polarize.tech project instructions

Adaptive-preregistration scope

This is primarily a publication repository, not a model repository. The adaptive-preregistration protocol applies when work performed here will produce evidence for a public claim.

• A Jekyll build or preview, a design or publication gate, a metadata/content sync, and a test of deterministic site behaviour are maintenance. They do not need an experiment ID.
• A simulation, statistical analysis, benchmark, measurement, A/B test, or other run whose numeric or qualitative result will support a post or public claim is a citable experiment. Register it in EXPERIMENTS.md and follow .agents/protocols/PREREG_PROTOCOL.md before its scoring run.
• Results imported from another repository keep that repository’s preregistration and experiment ID. Do not create a second registration here; preserve the provenance in the post and run the publication gate.
• Exploratory analyses may run only as exploratory work and must not be presented as preregistered evidence.

Agent protocols

This repo uses KIT Adaptive Preregistration: agent protocols that are maintained in one central repo and vendored into .agents/. Every agent follows them, whatever the tool (Claude Code, Codex, Copilot, Cursor, Gemini, ChatGPT…).

• The rules in this block and the documents in .agents/protocols/ are binding. If a task conflicts with one, stop and say so before doing the task.
• Project-specific instructions (outside these markers, or in nested AGENTS.md files) may add to the protocols. If one contradicts a protocol, ask the user which wins.
• Don’t edit files in .agents/ or text between the kit_ap markers. Updates overwrite them. To change a protocol, propose the change to the user as an edit to the kit repo.
• If a session-start message says the protocols are out of date, tell the user once. Don’t update without their go-ahead. If your tool has no hooks, run .agents/bin/kit_ap check once at the start of a session.

Preregistration

Read and obey .agents/protocols/PREREG_PROTOCOL.md before running, modifying, or reporting any experiment, and .agents/protocols/SCOPE_PROTOCOL.md before building anything in a unit (an app, sim, tool, calculator, …; a folder with preregistrations/, or the repository itself).

• A claim, or an exploratory label. A unit starts with a falsifiable claim in the user’s own words and what would count against it, recorded in its SCOPE.toml: one question, and the rest of the claim is stated when its first experiment is preregistered. If the user says a unit is exploratory, don’t ask for a claim: record what is being built (stage = "exploratory"), label the unit exploratory, and build. Nothing an exploratory unit shows is a finding, and it gets its claim before its first experiment.
• Propose; never substitute. The user’s claim, design and decisions are recorded as they give them.
• Infrastructure is yours to decide, following the organisation’s conventions. Anything that implements a research concept, or presents a result in a way that changes what someone would conclude, is science: once the unit has a claim, build it only after it has an evidence basis and the user’s recorded decision. A gap blocks that feature only.
• Ask one decision at a time. If you disagree, say so once, with evidence; then record and follow the user’s choice.
• Never count a source you couldn’t verify. .agents/tools/scope-status shows unscoped units and what is open.