Analysis integrity (social science)

Social science's decisive risk is not a crashing script — it is a confident, provocative misreading of a correct number, silent p-hacking, and results that don't replicate. Your job here is to run analyses and surface raw output, and to withhold interpretation the design doesn't support.

Pepinsky's rule: use the agent for tasks that follow rules; do not use it for tasks that generate answers, arguments, or interpretations.

Execute — don't interpret

• Report the estimate and its uncertainty — coefficients, standard errors, confidence intervals, test statistics, p-values, N — exactly as the software produced them.
• Do not volunteer causal or "provocative" claims. Regression and correlation are associational. Say "X is associated with Y", not "X causes / drives / leads to / increases Y", unless the design (RCT, IV, DiD, RDD, panel FE with a credible identification strategy) supports it — and then name the design.
• Do not tell the user what they want to hear. If the result is null or ambiguous, say so plainly.

Reproducible execution (fixed seeds + traceability)

• Any randomised step (bootstrap, permutation, train/test split, resampling, MCMC) must fix a seed: np.random.seed(...), random_state=..., or R set.seed(...).
• Every numeric claim in a report must be traceable to a script + line + output (provenance records this automatically when you write files).

Stata / SPSS / R round-trip

Read proprietary formats with real libraries — never transcribe numbers from memory. .dta and .sav round-trip through R (base foreign / haven) or pandas; use a fixed seed so estimates reproduce exactly:

df <- foreign::read.dta("data.dta")   # or haven::read_dta / haven::read_sav
set.seed(1)
m <- lm(y ~ x, data = df)
summary(m)                            # report coef + Std. Error verbatim

import pandas as pd
df = pd.read_stata("data.dta")        # or pd.read_spss("data.sav")

Report the coefficient and its standard error; if you compute the same model two ways (pandas vs R), confirm they match to the printed precision.

Run the integrity gate

The deterministic gate ships beside this SKILL.md. Run it on the workspace (or named files) before you report results:

python "$XDG_CONFIG_HOME/opencode/skills/stats-integrity/stats_integrity_check.py" [files...]

It prints one ```review fenced JSON block covering three risks:

• stats · interpretation — causal / provocative language over an association in a report.
• stats · prereg — a predictor or interaction the code runs that a preregistration plan (preregistration.md / analysis_plan.* / prereg.* in the workspace) never named — a HARKing path.
• stats · seed — a randomised analysis with no fixed seed.

Reporting

Copy the ```review block as the last thing in your message — the app renders it as dismissible reviewer cards. Never tell the user the analysis is "correct", "sound", or that a relationship is causal from observational data — the gate checks specific risks only.

Adding a check

Add a check_<name>(...) function in stats_integrity_check.py and call it from run(); each finding carries its own tag, so the app needs no change.