About

This is the research programme of Warren Smith, in the UK. The work is on whether claims about AI systems can be independently checked. Every result here is tied to a public artifact and a commit, and graded by how far the evidence actually reaches.

The artifacts are on github.com/repowazdogz-droid. Contact: warrensmith8@ymail.com.

Engineering work

The research runs alongside engineering practice rather than instead of it, and has done since 2023, though the public artifacts here date from 2026. Since 2024, Principal Engineer at WRKS Holdings Ltd in Bristol, where the work is eight applications on one shared Python engine, owned from the engine through React frontends to a packaged desktop build and the deploy, and a governed runtime for decision records with cryptographic record hashing over an append-only trail. On a shipped clinical product that meant removing a client-inlined model API key, routing the browser through a same-origin server-side proxy, capping request bodies, and putting both model-output paths behind one guarded send with a prebuild script that fails the build if the chokepoint is bypassed.

Before that: Head of Data for Clickout Media's Sports PR team from August 2024 to January 2026, Head of Finance at Block Labs, and three years at Deloitte in the internal finance function of Propel, its outsourced accounting and business-services firm for SMEs. The question this site turns on, what a passing check actually establishes, is an old one in accounting and assurance before it is a question about software.

Corrections to this programme's own results

The method includes attacking the instrument when the instrument is one of these. Where that has changed a published result, the change is recorded rather than absorbed:

• A released claim in mcp-authority-boundary stated that an observed effect matched the authorized operation. The shipped ledger executed neither tool for which independent read-back exists, so the claim was corrected downward on 14 August 2026 and a test now fails the build if that combination recurs.
• A security result was published as 12 of 12 attacks caught. A row-counting error had double-counted one attack that two gate layers each caught. The corrected figure, 11 of 11 in scope across 12 catch events, is the one now published, together with the 6 attacks that are structurally out of scope.
• In review on an external pull request, supporting evidence offered for the change was withdrawn after measurement showed it had sampled the wrong population.
• A pilot on 45 answers is kept on this site and superseded for anything about agreement and accuracy by the 600-item study.

How the work is stated

A claim and its limit sit next to each other, in the same weight. Numbers carry their denominators. The verb matches the grade. A result is said to be proved only when it is machine-checked inside a named frame, measured only with a stated sample, observed only in a single instance. Where a check has no grounded instance, that is stated rather than filled with a borrowed one. Decisions are first person; results are impersonal.

Licensing

The public artifacts are released under their own repositories' licences, predominantly MIT for code and permissive terms for accompanying documents; each repository states its own. The proofs, specifications and tools are open so that the claims resting on them can be reproduced from a clean clone, which is the standard every published finding here is held to.

Some work in the wider programme is not public. Where a result depends on an artifact that does not reproduce from a public clone, no page is built for it, and the dependence is recorded rather than hidden.