TS Review

ts-review

Step 1 — deterministic checks (always first, every section)

For each section file under {PAPER_WORKSPACE}/sections/, run (both {SKILLS_DIR} and {PAPER_WORKSPACE} are absolute paths given in your system prompt — pass {PAPER_WORKSPACE} exactly as shown, do not cd there instead):

python {SKILLS_DIR}/ts-review/scripts/lint_check.py <section_id> {PAPER_WORKSPACE}

It prints JSON with a violations list — empty means the section is clean. Violations cover: word count out of band, a number that doesn't match any real result (a hyperparameter/design constant like "a learning rate of 0.001" is exempt), a banned AI-writing phrase, a citation where none is allowed, a banned LaTeX command in prose (\textbf/\textit/ \documentclass/\usepackage/\begin{document}), a numbered heading, a markdown code fence, or a missing structural element for that section (the introduction's exactly-3 contribution \items, related work's exactly-3 \subsection themes, method's \subsection{Notation}, experiments' exactly-3 \subsections, discussion's required \subsection).

If violations are returned, rewrite that section to fix every one of them — trim or expand prose for length, remove or correct any fabricated number/citation, rephrase banned phrases — without changing what it claims otherwise. Save the fixed text back to the same file and re-run the check. Stop after 3 attempts on one section and move on, reporting it at the end, rather than looping forever.

Step 2 — adversarial review (only after every section passes Step 1)

Adapted from spark-to-paper-skills' ts-paper-review recipe. Its real implementation runs three fully isolated reviewers and prefers doing that via genuinely separate contexts ("Tier 2: subagents") whenever a task-delegation tool is available — it explicitly says this tier has "the SAME quality" as true parallel isolation, because each spawned agent cannot see its peers or any prior round. You have that tool: a paper-reviewer subagent is available via your task tool (see your system prompt's available-agents list). Use it — do not role-play the three angles yourself in one context; that is a strictly weaker approximation of the same algorithm.

Concatenate the full draft (every {PAPER_WORKSPACE}/sections/*.tex file, in document order) into one text blob in your own context — never write this blob to a file; it exists only to paste into each task-tool call's prompt below, and nothing else needs it on disk. Then call your task tool three times, once per lens below, delegating to the paper-reviewer subagent. Each call's task prompt must contain: the lens name and its mandate (verbatim from the list below), and the full draft blob. Nothing else — no summary of what you already think is wrong, no mention of the other lenses. Each call is a separate, isolated context by construction; do not let what one reviewer returns leak into another call's prompt.

1. Theory lens — "Judge whether the Method as described is sound and complete on its own terms. Flag any logical gap, unsupported causal claim, or claim that overstates what the rest of the draft actually establishes."
2. Empirical lens — "Judge whether every number, comparison, and claimed result in the draft is consistent with the Results section/table elsewhere in the same draft. Flag any mismatch between what the table shows and what prose elsewhere claims about it, including a correctly-real number used to support the wrong conclusion."
3. Applied lens — "Judge whether a reader who only sees this draft would come away with an accurate, non-misleading understanding of what was done and found. Flag anything technically true but stated misleadingly, or a caveat stated in one section that a later section silently drops or contradicts."

Each call returns a JSON array of issues (quote, section, severity, summary, close_criterion — see the subagent's own output contract). Collect all three arrays.

Step 3 — merge and survival gate

Merge: across the three returned arrays, collapse issues that quote the same sentence or make the same point into one, keeping every distinct issue. Drop anything without a usable close_criterion — not actionable.

Survival gate: for each remaining issue, check it against the other two lenses' mandates: would at least one of the other two lenses, applied to that same quote, also treat it as a real problem? If yes, it survives. If only the lens that raised it would defend it, discard it.

Step 4 — fix and repeat

Fix every surviving issue by editing the specific section it's in — change only what's needed to satisfy its close_criterion (no new numbers, no new citations, no claim beyond what the evidence already supports). Save the file, then repeat Steps 2-3 on the updated full draft (three fresh task-tool calls — do not reuse a prior round's reviewer output). Stop when one full pass finds zero surviving issues, or after 3 total review passes, whichever comes first — report any issue still open after 3 passes in your final summary rather than looping forever.

Step 5 — citation check (after the review passes)

python {SKILLS_DIR}/ts-review/scripts/check_citations.py {PAPER_WORKSPACE}

It prints hallucinated_keys — real citation keys used across all sections that don't actually exist in refs.bib. If any come back, find which section(s) use them and either replace with a real key from refs.bib or remove the citation entirely — a fabricated citation is a hard failure, never leave one in.