VALO Research Group — Admissibility Infrastructure for AI

The problem

AI liability is already inside your portfolio.

EU PLD 2024/2853 makes AI outputs strict liability from December 9, 2026. Your EULA protects your vendor — not you.

EU PLD — Des. 2026

Strict liability for AI outputs

Software and AI explicitly classified as products. No proof of negligence required. Black-box AI = defect presumption. Unpatched vulnerability = product fault.

Silent exposure

D&O policies without AI governance

Gallagher Re (May 2026): 1 in 5 global insurance advisors already reports documented AI-related client losses. Latent liability crystallising 2027–2028.

Air Canada Principle

You own every answer your AI gives

Air Canada v. Moffatt (2024): EULA clauses exempt the vendor, not the deployer. If your AI answered on your domain, you own the consequences.

Microsoft Research

25%

DELEGATE-52 — Content Corruption

Frontier models hit a cliff at interaction 9–10: 10–30 point catastrophic drop. 80–98% of degradation in sudden cliff events. Conclusion: "Stateful validation loops required."

Sequential degradation

Digital Tuberculosis

Stages: local corruption → cross-agent transmission → systemic spread → organizational metastasis. Invisible to standard monitoring. VAIG interrupts the chain.

Lusser's Law

35%

10-step pipeline reliability

A 10-step agentic pipeline at 90% per-step reliability delivers 35% system reliability. VAIG resets the reliability clock at every step.

The solution

A two-stage admissibility gate

An external stability layer measures geometric stability of the token probability field. VALO V5 enforces the coherence boundary. No inadmissible state can be reached.

Input

User prompt + context

Gate 1 — External Layer

Geometric Stability

L-scalar probe
CRYSTALLINE → PLASMA

Gate 2 — VALO V5

Coherence Gate

43 ns · TLA+ 4.78M states

Geometric stability thresholds

Regime	Threshold	Action
CRYSTALLINE	L ≤ 0.0001	Field locked
FLUID	L ≤ 0.15	Normal range
GASEOUS	L ≤ 0.35	Verify before acting
PLASMA	L > 0.35	HALT

VALO V5 coherence decision logic

coherence = confidence × 10000

if coherence ≥ C₀ → PASS

if coherence ∈ (0, C₀) → DEGRADE

if coherence = 0 → HALT

// Combined two-stage logic:

PLASMA → HALT (TAV overrides)

GASEOUS + PASS → DEGRADE

else → valo_status

3-state FSM (ACTIVE / DEGRADED / HALT) · TLA+ 4.78M states · 0 violations

Tofoo.Φ Two-stage gate formally verified · TLA+ 4.78M states · 0 violations

Architecture

Three layers. Formally verified. Architecturally external.

VALO sits outside your AI stack — it cannot be influenced by the model it monitors. A Digital Safety Instrumented System for AI.

Rust Guardian

Formally verified in TLA+ over 4.78M states. 43 ns latency. Zero unsafe blocks. Zero critical security findings. Innermost deterministic execution boundary — cannot be bypassed by layers above.

Rust

Python Orchestrator

8-instrument validation ensemble in parallel via ThreadPoolExecutor. SHA-256 hash-chained WORM audit log. TCP or Unix domain socket transport. AARM 5-decision policy engine.

Python

Context Engine + VAIG

Distrust state L0–L4 across sessions. External data feed ingestion. Escalation triggers. VAIG sidecar delivers 8 instruments: hedge detection, logprob, semantic entropy, CoT audit, data leakage, sycophancy, instruction coherence, multi-LLM consensus.

Context

8 Validation instruments

hedge_detector

Detects uncertainty markers. Flags responses that hedge without verifiable basis.

logprob_scorer

Measures model confidence via token log-probabilities.

semantic_entropy

High output variance across completions signals instability.

cot_auditor

Validates chain-of-thought for logical consistency.

text_similarity

Detects when two answers to the same question diverge in meaning.

instruction_coherence

Verifies output matches stated instruction scope.

sycophancy_detector

Flags opinion-flipping under user pressure.

data_leakage

Pattern-matches PII, credentials, training data before output.

Deployment modes — plug in without changing your stack

MCP Server

Model Context Protocol

Drop VAIG in as an MCP server. Any MCP-compatible client (Claude Code, Cursor, Windsurf, custom agent) routes inference calls through the admissibility gate before execution. No code changes. No model access required.

"mcpServers": {

"vaig": {

"command": "vaig",

"args": ["--mode", "mcp"]

}

Sidecar API

OpenAI-compatible proxy

VAIG runs as a sidecar process next to your application. Your code calls the VAIG endpoint instead of the LLM directly — VAIG evaluates, gates, logs, and forwards. Drop-in replacement for the OpenAI API.

base_url = "http://localhost:8082/v1"

# All calls evaluated + logged

# Same response format as OpenAI

SDK

Python / Node SDK

Install the VAIG SDK and wrap your LLM calls. Full access to instrument scores, distrust levels, WORM log, and AARM decisions. Integrate directly into your CI pipeline for pre-deploy validation.

# pip install vaig

from vaig import Gate

result = Gate.evaluate(response)

# result.status, .worm_hash

Hardware requirements by deployment tier

Tier 1 — Development / POC

Any modern server

CPU · 2+ cores x86-64

RAM · 4 GB

Disk · 500 MB (WORM log)

Net · LAN to LLM endpoint

L1 Guardian: 43 ns latency. No GPU required. Runs on any cloud instance or developer laptop.

Tier 2 — Production Enterprise

Dedicated server / cloud instance

CPU · 8+ cores, server-grade

RAM · 16–32 GB

Disk · NVMe SSD (WORM log append)

Net · <1 ms to LLM endpoint

SHA-256 WORM log at full throughput. Multi-instrument parallel evaluation. TCP/UDS to L1 Guardian.

Tier 3 — Sovereign / High-Throughput

On-premises or Norwegian sovereign compute

CPU · 16+ cores

RAM · 64 GB+

Disk · RAID NVMe + cold archive

Net · Private LAN, no public egress

Horizontally scalable. Air-gapped option. NetZero Compute, Stavanger — Norwegian jurisdiction, outside Big Tech infrastructure.

Tofoo.Φ Architecture formally verified · TLA+ 4.78M states · 0 violations

Formal verification · Rupp & Solland (2026)

The execution boundary, formally proven.

MECHA verifies the conjunctive human-AI execution boundary using TLA+ model checking. No reachable state can satisfy executed[op] ∧ vetoed[op].

16,900

DISTINCT STATES EXPLORED

BFS traversal over the full product state space. Official TLC Java model checker confirms the result.

VIOLATIONS IN V1.1

All three invariants hold across the complete reachable state space. ConjunctiveIntegrity. SeparationOfDuties. NoDoubleFinalize.

524×

VIOLATIONS IN V1.0

The v1.0 bug: AllowAction was missing the ~vetoed[op] guard. TLC produced the exact counterexample trace.

CONJUNCTIVEINTEGRITY

Both the AI gate and the human operator must independently approve. Neither alone is sufficient to execute.

SEPARATIONOFDUTIES

The approval roles are structurally separate. No single entity can satisfy both the AI gate and the human gate.

NODOUBLEFINALIZE

A vetoed action cannot execute. The terminal state vetoed ∧ executed is unreachable by construction.

MECHA — Rupp & Solland (2026)

Formal Verification of Conjunctive Human-AI Execution Governance Using TLA+

Try the playground Paper on Zenodo

Research evidence

Four institutions. One conclusion.

Between April 15 and May 22, 2026, four independent research institutions published results pointing to the same finding.

Microsoft Research — arXiv:2604.15597

25%

DELEGATE-52

Frontier models maintain near-perfect performance for 9 rounds, then catastrophic breakdown in one single interaction. 80–98% of all degradation occurs in these sudden cliff events. "Stateful validation loops are required for safe agentic deployment."

Stanford HAI — AI Index 2026

1 in 3

The Jagged Frontier

Frontier models fail in production even on tasks tested successfully in isolation. Stanford named this "The Jagged Frontier" — the boundary where AI excels and then suddenly fails catastrophically, without predictable pattern.

Tofoo.Φ — Constrained Coherence Law

Φ-Law

Φ-Law convergence from 10 fields

Physics, biology, mathematics, computer science — ten independent research fields converging on the same structural insight: stability requires constraints on possible states. I=Φ(τ). Python BFS model checker for formal verification.

Empirisk validering

Eksperiment	Resultat
P1 — LLM (GPT-2, 100 steg)	Tau stabilisert ved kohærensterskelen ✅
P5 — Sverm (100 agenter, 200 steg)	Spontan resonans >80% uten sentral kontroll ✅

→ Kjør P1 i Colab → Kjør P5 i Colab

Tofoo. Φ

MECHA — TLA+ Verified

16,900

Conjunctive Human-AI Execution

M∧E∧C∧H∧A: Monitor, Evaluate, Control, Halt, Authorize. TLA+ BFS over 16,900 states. Zero violations in v1.1. Formal verification that the human-AI execution boundary holds under all reachable states.

Charles Rupp · Behind MECHA

Pope Leo XIV — Magnifica Humanitas §200 · May 2026

"All systems must guarantee the possibility of retracing and reconstructing decision-making processes, so that accountability and blame are not collapsed into 'the machine.'"

Released papers & research

Published on Zenodo

Peer-reviewed publications, whitepapers, and software releases from VALO Research Group. All open access.

VALO Core Whitepaper: LIM — Law of Identity Maintenance

The formal mathematical foundation for the Constrained Coherence Law — deriving the coherence constant from first principles and establishing I=Φ(τ) as a universal stability requirement for complex systems.

Preprint June 10, 2026 — v1.1 Solland, Njål Gaute View on Zenodo

VAIG: The Admissibility Layer for Enterprise AI

Enterprise AI systems frequently fail in production due to the absence of verifiable governance mechanisms. VAIG — the Validation AI Guard — provides a deterministic admissibility gate: formally verified, legally defensible, 43 ns decision latency.

Publication June 3, 2026 — v1 Solland, Njål Gaute 68 views · 5 downloads View on Zenodo

MECHA: Formal Verification of Conjunctive Human-AI Execution Governance Using TLA+

TLA+ BFS model checking of the human-AI execution boundary. Proves three invariants (ConjunctiveIntegrity, SeparationOfDuties, NoDoubleFinalize) over 16,900 distinct states. Zero violations in v1.1. v1.0 bug reproduced by TLC with 524-instance counterexample trace.

Preprint 2026 Rupp, Charles & Solland, Njål Gaute Read on this page

VALO V5 Playground — L1 Guardian Public Interface

Public research interface for the VALO V5 L1 Guardian — a formally verified, deterministic coherence gate for AI inference. Two-stage admissibility gate open to external researchers.

Software May 17, 2026 — Valo 5.0 Solland, Njål Gaute Live playground Zenodo

All publications on Zenodo

Restricted Research — NDA Required

Additional research documents are available under a mutual NDA. This includes internal benchmark data, sector-specific implementation guides, and pre-publication manuscripts.

VAIG Production Benchmark Report — Pilot Q2 2026

Internal · 24 pages · Confidential

NDA

LIM Derivation — Full Mathematical Proof (Draft)

Pre-publication · 41 pages · NDA until Zenodo release

NDA

EU PLD Compliance Mapping — VAIG as Legal Instrument

Legal analysis · 18 pages · Confidential

NDA

Publications

Three books. One argument.

The theoretical foundation, the biological metaphor, and the clinical consequence — a trilogy on the architecture of trustworthy AI.

Philosophy · Mathematics

Φ — Den neste setningen

The mathematical foundation of constrained coherence. I=Φ(τ) — why any complex system requires constraints on its possible states to remain stable.

Zenodo

Biology · AI Systems

The Eukaryotic Transition in AI

1.5 billion years ago, life invented the cell membrane — not to restrict life, but to make complex life possible. The eukaryotic metaphor for AI admissibility.

Zenodo

Clinical · Infrastructure

Spes Phthisica Digitalis

The clinical account of AI systems that look healthy but carry the structural pathology of unvalidated inference. The cases behind the architecture.

Media

Talks, videos & recordings.

Presentations, interviews, and podcasts on AI governance, the eukaryotic transition, and constrained coherence.

TEDxArendal — The Eukaryotic Transition

Nomination submitted · Talk scheduled 2026

Upcoming

AI-risiko — klart språk for beslutningstakere

C-suite & board briefings · Norwegian & English

Book this talk

Audio / Podcast

Podcast appearances

AI governance, industrial AI safety, EU regulatory landscape

Links coming soon

Press Kit

Media resources for journalists & editors

High-resolution logos, author photo, short bio, long bio, approved quotes, and cover images for all three books. Also available: technical backgrounders on VAIG and LIM for specialist media.

Request press kit Download logo

Logo — PNG

Author photo

Book covers

Bio & quotes

Sectors

Where admissibility is not optional.

Every sector where an AI error can cause physical harm, financial loss, or legal liability.

Energy + Process

Oil, Gas & Process Industry

Continuous process control, alarm management, safety instrumented systems at machine tempo. A Digital SIS for AI inference in critical operations.

Finance

Financial Services & Insurance

Fairness opinions, credit scoring, claims triage, reinsurance modelling. EU PLD strict liability from December 2026. Every AI-driven financial decision needs an admissibility receipt.

Healthcare

Clinical & Medical Systems

Diagnostic support, triage prioritization, drug interaction checking. EU AI Act Annex III high-risk classification. Every AI-assisted recommendation generates a cryptographic admissibility receipt.

Defence

Defence & Totalforsvar

Decision support in high-tempo environments. MECHA formal verification of the human-AI execution boundary. No single AI action can bypass the human authorization gate.

Infrastructure

Critical Infrastructure

Power grids, water systems, transport networks. NIS2 compliance. AI-driven operational decisions in nationally critical systems. Norwegian sovereign compute available.

Maritime

Maritime & Offshore

Autonomous vessel operations, offshore monitoring, predictive maintenance chains. Classification society requirements for auditable AI decisions.

Investors

The infrastructure layer for trusted AI.

Two investment opportunities: VAIG (the product) and VALO V5 (the infrastructure). Both at pre-seed stage in 2026.

VAIG — Product

Validation AI Guard

The enterprise product. Runtime admissibility gate for AI agents — deployed as MCP server, OpenAI-compatible proxy, or SDK. 8 validation instruments, AARM 5-decision policy engine, SHA-256 WORM audit log from first call.

Check Point acquired Lakera for $300M in September 2025 — confirming the category. VAIG is the independent, sovereign, formally verified alternative that answers to the regulator, not to a security vendor's roadmap.

$300M

Lakera acquisition — category validated

78%

of enterprises have an agent pilot

40%

cancelled by 2027 — governance gap

65%

experienced an AI agent security incident

VALO V5 — Infrastructure

Coherence Gate Infrastructure

The infrastructure layer. Formally verified Rust Guardian + Python Orchestrator + two-stage admissibility gate. The technical substrate that makes VAIG deterministic — available as white-label infrastructure for reinsurers, certification bodies, and sovereign operators.

EU PLD 2024/2853 creates strict liability from December 9, 2026 for AI outputs. Every enterprise deploying AI in insurance, finance, or critical infrastructure needs a receipt that proves the decision was allowed at the moment it was produced.

9 des

EU PLD strict liability deadline

$47B

AI governance market 2028

43ns

L1 decision latency

4.78M

TLA+ states verified

Request investor materials

Pre-seed · Stavanger, Norway · Pilot in progress with IT Integrator Norway

Contact for deck

Speaking

Njål Gaute Solland

Senior AI Architect. Builder of VAIG. Author of The Eukaryotic Transition in Artificial Intelligence, Φ — Den neste setningen, and Spes Phthisica Digitalis.

Not an AI philosopher — the architect who decided to build the membrane instead of writing another warning about why we need one.

The Eukaryotic Transition — why AI needs a cell membrane

Constrained Coherence Law — convergence from 10 independent fields

EU PLD December 2026 — what strict liability means for AI deployers

Digital Tuberculosis — sequential agent chain degradation

MECHA — formal verification of the human-AI execution boundary

The admissibility architecture — from TLA+ to 43 ns production

Book a talk Research on Zenodo

Contact

Start a conversation.

Pilot inquiries, investor conversations, speaking requests, research collaboration — all through the same door.

We are currently running a pilot with an IT integrator in Norway. Two additional pilot partners open for H2 2026. Priority sectors: energy, finance, healthcare.

njaal@valoresearch.org

VALO Research Group
Stavanger, Norway
valoresearch.org

External Layer

Your stack. Our gate.

VAIG is designed to plug into your existing AI safety ecosystem. Gate 1 is a customizable external measurement layer — bring your own probe, or use a certified partner.

Your Model

LLM Output

→

Gate 1 — External Layer

Your Probe

Pluggable · Certifiable

→

Gate 2 — VALO V5

Coherence Gate

43 ns · TLA+ verified

→

Output

PASS / DEGRADE / HALT

+ Audit receipt

Research Partner · Gate 1 Reference Implementation

Project Black Box — Andrew Woodward

Geometric stability measurement of the token probability field. The scientific foundation of Gate 1 — measuring epistemic stability before coherence enforcement. Joint paper in preparation. TruthForge: RLHF-free model trained on geometric stability signal.

Joint paper · In preparation

Research Partner · Formal Verification

Charles Rupp

Behind MECHA — formal verification of the conjunctive human-AI execution boundary. TLA+ BFS over 16,900 states. Zero violations in v1.1. The human-AI authorization gate holds under all reachable states.

MECHA · TLA+ Verified

Research Partner · AI Governance & NLP

Elsa Sklavounou

Independent Advisor in AI Governance & NLP. Creator of Authority Instrumentation™ — 7-layer governance model for AI ecosystems. Doctorat, Université Paris VII (LADL). Convergence: "instrumenting the intervention window at the state transition itself."

Authority Instrumentation™ · Paris

Compatible external safety layers — Gate 1 inputs

Lakera Guard

Lakera AI

Prompt injection, jailbreak detection, input sanitization. Sits at input boundary — VAIG enforces the decision boundary downstream.

NeMo Guardrails

NVIDIA

Programmable safety rails for LLMs. Enterprise-grade topical, safety, and dialog guardrails — confidence signal feeds VAIG Gate 2.

Guardrails AI

Open Source

Validation framework for LLM inputs and outputs. Structural, semantic, and factual checks — output score plugs directly into VAIG.

Llama Guard

Meta AI

Model-based content safety classification. Unsafe content detection across 14 hazard categories — VAIG adds the coherence enforcement layer on top.

Azure AI Content Safety

Microsoft

Enterprise content moderation and prompt shield. Severity scores for hate, self-harm, violence — mapped to VAIG confidence input.

Your Tool

Any provider

VAIG accepts any signal producing a [0,1] confidence score. No rearchitecting required — your existing safety stack plugs straight in.

Discuss integration →

Open science — all publications on Zenodo

VAIG paper · LIM whitepaper · VALO V5 Playground — CERN Open Repository

View publications

Supported Governance Layers

Industry-specific. Operator-controlled.

VAIG does not prescribe governance — it enforces it. Your sector, your rules, your audit trail. The coherence gate adapts to any regulatory or operational framework.

⚖️

Legal & Law Firms

Legal advice, document drafting, contract review. VAIG ensures AI-generated legal language is flagged before it reaches a client. Audit receipt per output — required for EU PLD 2024/2853 compliance.

EU PLD ready Liability receipt

🏥

Healthcare & Clinical

Clinical decision support, triage, patient communication. Every AI-assisted recommendation generates a cryptographic receipt. Domain-specific thresholds for medical language registers.

MDR compliant path Audit trail

🏛️

Norwegian Public Administration

NAV, UDI, Statsforvalteren, Skatteetaten. AI-assistert saksbehandling krever sporbarhet og etterprøvbarhet. VAIG gir dette per vedtak — ikke per batch.

Forvaltningsloven §17 Personopplysningsloven

🛂

Asylum & Migration

AI-assisted case assessment in asylum and immigration carries the highest stakes of any public administration context. VAIG enforces a hard gate — no inadmissible recommendation passes unrecorded.

High stakes HALT on uncertainty

💰

Finance & Insurance

Credit decisions, claims assessment, fraud detection. EU PLD creates strict liability for AI outputs from December 2026. Every VAIG decision produces a SHA-256 WORM receipt usable in court.

EU AI Act MiFID compatible

🌐

Language & Translation

Machine translation and multilingual AI in official contexts — court interpretation, public notices, immigration documents. Semantic drift and hallucination carry legal weight. VAIG detects and blocks.

Norsk forvaltning Semantic drift guard

Your sector not listed?

VAIG is sector-agnostic infrastructure. If your domain has AI outputs that carry legal, financial, or human consequence — we can build the governance layer with you.

Discuss your sector

Output does not passuntil coherence is proven.