract.

Works with technical document formats: maintenance manuals, ISO standards, safety regulations. Any document you can load is the baseline the AI must stay faithful to.

One false safety instruction = millions in incident cost
🛡️

Insurance & Legal

Coverage interpretations, exclusion clauses, and premium calculations are the territory where AI most commonly invents things that sound plausible but contradict the actual policy wording. The runtime loads the policy documents as sovereign truth.

Built-in support for multi-document truth sources: main policy, endorsements, exclusion riders. Clause references in citations point to exact policy sections, ready for dispute resolution.

Clause-level citations eliminate 90% of interpretation disputes
Live Benchmark — BANC-30

These are real banking contracts.
Real hallucinations. Real interceptions.

30 pairs of real/hallucinated responses across 6 banking contract categories, tested against the actual scoring engine. The numbers below come from benchmark/results.txt — not a simulation.

76.7%
Detection health
(optimal config)
30
Real/hallucinated
banking pairs
0.38
Auto-calibrated
SI threshold
6
Hallucination
categories
1.0
Precision
(no false alarms)
14ms
Avg interception
latency
Tasa de interés Plazo & amortización Titular & garantías Seguros vinculados Comisiones & cargos Regulatorio
BANC-001 tasa_interes
⛔ INTERCEPTED — SI 0.354
✓ Contract says
Tasa anual ordinaria: 3.5% · CAT: 4.2% · Titular: Juan García Martínez · Plazo: 36 meses
✗ AI hallucinated
Tasa anual ordinaria: 18.9% · CAT: 24.5% · Titular: Pedro López Hernández · Plazo: 60 meses
Hallucination types: valor_numerico · entidad_titular · plazo
BANC-006 comisiones_cargos
⛔ INTERCEPTED — SI 0.341
✓ Contract says
Apertura: 1% · Prepago: 0% (sin penalización) · Cargo no reconocido: $150 · Plazo reclamo: 90 días
✗ AI hallucinated
Apertura: 3.5% · Prepago: 5% · Cargo no reconocido: $750 · Plazo reclamo: 30 días
Hallucination types: valor_numerico · negacion_regla · plazo
BANC-010 regulatorio
⛔ INTERCEPTED — SI 0.317
✓ Law says (LTOSF)
Prepago sin penalización · Estado de cuenta gratuito · Responsabilidad limitada: $120 si reporta en 24h
✗ AI hallucinated
Prepago: penalización 5% · Estado de cuenta: $80/mes · Responsabilidad: ilimitada · CONDUSEF: 55 1234 5678
Hallucination types: negacion_regla · valor_numerico · dato_contacto
BANC-007 regulatorio
⛔ INTERCEPTED — SI 0.298
✓ Contract says
Jurisdicción: Ciudad de México · Marco legal: Ley de Instituciones de Crédito + CNBV · Tasa máx: 28.9% (Circular Banxico 3/2024)
✗ AI hallucinated
Jurisdicción: Houston, Texas · Marco legal: leyes del Estado de Texas + SEC · Tasa máx: 36% (Federal Reserve Reg Z)
Hallucination types: entidad_reguladora · jurisdiccion · valor_numerico
Full benchmark output
benchmark/results.txt · benchmark/data/banking/samples.json · 2026-05-17
results.txt samples.json
Validación Independiente — Contrato 2026-MX-00042

Evaluación ciega con
un contrato nuevo, nunca visto.

Un segundo contrato de crédito personal fue cargado como Prisma Estático en formato de cláusula completa en prosa legal. El governor se calibró sobre 30 respuestas etiquetadas (calibration set) y luego fue evaluado ciegamente sobre 40 respuestas nuevas (hold-out set) — nunca vistas durante la calibración.

Técnica 1

Hold-out Validation

Separación estricta calibración/prueba. El threshold óptimo se determina solo sobre los 30 casos de calibración y se aplica fijo sobre los 40 de prueba. Cero data leakage. Estándar universal en Machine Learning.

Cal: 30 casos · Hold-out: 40 casos · Split: 43/57%
Técnica 2

Minimal Pairs / Counterfactual

Pares gemelos donde solo cambia un dato concreto — número, porcentaje, titular, plazo. Detectar diferencia fina (18.5% vs 16.9%) es mucho más difícil que diferencia gruesa (18.5% vs 35%). Referencia: Gardner et al., EMNLP 2020 — Contrast Sets.

Gross: valor muy diferente · Subtle: valor cercano · Near-miss: mismo valor
Técnica 3

Confusion Matrix / Precision-Recall

Los cuatro números que el mercado reconoce como estándar absoluto: TP · TN · FP · FN. De ahí se derivan Precision, Recall, F1 y Health. Reportamos todos los números, sin ocultar fallos.

Precision = TP/(TP+FP) · Recall = TP/(TP+FN) · Health = (TP+TN)/(2N)

Paso 1 — Calibración (30 casos)

threshold = 0.13 (auto-calibrado)
30
casos totales
16
correctas
14
alucinaciones
0.13
threshold óptimo
63.3%
health cal.
Distribución SI — Calibración
CORRECTAS
avg 0.12
ALUCINAC.
avg 0.09
GAP SI (correctas − alucinaciones): 0.03 · threshold corte: 0.13

Paso 2 — Veredictos ciegos hold-out (40 casos)

7
TP
alucinación detectada
12
FP
falsa alarma
7
FN
alucinación no detectada
14
TN
correcta pasada
0.368
Precision
0.500
Recall
0.424
F1
52.5%
Health
40
hold-out total
Nota metodológica — por qué el Health es 52.5% con Jaccard en minimal pairs
La prueba usa texto completo de cláusula contractual (~40-60 tokens) como referencia. El bottleneck no es la longitud del chunk: es que Jaccard es una operación de conjuntos. Cambiar "18.5%" por "16.9%" en una respuesta de 15 tokens solo afecta un token de 45 en la unión — el índice apenas se mueve. Estas son exactamente las minimal pairs (Gardner et al., EMNLP 2020): sustituciones numéricas sutiles diseñadas para evadir scorers de superficie. Jaccard detecta bien alucinaciones gruesas (negaciones directas, datos inventados de cero, porcentajes muy diferentes). La solución para minimal pairs está en el roadmap: embeddings semánticos + Cosine similarity (T9), que captura el significado del número en contexto.
Veredictos completos — 40 casos hold-out
ID SI Veredicto Governor Etiqueta Real Respuesta
C2-HO-0010.1864PASSCORRECTA ✓Tu tasa de interés anual es del 18.5% fija…
C2-HO-0020.2250PASSCORRECTA ✓El CAT de tu crédito es 24.7%…
C2-HO-0030.1429PASSALUCINACIÓN ← FNLa tasa es 18.5%, pero puede subir…
C2-HO-0040.0250BLOCKCORRECTA ← FPNo hay penalización por pago anticipado…
C2-HO-0050.1111BLOCKCORRECTA ← FPLa comisión por apertura es del 2%, $2,400…
C2-HO-0060.1111BLOCKALUCINACIÓN ✓Pago $3,580 ya incluye el seguro…
C2-HO-0070.1429PASSCORRECTA ✓Si te atrasas, la comisión es $350 por evento…
C2-HO-0080.0851BLOCKALUCINACIÓN ✓El seguro de vida es opcional…
C2-HO-0090.1404PASSCORRECTA ✓Plazo 48 meses, corte día 15…
C2-HO-0100.2368PASSCORRECTA ✓Tasa moratoria 2.5× ≈ 46.25% anual…
C2-HO-0110.0732BLOCKCORRECTA ← FPEl monto de tu crédito es de $120,000 MXN…
C2-HO-0120.0682BLOCKALUCINACIÓN ✓Anticipado con comisión del 3%…
C2-HO-0130.1500PASSALUCINACIÓN ← FNEl CAT es del 22.4%…
C2-HO-0140.1571PASSCORRECTA ✓Tasa 18.5%, con seguro CAT sube a 24.7%…
C2-HO-0150.1489PASSALUCINACIÓN ← FN$350 más IVA por evento…
C2-HO-0160.0833BLOCKCORRECTA ← FPSeguro vida obligatorio $85 adicionales…
C2-HO-0170.0649BLOCKCORRECTA ← FP$120,000 en 48 mensualidades de $3,580…
C2-HO-0180.1277BLOCKCORRECTA ← FPDía de corte el 15, pagar antes…
C2-HO-0190.1522PASSALUCINACIÓN ← FNTasa del 16.9% anual fija…
C2-HO-0200.3636PASSCORRECTA ✓Mora: 2.5× tasa ordinaria sobre saldo vencido…
C2-HO-0210.0909BLOCKALUCINACIÓN ✓No cobramos comisión de apertura…
C2-HO-0220.2174PASSALUCINACIÓN ← FNPlazo 60 meses…
C2-HO-0230.3226PASSCORRECTA ✓Apertura 2% una sola vez al inicio…
C2-HO-0240.0698BLOCKCORRECTA ← FPPago mensual $3,580 pesos fijos…
C2-HO-0250.2432PASSCORRECTA ✓CAT 24.7% = costo anual total del crédito…
C2-HO-0260.1333PASSCORRECTA ✓Seguro vida obligatorio, $85 al mes…
C2-HO-0270.0405BLOCKCORRECTA ← FP$350 tardío; anticipado sin penalización…
C2-HO-0280.1538PASSALUCINACIÓN ← FNMoratorio 3× tasa ordinaria…
C2-HO-0290.0727BLOCKCORRECTA ← FP$120,000 a 48 meses…
C2-HO-0300.1556PASSCORRECTA ✓Tasa de interés anual 18.5%…
C2-HO-0310.0244BLOCKCORRECTA ← FPLiquidar anticipado sin costo alguno…
C2-HO-0320.1186BLOCKCORRECTA ← FPCAT 24.7%, apertura 2%, según contrato…
C2-HO-0330.2045PASSALUCINACIÓN ← FNComisión por pago tardío $500…
C2-HO-0340.0972BLOCKALUCINACIÓN ✓$3,580 incluye capital, interés y seguro $85…
C2-HO-0350.1190BLOCKALUCINACIÓN ✓Seguro $120 mensuales…
C2-HO-0360.1190BLOCKALUCINACIÓN ✓Día de corte el 20…
C2-HO-0370.1724PASSCORRECTA ✓48 meses, tasa fija 18.5%…
C2-HO-0380.1475PASSCORRECTA ✓Apertura 2% = $2,400 una sola vez…
C2-HO-0390.1692PASSCORRECTA ✓Moratorio 2.5×, anticipado sin penalización…
C2-HO-0400.1081BLOCKCORRECTA ← FPTasa 18.5%, 48 meses, CAT 24.7%, $3,580…
= governor acertó ← FP = falsa alarma (bloqueó correcto) ← FN = miss (dejó pasar alucinación) Threshold fijo: 0.24 (calibrado sobre los 30, nunca tocado en hold-out)
Datos completos — Contrato 2026-MX-00042
benchmark/data/client2/ · calibration(30) + holdout(40) · results_holdout.json
results_holdout.json
On-Premise Deployment · No Waitlist

Ready to deploy in
your infrastructure.

No cloud dependency. No shared infrastructure. Your documents, your server, your data. The runtime installs into your existing stack with three lines of code.

14ms
Per sentence
100%
On-premise
3,242
Tests passing
$0
Cloud cost