Workflow 4: Rebuttal

Capability Routing

This is a first-layer entry skill. Keep it loaded as the user-facing route; when a request needs a specialized capability below, resolve the debuffer repo root from .debuffer_skills/installed-skills-codex.txt (repo_root) when available, read the referenced library SKILL.md, then follow that skill. Do not copy the whole library skill into this file.

• /auto-paper-improvement-loop: read ../library/review/auto-paper-improvement-loop/SKILL.md.
• /kill-argument: read ../library/review/kill-argument/SKILL.md.
• /paper-claim-audit: read ../library/review/paper-claim-audit/SKILL.md.
• /citation-audit: read ../library/review/citation-audit/SKILL.md.

Prepare and maintain a grounded, venue-compliant rebuttal for: $ARGUMENTS

Scope

This skill is optimized for:

• text-only rebuttal under strict character/word limits (e.g. ICML single-document)
• per-reviewer thread responses where each reviewer renders independently (e.g. OpenReview-style)
• multiple reviewers with shared and reviewer-specific concerns
• follow-up rounds after the initial rebuttal
• safe drafting with no fabrication, no overpromise, and full issue coverage

This skill does not:

• run new experiments automatically
• generate new theorem claims automatically
• edit or upload a revised PDF
• submit to OpenReview / CMT / HotCRP

If the user already has new results, derivations, or approved commitments, the skill can incorporate them as user-confirmed evidence.

Lifecycle Position

Workflow 1:   idea-discovery
Workflow 1.5: experiment-bridge
Workflow 2:   auto-review-loop (pre-submission)
Workflow 3:   paper-writing
Workflow 4:   rebuttal (post-submission external reviews)

Constants

• VENUE = ICML — Default venue. Override if needed.
• RESPONSE_MODE = TEXT_ONLY — v1 default.
• REVIEWER_MODEL = resolve from the shared model policy or an explicit per-run override; never hard-code a workflow default.
• REVIEWER_BACKEND = codex — Default: Codex MCP (xhigh). Override with — reviewer: oracle-pro for Oracle MCP, or — reviewer: manual for Manual Review MCP. If manual-review MCP is unavailable, stop and print the install command; do not fall back to Codex. See shared-references/reviewer-routing.md.
• MAX_INTERNAL_DRAFT_ROUNDS = 2 — draft → lint → revise.
• VENUE_MODE = single_document — single_document for one shared author response, or per_reviewer_thread when each reviewer thread renders independently. Confirm the venue/interface before drafting if unclear. Affects Phase 4/7 output shape.
• STRESS_TEST_ROUNDS_BASE = 1 — One external reviewer critique round on the full response set. Add focused rounds for reviewer_priority: pivotal responses, terminating when the reviewer returns no new substantive issues. Hard cap at 5.
• MAX_FOLLOWUP_ROUNDS = 3 — per reviewer thread.
• AUTO_EXPERIMENT = false — When true, automatically invoke /experiment-bridge to run supplementary experiments when the strategy plan identifies reviewer concerns that require new empirical evidence. When false (default), pause and present the evidence gap to the user for manual handling.
• QUICK_MODE = false — When true, only run Phase 0-3 (parse reviews, atomize concerns, build strategy). Outputs ISSUE_BOARD.md + STRATEGY_PLAN.md and stops — no drafting, no stress test. Useful for quickly understanding what reviewers want before deciding how to respond.
• REBUTTAL_DIR = rebuttal/
• RENDER_HTML = true — When true (default), auto-render rebuttal/REBUTTAL_DRAFT_rich.md (the detailed reviewer-facing draft) to HTML after Phase 6 / Phase 8 finalization. Uses full Codex review gate (final pre-submission deliverable — reviewer-facing content, render fidelity matters). The plain-text PASTE_READY.txt is NOT rendered (it's character-counted plain text by design). Set false to skip, or pass — render html: false.

Override: /rebuttal "paper/" — venue: NeurIPS, character limit: 5000

Reviewer Calling Convention

When calling the reviewer for stress-testing, branch on REVIEWER_BACKEND:

If REVIEWER_BACKEND = codex: Use mcp__codex__codex for new review threads. Use mcp__codex__codex-reply for follow-up rounds (reuse threadId).

If REVIEWER_BACKEND = manual: Use mcp__manual_review__review for new review threads with: prompt: [exact same prompt that would go to Codex] config: {"model_reasoning_effort": "xhigh"} Save the returned threadId. Use mcp__manual_review__review_reply for follow-up rounds with: threadId: [saved manual-review threadId] prompt: [follow-up prompt] config: {"model_reasoning_effort": "xhigh"}

Prompt fidelity: the manual prompt must be exactly the same text that Codex would receive. Review tracing applies equally to both backends.

Required Inputs

1. Paper source — PDF, LaTeX directory, or narrative summary
2. Raw reviews — pasted text, markdown, or PDF with reviewer IDs
3. Venue rules — venue name, character/word limit, text-only or revised PDF allowed, rendering mode (one shared response or independent reviewer threads)
4. Current stage — initial rebuttal or follow-up round

If venue rules, limit, or rendering mode are missing, stop and ask before drafting.

Safety Model

Three hard gates — if any fails, do NOT finalize:

1. Provenance gate — every factual statement maps to: paper, review, user_confirmed_result, user_confirmed_derivation, or future_work. No source = blocked.
2. Commitment gate — every promise maps to: already_done, approved_for_rebuttal, or future_work_only. Not approved = blocked.
3. Coverage gate — every reviewer concern ends in: answered, deferred_intentionally, or needs_user_input. No issue disappears.

Workflow

Phase 0: Resume or Initialize

1. If rebuttal/REBUTTAL_STATE.md exists → resume from recorded phase
2. Otherwise → create rebuttal/, initialize all output documents
3. Load paper, reviews, venue rules, any user-confirmed evidence

Phase 1: Validate Inputs and Normalize Reviews

1. Validate venue rules are explicit
2. Normalize all reviewer text into rebuttal/REVIEWS_RAW.md (verbatim)
3. Record metadata in rebuttal/REBUTTAL_STATE.md
4. If ambiguous, pause and ask

Phase 2: Atomize and Classify Reviewer Concerns

Create rebuttal/ISSUE_BOARD.md.

For each atomic concern:

• issue_id (e.g., R1-C2)
• reviewer, round, raw_anchor (short quote)
• issue_type: assumptions / theorem_rigor / novelty / empirical_support / baseline_comparison / complexity / practical_significance / clarity / reproducibility / other
• severity: critical / major / minor
• reviewer_stance: positive / swing / negative / unknown
• reviewer_priority: standard / pivotal
  • pivotal — a reviewer whose response is likely to affect the decision if addressed well: low or borderline rating, addressable concerns, and enough confidence/influence to matter. Phase 3 allocates extra drafting and stress-test budget here.
• response_mode: direct_clarification / grounded_evidence / nearest_work_delta / assumption_hierarchy / narrow_concession / future_work_boundary / structural_distinction
  • structural_distinction — for "your method reduces to X / is just generic Y / is subsumed by Z" attacks. Pattern: agree on the local reduction; show the structural feature your parameterization preserves that X/Y/Z does not capture, backed by a concrete mechanism (theorem dependency, derivation step, or empirical consequence). Never use rhetorically without the supporting mechanism.
• status: open / answered / deferred / needs_user_input

Phase 3: Build Strategy Plan

Create rebuttal/STRATEGY_PLAN.md.

1. Identify 2-4 global themes resolving shared concerns
2. Choose response mode per issue
3. Build character budget (10-15% opener, 75-80% per-reviewer, 5-10% closing) — applies in single_document mode; in per_reviewer_thread mode, set per-thread word/char targets instead
4. Identify pivotal reviewer(s) — reviewers whose vote or confidence shift would most affect the decision, especially when concerns are addressable rather than ideological. Mark them reviewer_priority: pivotal in ISSUE_BOARD.md. There may be more than one. Allocate disproportionate drafting + stress-test budget here.
5. Identify blocked claims (ungrounded or unapproved)
6. If unresolved blockers → pause and present to user

QUICK_MODE exit: If QUICK_MODE = true, stop here. Present ISSUE_BOARD.md + STRATEGY_PLAN.md to the user and summarize: how many issues per reviewer, shared vs unique concerns, recommended priorities, and evidence gaps. The user can then decide to continue with full rebuttal (/rebuttal — quick mode: false) or write manually.

Phase 3.5: Evidence Sprint (when AUTO_EXPERIMENT = true)

Skip entirely if AUTO_EXPERIMENT is false — instead, pause and present the evidence gaps to the user.

If the strategy plan identifies issues that require new empirical evidence (tagged response_mode: grounded_evidence with evidence_source: needs_experiment):

1. Generate a mini experiment plan from the reviewer concerns:
   • What to run (ablation, baseline comparison, scale-up, condition check)
   • Success criterion (what result would satisfy the reviewer)
   • Estimated GPU-hours
2. Invoke /experiment-bridge with the mini plan:

   /experiment-bridge "rebuttal/REBUTTAL_EXPERIMENT_PLAN.md"

3. Wait for results, then update ISSUE_BOARD.md:
   • Tag completed experiments as user_confirmed_result
   • Update evidence source for relevant issue cards
4. If experiments fail or are inconclusive:
   • Switch response mode to narrow_concession or future_work_boundary
   • Do NOT fabricate positive results
5. Save experiment results to rebuttal/REBUTTAL_EXPERIMENTS.md for provenance tracking.

Time guard: If estimated GPU-hours exceed rebuttal deadline, skip and flag for manual handling.

Phase 4: Draft Initial Rebuttal

Create the draft artifact(s) per VENUE_MODE:

• single_document mode → one rebuttal/REBUTTAL_DRAFT_v1.md
• per_reviewer_thread mode → one rebuttal/Reviewer_<ID>_response.md per reviewer (no top-level REBUTTAL_DRAFT_v1.md)

Structure depends on VENUE_MODE:

• single_document — one REBUTTAL_DRAFT_v1.md:
  1. Short opener — thank reviewers + 2-4 global resolutions
  2. Per-reviewer numbered responses — answer → evidence → implication
  3. Short closing — resolved / remaining / acceptance case
• per_reviewer_thread — one self-contained Reviewer_<ID>_response.md per reviewer:
  1. Brief acknowledgment of that reviewer's main thrust
  2. Numbered W#/Q# responses (answer → evidence → implication)
  3. Optional shared experimental-setup paragraph (see "Reusable setup block" below)
  • Each file must be readable standalone. No "see Reviewer X's response" references. No global opener.

Default reply pattern per issue:

• Sentence 1: direct answer
• Sentence 2-4: grounded evidence
• Last sentence: implication for the paper

Reusable setup block (per_reviewer_thread mode). If multiple reviewer-thread responses need the same experimental setup or metric definitions, write a canonical SETUP_METRICS_BLOCK.md. Reuse it consistently in each reviewer file that needs it. Target ≤ 150 words; expand only with genuinely reviewer-specific additions inline. Change-once-update-everywhere prevents drift across threads.

Heuristics from successful rebuttals (content):

• Evidence > assertion
• Global narrative first, per-reviewer detail second (single_document mode only)
• Concrete numbers for counter-intuitive points
• Name closest prior work + exact delta for novelty disputes
• Concede narrowly when reviewer is right
• For theory: separate core vs technical assumptions
• Answer friendly reviewers too

Reviewer-defensive moves:

• Minimum sufficient evidence per concern. Usually one numerical anchor: the metric that maps directly to that reviewer's specific ask. Cut metrics other reviewers care about — bloat dilutes the answer.
• Pre-registered calibration phrasing. When a threshold or hold-out was fixed before generated samples were inspected, say so explicitly with a phrase like "set on hold-out before any generated sample was inspected." Defuses cherry-pick attacks at near-zero word cost. Only use when actually true.
• Surface non-obvious design choices upfront. If the experimental setup has a non-obvious caveat (compute-matched ≠ epoch-matched, atypical seed protocol, restricted parameter subset, etc.), name it concretely with numbers where they clarify the design choice. Pre-empts adversarial reverse engineering.
• Structural distinction over denial. When a reviewer claims your work reduces to / is subsumed by a generic framework, do not deny the reduction. Identify the structural feature your parameterization preserves that the generic framework does not — see response_mode: structural_distinction.
• Concede without surrendering the claim. When the reviewer is partly right, explicitly accept the local point, then state what remains true and why it still supports the paper's contribution. Pair the concession with the preserved theorem, mechanism, empirical result, or scope condition.

Hard rules:

• NEVER invent experiments, numbers, derivations, citations, or links
• NEVER promise what user hasn't approved
• If no strong evidence exists, say less not more

single_document mode only: also generate rebuttal/PASTE_READY.txt (plain text, exact character count for the OpenReview/CMT paste field). In per_reviewer_thread mode skip this artifact — each Reviewer_<ID>_response.md is itself the paste target for its thread.

Also generate rebuttal/REVISION_PLAN.md — the overall revision checklist.

This document is the single source of truth for every paper revision promised (explicitly or implicitly) in the rebuttal draft. It exists so the author can track follow-through after the rebuttal is submitted, and so the commitment gate in Phase 5 has a concrete artifact to validate against.

Structure:

1. Header
   • Paper title, venue, character limit, rebuttal round
   • Links back to ISSUE_BOARD.md, STRATEGY_PLAN.md, REBUTTAL_DRAFT_v1.md

2. Overall checklist — a single flat GitHub-style checklist covering every revision item, so the author can tick items off as they land in the camera-ready / revised PDF:

   ## Overall Checklist
   
   - [ ] (R1-C2) Add assumption hierarchy table to Section 3.1 — commitment: `approved_for_rebuttal` — owner: author — status: pending
   - [ ] (R2-C1) Clarify novelty delta vs. Smith'24 in Section 2 related work — commitment: `already_done` — status: verify wording
   - [ ] (R3-C4) Add runtime breakdown figure to Appendix B — commitment: `future_work_only` — status: deferred, note in camera-ready
   - ...

   Checklist items must be atomic (one paper edit per line) and each must reference its issue_id so it maps back to ISSUE_BOARD.md.

3. Grouped view — the same items regrouped by (a) paper section/location and (b) severity, so the author can plan the revision pass efficiently.
4. Commitment summary — counts of already_done / approved_for_rebuttal / future_work_only, plus any needs_user_input items that are blocking.
5. Out-of-scope log — reviewer concerns that will not trigger a paper revision (e.g. deferred_intentionally, narrow_concession with no edit), with a one-line reason each. This keeps the checklist honest: nothing silently disappears.

Rules for REVISION_PLAN.md:

• Every checklist item must map to at least one issue_id from ISSUE_BOARD.md.
• Every promise in REBUTTAL_DRAFT_v1.md that implies a paper edit must appear as a checklist item — if it is not in the plan, it is a commitment-gate violation.
• Never add items that are not backed by the draft or by user-confirmed evidence.
• On rerun / follow-up rounds, update checkbox state in place rather than regenerating from scratch.

Phase 5: Safety Validation

Run all lints:

1. Coverage — every issue maps to draft anchor
2. Provenance — every factual sentence has source
3. Commitment — promises are approved AND every paper-edit promise in the draft appears as a checklist item in REVISION_PLAN.md (and vice versa — no orphan items in the plan)
4. Tone — flag aggressive/submissive/evasive phrases
5. Consistency — no contradictions across reviewer replies
6. Limit — exact character count, compress if over (redundancy → friendly → opener → wording, never drop critical answers)
7. Thread-local context (per_reviewer_thread mode only) — each reviewer file must be intelligible without reading any other reviewer file. Flag any "see Reviewer X" references or undefined terms that rely on cross-thread context.
8. Adversarial design-choice scan — for each experimental claim, ask: "Could a hostile reviewer find a non-obvious design choice (compute-match, frozen subset, sampling protocol) that I haven't disclosed?" If yes, add a one-line caveat in the Setup paragraph. Narrower than provenance; focused on design choices not factual sources.

Phase 6: External Reviewer Stress Test

Use the selected backend. For codex:

mcp__codex__codex:
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    Stress-test this rebuttal draft:
    [raw reviews + issue board + draft + venue rules]

    1. Unanswered or weakly answered concerns?
    2. Unsupported factual statements?
    3. Risky or unapproved promises?
    4. Tone problems?
    5. Paragraph most likely to backfire with meta-reviewer?
    6. Minimal grounded fixes only. Do NOT invent evidence.

    Verdict: safe to submit / needs revision

For manual: use mcp__manual_review__review with the same prompt and config: {"model_reasoning_effort": "xhigh"}.

Iterations. Run the base round on the full draft. Then run focused follow-up rounds on each reviewer_priority: pivotal response, terminating when the reviewer returns no new substantive issues. Hard cap at 5 rounds total. Save each round to rebuttal/MCP_STRESS_TEST_round<N>.md; the highest round number represents the final state. If any hard safety blocker remains → revise before finalizing.

Phase 7: Finalize

Outputs depend on VENUE_MODE:

single_document mode — produce two versions:

1. rebuttal/PASTE_READY.txt — strict version
   • Plain text, exact character count, fits venue limit
   • Ready to paste directly into the venue interface
   • No markdown formatting, no extras
2. rebuttal/REBUTTAL_DRAFT_rich.md — extended version
   • Same structure but with more detail: fuller explanations, additional evidence, optional paragraphs
   • Marked with [OPTIONAL — cut if over limit] for sections that exceed the strict version
   • Author can read this to understand the full reasoning, then manually decide what to keep/cut/rewrite
   • Useful for follow-up rounds — the extra material is pre-written

per_reviewer_thread mode — produce one file per reviewer:

1. rebuttal/Reviewer_<ID>_response.md — one self-contained file per reviewer, ready to paste into the corresponding reviewer thread
2. rebuttal/SETUP_METRICS_BLOCK.md — optional canonical setup/metrics text when reused across reviewer files
3. rebuttal/SUPPLEMENTARY_FIG_PDF/ (optional) — when the venue does not allow PDF revision but allows anonymous figure links, generate a venue-compliant supplementary PDF. Do not hard-code an anonymous-hosting platform or typesetting style; choose what the target venue accepts.

Both modes:

4. Update rebuttal/REBUTTAL_STATE.md
5. Refresh rebuttal/REVISION_PLAN.md so the overall checklist matches the final draft (add items, mark already_done as checked, carry forward any pending items)
6. Present to user:
   • For single_document: PASTE_READY.txt character count vs venue limit; REBUTTAL_DRAFT_rich.md for review
   • For per_reviewer_thread: list of per-reviewer files with word counts; SETUP_METRICS_BLOCK.md if used; supplementary PDF if generated
   • REVISION_PLAN.md checklist — counts of pending / approved / deferred
   • Remaining risks + lines needing manual approval

Phase 8: Follow-Up Rounds

When new reviewer comments arrive:

1. Append verbatim to rebuttal/FOLLOWUP_LOG.md
2. Link to existing issues or create new ones
3. Draft delta reply only (not full rewrite)
4. Update rebuttal/REVISION_PLAN.md in place — add any new checklist items introduced by the follow-up, tick off items the author has already completed, and keep existing items' status current
5. Re-run safety lints
6. Use the appropriate reply tool for continuity if useful (per Reviewer Calling Convention)
7. Rules: escalate technically not rhetorically; concede if reviewer is correct; stop arguing if reviewer is immovable and no new evidence exists

Phase 9: Render HTML view (auto, when RENDER_HTML = true, default)

After Phase 6 (initial rebuttal) or Phase 8 (follow-up rounds) finalize rebuttal/REBUTTAL_DRAFT_rich.md, invoke /render-html on the detailed draft:

/render-html "rebuttal/REBUTTAL_DRAFT_rich.md"

Uses full Codex review gate (reviewer-facing pre-submission deliverable — render fidelity matters; the rich draft contains author-controlled formatting that the reviewer should see exactly as intended). Output: rebuttal/REBUTTAL_DRAFT_rich.html with embedded source SHA256 and .review.json sidecar.

Do NOT render rebuttal/PASTE_READY.txt — it's exact-character-count plain text by design, not a structural artifact.

Non-blocking: if /render-html fails (helper missing, Codex MCP unavailable, file write error), log the failure and treat the rebuttal phase as complete — the PASTE_READY.txt and REBUTTAL_DRAFT_rich.md are the canonical outputs.

Skip if RENDER_HTML = false.

Key Rules

• Large file handling: If Write fails, retry with Bash heredoc silently.
• Never fabricate. No invented evidence, numbers, derivations, citations, or links.
• Never overpromise. Only promise what user explicitly approved.
• Full coverage. Every reviewer concern tracked and accounted for.
• Preserve raw records. Reviews and MCP outputs stored verbatim.
• Structure follows VENUE_MODE. single_document → global opener with shared concerns + per-reviewer numbered detail. per_reviewer_thread → no global opener; each reviewer file is self-contained.
• Answer friendly reviewers too. Reinforce supportive framing.
• Meta-reviewer closing. Summarize resolved/remaining/why accept.
• Evidence > rhetoric. Derivations and numbers over prose.
• Concede selectively. Narrow honest concessions > broad denials.
• Don't waste space on unwinnable arguments. Answer once, move on.
• Respect the limit. Character budget is a hard constraint.
• Resume cleanly. Continue from REBUTTAL_STATE.md on rerun.
• Anti-hallucination citations. Any reference added must go through DBLP → CrossRef → [VERIFY].

Review Tracing

After each reviewer call (mcp__codex__codex, mcp__codex__codex-reply, mcp__manual_review__review, or mcp__manual_review__review_reply), save the trace following shared-references/review-tracing.md (Policy C — forensic; never silently skip). Use save_trace.sh (resolved per the chain in shared-references/integration-contract.md §2) or write files directly to .debuffer_skills/traces/<skill>/<date>_run<NN>/. Respect the --- trace: parameter (default: full).