Field Notes

Field notes from agentic systems in regulated production.

Written by the engineers and principals who built the systems — not analysts summarising what others shipped. Technical depth. Regulatory precision. No vendor content.

Agentic ArchitectureEU AI ActAudit InfrastructureRegulated AIProduction Notes

Telemetry
Hub

Agent Efficiency+42%

Anomalies Detected0.001%

Threats Mitigated99.9%

Execution Accuracy100%

01 / Featured Insight

What changes when you read the EU AI Act as an engineering specification — not a compliance checklist.

EssayFinancial Services

22 min read

Article 9 through 15, applied to a working loan-origination agent — with the Layered AI architectural choices that satisfy each obligation.

The EU AI Act is 144 pages written by lawyers for lawyers. This piece translates Articles 9 through 15 — the high-risk system obligations — into concrete engineering decisions on a working loan-origination deployment. For each article, we state what it actually requires in plain language, then map it to the specific Layered AI layer that satisfies it, and describe the artefact a notified body would actually want to see.

Article 9 (risk management) maps to L03. Article 10 (data governance) maps to L02. Article 12 (logging) maps to L01. Article 14 (human oversight) maps to L04. This is not theoretical. These are the choices we made on a production deployment that has been examined by regulators.

Read the Full Piece

Access the complete architectural breakdown and layer mapping.

02 / Notes

Shorter pieces — technical, regulatory, and opinionated.

Click any title to read the full piece in-line. These are not summaries — they are substantive arguments drawn from our deployment experience.

Technical14 MIN READ

The retrieval problem in regulated AI: why we abandoned the open web.

What happens when you retrieve from Wikipedia and HMRC guidance PDFs in the same vector store — and why the answer forced us to redesign L02 entirely.

Engineering team · 2026

READ ARTICLE

Regulatory11 MIN READ

How HMRC examined our system: the things we did not expect.

A candid account of what a national tax authority actually spends time on when they examine an agentic system — and what it taught us about evidence pack design.

Principals · 2026

READ ARTICLE

Architecture9 MIN READ

Why the LLM is the least trusted component in an agentic system.

The architectural argument for treating your foundation model as a probabilistic sub-component — not the source of truth. Written for engineers who have shipped LLM features and are about to ship agents.

Quad Research · 2026

READ ARTICLE

Sector12 MIN READ

Consumer Duty evidencing as an emergent property, not a deliverable.

How an FCA Consumer Duty evidence pack falls out of L01 when the architecture is right — with specifics from a live UK lender deployment where the regulator never had to ask twice.

Financial services practice · 2026

READ ARTICLE

Sector10 MIN READ

Pharmacovigilance without the manual XML: what changes when E2B is generated, not drafted.

Schema-valid E2B(R3) reports as a mechanical output of the PV process — not a hand-validated XML exercise. With throughput, error rates, and MHRA gateway feedback from a live deployment.

Healthcare practice · 2026

READ ARTICLE

Opinion7 MIN READ

Stop asking AI vendors about ISO 42001. Ask them about evidence schemas.

A short polemic on AI procurement in regulated industries — the questions that distinguish vendors who have actually shipped from vendors who have built impressive presentations.

Principals · 2026

READ ARTICLE

03 / Field Notes Inbox

Get the next piece before it hits the site.

One email when a new piece drops. No newsletter cadence, no digest, no marketing. Just the next field note — when it is ready.

No cadence

No marketing

Unsubscribe instantly

Technical Briefings

Work with the engineers who built it

We run deep-dive sessions with engineering and risk leaders on a request basis. No sales engineers, no decks — just architecture and code review.