Research notes

Empirical investigations spun off from building Dazzle — short, reproducible studies of questions that came up while making an AI-collaborative software substrate. Each note states a question, runs a real experiment, and reports what held up (including what didn't, and any measurement mistakes caught along the way).

Notes

• Predicting task context from a code graph — how far a cheap static-graph baseline gets you at predicting which code must be in an agent's context before a task. A naive keyword seed (45%) + forward closure (73%) + bidirectional edges (91%), in ~50 deterministic lines — and why the residual is a retrieval problem, not a graph or database problem. Runnable eval scripts included.
• Do agent-written code biases persist across model generations? — an epoch-stratified study across four real model generations (Opus 4.5 → 4.8), attributed by commit trailer. A four-way taxonomy (model-shed / substrate-held / disciplined-rise / campaign-noise) and the finding that no measured construct worsens with newer models — including a rising-noqa regression that acquits under the same convict/acquit discipline the reflection programme uses on code.
• Can a cheap detector catch re-implemented capability before it's written? — a ~70-line Type-2 clone detector over the project's own agent-generated code. 3.2% of functions sit in near-duplicate clusters, including genuine re-derived capability (one availability check re-implemented in four files) — but also parallel-by-design families, so the cheap reuse check is a screen, not a verdict. Tests the B(ii) reuse half the context-bounds spike left open.

Related

• Agent self-reflection — the programme that motivated the context-prediction study: discovering counter-priors for agentic code production by having the agent adversarially interrogate its own assumptions.
• Model-driven failure modes — the 4GL/MDE/CASE failure-mode threat model these investigations operate against.