The records behind a study
Allagma gives a coding agent a research process it can carry between sessions. Its methods ask the agent to formulate a question, freeze a protocol, execute within limits, inspect uncertainty and tie claims to evidence. The helper retains the artifacts that let a researcher check what actually happened.
Three useful distinctions
Section titled “Three useful distinctions”A study is the continuing question; a campaign is one frozen plan. A study can contain several campaigns when a protocol changes. A campaign points to exact methods, scientific code, inputs and settings. An analysis revision gives a particular interpretation of included results without overwriting the old one.
A run is a planned observation; an attempt is an execution. A failed attempt and its retry are two attempts at one run. They are not two independent scientific observations. Pilots also do not become untouched confirmation data.
Completion is not assurance. Phase says where the workflow is, execution status says what the process did, and assurance says what kind of checking was performed. A completed session with a deterministic audit is not an independent scientific peer review.
How the pieces fit
Section titled “How the pieces fit”| Piece | Responsibility | Example |
|---|---|---|
| Method | Portable instructions with an input/output contract | Form a protocol, analyze results, audit claims |
| Recipe | Ordering, handoffs, branches and stopping rules | The default research workflow |
| Adapter | Connect a host, runner or reviewer to those contracts | Codex skill delivery or local execution |
| Bundle | Exact copies selected for a study | Methods plus the helper at a content digest |
| Lock | Identity, composition and configuration of that bundle | The campaign’s lock.yaml |
| Claim | A scoped statement linked to supporting/contradicting evidence | Supported average effect, contradicted universal claim |
Methods are canonical Agent Skills. Provider-specific behavior belongs in adapters. Hypotheses, training code, metrics and statistical choices belong in the study. A locked copy is generated evidence, not another editable source.
What the agent still has to do
Section titled “What the agent still has to do”Allagma does not supply the scientific answer, choose a universally valid statistical test, or replace expert judgment. An agent can still misunderstand a source, write a bad experiment or overinterpret data. The records make those failures more inspectable; deterministic tests cover only what they test.
For implementation detail, continue to architecture, contracts, study-owned programs, or versioning.
Canonical source · Generated from the maintained repository document at this build’s commit.