Zero Drift · white papers
Claims worth trying to break.
AI changes how software is made. These papers ask where that change helps, where it creates new risk, and which parts of a useful product still need exact rules and accountable operation. Pick a question below; you do not need to know our internal vocabulary.
Follow a question
Two papers for each question, in a suggested reading order. Or browse all eight papers.
All papers
Each paper includes its argument, sources, limits and a downloadable PDF.
Compile the specification
AI can reduce the effort of producing code, increasing the relative importance of resolving ambiguity before implementation. Specification-driven development turns business intent into versioned schemas, state, contracts and executable acceptance evidence before agents multiply the wrong assumption.
Read the paperThe depth signature
A feature is not complete because its happy path renders. Its required depth depends on the system it changes. This paper defines a risk-shaped signature for testing security, data, concurrency, recovery, performance, accessibility and operations at the right level.
Read the paperThe deterministic gap
Generative AI can propose, explain and build at extraordinary speed. It still cannot make a consequential answer true. This paper separates inference from computation and argues for systems in which models expand the option space while explicit rules, state and evidence close it.
Read the paperFive laws of agentic delivery
Agentic engineering fails when a persuasive session is mistaken for a delivery system. Five operating laws turn model capability into controlled production work by moving plans, units, depth, model choice and parallelism into inspectable artefacts and gates.
Read the paperThe local AI workstation
Frontier cloud models and controlled local models are complements. Use scarce frontier reasoning to compile specifications and resolve ambiguity; use open-weight models on owned or rented hardware for bounded engineering roles, with harnesses and CI/CD gates deciding what survives.
Read the paperMetamorphic oracles
AI may reduce the cost of authoring and maintaining behavioural challenges, giving teams reason to reassess the economics behind the test pyramid. Metamorphic relations, mutation and system-level observation can turn tests from examples of success into an adversarial evidence system.
Read the paperUtility economics
Software construction and operation are becoming cheaper, but much of SaaS still prices scarcity, lock-in and feature withholding. This paper argues for complete products sold like dependable utilities: low-friction, fairly margined and valuable because they work, not because leaving hurts.
Read the paperZero-divergence refactors
Legacy replacement does not require a cut-down successor and a decade of feature-parity work. If behaviour can be observed, specified and challenged, a modern system can be built behind an equivalence boundary and moved into service without forgetting its users.
Read the paper