light-venue-matching

# Venue matching · stage 12

Turn a delivered paper into an auditable venue decision. Read
`venue-resource-map.md` before a real run and
`references/workflow_contract.md` before producing or consuming JSON. Use
`references.md` to choose current sources. Start from
`templates/venue_input.json`; never start from model memory or a bundled venue
list.

## Non-negotiable boundaries

1. Consume typesetting's `venue-handoff.json`. Verify its PDF path/hash,
   `DELIVERED`, pages, page size, profile/source, compliance `PASS`, and zero
   critical findings. Preserve paper/figure/citation/typesetting provenance.
   Do not compile, reformat, inspect page boxes again, or treat stage-11
   `UNAVAILABLE` as compliance.
2. Consume paper-writing's manuscript/claim profile through
   `manuscript_profile.claims_delivery` with a safe relative path, schema, and
   SHA-256. Do not change claims, methods, results, article type, data scale,
   or evidence strength to make a venue fit. Citation owns reference
   authenticity; figure owns visual honesty.
3. Treat acceptance rate, review time, APC/OA, indexing, quartile, and CFP
   deadline as high-velocity fields. Require a source checked on the run date.
   Otherwise emit `UNKNOWN`, `UNAVAILABLE`, or `STALE`; never use memory.
   `AVAILABLE` sources must carry an auditable locator (`url`, `query`,
   `locator`, or `path`), `checked_at`, `access_tier`, and `authority`.
   Official rules/fees/deadlines require official/publisher/venue authority;
   indexing and quartile require index/registry authority.
4. Treat 403/429/5xx, missing key/login/subscription, robots denial, and source
   outage as `UNAVAILABLE`. They do not mean “not indexed,” “not in DOAJ,”
   “free,” or “risky.”
5. Keep predatory/hijacked signals as visible warnings pending current,
   multi-source human review. DOAJ absence, high APC, fast review, unsolicited
   email, unusual volume, or one archived list is not a final verdict.
6. Never estimate an acceptance percentage or convert fit into acceptance
   probability. Use official current acceptance figures only with source/date;
   otherwise `acceptance_likelihood.status=UNKNOWN`.
7. Keep `decision_point=true` and `chosen=null` through candidate discovery,
   evidence collection, ranking, and delivery. Stop and ask the user to choose.
   Only an explicit `light.venue_user_selection.v1` may create a selected
   handoff. Record a direct choice as `actor=user`; if the user explicitly
   delegates the choice, preserve the authorization verbatim and use
   `actor=agent_with_user_authorization`, `decision_authority=user`. Never
   submit. Every selection artifact must include timezone-aware `selected_at`
   and the user's stated trade-off in `because`. Bind the choice to the exact
   reviewed packet with `decision_sha256`; a changed packet or any changed
   registry/evidence/fit artifact requires a new review and selection.
8. Do not invent `STAGE_GATES[12]`, `ROUTES[12]`, a confirmation checkpoint,
   critical findings, or a back-edge. Stage 12 is a user decision point with no
   configured gate or route edge.

## Workflow

### 1. Consume the real submission artifact

Require `light.typesetting_venue_handoff.v1`. Run prepare only when:

- `status=DELIVERED`;
- `compliance_status=PASS` and `critical_count=0`;
- the PDF exists and its SHA-256 matches;
- the compliance report exists and agrees on page facts.

If any condition fails, return an input error and route the author to stage 11
without creating a stage-12 critical finding.

### 2. Capture the author constraints

Record, without filling gaps yourself:

- research direction, article type, methods, data scale, claims/evidence;
- author stage, region, required indexes, OA requirement, APC ceiling;
- hard submission deadline and acceptable review duration;
- reach/match/safety preference and transfer strategy;
- unacceptable venues, publishers, business models, or risks.

Mark each constraint as hard or soft. A hard author constraint can exclude; a
soft preference changes order and explanation.
Bind the claim/evidence profile to the current paper-writing artifact via
`claims_delivery.path + sha256 + schema`; a hand-typed manuscript profile is
not enough for stage 12.

### 3. Discover for recall, then verify for precision

Keep the unpublished manuscript local. Before any public or externally
authenticated search, translate it into author-approved broad field/method-family
terms and preflight the outgoing queries:

```powershell
python scripts/query_privacy_gate.py `
  --input templates/query-privacy.example.json
```

Do not send the unpublished title, abstract, exact hypotheses, unique method or
dataset names, result sentences, tables, or figures to a public search engine.
The preflight report intentionally retains only query hashes and match categories.
Its PASS detects supplied phrase overlap; it cannot prove anonymity or rule out
re-identification.

Use the author's candidate list, current official CFPs, publisher finders, and
`venue_discovery.py`:

```powershell
python scripts/venue_discovery.py `
  --query "author-approved broad field and method family" --rows 50 --out discovery.json
```

Crossref container frequency only discovers candidates. It does not establish
scope, rank, indexing, safety, or quality. Record every discovery query,
endpoint, access tier, status, and check time. Deduplicate by ISSN plus official
name; keep title conflicts for manual review.

### 4. Build field-level evidence

For every candidate, collect separate envelopes for:

- official Aims & Scope and article types;
- length/page, figure/table, supplement, anonymity, template, and format rules;
- method/data fit and recent comparable articles;
- OA/APC, timing, indexing/quartile, and current CFP deadline;
- risk/hijack checks and unresolved identity conflicts.

Prefer official venue/publisher instructions for rules, authoritative indexes
for index membership, and registration metadata for identity. Keep JCR,
Scopus, Cab