The Lab — ET-ERNITY

Yield simulation for analog AI (memristor-based neuromorphic) chips

TRL 4

A complete prototype, validated in the lab on stated parameters.

Towards zero spread. Not zero defects.

The problem: Manufacturing defects are unavoidable. On dense analog chips, one bad component can spread and take the whole array down. Designers usually learn where that breaking point sits after the chip is built.
What exists: A working simulation engine, an API and a dashboard, tested end to end. For a given design, it finds the defect rate at which fault tolerance breaks down, based on the device assumptions the designer provides.
What it is not yet: A target-matching model. It has not been calibrated against measured wafer data, and no outside team has run it. Its answer is only as good as the assumptions put in.
Built to stay blind: It accepts physical parameters only. Names, company details and project titles are rejected before any simulation runs.
What moves it to TRL 5: Measured defect or yield data from a research lab, used to calibrate the model and check it against real results.
The goal: Seeking a research group with measured defect or yield data. In return: a year of free access and co-authorship credit on published validation work.

Substrate Zero is part of the ET-ERNITY group of companies.

Incident records for small security teams and managed security providers

TRL 4

A complete prototype, validated in the lab on test signals.

Cause. Effect. Action.

The problem: When a session is stopped, people argue about what caused it. An account written by the same team that responded is hard for an outsider to trust.
What exists: Six working modules, an API with sign-in and a front end, tested together. 400 of 400 automated checks pass on our test machine. Each stop is recorded with its cause, and the cause is kept apart from the session that was stopped. The record is tamper-evident: edits and deletions inside the chain are caught, and people you choose can check it.
What it is not yet: It has no outside time-stamp service yet, so we say tamper-evident, not certified. It has not been tested on real security-team data. It has no customers and does not run on a real host.
Built to hold no personal data: No personal data by design. Sessions are identified by opaque codes.
What moves it to TRL 5: A trial on realistic data from a security team.
The goal: Looking for design partners: small security teams or managed security providers willing to try it on realistic data.

Guided conversations that help a person find the real question in a complex situation

TRL 4

A working prototype, tested on practice conversations. One outside professional has tried it on topics from their own field.

The problem: Most AI tools answer the question that was typed. A person in a complex situation often needs help finding the question first. Some situations also need a hand-over to a human, not more software.
What exists: A working system that guides a conversation step by step, from a tangled situation to one next step that the person chooses. It has a web chat interface. A first test round of five practice conversations, on topics from a childcare professional's own work, took place in March 2026. The set-up is designed so each client organisation gets its own separate environment and settings.
What it is not: Not an advice bot. Not therapy. Not a crisis service. Not a clinical tool. Not a replacement for a professional.
What it is not yet: It is not clinically validated. The step where it stops and refers a person to a human has not been tested in a recorded session. The five conversations were not scored and not compared with a plain AI chatbot. It has no paying clients yet.
What moves it up: A structured trial on real cases with outside users, scored on a fixed sheet and compared with a plain chatbot. A test of the stop-and-refer step on prepared scenarios. A legal and data-protection review.
The goal: Looking for a professional organisation (for example HR, care or training) to run a guided pilot on real cases. Also looking for a care or clinical partner to review the stop-and-refer step.

Measuring the quality of AI reasoning

TRL 3

A proof of concept that has run end to end on a live conversation. Its scores are not yet shown to be repeatable.

The problem: Monitoring tools usually report speed, cost and safety. This work looks at something else: whether an AI's reasoning holds together across a conversation.
What exists: A working loop that follows a live conversation with an AI model and scores its reasoning turn by turn. It has two layers. One asks whether the reasoning holds together. The other asks whether the AI keeps track of its own reasoning. A browser extension for live conversations is in development.
What it is not: Not a fact-checker. Not a safety filter. Not a consciousness detector. It rates how the reasoning holds up, not whether it is true.
What it is not yet: The scores come from a model-based judge. They can vary on near-identical answers, and they have not been compared with human reviewers.
What moves it up: Repeatable scores on the same conversation, and agreement with human reviewers.
The goal: Looking for teams that run long AI conversations and will test the scores on their own transcripts.

Computer modelling of cancer immunotherapy and immune balance

TRL 2

A written concept with one first computer demonstration. Not tested against patient or lab data.

The problem: Cancer immunotherapy can be powerful, but the immune response is hard to balance. Too weak, and the tumour escapes. Too strong, and it harms the patient.
What exists: A written research framework, and a simulation of one treatment cycle with forty tumour cell lines. In that run, a monitoring loop kept a simulated inflammation measure under its limit. It also adapted when one cell line changed its surface marker.
What it is not yet: The simulation was built from software designed for another field, not from immunology data. Its settings were not taken from patient or lab data. It ran once, and it was not compared with any real treatment record. It shows that the loop works as designed. It does not show that real immune systems behave this way.
What it is not: Not a treatment. Not medical advice. Not a clinical tool. Not a claim that cancer can be cured or controlled this way.
What moves it up: Settings fitted to published or trial data. Many runs with varied settings, compared with a version that has no monitoring loop. Review by an immunology or oncology group.
The goal: Looking for a research group with de-identified treatment data who will test whether the model's settings can be fitted to it.

Where philosophy becomes architecture.