[jerome-white]

Closes #42. Part of #38.

What this does

Replaces the information-retrieval words the package came with (topic, system, score, B) by the statistical ones, in the API, the code, the tests, the fixtures, the reference tooling, the benchmark scripts and the docs, using the decisions taken on #42:

Before	Now
systems=	treatments=
result fields system1, system2	treatment1, treatment2
scores (matrix argument)	responses
value= (table keyword)	response= (so the trio is block, treatment, response)
scores_matrix, wide_scores	tidy_matrix, wide_matrix (unexported)
topic, system, score everywhere else	block, treatment, response

(B to trials was #39/#45.) Breaking, deliberately: the package is 0.1.0-DEV, nothing depends on it, so there is no deprecation. Error messages changed with the words ("need at least two treatments", "responses contain N NaN values", "more than one response for block ... and treatment ...").

Also: a Vocabulary section in the README (a table mapping block/treatment/response/trials to information retrieval, to machine learning, and to the old words, so readers of old issues and PRs can find the new name), and a NEWS.md that records the renames and what was added, removed and changed since the first extraction.

Commits (six, each with a detailed message)

1. Use the statistical vocabulary in the library: blocks, treatments, responses (src/, plus the call sites in tests, snapshot script and example, so the suite passes)
2. Use the statistical vocabulary in the tests
3. Use the statistical vocabulary in the fixtures and the reference tooling
4. Use the statistical vocabulary in the benchmark scripts
5. Use the statistical vocabulary in the README and the example
6. Start NEWS.md with the renames of the vocabulary pass

The audit (the first one was not thorough enough)

Method, as agreed on #42: grep the whole tracked tree for topic, system, sys, score, and for query, relevan, retriev, TREC, NTCIR, IR, dataset; classify each hit as rename, keep or glossary; rerun at the end and list what remains.

Before: over a hundred hits in src/ and several hundred in the tests (counted per word, so a line with two words counts twice; the table is on #42), plus the reference scripts, the benchmark files, the example and the README. After (everything outside benchmark/README.md and NEWS.md, which are covered below), the remaining hits are all on purpose:

• README.md lines 4-5: the opening sentence says "treatments (systems, models, methods) over blocks (topics, datasets, folds)", to say what the words mean in the fields it is used in.
• README.md Vocabulary table (lines 27-29): the old words, in the "In information retrieval" and "Formerly" columns.
• README.md line 133: the example of an undefined block, "a topic with no relevant documents", after "a dataset where the metric is undefined for every model".
• test/reference/README.md line 24, test/reference/arrangements.jl line 3 and test/reference/sakai_reference.sh lines 4-5: one sentence each saying that Sakai's tool calls blocks topics and treatments systems, where the scripts describe the file format of his program.
• The NTCIR URL of Discpower (README and script) and the tool's name: a citation, not vocabulary.
• dataset, model, accuracy in the table tests and the README examples: neutral machine-learning column names, used on purpose as the non-IR example.
• benchmark/README.md and NEWS.md quote the old words: the first is a record of measurements and is kept as recorded, with a note at the top that maps its words (B, systems, topics, scores) to the new ones; the second is the rename history.
• Git history, issues and PRs keep the old words; that is what the Vocabulary table and NEWS.md are for.

Things the first audit missed and this one caught: the helper-function and test-set names, the block ids t1..t3 in the tests and fixtures (now b1..b3), the NTOPICS/NSYSTEMS variables of the shell script, the printed header of the benchmark, the header of Sakai's reference output (system1,system2,...), and the fixtures' own header (topic,system,score).

Decisions made (on #42, by the owner)

• Output fields: treatment1/treatment2.
• systems= becomes treatments=.
• The measured value is the response (responses, and response= in the table method).
• Start a NEWS.md.

Verification

• ] test: 136 pass + 1 deliberately broken, at 1 and at 4 threads: the same count as master, since only names changed.
• benchmark/snapshot.jl output is byte-identical to master's, so the computation did not change.
• The fixtures were renamed in place (headers and the ids t1..b1..; the numbers are untouched). To check that the stored *_sakai.csv files are what the renamed sakai_reference.sh writes, I ran the script on the three fixtures (it downloads and builds Discpower, a few seconds each): its output is identical to the stored files.
• The example script run on tiny.csv (500 trials, seed 1) writes the new field names and the same p-values as before.
• The README recipe for complete cases was run on a tidy file with a NaN, with the new column names: it kept 3 of 4 blocks and ran.
• 80 columns checked on every tracked text file except the fixtures and the recorded benchmark tables.
• One stale statement was fixed on the way, in test/reference/README.md: it said Random.seed! does not make multi-threaded runs reproducible (true before Run trials in fixed-size chunks so results do not depend on the thread count #25).

Not covered

• The README rendering on GitHub (the table, the anchor link #missing-responses).
• The Julia 1.9 CI job, which is the first run of this code there.
• The recorded benchmark tables were not re-run with the new column headers.

🤖 Generated with Claude Code