light-review-rebuttal

name

light-review-rebuttal

description

Build auditable peer-review revision and author-response packages for Light stage 13. Use after receiving reviewer comments, a decision or meta-review; when drafting a rebuttal or response letter; when triaging major/minor revisions; when simulating a pre-submission review; or when a rejection may require a user-chosen 13→3 novelty, 13→5 experiment, or 13→8 writing back-edge. Consumes the selected venue/context and real PDF facts, preserves reviewer wording, atomizes issues, binds claims/evidence/actions/provenance, separates PLANNED from DONE, checks current venue limits without borrowing another venue's rules, and emits the stage-13 gate without changing venue, manuscript, evidence, citations, figures, PDF, or passport automatically.

Review and rebuttal

Build a source-preserving review registry, issue matrix, revision plan, evidence/change map, response draft, commitment ledger, unknown/failure record, and delivery package. Treat prose generation as the last layer, not the first.

Read review-rebuttal-resource-map.md before a real run. Read references/workflow_contract.md before creating or consuming canonical JSON. Read references.md when selecting review/rule sources. The competitor evidence is ../../docs/competitors/review-rebuttal.md.

Non-negotiable boundaries

1. Consume venue-matching's selected handoff and review context. Verify selected identity, selected_at timezone, selection_basis, user/delegated authorization, chosen candidate ID, fit/risk row, unmodified rule envelopes, source evidence path/as-of/source IDs, manuscript profile, and PDF path/hash/pages/page size/profile/compliance. Never switch venue, reorder tiers, or turn venue UNKNOWN into a fact.
2. Keep reviewer, editor, decision, and meta-review text verbatim in the canonical registry. Atom labels, root causes, strategies, and generated prose are interpretation layers; they never replace source text.
3. Record fetch time, source URL/type, round, reviewer ID, attachments and AVAILABLE|UNKNOWN|UNAVAILABLE|STALE. A 401/403/429/5xx, timeout, login, private invitation or network failure is UNAVAILABLE, not “no review.”
4. Never invent an experiment, analysis, citation, change, line number, reviewer identity, venue rule or result. PLANNED and IN_PROGRESS may not be phrased as completed. DONE requires a real change locator; completed experiment/analysis additionally requires verifiable run provenance with a matching SHA-256, not merely a local path. Before marking a response package ready, run the atom/action contract gate so source spans, reconstruction hashes, policy/ethics authorization and perspective-specific self-review are machine-checked rather than trusted.
5. Paper-writing owns manuscript claims and edits. Result-analysis owns evidence strength. Citation owns new-reference identity and claim support. Figure owns visual honesty. Typesetting owns PDF rebuild/compliance. This skill records and routes work; it does not impersonate those producers.
6. Stage 13 critical is narrow: only a routable root cause (novelty|experiment|writing) explicitly marked rejection_driving=true with a complete decision/meta-review/reviewer evidence envelope may become critical. Major labels or an overall Reject alone do not make every comment critical.
7. reviewer_classify and reroute produce advice only. Stop after presenting evidence and alternatives. Run passport add-back-edge only after the user chooses the root cause/back-edge. Never mutate the passport automatically.

Canonical workflow

1. Verify upstream identity

Require:

• light.selected_venue_handoff.v1;
• light.review_rebuttal_venue_context.v1;
• light.paper_claims.v1;
• light.evidence_strength.v1;
• citation registry when any new citation is proposed;
• the actual delivered PDF and its stage-11 facts.

Run:

python scripts/review_workflow.py \
  --spec review-input.json \
  --outdir review-delivery

If venue identity, rule envelopes, PDF hash, compliance, claims or evidence IDs do not agree, stop and repair the producer artifact. Do not “normalize” a conflict away.

The selected handoff must also retain A32's audit fields: timezone-bearing selected_at that is not in the future, non-empty selection_basis, decision_authority=user, coherent selected_by/status, delegated user_authorization when applicable, unchanged fit_risk, and a readable source_evidence.path whose SHA-256 matches the selected handoff and whose as_of/source_ids cover every sourced venue rule.

2. Capture reviews and decisions

For user-provided/private material, copy the text into reviews[].raw_text without correction and record reviews[].raw_sha256 plus a timezone-aware captured_at; the workflow re-computes the hash and blocks future/naive capture times. For a public OpenReview forum:

python scripts/fetch_openreview.py \
  --forum <forum-id> \
  --out openreview-capture.json

If the live API is unavailable but a fixed public PeerRead/OpenReview snapshot is the declared evidence source, capture that exact commit-pinned JSON instead:

python scripts/fetch_openreview.py \
  --peerread-url <raw-fixed-commit-json-url> \
  --out peerread-capture.json

The capture is calibration/source evidence. Do not redistribute restricted reviews. If capture is unavailable, continue only with material the user provided and retain the failure record.

3. Atomize without erasing source

Create one atom for each distinct request, claim, question, misunderstanding, or editorial item. Each atom must contain an exact contiguous source span copied from raw_text, with start/end offsets, span text and SHA-256. Also create addressable coverage units and a reconstruction hash for the reviewer units that require a response.

Assign one root cause such as novelty, experiment, writing, clarification, citation, ethics, scope, or editorial. Add a separate interpretation explaining the inferred concern. If a sentence contains two independent asks, create two atoms pointing to the same or overlapping source span; do not paraphrase the reviewer into a new source quote.

Run the stricter losslessness/response-action gate before drafting:

python scripts/review_response_contract.py \
  --input templates/review-response-contract.example.json

Replace the template with the real contract. The example is intentionally non-passing until current venue policy, ethics state and user authorization are verified. This gate catches missing reviewer units, duplicate atom/action coverage, fake DONE wording, incomplete evidence kinds, policy-forbidden reviewer requests, missing reviewer competence/conflict cards, and missing domain|method|statistics|ethics|cold_reader self-review perspectives.

4. Bind issues to owned evidence and actions

For every atom:

• bind exact paper claim_id values or leave the list empty;
• bind only real result-analysis evidence IDs;
• choose one strategy: acknowledge_and_fix, rebut_with_evidence, clarify, downgrade_claim, or request_editor_ruling;
• assign an owner and PLANNED|IN_PROGRESS|DONE|DECLINED|NOT_APPLICABLE;
• add change locator only after paper-writing actually edits the manuscript;
• add run provenance only after an experiment/analysis actually runs; for DONE experiment/analysis the artifact path must exist and match run_provenance.sha256;
• route new references through citation and keep them out until CONFIRMED.

Reviewer error is not permission to ignore a comment. Clarify with manuscript locator and evidence, or request editor ruling when the disagreement is material.

When a reviewer request itself conflicts with venue policy, ethics approval, data rights, consent, budget authorization or editor instructions, do not silently comply. Mark the action as DECLINED or REQUEST_RULING, bind the policy/ethics evidence, and keep the reviewer wording intact.

5. Budget from the selected venue only

Run:

python scripts/rebuttal_budget.py \
  review-delivery/response-draft.md \
  --context review-rebuttal-context.json

An AVAILABLE current authoritative rule can yield PASS/FAIL. UNKNOWN, UNAVAILABLE, STALE, or a page-only rule stays non-passing and explicit. Never apply an ICLR/CVPR/other venue preset to JORS or vice versa.

For every reviewer request for new numbers/experiments, classify it before any run as reanalysis/minimal/adapted/new-data plus feasibility and intended action:

python scripts/experiment_request_gate.py \
  --input templates/experiment_request.example.json

The template is intentionally UNKNOWN and non-passing until current official rules and a real user authorization replace its placeholders. RUN requires a VERIFIED current venue rule with source_type=OFFICIAL, a real source and ISO check date that allows new results, plus feasible scope, protocol, budget and user authorization. Tier-D new data/human study/ large sweep needs separate authorization. This gate only permits a run; DONE still requires run manifest + result artifact hash, and only then may the response use completed tense.

6. Check commitment truth

Run:

python scripts/check_commitments.py \
  --ledger review-delivery/commitment-ledger.json \
  --issues review-delivery/issue-matrix.json \
  --change-map review-delivery/evidence-change-map.json

Repair every critical finding. A valid locator proves only that a claimed change is traceable, not that the scientific response is adequate; perform a human re-review against the actual revised artifact.

7. Gate and pause

Run:

python scripts/reviewer_classify.py \
  --issues review-delivery/issue-matrix.json \
  --out reviewer-findings.json
python ../light-orchestrator/scripts/run_checkpoint.py \
  --file .light/passport.yaml --stage 13 \
  --findings reviewer-findings.json --ts <ISO-8601> --write
python ../light-orchestrator/scripts/reroute.py \
  --findings reviewer-findings.json --stage 13 \
  --passport .light/passport.yaml

If the gate fails, present each evidenced option:

• 13→3 for rejection-driving novelty;
• 13→5 for rejection-driving experiment/design;
• 13→8 for rejection-driving writing/presentation;
• rebut with evidence;
• downgrade the claim or record a limitation;
• request editor ruling.

Stop for the user's choice. Only then run:

python ../light-orchestrator/scripts/passport.py add-back-edge \
  --to <3|5|8> --from 13 --root-cause "<user-approved reason>" \
  --evidence-ptr <issue/evidence locator> --file .light/passport.yaml

Resource ownership

Resource	Responsibility
review-rebuttal-resource-map.md	execution order, source/access layers, cross-skill routing
references/workflow_contract.md	schemas, statuses, invariants and artifact semantics
references.md	live source policy, OpenReview/JORS caveats, verification guidance
templates/*	blank author inputs and human-readable response/re-review shapes
scripts/review_workflow.py	canonical validation and package emission
scripts/review_response_contract.py	source-span/reconstruction, atom coverage, response-action evidence, policy/ethics, reviewer-card and self-review gate
scripts/fetch_openreview.py	live API or fixed public snapshot capture; honest unavailable state and duplicate accounting
scripts/reviewer_classify.py	only evidenced rejection-driving stage-13 findings
scripts/check_commitments.py	coverage and PLANNED/DONE/provenance gate
scripts/rebuttal_budget.py	selected-context budget assessment; no venue presets
scripts/experiment_request_gate.py	venue-policy/feasibility/tier/authorization gate; never runs experiments

Self-check

• Selected venue/context, user-selection audit fields, source-evidence hash and actual PDF facts agree byte-for-byte.
• Reviewer/editor/meta-review source text is preserved and every atom points to an exact source span with offset and hash.
• Every available review source has timezone-aware captured_at and raw_sha256; the registry hash matches the exact raw_text.
• Addressable reviewer units reconstruct to the expected hash; no unit is omitted or duplicated unless explicitly justified.
• Every issue has type, root cause, strategy, owner and truthful status.
• Every response action declares evidence kind; DONE rows have real locators; completed scientific actions have run provenance path + SHA-256 match; proposed citations are citation-confirmed.
• Reviewer competence/conflict cards are recorded but never used to dismiss comments without editor ruling.
• Domain, method, statistics, ethics and cold-reader self-review all pass or have unresolved blockers surfaced before package-ready.
• Current venue response limits are sourced, or remain UNKNOWN.
• Commitment coverage passes and no planned action is written as done.
• Every experiment ask has tier/feasibility/action; RUN is venue-verified and user-authorized.
• Critical findings have explicit rejection-driving evidence.
• Reroute advice was shown to the user and no passport edge was added before the user's choice.

Honest capability boundary

Atomization, triage, point-by-point drafting, tone guidance, reviewer priority, budgeting, change locators, promise tracking and venue adaptation are common in peer skills; do not claim them as unique. Light's narrower machine contribution is verified consumption of upstream venue/PDF/claim/evidence/citation contracts, immutable source versus interpretation layers, strict PLANNED/DONE/run-provenance checks, and evidence-gated stage-13 routing that cannot execute without a user decision. Classification and response quality still require expert judgment.