Strategic Revision

Strategic Revision — Feedback to DAG-Validated Master Plan

Convert feedback on the author's own paper into a computationally validated revision plan. The shared analytical core produces atomic tasks, a dependency DAG, execution blocks A--E, a critical path, and bottleneck analysis. Provenance determines which of two modes supplies and stores the surrounding artifacts:

1. External R&R mode preserves genuine venue correspondence and prepares response-oriented tracking.
2. Internal revision mode consumes AI reviews, review syntheses, or informal collaborator feedback without representing them as venue correspondence.

Provenance: DAG validation + critical-path architecture adapted from Jukka Sihvonen's strategic-revision skill (https://github.com/jusi-aalto/strategic-revision). the user's ingestion layer (correspondence scaffolding, LaTeX verbatim, R&R routing, venue strategy, coaching) retained.

When to Use

• Received reviewer or editor reports from a journal or conference
• Need to turn internal pre-submission reviews into an executable revision sequence
• Have a review-cluster or synthesise-reviews output that needs dependency mapping and critical-path validation
• Need to extend an existing revision DAG with another review of the same draft

When NOT to Use

• Writing the actual response letter (use generated response blocks as a starting point, then write manually)
• Reviewing someone else's paper (use proofread or peer-reviewer agent)
• Combining overlapping internal reports without execution planning — use synthesise-reviews first
• Applying fixes directly without first producing and approving a plan

Modes and Provenance Gate

Select the mode from who authored the source feedback, not from whether it sounds like a referee report. Full routing and cross-mode rules: references/modes.md.

An auto-generated "Reviewer 2" report is internal mode. A human venue report pasted by the user is external mode. If provenance is unclear, ask before creating files. Never copy AI review text into the venue-correspondence record.

External mode — reviewer-gap extraction (end of run). After the plan is built, run skills/_shared/reviewer-gap-extraction.md on the venue referee reports: capture any generalizable reviewing pattern the human referees raised that our own review agents don't already cover, as reviewer-gap signals for the feedback-review skill to propose as agent-contract refinements (Phase 2 of the review-agent calibration loop). Internal mode does NOT emit — those sources are our own agents.

Inputs

Gather the shared inputs first:

1. Feedback path(s) — one or more reports, or a synthesise-reviews output
2. Project path — auto-detect from the current research project when possible
3. Paper scope — e.g., paper-eacl; _project only for project-level feedback
4. Coordinating author — who owns revision decisions

External mode also requires venue slug, revision round (default 1), response deadline, and the original venue-supplied review/decision files. Internal mode records the draft identity and review dates; it does not invent a venue round or response deadline.

External PDF Auto-Discovery

Search for the reviews PDF in this order. Use the first match; if multiple PDFs found at a location, list them and ask the user to pick.

1. to-sort/*.pdf — most likely landing spot after download
2. correspondence/referee-reviews/{venue}-round{n}/*.pdf
3. correspondence/referee-reviews/*.pdf
4. Ask the user for the path if nothing found

Output Locations

External R&R mode

correspondence/referee-reviews/{venue}-round{n}/
├── {venue}-round{n}-reviews.pdf      (copy of input PDF — source NEVER moved; taxonomy name per rules/submission-file-archive.md, 2026-07-03)
├── {venue}-round{n}-reviews.md       (searchable parsed transcription — required pair with PDF)
├── {venue}-round{n}-rebuttal.md      (empty — for response draft)
├── reviews/                          (individual reviewer files)
│   ├── reviewer-1.md
│   └── ...
├── analysis/
│   ├── comment-tracker.md            (atomic comment matrix with R&R + 5-cat classification)
│   ├── review-analysis.md            (strategic overview + venue strategy)
│   └── reviewer-comments-verbatim.tex (LaTeX transcription)
└── plan/
    ├── REVISION_MASTER_PLAN.md       (six-phase Sihvonen plan — atomic tasks, DAG, blocks, risks)
    ├── revision_tasks.json           (DAG input — see references/task-schema.md)
    └── revision_dag_analysis.json    (computational output — parallel batches, critical path, bottlenecks)

Source PDF preservation: The original PDF is only ever copied to {venue}-round{n}-reviews.pdf. Never move, rename, or delete the source.

Principle: correspondence/ holds human exchanges; AI-derived review and planning artifacts stay under reviews/. Informal human collaborator feedback may remain in correspondence/internal/, while the derived internal plan goes under reviews/{scope}/strategic-revision/.

Submission-history stamp: ingesting a reviews PDF is a submission event — append a history: row (event: reviews-in, round: r{n}, files: pointing at {venue}-round{n}-reviews.pdf) to the paper's vault submission entry per rules/submission-file-archive.md § history. When the revision is later resubmitted, that resubmission appends its own response-sent + submitted rows (usually via session-close sync).

No-overwrite rule: If outputs already exist, version them (comment-tracker-v2.md, REVISION_MASTER_PLAN-v2.md). Always flag before writing.

Internal revision mode

reviews/{scope}/strategic-revision/{YYYY-MM-DD-HHMM}/
├── source-manifest.md                 (source paths, provenance, dates, hashes)
├── analysis/
│   ├── comment-tracker.md             (atomic findings + verification anchors)
│   └── review-analysis.md             (cross-review picture + readiness decision)
├── plan/
│   ├── REVISION_MASTER_PLAN.md
│   ├── revision_tasks.json
│   └── revision_dag_analysis.json
├── bundle/                            (created at successful closeout)
│   ├── strategic-revision.bundle.zip
│   ├── strategic-revision.manifest.json
│   └── strategic-revision.sha256
└── snapshots/                         (token-conservation baselines — see below)
    └── {label}/                       (e.g. pre-revision/, post-block-C/: main.tex + sections/ + .bib)

Internal mode links to its source reports in place and does not duplicate them. It creates no rebuttal file, venue-response transcription, submission-history event, or alternative-venue analysis unless the user separately requests one. Use a new timestamped package for a materially changed draft; use fold-in versioning for another review of the same draft.

When the source feedback was produced from a sealed review packet, source-manifest.md records that packet's path, archive SHA-256, canonical artifact SHA-256, and verification receipt. At successful closeout, invoke review-packet in gate-closeout mode so the exact reviews, analysis, adjudication, DAG, snapshots, and completion receipt form one decision-ready packet. A closeout packet is provenance, not a review verdict, and receives no separate reviews/INDEX.md row.

Snapshot location (mandatory, 2026-08-02). Token-conservation baselines and any other source snapshots taken during a revision live in the plan package's snapshots/{label}/ — git-tracked, tied to the plan that uses them. NEVER store revision snapshots under paper-*/backup/: that directory is the Overleaf backup automation's mirror surface (rsync --delete), and on 2026-07-31/2026-08-02 an unprotected mirror pass in daily-maintenance.sh gutted dated snapshot dirs stored there (both the P108 round-1 and round-2 baselines, plus two AIES submission-era archives). The automation's filters have been hardened, but the ownership boundary stands: paper-*/backup/ belongs to the automation; revision provenance belongs to the plan package.

Protocol — 11 Mode-Aware Phases

The skill runs 11 sequential phases. Detailed step-by-step procedure: references/phases.md.

JS = Sihvonen phases. FB = the user phases. Both = merged.

Phase Dependencies

• Phases 1-4 (mode-aware ingestion) must complete before Phase 5 begins
• Phase 8 is a hard gate: cycles → return to Phase 7, do not advance to Phase 9
• Phase 11 requires the Phase 9 block assignments before running full DAG analysis
• Phase 10 author decisions run only on Major/Critical findings (skip Minor/Editorial)

DAG Validation Quick Reference

Phase 8 (structural gate):

cd correspondence/referee-reviews/{venue}-round{n}/plan
uv run --with networkx python dag_validator.py revision_tasks.json --validate-only

Phase 11 (full analysis):

uv run --with networkx python dag_validator.py revision_tasks.json
# Produces revision_dag_analysis.json

Copy dag_validator.py from skills/strategic-revision/scripts/ into the plan directory before running. Invoke it through uv run --with networkx so the project environment is not mutated.

Full validator documentation: references/dag-validation.md. Task JSON schema: references/task-schema.md. R&R routing rules: references/rr-routing.md.

Classification: Two Axes

Every atomic task gets both classifications recorded in the comment tracker:

Category (5-cat, Sihvonen): structural intent of the change

• 🔴 STRUCTURAL — moves, cuts, reorganizations
• 🟠 ARGUMENTATIVE — theory, narrative, logical framing
• 🟡 EMPIRICAL — new regressions, data work, robustness
• 🟢 CLARIFICATION — definitions, justifications, explanations
• 🔵 EDITORIAL — formatting, typos, figures, terminology

R&R Routing (FB): workflow routing

• NEW ANALYSIS — needs code/data work
• CLARIFICATION — textual fix only
• DISAGREE — author adjudication; external mode may produce push-back in the later response, internal mode records Reject/Adapt/Defer
• MINOR — trivial or editorial

The two axes are orthogonal: a STRUCTURAL task is usually a CLARIFICATION routing; an EMPIRICAL task is usually NEW ANALYSIS; ARGUMENTATIVE with DISAGREE signals a high-stakes author decision and, in external mode only, a potentially rebuttal-heavy revision.

Strategic Coaching (Phase 10)

External mode: for each Major or Critical referee comment, walk the user through:

1. Understanding: reviewer's core concern — methodology, theory, or framing?
2. Position: Agree / Partially agree / Disagree (with core rebuttal if Disagree)
3. Risk: likelihood of reviewer escalation if pushed back on
4. Response sketch: one-sentence strategy (not full response)

Record in the comment tracker with two new columns: Position + Strategy.

Rules:

• Only Major/Critical get coaching. Minor/Editorial auto-classify as Agree.
• the user can say "skip coaching" to auto-classify all remaining as Agree.
• Maximum 2 rounds of dialogue per comment.
• Do not write the actual response letter — that remains the user's job.

Internal mode uses the same risk analysis but records Decision (Adopt / Adapt / Reject / Defer) and Rationale rather than simulating agreement with a referee. Internal findings never become response-letter claims.

Strategy (Phase 11)