Every lab's starter zip, guide, and environment file; the six reference
subagents and the referee-review skill; and the shared infrastructure that
keeps API keys and the open web out of participants' hands. All links point
at the four-lab-redesign branch on GitHub unless otherwise noted.
Each row is a starter folder — download the zip, or browse the source (which also has the instructor-only manifest, never included in the zip).
| Lab | Guide / CLAUDE.md | Starter zip | Notes |
|---|---|---|---|
| 1 · Literature review as a pipeline Specify |
GUIDE.md · CLAUDE.md · FILES.md | starter.zip | ~440-record corpus, dedup log, 18 OA full-text stand-ins, deliberately imperfect search query. Semantic Scholar proxy is live — see infrastructure. |
| 2 · From the web to a dataset Plan + Specify |
GUIDE.md · CLAUDE.md · FILES.md | starter.zip | 18 real, curated European companies (7 clear AI-governance / 6 vague / 5 none). Mirror status: serving the fully-synthetic fallback set at /mirror/ until the real site capture is done — see the lab's FACILITATOR_NOTES.md. |
| 3 · Analysis under supervision Supervise |
GUIDE.md · CLAUDE.md · FILES.md | starter.zip | 2,500-row synthetic survey; a naive merge silently drops exactly 417 rows (the mid-lab crisis). Participants may not write or read code. |
| 4 · Reproducing and stress-testing a paper Audit |
GUIDE.md · CLAUDE.md · FILES.md | starter.zip | Self-contained synthetic replication package. Headline effect is significant without sector controls, insignificant with them — a genuine robustness-vs-fragility debate. |
Six specialist subagents used across the labs — delegate a narrow review task to the matching one rather than doing everything in one pass.
| Subagent | Description | Used in |
|---|---|---|
literature-scout |
Finds and verifies literature; checks whether a claim has support in the actual literature, or finds prior work on a sub-question. | Lab 1 |
identification-skeptic |
Hunts for confounds, pre-trend violations, and spurious causality in a research design or reported results. | Labs 3, 4 |
data-integrity-checker |
Checks data-wrangling and analysis code for silent bugs — dropped rows, wrong joins, unit mismatches, off-by-one windows. | Lab 3 |
stats-reviewer |
Checks whether the statistical method matches the question being asked — right model, right standard errors, right sample. | Labs 3, 4 |
writing-editor |
Tightens manuscript prose while preserving the author's own voice — editing passes, not new content from scratch. | Optional writing wrap-up |
disclosure-auditor |
Checks that every AI-assisted step has a matching entry in the governance disclosure checklist before a manuscript is submitted. | Optional writing wrap-up |
| Skill | Description | Used in |
|---|---|---|
referee-review |
A critical peer-review pass over a manuscript draft — a skeptical referee check before treating a draft or results section as final. | Lab 4 critic/audit roles; optional writing wrap-up |
Both hard constraints of the workshop in one line: participants direct Claude Code themselves, and nobody ever handles an API key. Design notes: infra/README.md.
| Component | Status | Description |
|---|---|---|
| Semantic Scholar proxy | live | Reverse proxy at https://uva-ete.strezoski.com/s2/,
forwarding to the Semantic Scholar API. Injects the institutional key
server-side (or runs keyless at classroom-scale rate) — participants
never see a key. Source: infra/s2-proxy. |
| Lab 2 website mirror | fallback live | Serving the fully-synthetic 12-company fallback at
https://uva-ete.strezoski.com/mirror/ today. The real,
curated 18-company mirror is the next step — see the lab's
FACILITATOR_NOTES.md. |
| Claude access | decided | Claude Team/Enterprise seats — participants log in once inside the FEB workspace, no keys, central billing. |
| Uniform environment | available | Devcontainer with Claude Code, pinned Python, R, git — see .devcontainer/. |
| Replication Docker image (Lab 4) | reference only | Pinned image for the synthetic replication package — see instructor/Dockerfile. |