[arena-ai-coding-agent]

The problem this fixes

Execution.run_analysis returned, per feature:

p_val   = 0.01 + (h % 100) / 1000.0     # h = hash(taxon_id)
padj    = p_val * 1.5                   # "mock BH"
log2fc  = (h % 20) / 10.0 - 1.0
baseMean= 100.0 + (h % 1000)

Every number is a deterministic function of the feature's name. Nothing in the
run touches the counts. The rows render in a table as p-values, fold changes and
means, and a caller has no way to tell them from a fit. The source called it a stub;
the caller received it as a result.

What replaces it

src/analysis/estimation.jl (catalogue item 2 — conditions published before the
implementation, in docs/statistics/method-conditions/parametric-fits.md):

method	fit	note
nb_glm	MASS::glm.nb, per feature	offset required; glmGamPoi/local/mean/pooled refused by name (issue #21)
clr_lm, ilr_lm	stats::lm, per feature	offset refused: it means nothing to a Gaussian fit on CLR units
logistic	stats::glm(binomial), per feature	0/1 response required; proportions refused rather than given invented weights

• Unsuccessful states are states. A failed fit gets status = "failed", a note
  saying why and null for every statistic; it is excluded from the BH family and
  counted. Nothing that cannot be fitted gets a number.
• A run that cannot happen says so. No design, no metadata column, single-level
  grouping, RBusyError past the timeout, R unreachable → status = "not_run" with a
  reason and an empty result set, and a warning on the diagnostics. An empty result is
  never dressed up as "no significant features".
• Refusals, not substitutions: interactions/transformations/random effects/nesting
  in a formula, $/backticks/; in a formula, offsets on non-count models, count models
  without one, binomial fits on proportions.
• Provenance: R and MASS versions, formula terms and the contrast actually tested,
  offset kind and its SHA-256, BH family size, exclusion count, and the SHA-256 of the
  fits and coefficients tables as they crossed the R boundary.
• BH implemented in Julia and tested against R's p.adjust(method = "BH").

Evidence (test/unit/test_estimation.jl)

• Known answers written into the data (four-fold up, no effect, ten-fold down) rather
  than recorded from a previous run.
• Independent reference: the same counts fitted directly in R, compared coefficient by
  coefficient, SE by SE, p by p.
• Negative controls for every refusal, and for not_run/failed shapes.
• A source-level guard that hash(taxon_id), hash( and rand( are gone from the
  execution path.

Refs #1 (catalogue item 2 delivered; item 3's nonparametric tests and the independent
statistical review are still outstanding there). Issue #21 stays open: glmGamPoi is
refused rather than aliased, and the refusal says so.

[coderabbitai]

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Advanced

Run ID: cf416472-674e-45e0-b8b3-81527cfea448

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}