light-review-rebuttal

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

```bash
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:

```bash
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:

```bash
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:

```bash
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 catch