TS Paper Plan

ts-paper-plan — proposal → blueprint.json

Produce the entire blueprint in ONE reasoning pass (the original split this into 3 calls; you don't need to). Write it to blueprint.json in the working dir. This is a proposal — design the experiments but commit to placeholder results only; never plan a fabricated number.

Step 0 — load the template (the suite is template-agnostic)

Before planning, pick the template (default ts_iieta; or the user's template=<name>). Templates live in ../ts-paper/templates/<name>/. Validate and copy it into the workdir:

python scripts/template_lint.py ../ts-paper/templates/<name>/      # must be ok:true
cp ../ts-paper/templates/<name>/*  <workdir>/                       # template.json + .sty/.cls + main.tex.tmpl + assets

Set results_mode now if the router classified this run as data_aware. The cp just overwrote <workdir>/template.json with the bundled template, which ships "results_mode": "proposal" — so any mode the router set earlier is gone. If the router classified this as data_aware (class c, or class d with real results), edit <workdir>/template.json to set "results_mode": "data_aware" (the cp reset it to the bundled proposal default). Leave it as proposal otherwise. template.json is the single source of truth every downstream reader keys on; record the chosen mode in logs/1_plan.io.md.

Then read <workdir>/template.json and let it drive the blueprint: the section list + ids + titles, the per-section word bands (sections[].words), the contributions count (contributions.count — may be 0 = not enforced, e.g. NeurIPS), the result-table set (experiments.recipe.result_tables), the title/keyword limits, and the citation types/floor all come from the spec. The TS values below are the default template's values; for another template, use that template's sections/recipes instead. Everything downstream (draft_lint, assemble, citations_lint) reads the same template.json.

blueprint.json schema

{
  "paper_title": "8–14 word technical title, no filler, ≤1 colon",
  "keywords": "4–6 lowercase comma-separated index terms (proper nouns/acronyms keep case)",
  "contributions": ["C1: ...", "C2: ...", "C3: ..."],
  "terminology": {"method_name":"X","component_1":"...","dataset_1":"...","baseline_1":"..."},
  "notation": {"x_s":"input skeleton sequence","T":"number of frames"},
  "experiment_design": {
    "datasets":[{"name":"","description":""}],
    "baselines":[{"name":"","description":""}],
    "metrics":[{"name":"","description":""}],
    "main_table":{"columns":["Method","..."],"rows":["Baseline A","...","<Method> (Ours)"]},
    "ablation_table":{"columns":["Variant","..."],"rows":["Full","w/o <component>"]}
  },
  "section_order": ["introduction","related_work","method","experiments","analysis","conclusion"],
  "sections": {
    "<id>": {"title":"","target_words":[min,max],"key_arguments":[".."],
             "citation_types":["CORE","CONTEXT"],
             "paragraph_outline":[{"para":1,"topic":"","key_points":[".."],"citations":"2-3","sentences":4}],
             "figures":[{"id":"system_overview","type":"architecture","caption":"","label":"fig:.."}],
             "tables":[]}
  }
}

Canonical sections, default word targets (min,max), dependencies — ts_iieta default

(For another template, use its sections/words from template.json instead of this table.)

(These ts_iieta bands target a substantial ~10–12 page two-column journal article. Copy them into blueprint.json target_words as-is unless the template differs; the band is reached by real substance, never padding.)

(Optional limitations (200,400) only if the user wants it; TS samples usually omit it.) You may shift a section's target_words to match how much source content it has, but keep [min,max] pairs.

Rules

• Title: technical, specific, 8–14 words; avoid filler ("comprehensive/framework/study/toward/universal"); key terms should overlap the method/problem terms.
• Keywords: 4–6, lowercase, comma-separated; specific topical terms (task, method, signal/data type, key technique); no connectors ("and/of/based").
• Contributions: exactly template.contributions.count (TS=3; C1:/C2:/C3:), each a concrete technical statement from the proposal's innovation claims; the last may describe the planned evaluation (never a result number). If the count is 0 (e.g. NeurIPS), contributions are optional/free-form.
• Terminology glossary: fix the exact canonical names (method, components, datasets, baselines) that EVERY section must reuse — prevents drift.
• Tables live ONLY in experiments, exactly the set in template.experiments.recipe.result_tables (TS: main_results, secondary_results, ablation_results; NeurIPS: main_results, ablation_results). Every other section has "tables":[]. Quantitative results belong in tables, NEVER in figures.
• Figures: plan at least template.figures.min (TS = 5), distributed across method / experiments / analysis (the architect's rule). These are schematic / conceptual / qualitative diagrams that depict the method or an idea, each standing on its own merit (not padding): e.g. (1) the architecture/pipeline overview (method); (2) a component-detail diagram (e.g. how the trajectory descriptors are computed) (method); (3) a concept diagram (e.g. formation-graph construction + its Laplacian-spectrum signal) (method/analysis); (4) a qualitative scenario (e.g. a crossing event where appearance fails but the formation prior disambiguates — illustrative, NO numbers) (experiments/analysis); (5) a protocol/diagnostic concept (e.g. how the stratification by occlusion/density is defined) (experiments). Each gets {id, type, caption, label} with type in the drawable set {architecture, pipeline, framework, concept, schematic, overview, qualitative, diagram, flow}.
  • Results figures depend on template.results_mode: in proposal mode (no real data) do NOT plan any results figure — there is nothing to plot; numbers live only in the (blank --) tables. Do not invent a stand-in figure for the missing results. In data_aware mode (the proposal ships real, validated results) DO plan results figures (metric curves/bars/comparison plots) drawn from that data, in addition to the schematic ones.
• citation_types: only CORE, CONTEXT, BASELINE, METRIC, DEFINITION; don't request more CORE cites than real papers will exist. Abstract & conclusion carry no citations.
• paragraph_outline: 3–6 paragraphs/section (2–4 for abstract/conclusion), each specific to THIS paper, never a generic template.
• Writing order is Method-First (method → related_work → introduction → experiments → analysis → conclusion → abstract); section_order above is the reading order for the PDF.
• Honesty: results are blank placeholders; contributions/abstract use proposal language ("we evaluate", "expected to improve"), never achieved numbers.

DATA-AWARE branch (only when template.results_mode == "data_aware")

When the router set results_mode == "data_aware" (real results present — see ts-paper-data, which read the data), plan against the real data, not a generic shape:

• Plan the result tables with the REAL keys Claude read from the data: each table's rows = the actual method/variant names, columns = the actual metric names (not invented placeholders). Plan the table ids (main_results, secondary_results, ablation_results) and honor the tables.min floor; the write stage fills the cells itself with the real numbers (no markers, no auto-filler).
• Force experiment_complexity="full" (real data warrants the full experimental treatment).
• Enable the analysis section and plan results figures (metric bar/curve/comparison plots) drawn from the data, in addition to the schematic figures — these are rendered later via matplotlib (ts-paper-data / plot_results.py).
• Everything else (title, keywords, contributions, notation, terminology, per-section outlines) is planned exactly as in proposal mode. blueprint_lint accepts the exact-key data-aware table plan and only enforces tables.min (it does not force the proposal table count).

Validate (enforced) + log

Run python scripts/blueprint_lint.py <workdir> --fix, then re-run without --fix until ok=true. It is now template-driven (reads template.json from the workdir): citation_types must be a subset of the template's citations.types (remaps aliases — DATASET→METRIC, METHOD→CORE, …), all of the template's sections present (self-heal with the template's titles/word-targets), every target_words a 2-tuple, the template's result-table count in experiments (mode-dependent: in proposal mode the experiments section plans either 0 or exactly template.experiments.recipe.result_tables result tables — 0 is allowed because a proposal has no measured results to tabulate; in data_aware mode it must plan at least tables.min filled tables), title within the template's title limits, contributions == the template's contributions.count (skipped when 0). Do not proceed on a failing blueprint.

Then write logs/1_plan.io.md — three blocks: INPUT (the proposal text/path), DECISIONS (why this title / these word targets / these keywords), OUTPUT (blueprint.json one-line summary). Hand off to ts-paper-cite.