Recall Failure Triage — turn recall@5 = X into a ranked worklist

A single recall number tells you nothing about what to fix. Five distinct failure modes produce identical recall drops, and the fix is completely different per mode. This skill is the bridge from "recall is low" to "open parsers/cell_parser.py line N, here is the bug".

Pair with excel-extraction-pipeline-improver — that skill implements fixes; this one decides what to fix.

When this skill fires

• User says "recall is low / dropped / not improving"
• New benchmark run produced a failures.ndjson or summary.json
• PR-time benchmark sample regressed on the CI step summary
• You're picking the next parser/chunking PR and need to choose what's highest leverage

Inputs you should expect

The decision procedure

0. Apply the in-scope filter — most "misses" are unfixable

Before counting buckets, run scripts/enrich_failures.py to mark instances flagged instruction_requires_execution — those where the benchmark's answer_position is empty in input.xlsx because the question literally asks the system to write the answer. ~63% of the SpreadsheetBench corpus is this class on the 200-sample. They are NOT parser bugs and CANNOT be fixed by parser work. Exclude them.

python scripts/enrich_failures.py tests/benchmarks/reports/retrieval
jq -c 'select((.flags | contains(["instruction_requires_execution"])) | not)' \
    tests/benchmarks/reports/retrieval/*/enriched_failures.ndjson

1. Read the bucket histogram FIRST

python scripts/triage_recall.py tests/benchmarks/reports/retrieval

The output ranks buckets by count. Spend your time on the largest bucket only. The next-largest is the next PR.

2. Read 5 example failures from the top bucket

python scripts/triage_recall.py tests/benchmarks/reports/retrieval \
  --bucket <largest> --examples 5

For each example, the script prints:

• the natural-language question
• answer_position (the ground-truth cell)
• the ground-truth string values that should appear in a chunk
• the top-8 ranked chunks with sheet, range, and whether each contains the answer

Pattern-match. What do the 5 examples have in common? E.g.

• All answer cells are formulas → H2 confirmed: render the cached value
• All chunks span the entire sheet → H1 confirmed: cap chunk size
• All answer sheets are hidden → wrong_sheet from hidden-sheet skip

3. State the hypothesis before changing code

In your reply, write the hypothesis as one sentence:

"Bucket X dominates because [observable common pattern in examples]. Expected fix: [specific file + change]. Expected lift on recall@5: [number]."

The expected lift = (bucket count) ÷ (scoreable instances). If a fix is supposed to halve the biggest bucket but recall barely moves on the next run, the hypothesis was wrong — go back to step 1, don't pile on.

4. Wire the fix into the existing pipeline

Pass the hypothesis to the excel-extraction-pipeline-improver skill (or implement directly) with TDD: write a failing crossval/invariant test that captures the bug from one example, fix, run make test, then re-run make bench-track to confirm the bucket shrank.

5. Verify on the history

tail -5 tests/benchmarks/reports/history.jsonl

The script prints a row-over-row delta. Expected behaviour:

• The targeted bucket count drops
• recall_text@5 rises by approximately target_bucket_drop / scoreable_instances
• recall_text@1 improves proportionally OR more (better chunks rank higher too)
• Other buckets shouldn't grow — if they do, the fix had a side effect

Anti-patterns

• Don't chase the headline number. "Recall went up 1 pp" without a bucket explanation is suspicious — it might be benchmark noise.
• Don't fix two buckets at once. You won't know which fix worked.
• Don't tune chunk size as the first move. It only helps present_but_ranked_low. If answer_absent_from_chunks is bigger, chunk tuning is wasted effort.
• Don't trust geometric and text recall independently. A docling-style parser with zero geometric overlap can still have high text recall (because it serializes everything to plain text). That's not the same thing as good — citation overlays need the geometric metric. Always report both.

Quick reference — running it from scratch

make corpus-download                                    # one-time, large
make bench-track                                        # runs + triages
python scripts/triage_recall.py tests/benchmarks/reports/retrieval --examples 5
cat tests/benchmarks/reports/history.jsonl | tail -10   # commit-over-commit

CI runs a 60-instance sample on every PR (.github/workflows/benchmark.yml) and posts the bucket histogram to the GitHub job summary. Weekly schedule runs the full 912-instance corpus.

Hand-off contract

When you finish a triage round, your final message MUST include:

1. The bucket histogram (counts + percentages).
2. The dominant bucket + one-sentence root cause hypothesis.
3. The 1–2 file paths where the fix lives.
4. The expected lift on recall_text@5 if the fix lands.
5. A pointer to the example instance IDs you read.

Anything less and the next agent will have to re-do the analysis.