Audit cross-venue submission histories without inferring chronology

Reviewed-on: #16

.gitea/workflows/check.yml

@@ -52,8 +52,11 @@ jobs:
               subprocess.run([python,'-I','-B','-m','paper_atlas','import','records','--help'],cwd=outside,check=True)
               subprocess.run([python,'-I','-B','-m','paper_atlas','import','arxiv','--help'],cwd=outside,check=True)
               subprocess.run([python,'-I','-B','-m','paper_atlas','maintain','backfill','--help'],cwd=outside,check=True)
+              subprocess.run([python,'-I','-B','-m','paper_atlas','audit','public-openreview','--help'],cwd=outside,check=True)
+              subprocess.run([python,'-I','-B','-m','paper_atlas','audit','histories','--help'],cwd=outside,check=True)
               subprocess.run([python,'-I','-B','-c','from paper_atlas.postgres.writer import PostgresStore, WriteConnection'],cwd=outside,check=True)
               smoke = "from paper_atlas.db.store import Store; from paper_atlas.config import DEFAULT_REGISTRY; import json; assert {'COLM','MM','MLSys'} <= {v['acronym'] for v in json.loads(DEFAULT_REGISTRY.read_text())['venues']}; s=Store('installed.sqlite'); s.conn.close()"
               subprocess.run([python,'-I','-B','-c',smoke],cwd=outside,check=True)
               subprocess.run([python,'-I','-B','-m','paper_atlas','query','--db','installed.sqlite','--stats'],cwd=outside,check=True)
+              subprocess.run([python,'-I','-B','-m','paper_atlas','audit','histories','--db','installed.sqlite','--examples','0'],cwd=outside,check=True)
           PY

docs/operations.md

@@ -38,7 +38,11 @@ SQLite 允许文本包含 `U+0000`（NUL），PostgreSQL `text` 不允许。迁
 
 论文身份匹配、跨来源合并、投稿/决定来源和标签维护复用同一套领域逻辑。PostgreSQL 使用原生 RETURNING 和 identity 序列分配 ID，继承迁移时设定的历史 ID 高水位；失败和冲突尝试可以留下序列空洞，不能把最大 ID 当作行数。绑定文本按 `nul-soh-v1` 编码，结果解码，GIN 随论文写入自动更新。
 
-会议文件的数据变化在同一事务内完成，manifest 数量不符等错误会回滚整份文件，`ingest_runs` 另保留 failed、已读数和零写入数；身份回填整个队列一次提交。arXiv 每 500 行提交数据和 checkpoint，异常只回滚尚未提交部分，重试必须仍为同一压缩包。大批次开始前核对独立审计清单、当前主库冲突和磁盘空间；完成后对数量、实际来源/决定关联及检索做数据验收，代码测试成功不代表生产导入成功。
+会议文件的数据变化在同一事务内完成，manifest 数量不符等错误会回滚整份文件，`ingest_runs` 另保留 failed、已处理数和零写入数；身份回填整个队列一次提交。arXiv 每 500 行提交数据和 checkpoint，异常只回滚尚未提交部分，重试必须仍为同一压缩包。大批次开始前核对独立审计清单、当前主库冲突和磁盘空间；完成后对数量、实际来源/决定关联及检索做数据验收，代码测试成功不代表生产导入成功。
+
+`audit public-openreview --pg-service ...` 在只读快照中核验两个受支持公开负面决定来源的固定 revision、逐文件哈希、manifest、原始出处标记和既有同届投稿。最新决定排序与查询端一致：观察时间、决定日期（为空时回退观察时间）、ID。实际导入持有写锁后再次复核该投稿的最新状态，阻止过期预检或弱来源标签覆盖冲突决定；冲突文件整体回滚并等待进一步证据核查，其他会议录用不构成同一次投稿的冲突。
+
+会议导入按原始字节流计算 SHA-256，在提交前与 manifest 提供的 `file_sha256` 比较；不符则回滚整份文件。成功批次的 checkpoint 会记录实际读取内容的哈希，便于复核导入时使用的文件版本。
 
 代码以固定 commit 部署，已经启动的任务继续使用其原版本。目录重构不要求迁移数据库，也不更改已有 JSONL 和 manifest 格式。
 

docs/usage.md

@@ -41,6 +41,16 @@ paper-atlas audit public-openreview --help
 
 其他会议的公开拒稿使用 `collect public-openreview`，再以 `audit public-openreview` 核对既有投稿身份和决定冲突。来源版本、逐文件状态盘点和未解决记录分别保存在 inventory 与 held 清单中，具体命令见[数据源说明](data-sources.md#其他会议的公开拒稿)。这类公开子集不能当作会议完整投稿名单。
 
+主库切换后，预检须直接连接当前 PostgreSQL，不能用冻结的 SQLite 备份推断当前决定：
+
+```sh
+paper-atlas audit public-openreview --pg-service foundry-research \
+  --inventory /data/paper-index/collected/public-openreview-v1/public-openreview-inventory.json \
+  --out /data/paper-index/collected/public-negative-pg-preflight.json
+```
+
+预检使用只读一致性快照，报告会记录实际数据库、schema、文件哈希和同届次/轨道的决定冲突；其他会议的录用保留为独立历史。预检不预留写锁，实际导入这两类公开负面决定时会在写事务中再次复核同一次投稿的最新状态；过期预检报告不能覆盖已经变化的决定。
+
 ## 3. 串行导入
 
 PostgreSQL 主库逐文件导入已经核验的资料：
@@ -69,10 +79,10 @@ paper-atlas maintain merge --db /data/paper-index/samples/example.sqlite \
 ## 4. 查询与分析前检查
 
 ```sh
-paper-atlas query --db /data/paper-index/papers.sqlite --query '"language model"' --limit 20
-paper-atlas query --db /data/paper-index/papers.sqlite --tag arxiv:category=cs.AI
-paper-atlas query --db /data/paper-index/papers.sqlite --venue ICLR --year 2025 --status rejected
-paper-atlas query --db /data/paper-index/papers.sqlite --paper 123
+paper-atlas query --pg-service foundry-research --query '"language model"' --limit 20
+paper-atlas query --pg-service foundry-research --tag arxiv:category=cs.AI
+paper-atlas query --pg-service foundry-research --venue ICLR --year 2025 --status rejected
+paper-atlas query --pg-service foundry-research --paper 123
 paper-atlas coverage --db /data/paper-index/papers.sqlite --out /data/paper-index/collected/coverage.json
 ```
 
@@ -81,10 +91,26 @@ paper-atlas coverage --db /data/paper-index/papers.sqlite --out /data/paper-inde
 例如，查找在 ICLR 有拒稿记录、但在其他已知投稿中已有录用的论文：
 
 ```sh
-paper-atlas query --db /data/paper-index/papers.sqlite \\
+paper-atlas query --pg-service foundry-research \
   --venue ICLR --status rejected --outcome accepted
 ```
 
 `known_acceptances` 列出目前仍为录用状态的各次投稿及会议、年份、track 和决定证据，`publication_outcome` 汇总为 accepted 或 no_known_acceptance。汇总会检查所有会议和 track，不把某次拒稿覆盖到整篇论文；同一次投稿的后续撤稿也会更新该次状态并保留原决定。no_known_acceptance 只表示现有来源中没有查到录用，不能称为“从未录用”。
 
 `--stats` 中 papers 是当前规则下的论文实体数；source_records 是来源记录数；decisions 是决定观察数。一个投稿可能有多个来源和多条观察，不能相加当作论文数。
+
+## 5. 投稿历史关系审计
+
+```sh
+PGSERVICEFILE=/etc/paper-atlas/pg_service.conf paper-atlas audit histories \
+  --pg-service foundry-research --pg-schema research --examples 20 \
+  --out /data/paper-index/collected/history-audit.json
+```
+
+使用一个只读一致性快照，区分论文实体、投稿和决定观察；分别统计同次投稿曾出现负面与录用观察、不同投稿共存负面与当前已知录用、不同会议之间的共存。workshop/main 是不同 track，属于同一会议时不算跨会议；会议别名也不会增加一个会议。当前已知录用使用与 `query --outcome` 相同的“每次投稿最新观察”规则，后续撤稿的该次投稿不再算当前录用，但旧录用决定仍保留。
+
+报告不凭观察时间、届次年份或行顺序断言“拒稿发生在录用之前”。样例保留真实 `decision_date`、`observed_at`、原始标签、证据 URL 和决定来源，供进一步核查；日期缺失仍保持未知。`negative` 统计包含 rejected、desk_rejected 和 withdrawn，其原始状态不互相替换。
+
+总数覆盖快照内全部投稿，样例由 `--examples` 限定为 0–100 篇；每篇最多展示 20 次投稿、每次最多 50 条决定、每条决定最多 100 个来源，并明确标记截断。样例优先保留负面/录用的不同投稿对。四项来源关系检查报告缺失关联和来源论文归属不一致，不自动修复。
+
+这些统计依赖现有论文身份关联，不证明实体合并正确，也不证明原始决定签名、公开拒稿全集、标签类型或最终全库去重已完成。SQLite 独立样本可使用 `--db`；冻结备份不能代表当前 PostgreSQL 的历史情况。

paper_atlas/audits/histories.py

@@ -0,0 +1,196 @@
+"""Audit linked submission histories without inferring event chronology."""
+import argparse
+from collections import Counter
+import json
+from pathlib import Path
+import sqlite3
+
+from paper_atlas.db.backend import database_arguments, open_read_connection
+from paper_atlas.db.submissions import timestamp
+from paper_atlas.querying import LATEST_DECISION
+
+
+ATTEMPTS = '''SELECT s.paper_id,s.id AS submission_id,e.venue_id,v.acronym AS venue,e.year,e.track,
+    count(d.id) AS decision_observations,
+    MAX(CASE WHEN d.status IN ('rejected','desk_rejected','withdrawn') THEN 1 ELSE 0 END)
+        AS historical_negative,
+    MAX(CASE WHEN d.status='accepted' THEN 1 ELSE 0 END) AS historical_acceptance,
+    COALESCE(MAX(CASE WHEN ''' + LATEST_DECISION + ''' THEN d.status END),'unknown')
+        AS latest_observed_status
+    FROM submissions s JOIN editions e ON e.id=s.edition_id JOIN venues v ON v.id=e.venue_id
+    LEFT JOIN decisions d ON d.submission_id=s.id
+    GROUP BY s.paper_id,s.id,e.venue_id,v.acronym,e.year,e.track
+    ORDER BY s.paper_id,e.venue_id,s.id'''
+
+
+def _paper_counts(attempts):
+    negative = sum(row['historical_negative'] for row in attempts)
+    accepted = sum(row['latest_observed_status'] == 'accepted' for row in attempts)
+    overlap = sum(row['historical_negative'] and row['latest_observed_status'] == 'accepted'
+                  for row in attempts)
+    venues = {}
+    for row in attempts:
+        counts = venues.setdefault(row['venue_id'], [0, 0])
+        counts[0] += row['historical_negative']
+        counts[1] += row['latest_observed_status'] == 'accepted'
+    return {
+        'papers_with_submissions': 1,
+        'papers_with_multiple_submissions': int(len(attempts) > 1),
+        'papers_with_historical_negative': int(negative > 0),
+        'papers_with_current_known_acceptance': int(accepted > 0),
+        'papers_with_negative_and_acceptance_in_distinct_attempts': int(negative * accepted > overlap),
+        'papers_with_negative_and_acceptance_in_distinct_venues':
+            int(negative * accepted > sum(n * a for n, a in venues.values())),
+    }
+
+
+def _example(connection, attempts):
+    # Keep a negative/accepted pair visible even when a paper has many attempts.
+    accepted = [row for row in attempts if row['latest_observed_status'] == 'accepted']
+    first = accepted[0] if accepted else None
+    other_venue = next((row for row in accepted if row['venue_id'] != first['venue_id']), None)
+    pair = None
+    for negative in attempts:
+        if negative['historical_negative'] and first:
+            candidate = first if first['venue_id'] != negative['venue_id'] else other_venue
+            if candidate:
+                pair = (negative, candidate)
+                break
+    if pair is None:
+        for negative in attempts:
+            if negative['historical_negative'] and first:
+                candidate = first if first['submission_id'] != negative['submission_id'] else (
+                    accepted[1] if len(accepted) > 1 else None)
+                if candidate:
+                    pair = (negative, candidate)
+                    break
+    selected = {}
+    for row in list(pair or ()) + attempts:
+        if len(selected) < 20:
+            selected.setdefault(row['submission_id'], dict(row))
+    for row in selected.values():
+        row['decisions'] = []
+        for decision in connection.execute('''SELECT id,status,raw_value,evidence_url,
+                decision_date,observed_at FROM decisions WHERE submission_id=?
+                ORDER BY observed_at DESC,COALESCE(decision_date,observed_at) DESC,id DESC LIMIT 50''',
+                (row['submission_id'],)):
+            entry = dict(decision)
+            sources = [dict(source) for source in connection.execute('''
+                SELECT so.name,r.external_id,ds.source_decision_id FROM decision_sources ds
+                JOIN source_records r ON r.id=ds.source_record_id JOIN sources so ON so.id=r.source_id
+                WHERE ds.decision_id=? ORDER BY so.name,r.external_id LIMIT 101''', (entry['id'],))]
+            entry['sources'] = sources[:100]
+            entry['sources_truncated'] = len(sources) > 100
+            row['decisions'].append(entry)
+        row['decisions_truncated'] = row['decision_observations'] > len(row['decisions'])
+    paper_id = attempts[0]['paper_id']
+    return {'paper_id': paper_id,
+            'title': connection.execute('SELECT title FROM papers WHERE id=?', (paper_id,)).fetchone()[0],
+            'counts': _paper_counts(attempts), 'attempt_count': len(attempts),
+            'attempts': list(selected.values()), 'attempts_truncated': len(attempts) > len(selected),
+            'chronology': 'Not inferred from observed_at, edition year, or row order.'}
+
+
+def report(db=None, *, pg_service=None, pg_schema=None, examples=20):
+    if not 0 <= examples <= 100:
+        raise ValueError('Examples must be between 0 and 100')
+    started_at = timestamp()
+    connection = open_read_connection(db, pg_service=pg_service, pg_schema=pg_schema)
+    if not pg_service:
+        connection.row_factory = sqlite3.Row
+    try:
+        if pg_service:
+            identity = connection.execute('SELECT current_database(),current_schema()').fetchone()
+            database, schema = identity[0], identity[1]
+        else:
+            database, schema = str(Path(db).resolve()), None
+        names = ['papers_with_submissions', 'papers_with_multiple_submissions',
+                 'papers_with_historical_negative', 'papers_with_current_known_acceptance',
+                 'papers_with_negative_and_acceptance_in_distinct_attempts',
+                 'papers_with_negative_and_acceptance_in_distinct_venues']
+        totals = Counter({key: 0 for key in names})
+        totals.update(submissions=0, decision_observations=0,
+                      attempts_with_historical_negative=0,
+                      attempts_with_historical_negative_and_acceptance=0)
+        latest = Counter()
+        sample_attempts = []
+
+        def consume(rows):
+            if not rows:
+                return
+            counts = _paper_counts(rows)
+            totals.update(counts)
+            if counts['papers_with_negative_and_acceptance_in_distinct_attempts'] and len(sample_attempts) < examples:
+                sample_attempts.append(rows)
+
+        current = []
+        for raw in connection.execute(ATTEMPTS):
+            row = dict(raw)
+            if current and row['paper_id'] != current[0]['paper_id']:
+                consume(current)
+                current = []
+            current.append(row)
+            totals['submissions'] += 1
+            totals['decision_observations'] += row['decision_observations']
+            totals['attempts_with_historical_negative'] += row['historical_negative']
+            totals['attempts_with_historical_negative_and_acceptance'] += int(
+                row['historical_negative'] and row['historical_acceptance'])
+            latest[row['latest_observed_status']] += 1
+        consume(current)
+        # These are relationship checks, not a substitute for the final entity/FK audit.
+        integrity = {}
+        checks = {
+            'submissions_without_source_links': '''SELECT count(*) FROM submissions s
+                WHERE NOT EXISTS(SELECT 1 FROM submission_sources ss WHERE ss.submission_id=s.id)''',
+            'decisions_without_source_links': '''SELECT count(*) FROM decisions d
+                WHERE NOT EXISTS(SELECT 1 FROM decision_sources ds WHERE ds.decision_id=d.id)''',
+            'submission_source_paper_mismatches': '''SELECT count(*) FROM submission_sources ss
+                JOIN submissions s ON s.id=ss.submission_id JOIN source_records r ON r.id=ss.source_record_id
+                WHERE s.paper_id<>r.paper_id''',
+            'decision_source_paper_mismatches': '''SELECT count(*) FROM decision_sources ds
+                JOIN decisions d ON d.id=ds.decision_id JOIN submissions s ON s.id=d.submission_id
+                JOIN source_records r ON r.id=ds.source_record_id WHERE s.paper_id<>r.paper_id''',
+        }
+        for name, query in checks.items():
+            integrity[name] = connection.execute(query).fetchone()[0]
+        return {'report_started_at': started_at, 'checked_at': timestamp(),
+                'engine': 'postgresql' if pg_service else 'sqlite', 'database': database,
+                'schema': schema, 'pg_service': pg_service,
+                'status': 'observed_history_audit_not_final_dedup', 'counts': dict(totals),
+                'latest_attempt_status_counts': dict(sorted(latest.items())),
+                'evidence_relationship_checks': integrity,
+                'examples': [_example(connection, rows) for rows in sample_attempts],
+                'example_limit': examples,
+                'interpretation': [
+                    'Counts use existing canonical paper links, not independently verified real-world identity.',
+                    'Negative includes rejected, desk_rejected and withdrawn observations; these remain distinct decisions.',
+                    'Known acceptance follows the latest observed decision per attempt, matching query --outcome.',
+                    'Distinct attempts/venues require separate submission/venue IDs; aliases do not create new venues.',
+                    'Coexisting evidence does not prove rejection happened before acceptance. observed_at is collection time.',
+                    'No known acceptance is not proof of never accepted. Source signatures and completeness are not established here.',
+                    'Examples are bounded; counts include all attempts and observations in one read-only snapshot.',
+                    'This does not perform entity merging, final deduplication, or label namespace/type validation.',
+                ]}
+    finally:
+        connection.close()
+
+
+def main(argv=None, *, prog=None):
+    parser = argparse.ArgumentParser(prog=prog)
+    database_arguments(parser)
+    parser.add_argument('--examples', type=int, default=20)
+    parser.add_argument('--out')
+    args = parser.parse_args(argv)
+    result = report(args.db, pg_service=args.pg_service, pg_schema=args.pg_schema, examples=args.examples)
+    text = json.dumps(result, ensure_ascii=False, indent=2)
+    if args.out:
+        path = Path(args.out)
+        stage = path.with_suffix('.part')
+        stage.write_text(text + '\n')
+        stage.replace(path)
+    else:
+        print(text)
+
+
+if __name__ == '__main__':
+    main()

paper_atlas/audits/public_openreview.py

@@ -4,14 +4,14 @@ from contextlib import closing
 import hashlib
 import json
 from pathlib import Path
-import sqlite3
 
 from paper_atlas.collectors.public_openreview import SOURCE, REVISION
 from paper_atlas.collectors.icml_public_archive import SOURCE as ARCHIVE_SOURCE, REVISION as ARCHIVE_REVISION
-from paper_atlas.db.submissions import timestamp
+from paper_atlas.db.submissions import timestamp, latest_status
+from paper_atlas.db.backend import database_arguments, open_read_connection
 
 
-def audit(db, inventory, out):
+def audit(db, inventory, out, *, pg_service=None, pg_schema=None):
     inventory = Path(inventory).resolve()
     result = json.loads(inventory.read_text())
     specs = {SOURCE: (REVISION, 'curated_public_status_assertion'),
@@ -20,17 +20,18 @@ def audit(db, inventory, out):
     if source not in specs or result.get('revision') != specs[source][0]:
         raise ValueError('Unsupported public source inventory')
     revision, evidence_kind = specs[source]
-    db = Path(db).resolve()
-    if not db.is_file():
-        raise ValueError('Existing database required for identity and status audit')
     conflicts = []
     seen = set()
     checked = 0
     existing_attempts = 0
     files = []
     checksums = {}
-    with closing(sqlite3.connect(db.as_uri() + '?mode=ro', uri=True)) as c:
-        c.execute('BEGIN')
+    with closing(open_read_connection(db, pg_service=pg_service, pg_schema=pg_schema)) as c:
+        if pg_service:
+            identity = c.execute('SELECT current_database(),current_schema()').fetchone()
+            database, actual_schema = identity[0], identity[1]
+        else:
+            database, actual_schema = str(Path(db).resolve()), None
         for manifest in result['files']:
             path = Path(manifest['file'])
             if hashlib.sha256(path.read_bytes()).hexdigest() != manifest['file_sha256']:
@@ -62,12 +63,11 @@ def audit(db, inventory, out):
                     raise ValueError('Multiple canonical attempts for one forum')
                 if existing:
                     existing_attempts += 1
-                    latest = c.execute('''SELECT status FROM decisions WHERE submission_id=?
-                        ORDER BY observed_at DESC,COALESCE(decision_date,'') DESC,id DESC LIMIT 1''', existing[0]).fetchone()
-                    if latest and latest[0] not in ('unknown', 'under_review', row['status']):
+                    latest = latest_status(c, existing[0][0])
+                    if latest and latest not in ('unknown', 'under_review', row['status']):
                         conflicts.append({'forum': forum, 'venue': row['venue'], 'year': row['year'],
                                           'track': row['track'], 'submission_id': existing[0][0],
-                                          'existing_status': latest[0], 'curated_status': row['status']})
+                                          'existing_status': latest, 'curated_status': row['status']})
                 count += 1
             if count != manifest['count']:
                 raise ValueError('Collected record count differs from inventory')
@@ -76,7 +76,8 @@ def audit(db, inventory, out):
             checksums[str(path)] = manifest['file_sha256']
     if checked != result['normalized_records']:
         raise ValueError('Normalized inventory total differs from checked records')
-    report = {'source': source, 'revision': revision, 'database': str(db),
+    report = {'source': source, 'revision': revision, 'database': database,
+              'engine': 'postgresql' if pg_service else 'sqlite', 'schema': actual_schema, 'pg_service': pg_service,
               'inventory': str(inventory), 'checked_at': timestamp(), 'records_checked': checked,
               'existing_same_edition_attempts': existing_attempts, 'conflicts': conflicts,
               'status': 'blocked' if conflicts else 'ready_for_serial_import',
@@ -94,11 +95,11 @@ def audit(db, inventory, out):
 
 def main(argv=None, *, prog=None):
     parser = argparse.ArgumentParser(prog=prog)
-    parser.add_argument('--db', required=True)
+    database_arguments(parser)
     parser.add_argument('--inventory', required=True)
     parser.add_argument('--out', required=True)
     args = parser.parse_args(argv)
-    report = audit(args.db, args.inventory, args.out)
+    report = audit(args.db, args.inventory, args.out, pg_service=args.pg_service, pg_schema=args.pg_schema)
     if report['status'] == 'blocked':
         raise SystemExit('Conflicting existing decisions require review before import')
 

paper_atlas/cli.py

@@ -35,6 +35,7 @@ COMMANDS = {
     ('audit','arxiv'): ('paper_atlas.audits.arxiv', {}, '核验 arXiv 原始编号、重复行与 ZIP 完整性'),
     ('audit','arxiv-tags'): ('paper_atlas.audits.arxiv_tags', {}, '逐记录核验 arXiv 类别及出处完整性'),
     ('audit','public-openreview'): ('paper_atlas.audits.public_openreview', {}, '核对公开负面决定与已有投稿历史'),
+    ('audit','histories'): ('paper_atlas.audits.histories', {}, '只读核对跨会议投稿历史及决定来源关系'),
 }
 
 

paper_atlas/db/backend.py

@@ -19,6 +19,25 @@ def open_store(db=None, *, pg_service=None, pg_schema=None):
     return Store(db)
 
 
+def open_read_connection(db=None, *, pg_service=None, pg_schema=None):
+    if bool(db) == bool(pg_service) or (pg_schema and not pg_service):
+        raise ValueError('Choose exactly one of SQLite db or PostgreSQL service; schema requires PostgreSQL')
+    if pg_service:
+        from paper_atlas.postgres.client import ReadConnection
+        return ReadConnection(pg_service, pg_schema)
+    import sqlite3
+    path = Path(db).resolve()
+    if not path.is_file():
+        raise ValueError('Existing database required for read-only audit')
+    connection = sqlite3.connect(path.as_uri() + '?mode=ro', uri=True)
+    try:
+        connection.execute('BEGIN')
+    except BaseException:
+        connection.close()
+        raise
+    return connection
+
+
 def inserted_id(connection, sql, params=()):
     if getattr(connection, 'postgres', False):
         return connection.execute(sql + ' RETURNING id', params).fetchone()[0]

paper_atlas/db/store.py

@@ -5,7 +5,7 @@ import re
 import sqlite3
 import unicodedata
 from pathlib import Path
-from paper_atlas.db.submissions import attach_submission,migrate_sources,write_decision
+from paper_atlas.db.submissions import attach_submission,migrate_sources,write_decision,latest_status
 from paper_atlas.db.backend import require_sqlite_writable, inserted_id
 
 
@@ -142,7 +142,7 @@ class Store:
             tag=c.execute('SELECT id FROM tags WHERE namespace=? AND value=?',(namespace,value)).fetchone()[0]
             c.execute('INSERT INTO paper_tags VALUES(?,?,?,?) ON CONFLICT DO NOTHING',(paper,tag,record,evidence))
 
-    def decision(self,paper,record,venue,year,track,external_id,url,status,raw_value,aliases=None,identities=None,observed_at=None,decision_date=None,source_decision_id='',submitted_at=None):
+    def decision(self,paper,record,venue,year,track,external_id,url,status,raw_value,aliases=None,identities=None,observed_at=None,decision_date=None,source_decision_id='',submitted_at=None,guard_status=False):
         c = self.conn
         existing=c.execute('SELECT id FROM venues WHERE lower(acronym)=lower(?)',(venue,)).fetchone()
         if existing:venue_id=existing[0]
@@ -154,6 +154,10 @@ class Store:
         # A previous event in this row may already have merged the input paper.
         paper=c.execute('SELECT paper_id FROM source_records WHERE id=?',(record,)).fetchone()[0]
         submission=attach_submission(self,edition,paper,record,external_id,identities or [])
+        if guard_status and status in ('rejected','desk_rejected','withdrawn'):
+            current = latest_status(c, submission)
+            if current and current not in ('unknown','under_review',status):
+                raise ValueError('Curated decision conflicts with current canonical submission status: ' + current)
         if submitted_at:
             from paper_atlas.db.submissions import timestamp
             c.execute('UPDATE submissions SET submitted_at=COALESCE(submitted_at,?) WHERE id=?',(timestamp(submitted_at),submission))

paper_atlas/db/submissions.py

@@ -70,6 +70,12 @@ def next_id(c,table,audit,former):
     return max(c.execute('SELECT COALESCE(max(id),0) FROM '+table).fetchone()[0],c.execute('SELECT COALESCE(max('+former+'),0) FROM '+audit).fetchone()[0])+1
 
 
+def latest_status(c, submission):
+    row = c.execute('''SELECT status FROM decisions WHERE submission_id=?
+        ORDER BY observed_at DESC,COALESCE(decision_date,observed_at) DESC,id DESC LIMIT 1''', (submission,)).fetchone()
+    return row[0] if row else None
+
+
 def merge_submissions(c,target,other):
     if target==other:return
     editions=c.execute('SELECT id,edition_id FROM submissions WHERE id IN (?,?)',(target,other)).fetchall()

paper_atlas/ingestion/records.py

@@ -1,5 +1,6 @@
 """Import normalized conference JSONL files with coverage evidence."""
 import argparse
+import hashlib
 import json
 import time
 from pathlib import Path
@@ -7,6 +8,8 @@ from pathlib import Path
 from paper_atlas.db.backend import database_arguments, open_store, inserted_id
 from paper_atlas.db.submissions import identities_for_row,timestamp
 from paper_atlas.collectors.pmlr_tracks import require_track_manifest, require_track_record
+from paper_atlas.collectors.public_openreview import SOURCE as PUBLIC_SOURCE
+from paper_atlas.collectors.icml_public_archive import SOURCE as ARCHIVE_SOURCE
 
 
 def ingest_file(db,path, *, pg_service=None, pg_schema=None):
@@ -30,10 +33,12 @@ def ingest_file(db,path, *, pg_service=None, pg_schema=None):
     editions = set()
     try:
         source_id = store.source(source_name,manifest['list_url'])
-        run = inserted_id(c,'INSERT INTO ingest_runs(source_id,checkpoint_json) VALUES(?,?)',(source_id,json.dumps({'file':str(path)})))
+        run = inserted_id(c,'INSERT INTO ingest_runs(source_id,checkpoint_json) VALUES(?,?)',(source_id,json.dumps({'file':str(path),'expected_sha256':manifest.get('file_sha256')})))
         c.commit()
-        with path.open() as stream:
+        digest = hashlib.sha256()
+        with path.open('rb') as stream:
             for line in stream:
+                digest.update(line)
                 row = json.loads(line)
                 if row['source'] != source_name:
                     raise ValueError('Source differs from manifest')
@@ -60,7 +65,7 @@ def ingest_file(db,path, *, pg_service=None, pg_schema=None):
                         status=event['status']
                         if events and status is None:raise ValueError('Explicit decision event requires a status')
                         if status is not None and status not in ('accepted','rejected','desk_rejected','withdrawn','under_review','unknown'):raise ValueError('Unsupported decision event')
-                        edition=store.decision(paper,record,row['venue'],row['year'],row['track'],external,event.get('evidence_url',row['url']),status,event.get('raw_value',''),row.get('venue_aliases'),identities=identities,observed_at=event.get('observed_at') or row_observed,decision_date=event.get('decision_date'),source_decision_id=event.get('source_decision_id',''),submitted_at=row.get('submitted_at'))
+                        edition=store.decision(paper,record,row['venue'],row['year'],row['track'],external,event.get('evidence_url',row['url']),status,event.get('raw_value',''),row.get('venue_aliases'),identities=identities,observed_at=event.get('observed_at') or row_observed,decision_date=event.get('decision_date'),source_decision_id=event.get('source_decision_id',''),submitted_at=row.get('submitted_at'),guard_status=source_name in (PUBLIC_SOURCE,ARCHIVE_SOURCE))
                     editions.add(edition)
                 else:
                     raise ValueError('This importer requires an explicit supported decision status')
@@ -68,6 +73,8 @@ def ingest_file(db,path, *, pg_service=None, pg_schema=None):
                 missing_abstracts += not bool((row.get('abstract') or '').strip())
         if not count:
             raise ValueError('Empty metadata file')
+        if manifest.get('file_sha256') and digest.hexdigest() != manifest['file_sha256']:
+            raise ValueError('Collected file hash differs from manifest')
         if len(editions) != 1:
             raise ValueError('A manifest must describe exactly one conference edition and track')
         edition = next(iter(editions))
@@ -76,7 +83,7 @@ def ingest_file(db,path, *, pg_service=None, pg_schema=None):
             raise ValueError(f'Manifest count {expected} differs from records {count}')
         complete=expected==count and manifest.get('status') not in ('partial','blocked') and not manifest.get('missing_abstract_count',0) and not missing_abstracts
         c.execute('INSERT INTO edition_sources(edition_id,source_id,list_url,status,expected_count,imported_count,last_success_at) VALUES(?,?,?,?,?,?,?) ON CONFLICT(edition_id,source_id,list_url) DO UPDATE SET status=excluded.status,expected_count=excluded.expected_count,imported_count=excluded.imported_count,last_success_at=excluded.last_success_at,last_error=NULL',(edition,source_id,manifest['list_url'],'complete' if complete else 'partial',expected,count,time.strftime('%Y-%m-%dT%H:%M:%SZ',time.gmtime())))
-        c.execute("UPDATE ingest_runs SET status='complete',finished_at=?,records_seen=?,records_written=? WHERE id=?",(timestamp(),count,count,run))
+        c.execute("UPDATE ingest_runs SET status='complete',finished_at=?,records_seen=?,records_written=?,checkpoint_json=? WHERE id=?",(timestamp(),count,count,json.dumps({'file':str(path),'file_sha256':digest.hexdigest()}),run))
         c.commit()
         print(json.dumps({'file':str(path),'status':'imported','records':count,'missing_abstract_count':missing_abstracts}),flush=True)
     except BaseException as exc:

tests/test_histories.py

@@ -0,0 +1,114 @@
+from pathlib import Path
+import tempfile
+import unittest
+
+from paper_atlas.audits.histories import report
+from paper_atlas.db.store import Store
+
+
+def add_history_fixture(store):
+    def event(work, forum, venue, year, track, status, observed):
+        paper, record = store.record('history_fixture', work + '/' + forum,
+            'https://example.org/' + forum, work + ' title\x00 preserved', 'Abstract', ['Jane Doe'],
+            {'arxiv': work})
+        store.decision(paper, record, venue, year, track, forum,
+            'https://example.org/' + forum + '/' + str(status), status, status or '',
+            identities=[('openreview', forum)], observed_at=observed)
+        return paper
+
+    # The negative observation was collected later: observation order is not chronology.
+    event('2501.00001', 'HistoryA01', 'ICLR', 2024, 'main', 'rejected', '2026-02-01T00:00:00Z')
+    event('2501.00001', 'HistoryA02', 'ICML', 2025, 'main', 'accepted', '2026-01-01T00:00:00Z')
+    event('2501.00002', 'HistoryB01', 'ICLR', 2024, 'main', 'rejected', '2026-01-01T00:00:00Z')
+    event('2501.00002', 'HistoryB01', 'ICLR', 2024, 'main', 'accepted', '2026-02-01T00:00:00Z')
+    event('2501.00003', 'HistoryC01', 'ICLR', 2024, 'main', 'desk_rejected', '2026-01-01T00:00:00Z')
+    event('2501.00003', 'HistoryC02', 'ICLR', 2024, 'workshop_example', 'accepted', '2026-02-01T00:00:00Z')
+    event('2501.00004', 'HistoryD01', 'ICML', 2025, 'main', 'accepted', '2026-01-01T00:00:00Z')
+    event('2501.00004', 'HistoryD01', 'ICML', 2025, 'main', 'withdrawn', '2026-02-01T00:00:00Z')
+    event('2501.00004', 'HistoryD02', 'ICLR', 2024, 'main', 'rejected', '2026-01-01T00:00:00Z')
+    event('2501.00005', 'HistoryE01', 'COLM', 2026, 'main', None, '2026-01-01T00:00:00Z')
+    store.conn.commit()
+
+
+class HistoryTests(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.addCleanup(self.tmp.cleanup)
+        self.db = Path(self.tmp.name) / 'sample.sqlite'
+        self.store = Store(self.db)
+        self.addCleanup(self.store.conn.close)
+
+    def test_distinct_attempts_venues_and_latest_acceptances_are_not_observation_counts(self):
+        add_history_fixture(self.store)
+        result = report(self.db)
+        self.assertEqual(result['counts'], {
+            'papers_with_submissions': 5, 'papers_with_multiple_submissions': 3,
+            'papers_with_historical_negative': 4, 'papers_with_current_known_acceptance': 3,
+            'papers_with_negative_and_acceptance_in_distinct_attempts': 2,
+            'papers_with_negative_and_acceptance_in_distinct_venues': 1,
+            'submissions': 8, 'decision_observations': 9,
+            'attempts_with_historical_negative': 5,
+            'attempts_with_historical_negative_and_acceptance': 2,
+        })
+        self.assertEqual(result['latest_attempt_status_counts'],
+                         {'accepted': 3, 'desk_rejected': 1, 'rejected': 2, 'unknown': 1, 'withdrawn': 1})
+        self.assertEqual(len(result['examples']), 2)
+        first = result['examples'][0]
+        self.assertIn('\x00', first['title'])
+        self.assertEqual({r['venue'] for r in first['attempts']}, {'ICLR', 'ICML'})
+        self.assertTrue(all(r['decisions'][0]['sources'] for r in first['attempts']))
+        self.assertTrue(all(r['decisions'][0]['decision_date'] is None for r in first['attempts']))
+        self.assertIn('Not inferred', first['chronology'])
+        self.assertEqual(result['status'], 'observed_history_audit_not_final_dedup')
+        self.assertTrue(all(v == 0 for v in result['evidence_relationship_checks'].values()))
+
+    def test_missing_and_mislinked_source_evidence_is_reported_without_repair(self):
+        add_history_fixture(self.store)
+        self.store.conn.execute('DELETE FROM decision_sources WHERE decision_id=1')
+        self.store.conn.execute('DELETE FROM submission_sources WHERE submission_id=1')
+        # Record3 belongs to paper2; move it to paper1 to create a real mismatch.
+        self.store.conn.execute('UPDATE source_records SET paper_id=1 WHERE id=3')
+        self.store.conn.commit()
+        result = report(self.db, examples=0)
+        checks = result['evidence_relationship_checks']
+        self.assertEqual(checks['decisions_without_source_links'], 1)
+        self.assertEqual(checks['submissions_without_source_links'], 1)
+        self.assertEqual(checks['submission_source_paper_mismatches'], 1)
+        self.assertEqual(checks['decision_source_paper_mismatches'], 2)
+        self.assertEqual(self.store.conn.execute('SELECT paper_id FROM source_records WHERE id=3').fetchone()[0], 1)
+
+    def test_example_limit_does_not_change_global_counts(self):
+        add_history_fixture(self.store)
+        none, one = report(self.db, examples=0), report(self.db, examples=1)
+        self.assertEqual(none['counts'], one['counts'])
+        self.assertEqual(none['examples'], [])
+        self.assertEqual(len(one['examples']), 1)
+
+    def test_empty_snapshot_and_invalid_limit_do_not_create_missing_database(self):
+        empty = report(self.db)
+        self.assertTrue(all(v == 0 for v in empty['counts'].values()))
+        self.assertEqual(empty['latest_attempt_status_counts'], {})
+        absent = self.db.parent / 'absent.sqlite'
+        with self.assertRaisesRegex(ValueError, 'Examples'):
+            report(absent, examples=-1)
+        with self.assertRaisesRegex(ValueError, 'Existing database'):
+            report(absent)
+        self.assertFalse(absent.exists())
+
+    def test_truncated_example_keeps_the_cross_venue_acceptance_visible(self):
+        paper, record = self.store.record('fixture', 'many', 'https://example.org/many',
+                                          'Many attempts', 'Abstract', [])
+        for index in range(26):
+            accepted = index == 25
+            forum = f'BoundedForum{index}'
+            self.store.decision(paper, record, 'ICML' if accepted else 'ICLR', 2025, 'main', forum,
+                                'https://example.org/' + forum, 'accepted' if accepted else 'rejected',
+                                'Accept' if accepted else 'Reject', identities=[('openreview', forum)])
+        self.store.conn.commit()
+        result = report(self.db)
+        example, = result['examples']
+        self.assertEqual((example['attempt_count'], len(example['attempts'])), (26, 20))
+        self.assertTrue(example['attempts_truncated'])
+        self.assertEqual({row['latest_observed_status'] for row in example['attempts']},
+                         {'accepted', 'rejected'})
+        self.assertEqual(result['counts']['submissions'], 26)

tests/test_postgres.py

@@ -175,6 +175,22 @@ class PostgreSQLIntegrationTests(unittest.TestCase):
         finally:
             reader.close()
 
+    def test_history_audit_matches_sqlite_for_actual_postgresql_snapshot(self):
+        from paper_atlas.audits.histories import report
+        from tests.test_histories import add_history_fixture
+        store = Store(self.db)
+        try:
+            add_history_fixture(store)
+        finally:
+            store.conn.close()
+        expected = report(self.db)
+        self.assertEqual(self.migrate()['status'], 'complete')
+        actual = report(pg_service=self.service, pg_schema=self.target)
+        self.assertEqual(actual['engine'], 'postgresql')
+        self.assertEqual(actual['schema'], self.target)
+        for key in ('counts', 'latest_attempt_status_counts', 'evidence_relationship_checks', 'examples'):
+            self.assertEqual(actual[key], expected[key], key)
+
     def interrupt_after_first_table(self):
         calls = 0
         def interrupt(*args):

tests/test_postgres_ingestion.py

@@ -11,7 +11,7 @@ from unittest.mock import patch
 import uuid
 import zipfile
 
-from paper_atlas.db.backend import open_store
+from paper_atlas.db.backend import open_store, open_read_connection
 from paper_atlas.db.store import Store
 from paper_atlas.ingestion import arxiv
 from paper_atlas.ingestion.backfill import backfill
@@ -27,6 +27,13 @@ class BackendSelectionTests(unittest.TestCase):
             with self.assertRaisesRegex(ValueError, 'Choose exactly one'):
                 open_store(**kwargs)
 
+    def test_readonly_audit_does_not_create_an_absent_sqlite_database(self):
+        with tempfile.TemporaryDirectory() as tmp:
+            path = Path(tmp) / 'absent.sqlite'
+            with self.assertRaisesRegex(ValueError, 'Existing database'):
+                open_read_connection(path)
+            self.assertFalse(path.exists())
+
 
 @unittest.skipUnless(os.environ.get('PAPER_ATLAS_TEST_PG_SERVICE'), 'Requires the approved remote PostgreSQL test service')
 class PostgreSQLIngestionTests(unittest.TestCase):
@@ -138,6 +145,21 @@ class PostgreSQLIngestionTests(unittest.TestCase):
         finally:
             store.conn.close()
 
+    def test_changed_file_hash_rolls_back_data_before_commit(self):
+        path = self.file('changed', [self.row()])
+        manifest = json.loads(path.with_suffix('.manifest.json').read_text())
+        manifest['file_sha256'] = '0' * 64
+        path.with_suffix('.manifest.json').write_text(json.dumps(manifest))
+        with self.assertRaisesRegex(ValueError, 'hash differs'):
+            self.ingest(path)
+        store = self.store()
+        try:
+            self.assertEqual(store.conn.execute('SELECT count(*) FROM papers').fetchone()[0], 1)
+            self.assertEqual(store.conn.execute('SELECT count(*) FROM source_records').fetchone()[0], 1)
+            self.assertEqual(store.conn.execute('SELECT status,records_written FROM ingest_runs').fetchone(), ('failed', 0))
+        finally:
+            store.conn.close()
+
     def test_writer_lock_survives_commit_and_rollback_but_readers_continue(self):
         first = self.store()
         try:
@@ -165,6 +187,62 @@ class PostgreSQLIngestionTests(unittest.TestCase):
         with connect(self.service) as pg:
             self.assertEqual(pg.execute('SELECT count(*) FROM information_schema.tables WHERE table_schema=%s', (staging,)).fetchone()[0], 0)
 
+    def test_negative_preflight_uses_current_pg_attempt_and_blocks_stale_import(self):
+        import hashlib
+        from paper_atlas.collectors.public_openreview import SOURCE, REVISION
+        from paper_atlas.audits.public_openreview import audit
+        row = self.row(SOURCE, 'negativeForum')
+        row['identifiers']['openreview'] = 'negativeForum'
+        row['metadata'] = {'evidence_kind': 'curated_public_status_assertion',
+                           'source_revision': REVISION, 'original_decision_signature_verified': False}
+        path = self.file('negative', [row])
+        manifest = json.loads(path.with_suffix('.manifest.json').read_text())
+        manifest.update(file=str(path), file_sha256=hashlib.sha256(path.read_bytes()).hexdigest())
+        path.with_suffix('.manifest.json').write_text(json.dumps(manifest))
+        inventory = self.root / 'negative-inventory.json'
+        inventory.write_text(json.dumps({'source': SOURCE, 'revision': REVISION,
+                                        'normalized_records': 1, 'files': [manifest]}))
+        # An acceptance in another conference does not block this rejection.
+        self.ingest(self.file('other-conference', [self.row('official', 'acceptedElsewhere', 'accepted', 2026, 'COLM')]))
+        def preflight(name):
+            with redirect_stdout(io.StringIO()):
+                return audit(None, inventory, self.root / (name + '.json'), pg_service=self.service, pg_schema=self.schema)
+        first = preflight('before')
+        self.assertEqual(first['status'], 'ready_for_serial_import')
+        self.assertEqual(first['engine'], 'postgresql')
+        self.assertEqual(first['schema'], self.schema)
+        self.ingest(path)
+        self.assertEqual(preflight('existing')['existing_same_edition_attempts'], 1)
+        store = self.store()
+        try:
+            paper, record = store.conn.execute('''SELECT r.paper_id,r.id FROM source_records r
+                JOIN sources s ON s.id=r.source_id WHERE s.name=? AND r.external_id=?''', (SOURCE, row['external_id'])).fetchone()
+            store.decision(paper, record, 'ICLR', 2025, 'main', row['external_id'],
+                           'https://example.org/formal', 'accepted', 'Accept', observed_at='2027-01-01T00:00:00Z')
+            # Equal observation time, but an explicitly earlier decision date.
+            # NULL dates must fall back to observed_at, not the empty string.
+            store.decision(paper, record, 'ICLR', 2025, 'main', row['external_id'],
+                           'https://example.org/older', 'rejected', 'Reject', observed_at='2027-01-01T00:00:00Z',
+                           decision_date='2026-12-31T00:00:00Z')
+            store.conn.commit()
+        finally:
+            store.conn.close()
+        blocked = preflight('blocked')
+        self.assertEqual(blocked['status'], 'blocked')
+        self.assertEqual(blocked['conflicts'][0]['existing_status'], 'accepted')
+        self.assertEqual(blocked['files'], [])
+        # The earlier ready report cannot authorize overwriting current status.
+        with self.assertRaisesRegex(ValueError, 'conflicts with current'):
+            self.ingest(path)
+        reader = open_read_connection(pg_service=self.service, pg_schema=self.schema)
+        try:
+            self.assertEqual(reader.execute('SHOW transaction_read_only').fetchone()[0], 'on')
+            from paper_atlas.db.submissions import latest_status
+            sid = reader.execute("SELECT submission_id FROM submission_identifiers WHERE scheme='openreview' AND value=?", ('negativeForum',)).fetchone()[0]
+            self.assertEqual(latest_status(reader, sid), 'accepted')
+        finally:
+            reader.close()
+
     def test_backfill_merges_evidence_without_reimporting_or_losing_history(self):
         files = []
         for source, external, title in [('official', 'legacy', 'First bibliographic name'), ('openreview_snapshot', 'forum', 'Renamed submission')]: