Skills pack v0.1
Fifteen Agent Skills covering econometric practice, R and Stata, economic reasoning, and number provenance — with the three method standards behind them.
Read announcementOpen-source econometrics for AI
AI that does econometrics properly.
Language models can write R and Stata. They cannot reliably do econometrics, and they carry economic priors they never declare. Equilibrium Labs is the open layer that supplies the practice — skills, standards, and tooling, with nothing held back.
What exists
Available now
Shipped
In the repository and usable today. Clone it, copy the skills into your agent, run the linter over your analysis, and read the standards that govern both.
- Skills pack
Fifteen Agent Skills covering econometric practice, R and Stata, economic reasoning, and number provenance. Install by copying directories.
- Econometrics linter
32 rules over R, Stata and Python. Reads a script and reports the wrong clustering level, the silent default, the reflex transformation. No dependencies, no network, no model.
- Method standards
The three written standards the skills enforce: econometric reporting, economic reasoning, and provenance. Argue with these, not the skills.
- Skill validator
A dependency-free Python checker that holds every skill to the format contract, run in CI on every push.
- econ-eval
24 frozen econometric tasks with fully programmatic scoring — no LLM judge. Scores the absence of a wrong reflex, not only the presence of a right answer.
- MCP server
Four tools over the same engine, so an agent can lint its own regression before reporting a number. Offline and deterministic; it fetches no data.
- Harness profile
Wiring that puts the skills, the linter and the MCP server into whatever agent you already run. Deliberately not an agent — thin harness, fat skills.
- RL reward adapter
Turns an econ-eval score into a scalar a training loop can optimise. A rule-engine reward, because text similarity rates a correct answer in the wrong dialect as a failure.
Not built yet
Planned
Committed to but not built — listed here rather than described as finished. The roadmap says why this order, and what was deliberately ruled out.
- Series-identity CLI
The record and the comparability check exist as a library and as an MCP tool. There is no eqlabs identity subcommand yet.
- Worked examples
Complete analyses from question to write-up, in R and Stata, as plain files any agent can execute through a shell.
- PyPI release
The package builds and a Trusted Publishing workflow exists. It has never been run, and the one-time publisher setup is a manual step not yet done.
- Training a local model
Gated on the evaluation set discriminating and showing a measurable gap. An honest negative result is a likely outcome and a publishable one.
Skills
Fifteen skills, four groups.
The pack covers designing and estimating (identification, causal designs, panel data, time series, inference, diagnostics), writing it up in R and Stata reproducibly, reasoning about economics without inherited priors, and giving every number its identity. Each skill states the specific model failure it prevents — and 32 of those failures are now executable rules the linter checks.
Practice
Designing and estimating
econometric-workflowidentification-strategycausal-designspanel-datatime-series-econometricsstandard-errors-and-inferenceregression-diagnostics
Software
Writing it up in R and Stata
econometrics-in-reconometrics-in-statareproducible-analysis
Reasoning
Economics without inherited priors
economic-reasoningparadigm-pluralismbias-audit
Numbers
Provenance and series identity
series-identitynumber-hygiene
See what each one does
Agent Skills format — a directory with a SKILL.md. Install by copying into .claude/skills, or read them as documentation.
Latest announcements
All announcementsWhat we are
We build the layer that makes a model show its work: the identification, the assumption, the series, the uncertainty.
Every rule in the pack exists because models fail in a specific, repeated way — a default standard error on clustered data, a contested literature reported as settled, a nominal figure compared against a real one. Naming the failure is what makes the rule hold. Where the honest answer is that a question cannot be identified from the data available, the model is instructed to say so and stop.
Fully open
Everything published, nothing withheld. Code under MIT, written material under CC BY 4.0. No paid tier, no hosted edition, no gated component.
Checkable
Rules carry citations where they are contestable and machine validation where they are testable. Disagree with one and the standard behind it is in the repository to argue with.
Neutral on outcomes
The project has views about method and none about policy. A skill that acquires an opinion about what should be done is a bug, not a feature.