R&D · Local AI

We do the research so the product can be trusted.

Our mission is to make technology accessible to everyone. That starts by fixing what's broken in today's AI: applications that fail silently, models that only run in someone else's cloud, and systems nobody is accountable for. We take what this research proves and put it into our products, to make them more reliable, more local, and more accountable.

The principle behind the research

This is what the research is for.

Every result on this page feeds one goal: answers you can hold to account. The model's own output is never deterministic — what's deterministic is the layer that checks it before a human ever sees the result.

  1. 1

    Grounded

    Answers are drawn from your own documents and data — retrieval, not memory.

  2. 2

    Checked

    Deterministic rules, lookups, and reconciliation with tolerances check the result before anyone sees it.

  3. 3

    Held for a human

    Anything uncertain, disputed, or out of policy waits for a person instead of shipping unapproved.

  4. 4

    Recorded

    An audit-ready record of what was decided, and why — so it can be checked later, not just trusted now.

What every product is held to

Our R&D program exists to earn these three commitments, in this order — not as slogans, but as research we hold ourselves to with measured, honest results.

The problem

AI applications that fail silently — a confident, wrong answer looks identical to a correct one, and nobody is told the difference.

Our approach

We make AI-native applications reliable at the core. Our verification-anchored research cut silent errors from 36% to 4% in our research.

“Fails loudly, never silently.”

Where it shows up
  • Orchestrator Studio

    A second-opinion reviewer model holds a run for humans whenever it disagrees; confidence-gated review queues catch what a single model would have missed.

  • Vexilo

    Tested against a real persona task set with zero silent wrong answers.

Verification-Anchored Federation (VAF)

The research program behind "reliable at the core." A router hands each request to a small specialist model, and nothing ships until it has been verified and calibrated.

Verification-Anchored Federation pipelineInput flows to a router, which hands off to four LoRA specialists sharing a 1.2B backbone, then to a tiered verifier, a conformal abstention gate calibrated for 95% coverage, and an out-of-distribution rejection stage, ending in an answer, an abstention, or an escalation to a human.InputRouterToolLoRA specialistRAGLoRA specialistSoft-skillLoRA specialistExtractLoRA specialistShared backbone — LFM2-1.2BTiered verifierConformal abstention gate95% coverage targetOut-of-distributionrejectionAnswerAbstainEscalate to human
  1. 1

    Input

    A request arrives — a question, a document, a spreadsheet task.

  2. 2

    Router

    The router reads it and decides which kind of work this is.

  3. 3

    Four specialists, one backbone

    Four LoRA specialists — tool, RAG, soft-skill, and extract — sit on a single shared 1.2B-parameter backbone (LFM2-1.2B). No 7–8B-scale model required.

  4. 4

    Tiered verifier

    Before an answer ships, a tiered verifier checks it. Self-reported model confidence is discarded — it isn't trusted as evidence.

  5. 5

    Conformal abstention gate

    Calibration comes from conformal prediction, tuned for 95% coverage — not from the model grading its own work.

  6. 6

    Out-of-distribution rejection

    Requests that fall outside what the system was calibrated on are rejected rather than guessed at.

  7. 7

    Outputs

    Every request ends one of three ways: an answer, an abstention, or an escalation to a human.

Latest results — August 2026

What the research produced

Measured on our own hardware, not projected from a slide.

36% 4%

Silent-error rate, down from an unverified 7–8B baseline, on our internal benchmark.

2.2×

Tokens per second vs. baseline on the same host (19.5 vs. 9.0 tok/s).

4.5 GB

Active memory, vs. 13 GB for the baseline.

Q4_K_M

No measurable quality loss vs. F16 in our quantisation sweep.

Still open
24%
Abstention on solvable tasks
vs. a <15% target

We are driving abstention down without raising silent errors — we do not claim this is solved.

How we measure

We measure honestly, including where the result falls short of our target. A model's own confidence in its answer is never treated as evidence — every number here is checked against held-out data, not the model's opinion of itself.

What didn't work

We publish what doesn't work, not only what does. A negative result, measured honestly, is still research.

We tried a training-free "context compiler" — turning a document directly into a LoRA adapter, with no training step at all. We tested it four different ways, and it didn't clear the bar we set for it in any of them.

The reason lines up with a simple trilemma: quality, zero training, and zero context tokens don't come as a set of three. A system can have any two.

The pick-two trilemmaA triangle with quality, zero training, and zero context tokens at its three corners, labelled "pick two" at the centre — a system gets two of the three, never all three.pick twoQualityZero trainingZero context tokens

Open for collaboration

Our mission is to make technology accessible to everyone — that's easier with partners who share the problem.

Universities & labs

Joint work on verification, conformal calibration, and evaluation methodology — including the negative results we publish internally.

Email us

Hardware & model partners

We build for consumer-grade hardware. If you make GPUs, NPUs, or open-weight models, there's a shared interest in what runs well where.

Email us

Design partners (pilots)

Teams running Orchestrator Studio or Vexilo pilots shape what ships next. Early access, direct line to the people who built it.

Email us

Research you can hold the product to.

Talk to the team doing the research, or see what it has already shipped.