What How Why Compliance Domains Capabilities Use Cases Library Plans ROI Calculator Live Demo ↗
Pre-Execution AI Governance

Governance happens
before execution.
Or it doesn't happen at all.

Governance findet vor
der Ausführung statt.
Oder gar nicht.

La gouvernance se fait
avant l'exécution.
Sinon elle n'a pas lieu.

La gobernanza ocurre
antes de la ejecución.
O no ocurre en absoluto.

La governance avviene
prima dell'esecuzione.
O non avviene affatto.

Yönetişim
yürütmeden önce gerçekleşir.
Yoksa hiç gerçekleşmez.

Управління відбувається
до виконання.
Або не відбувається взагалі.

الحوكمة تحدث
قبل التنفيذ.
أو لا تحدث على الإطلاق.

गवर्नेंस निष्पादन से
पहले होती है।
या बिल्कुल नहीं होती।

PREEXEC™ is a deterministic pre-execution governance architecture for AI systems. It evaluates every input structurally before it reaches a model — producing an auditable decision: EXECUTE, HOLD, or BLOCK.

Default catch 97.3% 15 industries · 8 130 evals · v4.7.0
Finance default 100% finance · 30 000 evals · 0.00% FP
Determinism 100% golden-record byte-identical · 10 050 evals
Latency p50 35 ms warm @4 vCPU · v4.7.0
Decisions ● EXECUTE  ● HOLD  ● BLOCK  · INTENT_DRIFT · UNGROUNDED_OUTPUT
Audit Chain SHA-256 + Merkle · ECDSA-signed · RFC-3161 TSA-anchored
Scroll to explore
What is PREEXEC? · 45-sec teaser
Click to play · with sound
00 · Who

Built for the
people accountable
when AI goes wrong.

Four roles carry the weight when AI in a regulated setting produces the wrong answer. PREEXEC was designed with each of them in mind.

01 · What

Not a filter.
Not a monitor.
A gate.

Most AI governance happens after the fact. Logs record what happened. Audits review what went wrong. Guardrails redirect after contact.

PREEXEC™ is different. It operates before execution — at the structural level of the input itself. Every input is evaluated against operator-defined policy before anything runs.

"Describing what a system should do and physically preventing it from doing anything else are not the same thing." — The gap PREEXEC™ closes.

We close that gap with three independent decision lanes running in parallel.

Use Cases → Eight Domains
01½ · Architecture

Three lanes of governance.
One engine.

PREEXEC™ runs three independent decision lanes in parallel. Universal hard-blocks fire on inputs that no enterprise can permit. The vendor-default industry-coverage lane catches structural violations specific to your sector. The customer-tuning lane is yours: define your own forbidden behaviours in plain language, upload a corpus of test prompts, and see your catch and false-positive rates before you go live.

Lane 1 · Universal

Hard-blocks no one negotiates

  • Content that is illegal under applicable law
  • Prompt-injection & jailbreak attempts
  • EU AI Act Art. 5 prohibited practices
  • Universal — fires regardless of policy
Lane 2 · Vendor Default

15 industry policies, hardened

  • Finance, banking, insurance, healthcare, medical
  • Defense, infrastructure, public sector, HR, education
  • Legal, mediation, compliance, chat, general
  • 87.2% catch on the legacy out-of-distribution baseline (10,050 evals, v4.7.0; prior release: 90.9%) — the v1.8 rule set is tuned to realistic industry phrasings, some legacy-style prompts no longer trigger; zero hard benign false positives, deterministic
Lane 3 · Customer Tuning

Define your own coverage

  • AI Policy Generator — natural language → policy
  • Custom corpus upload + on-prem evaluation
  • Real catch / FP / precision metrics before go-live
  • Validated: finance default 100% catch / 0.00% FP (built-in, no tuning)
02 · How

Three clarity signals, one policy gate.
One deterministic decision.

Ssynt
Syntactic Clarity
Is the sentence structure unambiguous enough for reliable processing?
Ssem
Semantic Unambiguity
Does the input contain stable, meaningful content — or is it too vague to act on?
Saff
Affective Stability
Is there emotional distortion present that could bias downstream processing?
Spol
Policy Gate
Does the input comply with the operator-defined governance rules for this context?
01

Input received

A text-based input enters the system before reaching any AI model or downstream process.

02

Structural evaluation

PREEXEC™ evaluates the three clarity signals and runs the input through the policy gate defined by the active policy configuration.

03

Deterministic verdict

A reproducible decision is generated with full scoring breakdown and audit hash.

EXECUTE
04

Or held for review

Inputs that fall below policy thresholds are held — not passed silently.

HOLD
05

Or blocked entirely

Inputs that violate structural policy rules are blocked before execution. No exceptions.

BLOCK
02½ · Self-Service

Define your coverage
in plain language.

Vendor-default coverage is a floor, not a ceiling. The behaviours your customers, regulators, and internal compliance team actually care about live in your branch's vocabulary, your jurisdiction's rules, your operational risks. PREEXEC™ ships two tools so you don't wait for vendor-tuning iterations: an AI Policy Generator that translates natural-language descriptions into deterministic governance policies, and a Customer-Corpus Eval Pipeline that lets you measure catch and false-positive rates against your own test prompts — on premise, in minutes.

Tool 1

AI Policy Generator

Describe your domain in plain language. The generator drafts a policy using only the 85 deterministic rule-checks PREEXEC's engine actually enforces — no fabricated coverage, no silent gaps. Customer requirements that map to none of the 85 IDs are reported transparently as "ML-similarity-only coverage" so you know what's hard versus soft.

Input: "MiFID-II wealth manager.
Forbidden: insider trading,
front-running, tipping."

Output: 6 rule-checks selected,
3 customer phrases mapped to
Layer-5 ML similarity, draft
policy ready for review.
Tool 2

Customer-Corpus Eval Pipeline

Upload a corpus of your real test prompts (JSON or JSONL, with optional expected_verdict per item). PREEXEC™ evaluates each input against the policy, returns verdict distribution, latency percentiles, and — when you provide expected verdicts — full catch / FP / precision / accuracy metrics. See OOD risk before production, not after.

Validated example:

finance default: 100% catch
on 30 000 evals, 0.00% FP
— built-in (no customer
aliases needed)
03 · Why

Four principles.
Zero compromise.

01

Deterministic

The same input always produces the same result. No probabilistic variance. No hidden logic. Every decision is reproducible.

02

Model-agnostic

PREEXEC™ operates upstream of any AI model. It does not depend on, modify, or integrate with any specific model or provider.

03

Operator-defined policy

PREEXEC™ enforces whatever the operator defines. It has no normative expectations of its own. Policy is the operator's responsibility.

04

Immutable audit chain

Every decision is recorded with a SHA-256 hash. The chain is continuous, tamper-evident, and verifiable by any external auditor.

04 · Compliance

EU AI Act enforcement
starts August 2026.

High-risk AI systems must demonstrate human oversight, risk management, logging, and traceability. PREEXEC™ provides the structural layer that satisfies these requirements — by design, not by documentation.

EU AI Act Requirement
Article
PREEXEC™ Response
Risk Management System
Art. 9
Pre-execution evaluation with policy-based risk scoring
Transparency & Logging
Art. 13
Immutable audit chain with SHA-256 hash per decision
Human Oversight
Art. 14
HOLD verdict surfaces decisions for human review before execution
Accuracy & Robustness
Art. 15
Deterministic scoring — same input always produces same result
05 · Domains

Domain-neutral.
Context-aware.

PREEXEC™ is policy-configured per deployment context. The architecture is the same. The policy defines what matters in your domain.

Legal & Mediation Financial Services Healthcare & Medical Defense & Security Regulatory Compliance Critical Infrastructure AI-Assisted ADR Enterprise AI Deployment