| name | compare-literature |
| description | Compare task results against published literature and write the comparison output. |
Compare Literature
Version: 4
Goal
Compare task results against published results from the literature and write
results/compare_literature.md following the specification.
Inputs
$TASK_ID — the task folder name (e.g., t0019_mfs_baseline_raganato)
Context
Read before starting:
-
arf/specifications/compare_literature_specification.md — authoritative format for
results/compare_literature.md
-
tasks/$TASK_ID/task.json — task objective, description, expected assets
-
tasks/$TASK_ID/results/metrics.json — quantitative results; this may be a legacy flat metrics
object or an explicit multi-variant file
-
tasks/$TASK_ID/results/results_detailed.md — methodology and detailed numbers
-
tasks/$TASK_ID/results/results_summary.md — headline metrics
Steps
Phase 1: Gather Task Context
-
Read tasks/$TASK_ID/task.json to understand the task objective and scope. If it contains
long_description_file, also read the referenced markdown file.
-
Read tasks/$TASK_ID/results/metrics.json for quantitative results.
- If the file uses explicit variants, list all variants and their dimensions first.
- Do not assume the task has one canonical structured metrics block. Choose the relevant variant
or variants for each literature comparison row based on the compared conditions.
-
Read tasks/$TASK_ID/results/results_detailed.md for methodology and detailed numbers.
-
Read tasks/$TASK_ID/results/results_summary.md for headline metrics.
-
Read all available research files (skip any that do not exist):
tasks/$TASK_ID/research/research_papers.md
tasks/$TASK_ID/research/research_internet.md
tasks/$TASK_ID/research/research_code.md
-
Scan step logs in tasks/$TASK_ID/logs/steps/ for implementation details that affect
comparability (e.g., preprocessing choices, model configuration).
Phase 2: Identify Published Results
-
From the research files, extract published performance numbers relevant to this task's metrics.
-
Identify the specific metrics reported (F1, accuracy, etc.) and the evaluation conditions (test
set, preprocessing, model size).
-
Note which papers report results on the same datasets and metrics as this task.
-
List at least 2 comparable published results. If fewer than 2 exist in the research files, state
this explicitly in the Limitations section.
Phase 3: Build Comparison Table
-
For each published result, record: Method / Paper, Metric, Published Value, Our Value, Delta,
Notes.
- If your task has multiple variants, include the variant label or dimensions in the method name
or Notes column so the comparison is unambiguous.
-
Compute deltas as Our Value minus Published Value (positive means this task outperforms the
published result).
-
Flag any non-comparable conditions in the Notes column (e.g., "different test split", "few-shot
vs zero-shot", "different training data").
-
Include at least 2 data rows in the comparison table.
Phase 4: Analyze and Write
-
Write tasks/$TASK_ID/results/compare_literature.md with all mandatory sections per the
specification:
-
## Summary — 2-4 sentences: what was compared, headline finding
-
## Comparison Table — the comparison table from Phase 3
-
## Methodology Differences — bullet list of key differences between this task and the
published methods
-
## Analysis — interpret the comparison honestly: explain gaps, note agreements, highlight
surprising results
-
## Limitations — note missing references, different conditions, or non-comparable metrics
-
Include YAML frontmatter with spec_version, task_id, and date_compared (today's date in
ISO 8601 format).
-
Use citation keys (e.g., Raganato2017) when referencing published results. Verify each
citation key exists in the research files. Every published metric value in the comparison table
must cite the specific table, figure, or page from the source paper. Use format:
[AuthorYear, Table N] or [AuthorYear, p. N]. If a value cannot be traced to a specific table
in a specific paper, omit it rather than estimating.
-
Use **bold** for specific quantitative values.
-
When the plan cites specific results from prior project tasks as motivation or baselines,
include a ### Prior Task Comparison subsection. Compare the current task's results against
those cited values. If results contradict prior-task conclusions, highlight this as a finding —
it is just as important as the literature comparison.
Phase 5: Verify
-
Run the verificator:
uv run python -u -m arf.scripts.verificators.verify_compare_literature \
$TASK_ID
-
Fix all errors. Re-run until zero errors.
Done When
tasks/$TASK_ID/results/compare_literature.md exists and follows the specification
- Verificator passes with zero errors
- At least 2 published results compared with specific numbers
- Methodology differences documented for each comparison
- Analysis is honest — negative results (where this task underperforms) are not omitted
Forbidden
-
NEVER run prestep or poststep — the orchestrator handles the step lifecycle
-
NEVER commit — the orchestrator handles all commits
-
NEVER modify step_tracker.json — the orchestrator manages step state
-
NEVER write step_log.md — the orchestrator writes it after this skill completes
-
NEVER fabricate published numbers. Every published value must come from a paper in the research
files. If a number cannot be verified, omit it and note this in Limitations.
-
NEVER omit negative results. If this task underperforms a published result, report the delta
honestly and explain possible causes.
-
NEVER compare against published results without noting condition differences in the Notes column
or Methodology Differences section.
-
NEVER use vague qualifiers ("similar", "comparable", "competitive") in place of specific numbers.
-
NEVER skip the verificator step.
-
NEVER cite a published metric without specifying the exact table, figure, or page number.