[arena-ai-coding-agent]

Two things in one PR, because they have one cause: the layer that claimed to compute
TSS/CSS/RSS offsets did not, and the layer that accepted them refused them.

1. The refusal — main is red because of this

NormalizationConfig canonicalises its method to lower case ("TSS" → "tss"), while
VALID_NORMALIZATION_FOR_METHOD listed "TSS", "CSS", "RSS". The membership test
therefore failed for those three spellings and AnalysisConfig raised
normalization.method 'tss' incompatible with method 'nb_glm' — including for the
existing test/unit/test_execution.jl testset TSS offset for NB_GLM, which constructs
method="TSS" exactly as the documentation instructs. That is the failure behind the red
Julia tests job on the TSS-offsets commits; it is not a flaky test and it is not the
new estimation layer.

The table and both comparisons (constructor, validate_config) are lower case now, the
JSON-schema enum still accepts the upper-case spellings for existing documents, and
test/unit/test_scaling.jl asserts that every admissible spelling of every method name
is accepted.

2. The wiring — issue #16's offsets are computed, not aliased

prepare_analysis_table now computes scaling factors through src/analysis/scaling.jl
and hands the offset to the fit:

declared	response	offset
tss / none (count model)	counts	log(library size)
css	counts	log cumulative sum at the declared css_quantile
rss (TMM)	counts	log trimmed weighted mean of log-ratios to a reference sample
size_factors	counts	log median-of-ratios (RLE) — the estimator the name claimed
relative	proportions	none (a transform, not an offset)

Previously css/rss were aliased to relative under @warn, tss did that for every
method but nb_glm, and size_factors computed library-size centring wearing DESeq2's
name. css/rss are now refused for a response with no counts to offset — in the
configuration layer and again in the execution path.

css_quantile (0.75), tmm_ref_column (chosen the way edgeR chooses it, and recorded),
tmm_log_ratio_trim (0.3) and tmm_sum_trim (0.05) are validated in the constructor,
included in the config hash and canonical JSON, round-tripped through
to_json/from_json/Nickel/DEED, mirrored in the JSON schema and Nickel contracts,
exposed through the server's configuration route, and explained in context_help and the
frontend help. A run that declared a different quantile is a different run.

diagnostics.checks["scaling"] and the manifest provenance record the kind, the
definition, the parameters, the raw quantities before centring and a SHA-256 of the
offset vector.

Evidence

• test/unit/test_scaling.jl: known answers with the arithmetic written beside them;
  properties a wrong implementation fails (feature permutation, a sample that is an exact
  multiple of another, invariance to a shared constant); the CSS outlier-robustness
  property; negative controls for every refusal (zero-total sample, zero CSS cumulative
  sum, unknown reference, disjoint supports, out-of-range trims); the integration path
  through prepare_analysis_table; and a base-R transcription of the same conditions,
  which skips loudly by name where R is unreachable.
• Conditions document, published before the implementation and merged in feat(analysis): exact TSS/CSS/RSS offsets, and no silent substitutions #61:
  docs/statistics/method-conditions/scaling-and-offsets.md.
• Local gates: check-spdx.sh OK (284 files), check-format.sh OK (331 files),
  check-lint.sh OK (tsc --noEmit clean over frontend/src).

Not claimed

Parity with metagenomeSeq::cumNorm and edgeR::calcNormFactors at 1e-6 is not
verified: neither package is in renv.lock, and this PR does not add an unpinned R
dependency to make a test pass. The base-R transcription catches implementation slips, not
specification errors. That outstanding condition stays recorded in issue #16 and in the
conditions document rather than being asserted as met.

Closes #16.

[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: 9bc858b1-5d6b-4943-a683-5f7aeec1debd

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.}