What problem was I trying to solve?
In staged computational research, the sequence — question, protocol freeze, execution, audit, reporting, handoff — comes apart quietly. By the time a later session picks the work up, what was actually run, what was decided, and on what basis is no longer recoverable from the artifacts. This is the failure I had already watched cost me a project, documented in Agent Harness Anatomy.
What did I build?
An Agent Skill for Codex and Claude Code. It adds one canonical, machine-readable research status record; prospective, exploratory and confirmatory study labels; explicit verification-versus-validation language; four audit levels that stop an automated check from being written up as independent review; preservation of failed, null, excluded and inconclusive runs; and plain-language handoffs that carry exact evidence paths and the conditions for re-entry.
State lives in two files: a canonical status record, and an append-only event log. A dependency-free helper CLI initializes, freezes, validates and summarizes the workflow. Two operating profiles differ in how much ceremony they impose, but not in the scientific invariants — so “I was in a hurry” cannot become “I skipped the gate.”
What actually happened?
The skill is released and its structure is inspectable. It has not yet been carried through a full research cycle of my own from question to handoff, and this page will say so until it has been. A gate that has never refused anything has not been observed working.
What did I learn?
Discipline has to be a gate that can refuse, not a document that can be read. And the invariants have to be separated from the operating intensity, or the intensity setting becomes the escape hatch.
What can be verified?
SKILL.md states the invariants and the state schema; both can be read without running
anything. Tests cover the operating profiles, platform handling and gate behaviour. The
repository is also explicit about what it does not do: it does not judge whether a
scientific design is substantively correct, and it does not replace domain review, ethics
review, statistical expertise, or regulatory controls.