We build ML systems for regulated industries where somebody downstream has to check the output without taking our word for it.
That reaches you three ways: an instrument we operate on your data, a system we build and hand over, or a tool you run yourself.
When did the training data change? What was the model’s confidence on that decision? Can you replay last Tuesday’s inference pipeline exactly?
If a regulator, auditor, or client asks, silence is expensive.
We build the infrastructure that makes the answers automatic.
Article 50 has applied since 2 August 2026 and is enforceable by national market surveillance authorities now. Generative systems placed on the market before that date have until 2 December 2026 to meet the machine-readable marking requirement. Annex III high-risk obligations apply from 2 December 2027.
Provenance cannot be backfilled. Records that were not captured when a dataset was built cannot be reconstructed at audit time. Compliance work on training data therefore starts at dataset construction, not at conformity assessment.
The difference is who operates the system. In all three, the output carries the evidence for what it claims.
You send raw data. We run the instrument and return a versioned dataset with four layers of provenance, an EU AI Act evidence pack, and recipes that check our claims against tools we do not maintain.
Industrial vision, capture to deployment: the rig, the taxonomy derived from policy, the model, the operator interface, and the evidence the system produces in production. Built to abstain rather than guess when a decision is not supported.
Point it at a prediction file and a claim about a model’s performance. It returns what the evidence defensibly supports, and everything it does not. Runs on your machine; nothing is uploaded.
The first of the three, in depth
We built this instrument and we run it ourselves. There is nothing to install, licence, or learn. Your data goes in; a dataset that carries its own evidence comes out.
Every delivery ships with recipes that run against tools we do not maintain. Acceptance means you ran one and the numbers came back the same.
Nineteen constants in the instrument can refuse a sample or a build. Each is registered with a verdict, and findings that rest on an uncalibrated one say so in your report.
Not quietly kept and counted. Measured, found wanting, and marked non-gating, which is why we can tell you four of six tiers gate instead of advertising six.
Drag the strip, or use the arrow keys.
That column describes a category, not a competitor
You describe the problem, the volume and the deadline. We reply with a tier and a range. An NDA is in place before any data moves.
A signed scope pins the schema, the acceptance criteria and the exact instrument versions used.
Six tiers of validation run before anything is constructed. Four can stop the run, and a refusal is written down.
Four layers of provenance are attached. A signing attempt that cannot proceed writes nothing at all.
You run a verification recipe yourself. That is what acceptance means. Raw inputs are deleted within 30 days.
You describe the problem, the volume and the deadline. We reply with a tier and a range. An NDA is in place before any data moves.
A signed scope pins the schema, the acceptance criteria and the exact instrument versions used.
Six tiers of validation run before anything is constructed. Four can stop the run, and a refusal is written down.
Four layers of provenance are attached. A signing attempt that cannot proceed writes nothing at all.
You run a verification recipe yourself. That is what acceptance means. Raw inputs are deleted within 30 days.
Engagement model: Written scope, milestones, acceptance criteria, and verification plan are agreed before work begins. Four tiers define scope and price; every tier produces the same four artefacts.
Vision-transformer-based computer vision for automated laptop quality assessment. A defect taxonomy derived from manufacturer warranty policy (Customer Induced Damage). PASS/FAIL classification with explicit confidence scoring, abstention threshold, and evidence capture.
Judged on the record.
Most companies publish what worked. A record only means something if it also holds what did not.
Four things on this site are here because they are unflattering.
Four of Paradigm’s six validation tiers gate. Two do not. When a check was measured and found unable to separate good input from bad, it was demoted rather than counted.
Nineteen constants in the instrument can reject your data. Some are plain configuration derived from no measurement, and a finding that rests on one is marked uncalibrated in your report.
PIRX was run against our own detection model on public benchmark data. It returned 59.6% recall at a 1% false-alarm rate. That figure is on the PIRX page.
The accessibility statement names the independent audit we have not commissioned rather than implying one.
We work with teams in regulated industries — logistics, manufacturing, finance, healthcare — who need AI systems that survive audit.
Static Signal LTD
Company No. 16826101
71–75 Shelton Street, Covent Garden
London WC2H 9JQ