ai4s-research / domain-check

Domain-correctness gate

Across every field the top complaint is code that **executes cleanly but is
scientifically wrong**. This gate intercepts that field's classic error classes
deterministically — by analysing the code you actually wrote, not by
recalling rules. It verifies specific error classes; it never proves correctness.

It is fast, offline and stdlib-only — but it is an on-request gate, not a
background one.

When to run

• When the user asks to check, validate, review or audit an analysis for
  correctness.

• When you are the reviewer agent, running the checks that agent's prompt
  lists (the app's auto-review turn, which the user switches on in Settings).

Do NOT run it on your own initiative while writing or executing analysis code.
A clean run still emits a review block, so an unprompted pass puts a card with
no findings in the user's conversation — noise they did not ask for.

How to run

The gate ships beside this SKILL.md. Run it on the code files in play (or with
no arguments to scan the workspace):

python "$XDG_CONFIG_HOME/opencode/skills/domain-check/domain_check.py" <file.py|notebook.ipynb|analysis.R ...>

It prints exactly one review ` fenced JSON block on stdout.

What it catches (one rule set per discipline)

• physics · units — adding/subtracting/comparing quantities of different
  dimensions (e.g. t_seconds + d_meters); trig on a degree-valued angle.

• earth · crs — Euclidean/Pythagorean distance on latitude/longitude
  (sqrt((lat1-lat2)**2 + (lon1-lon2)**2)); a geopandas geometric op with no
  CRS ever set.

• biology · coords / strand — off-by-one on BED intervals (0-based
  half-open, so length is end - start, never +1); a sequence sliced from a
  stranded feature file (GFF/GTF/BED) with no reverse-complement for the -
  strand.

• chem · valence — a SMILES string literal (assigned to a smiles/smi
  variable, or passed to MolFromSmiles/MolFromSmarts) that cannot be a real
  molecule. If RDKit is installed it is used as the authoritative judge —
  Chem.MolFromSmiles sanitizes the parse, so it catches far more than a
  five-bond carbon (bad ring closures, impossible aromaticity, over-valent
  N/O/S) and, being authoritative, clears molecules a heuristic would
  wrongly flag. Without RDKit it falls back to a stdlib bond-counter (carbon

  4, over-bonded halogen; bails on bracket atoms for precision).

• social · multiple-comparisons — a significance test (ttest_ind,
  pearsonr, f_oneway, chi2_contingency, …) run inside a loop or ≥3 times
  with no multipletests/FDR/Bonferroni correction anywhere — the inflated
  family-wise false-positive rate that silent p-hacking produces.

• social · categorical — a numeric reduction (.mean()/.median()/.std()
  …) taken directly on a nominal category code (gender, race, region,
  condition, …), treating an unordered label as an interval quantity. A
  groupby('gender') key is correct usage and is not flagged.

• bioprocess · unconstrained-kinetics — curve_fit fitting a local model
  function whose parameters are classic non-negative kinetic constants
  (Monod/Haldane mu_max/Ks/Ki, Pirt/yield Yxs/Yps, Luedeking-Piret
  alpha/beta, …) with no bounds=, letting a noisy or sparse fit converge
  to a physically impossible negative value. Names that belong to any curve at
  all (alpha, beta, kd, ka) count only beside an unambiguous one or in
  a file that otherwise reads as a fermentation, so a plain power law, whose
  exponent is routinely negative, is left alone.

• bioprocess · kla-driving-force — in a file that computes kLa, taking
  log() of the raw dissolved-oxygen reading instead of the (C* - C)
  driving force the dynamic gassing-out method requires (dC/dt = kLa(C*-C),
  so the regression is on ln(C* - C), not ln(C)).

• bioprocess · cfu-log-scale — a numeric reduction (.mean()/.std()/…)
  taken directly on a raw CFU (colony-forming-unit) plate count. Microbial
  counts are approximately log-normal; the convention is to average
  log10(CFU), not the raw count. A variable already named as the log
  quantity (log_cfu) or a .mean() taken after np.log10(...) is not
  flagged.

• bioprocess · anova-assumptions — f_oneway/anova_lm/
  pairwise_tukeyhsd/tukey_hsd run in a file with no normality
  (shapiro/normaltest/anderson) or variance-homogeneity
  (levene/bartlett/fligner) check anywhere in it — the two assumptions
  the test's stated false-positive rate depends on. Scoped to files that read
  as a fermentation (biomass, bioreactor, CFU, fed-batch, …): the
  statistics generalize, but a finding tagged bioprocess on a three-arm
  survey does not.

• bioprocess · rsm-first-order-fit — a Box-Behnken/central-composite
  design (bbdesign/ccdesign/box_behnken/central_composite in the
  file) fit through a statsmodels formula with no quadratic (I(x**2)) or
  interaction (x1:x2) term — the design was built to estimate curvature, so
  a first-order model wastes it and cannot locate an interior optimum.

• bioprocess · cross-scale-fit — a real model fit (.fit(...), guarded to
  a file that imports scikit-learn, XGBoost, statsmodels, PyTorch,
  TensorFlow, or Keras — a plain scipy.curve_fit never counts) in a file
  that mentions both a small-scale (flask, bench-scale) and a large-scale
  (bioreactor, fermenter, fed-batch, pilot-scale) process vocabulary, e.g.
  trained on flask data and applied to a bioreactor. Advisory: a warn, not a
  defect. Silent once a calibration / scaling-factor / cross-validation term is
  in the file, and silent on batch_size and friends — bare batch matched the
  training knobs of the very libraries this rule requires.

Rules favour precision: an unrecognized unit, arithmetic with no discipline
signal, a SMILES using bracket atoms (which carry their own valence/charge), a
single significance test, or a categorical used only as a groupby key is left
silent rather than flagged.

Reporting findings

Copy the review ` block the tool prints as the last thing in your
message — the app renders it as dismissible reviewer cards. Do not paraphrase
the findings into prose and drop the block; the structured block is the
contract. If the gate found nothing, say so plainly and keep the block (its
note states that no findings is not a guarantee of correctness).

Never tell the user the code is "correct" or "error-free" — the gate checks
known error classes only.

Adding a discipline

Add a check_<field>(ctx) function in domain_check.py and append it to
VALIDATORS. No other change is needed — the review contract and the app's
rendering are discipline-agnostic (each finding carries its own tag).