Loading signals...
| Model | Sources | Signal profile | EV pressure | Bias style | ELO risk terrain |
|---|
RAILS
Report Cards
Research
About Us
Get started
The trust engine for autonomous AI.
Every agent.Every output.Every transaction.
Your agents get corrected as they work. No payment moves until the output is verified. RAILS is the independent clearinghouse for agentic commerce.
Start with Mindlas
View the Report Cards
WHY RAILS
AI agents fail, confidently.
RAILS catches it in real-time, so you don't have to.
The Agent Era is Here
AI agents are starting to transact.
Software agents already work without a human watching. Now they work in fleets: finding each other, hiring each other, and paying each other at machine speed.
The autonomous future
Agent to Agent
Hiring
One agent hires another expert to draft a legal contract.
Agent to Merchant
Booking
An agent books the flights, the hotel, and pays.
Agent to Market
Markets
Agents trade grid energy the instant prices spike.
Machine to Machine
Autonomous
Self-driving cars buy and clear priority-lane access between themselves.
Authorization is not verification.
The expert agent hallucinates a clause in the contract. The booking agent pays for the wrong flight and the wrong room. Each one is authorized to act and to pay. Nothing checks whether the work is right.
LLM-as-a-judge fails.
Current LLM verification is models grading their own homework. The lenient ones wave defective work through. The strict ones reject legitimate work. Neither is a check. RAILS closes that gap.
How the clearinghouse works
RAILS is the integrity check between authorization and settlement. Authentication proves who the agent is. RAILS proves the output is worth paying for.
Watch a transaction clear, end to end.
Agent buys from a merchant · A user agent books the flights and the hotel.
Settlement record
Simulated outputs for illustration. RAILS is designed to work with most agentic commerce protocols, including the examples shown.
Frameworks it comes from
MCPA2AScored by
Networks it settles to
A trusted output clears to the counterparty. A wrong one never leaves the gate.
Payment networks prove an agent can pay. RAILS proves it should.
/ two worked examples
The same gate, end to end.
Agent → Agent · settlement
An orchestrator hires an expert agent, internal or from a marketplace, to draft an IP contract. RAILS verifies the work before escrow releases.
Agent → Merchant · settlement
A buyer agent places an order with a merchant. RAILS checks the order matches the user's intent before the payment clears.
Payment rails prove an agent can pay. RAILS proves it should, scoring the output before anything settles. Pass, and it clears to the counterparty. Wrong, and it never leaves the gate.
Logos denote the major frameworks, rails, and protocols RAILS is built to interoperate with. All marks are the property of their respective owners; no partnership or endorsement is implied.
The RAILS Platform.
Real-Time Agent Integrity & Ledger Settlement
Every agent output gets a real-time RAILS Score, built from a mesh of different trust dimensions. The score does two jobs. It catches an agent drifting mid-task, which ships as Mindlas for coding agents. And it decides whether an autonomous transaction should settle through the RAILS Clearinghouse for agentic commerce.
/ The RAILS platform
Real-time integrity infrastructure
01The rating
RAILS Report Cards
Independent, research-grounded integrity ratings of the frontier models.
Illustrative
A94.2
A91.8
B89.4
B86.5
02The composite
RAILS Score
The single inline score the mesh composes into, returned inline on every agent output.
Live · inline
96.7▲ 0.3
One number, returned on every agent output.
03The settlement
RAILS Clearinghouse
The integrity gate every agent transaction clears through. No payment moves until the output behind it is scored.
output 96.7PASS
↓ clears to counterparty
output 41.0HOLD
× never leaves the gate
Powering four institutional surfaces:
Developers · Enterprises · Insurers · A2A commerce
Independent model report cards for every major lab
Model logos are property of their respective owners. Inclusion indicates models tracked in RAILS Report Cards. RAILS is independent; no partnership or endorsement is implied.
Inside RAILS.
The verification mesh has three parts today: Mindlas catches coding session drift as it happens, the Report Cards track robustness of frontier models, and the Verification Inputs stress-test every agent output. Combined, they form a unified RAILS Score. The score provides feedback for agents and a clearing threshold for agentic commerce transactions.
Open source · Apache 2.0
/Mindlas catches your coding agent drifting.
A coding agent rarely fails at turn one. It fails at turn forty. The agent says done; the tests disagree. Mindlas (Mind-Atlas) reads the session live, catches the drift as it builds, and hands the agent a correction with the effect measured before and after.
Open and local. Nothing leaves your machine.
View on GitHub ↗Works withClaude CodeLiveCodexComing soonCursorComing soon
DRIFT ALERTVERIFY 74 · BLAST 78
MINDLASFable 5 (1M context)|converter-app2%
RotStable2
VerifyStable0
BlastStable0
LoopStable0
◈ All gauges stable · session healthy.
RAILS Report Cards.
Independent integrity ratings of the frontier models.
Research snapshot
·
Methodology: arXiv:2605.23952
·
Expanding as the score matures
Explore all Report Cards
Read the methodology
Who uses RAILS.
A guardrail for developers, an audit trail for enterprises, loss data for insurers, a settlement gate for commerce.
01
Runtime guardrails
Developers
The RAILS Score grades every agent output on multiple dimensions, one call, milliseconds of latency. Start with Mindlas, the behavioral instrument that catches drift.
Start with Mindlas02
AI governance
Enterprises
Continuous monitoring, logging, and an immutable audit trail on every agent output. The compliance evidence AI regulations now require, generated automatically as RAILS runs.
03
Risk intelligence
Insurers
The empirical failure-frequency data that makes AI E&O underwriteable. Continuous, comparable, defensible.
04
The clearing layer
A2A Commerce
Trust every transaction before funds move. The integrity gate agentic commerce needs to settle at scale.
The team.
Agent commerce won't clear at credit-card speed. It will clear at machine speed.
RAILS comes from a team that has built and deployed adversarial AI at high-frequency latencies since 2015, in environments where mistakes cost millions and nanoseconds matter. Scientific depth: a team with multiple PhDs, including a Lead Scientist with twenty-two issued patents. A decade of shared engineering history at the speed required by agentic commerce.
Adrian de Valois-Franklin
/ Chief Executive Officer
Co-founder of one of the industry's first AI-native investment platforms. For over a decade, the firm has built and deployed fleets of autonomous AI agents generating over ten billion dollars in annual trading turnover, in production at machine speed. Previously at CPP Investments, Accel-KKR, Goldman Sachs. Frequent speaker at AI and quantitative technology conferences, and serves as an Advisor to NextAI, a global innovation hub and accelerator for artificial intelligence.
Alex Bogdan, Ph.D.
/ Chief Scientific Officer
Forty years in AI research. Twenty-two issued patents. Dual PhDs in Artificial Intelligence and Electrical Engineering. Inventor of the Geno-Synthetic Algorithm and the Ranking Inference primitive. Earlier in his career, Alex designed orbital satellite systems, work for which he was awarded the Lenin Komsomol Prize, the highest government award for a scientist.
Anas Ibrahim, Ph.D.
/ AI Scientist
PhD in Electrical and Computer Engineering. Research in evolutionary computation, reinforcement learning, signal extraction, and biomedical sensor systems. Lead developer of hybrid AI engines combining deep learning with evolutionary computation.
Ibrahim Shaer, Ph.D.
/ AI Scientist
PhD in Software Engineering. Affiliated with the Vector Institute, the Toronto-based AI research hub co-founded by Geoffrey Hinton. Research in explainable AI, federated learning, and anomaly detection.
Tamaz Andguladze, CFA
/ Chief Technology Officer
Lead architect of distributed AI infrastructure for high-frequency, machine-speed environments. Fifteen years building real-time decision systems at the intersection of applied mathematics and high-performance computing at institutions like RBC, Bank of America/Merrill Lynch, and SS&C Technologies as Lead Financial Engineer.
Edwin Li
/ Chief Commerce Officer
Former Managing Director at JPMorgan, where he led Corporate Derivatives and Cash Trading at institutional-commerce scale. Decade-plus on the Global Equity Management Committee.
Michael Petruzella, CIM
/ Chief Operating Officer
Thirty years operating financial services platforms across multiple regulated jurisdictions, including Portfolio Management, Risk Management, and Fund Accounting.
Jason Bernstein
/ VP, Operations
Manages product execution, vendor coordination, and the day-to-day infrastructure that keeps an AI-native firm running. Background in financial services operations and platform delivery.
Read Our Story
→
Research.
The science behind the RAILS Score.
Paper
RAILS
Verification-Native Clearing for Agentic Commerce
Adrian de Valois-Franklin · Alex Bogdan, PhD
Payment checks whether an agent can pay. RAILS checks whether it should. It verifies the work itself and blocks any settlement the evidence cannot support, the missing clearing layer for agentic commerce.
Paper
The Surprising Universality of LLM Outputs
A Real-Time Verification Primitive
Alex Bogdan, PhD · Adrian de Valois-Franklin · Anas Ibrahim, PhD · Ibrahim Shaer, PhD · Tamaz Andguladze, CFA
Up to 100,000x cheaper than running an LLM as a judge, and it runs CPU-only. Frontier models from different vendors all leave the same statistical signature in their output, which turns two hard jobs into fast ones: a triage layer for flagging likely hallucinations, and fingerprinting which model actually wrote a given piece of text.
Paper
Machine Psychometrics
A Mathematical Psychology of Artificial Intelligence
Alex Bogdan, PhD · Adrian de Valois-Franklin
Models have psychological structure that capability scores and traditional benchmarks never touch. Machine Psychometrics is the measurement science for it: profiling a model’s hidden dispositions from behavior and turning that into a Mindprint you can act on.
Essay
First Contact
The Most Important Question in AI
Adrian de Valois-Franklin
The most consequential question in artificial intelligence is not “when will machines become superintelligent?” It is: when they do, how will we know what they are actually thinking?
Paper
The Geno-Synthetic Algorithm
Type-Factored Coevolutionary Optimization for Heterogeneous Genotypes and Assembled Phenotypes
Alex Bogdan, PhD
A better kind of evolutionary computing. Where standard genetic algorithms force every parameter into one number format and break on the rest, Geno-Synthetic algorithms simultaneously optimize different input types. That makes it a direct optimizer for the prompts and embeddings inside LLM systems.
Paper
Free Energy Heuristics
Fast-and-Frugal Cognition as Active Inference Under Uncertain Precision
Alex Bogdan, PhD
More reasoning is not always better. On problems a model can check, chain-of-thought helps; on judgment calls it cannot verify, extra steps just make it confidently wrong. This is the basis for an orchestrator that makes frontier models smarter by reasoning only when it helps.
Explore all research →
Build on RAILS.
Developers, install Mindlas. Enterprises, insurers, and commerce infrastructure teams work with us to shape how RAILS fits your stack.
Investors and press:
The trust engine for autonomous AI.
Every agent. Every output. Every transaction.
Independent, real-time integrity scoring for AI agents.
Toronto · New York
© 2026 Evolutionairy AI · All rights reserved
