Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions DECISIONS.md
Original file line number Diff line number Diff line change
Expand Up @@ -1444,3 +1444,9 @@
- 目标线沿用 11 月 12 日候选工作树记录的两项隐藏集门槛结论。迁移 042 给新库和从未保存过目标设置的旧库开启两个开关。已保存过设置的旧库保留之前的有效值。另用 v0.0.3 创建数据库,验证了迁移顺序、Persona 设置和三种旧库目标设置的升级行为。
- G1 在冻结公开集上重新评测。GLM high 的 27 次调用均正常,30/30 正确,应关闭目标识别 12/12,误关闭和外部目标被关闭均为 0。完整报告留在仓库外。后续发布以审查过的候选提交为准。
- 人物线预检核对了冻结公开集的 70 条待复核同一人联系。24 条应合并联系的标注都能找到跨本地日期的独立依据,保留主体顺序与创建时间一致。现有消息入库使用合并后的主体 ID。人物线实现拆分时需保留合并后收到的消息原始归属,以便还原。

## 2026-11-17 人物自动合并与拆分候选

- 每日维护复核已存在的可能联系。代码核对跨日原始发言、独立发言者、永久否认和人物关系后,以维护身份合并至创建较早的人物。开关默认关闭,等待冻结隐藏集验收。
- 从迁移 043 起,合并记录保存拆分需要的身份和关联变化。拆分恢复原有归属,沿用管理员修订号检查,并永久否认该联系。合并后新收到的账号消息及明确的引用保留原始主体。后续合并涉及同一保留人物时,先拆后一次合并。
- 冻结公开集第一轮两次超时,运行作废。两处误合并都把跑团或剧情账号视为操作者的现实身份。补充原始发言中的明确扮演护栏和提示词后,全量复跑 30 例、70 条联系,59 次模型调用均完成。GLM high 误合并 0,正确合并 23/24(95.8%),余下一条保持可能。完整报告放仓库外。
4 changes: 4 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -670,6 +670,10 @@ GD2:`pending`/运行中的目标在补判结束前不进入 prepare 和 sear

人物管理接口沿用管理员会话、CSRF、JSON Content-Type 和本机 Host 校验;人物页提供列表、详情、否认、确认合并和别名管理,有待确认联系时导航和试用页会提示。

全自动阶段 I 在生命周期设置提供“每日维护自动合并同一人”,默认关闭。开启后,每次维护最多复核 32 条待确认联系。模型核对身份及反证,代码核对候选的原始发言、角色时区日期、发言者和合并链。符合条件时保留创建较早的主体,同一时刻取较小的主体 ID。维护记录包含判断结果,人物详情显示合并记录及已核对的依据消息。

从迁移 043 起记录可拆分的合并。人物详情可以拆分,管理接口为 `POST /admin/api/people/merges/{id}/split`,传保留人物的 `expected_revision`。拆分恢复原有身份和相关记录,把联系标为永久否认。同一人物先拆后发生的合并;涉及资料有冲突时返回 409。账号和明确识别出的引用保留原始人物归属,拆分时恢复。合并后新学习的记忆只有在原始来源明确归属一方时才调整说话人。

| 接口 | 输入与行为 |
| --- | --- |
| `GET /admin/api/people` | `text` 搜索名字/有效别名;`pending_only=true` 只列有待确认联系的人;`include_merged=true` 包含占位。支持 `limit`、`offset`,返回 `pending_links`、别名、记忆数与主体 `revision`。默认不列 self/scene。 |
Expand Down
2 changes: 1 addition & 1 deletion companion_memory_cognition_system_design_integrated.md
Original file line number Diff line number Diff line change
Expand Up @@ -485,7 +485,7 @@
- 各平台账号先作为独立主体保存。昵称相同不足以判定是同一个人。
- 系统可以从对话中学到“小林=林同学”这样的联系,并记录相信程度和来源。召回某人时可以附带与其“可能是同一人”的主体的要点,并标明相信程度。管理员可以确认合并,也可以否认这条联系,否认永久有效。

全自动阶段(people_merge_hidden_v1 门槛通过后生效,此前仍只由管理员确认合并):每日维护复核“可能是同一人”的联系,至少两条独立证据(不同日期、不同来源消息,至少一条来自本人或两条来自不同他人,角色自己的发言不算)、没有反证、两主体之间没有被否认过时才自动合并;合并可以拆分还原。
全自动阶段:每日维护复核“可能是同一人”的联系,至少两条独立证据(不同本地日期、不同原始消息,至少一条来自本人或两条来自不同他人,角色自己的发言不算)。没有反证,双方合并链无永久否认、彼此不涉及扮演或祖先关系时,保留创建较早的主体。合并可拆分还原,拆分记为永久否认。people_merge_hidden_v1 门槛通过前默认关闭自动合并。
- 本人声明的昵称或新账号名(例如“叫我软糖就行”“我新号叫杰克船长”)直接记为这个主体的别名,不新建主体,也不需要确认。只有他人推测或无法确认时,才建立带相信程度的“可能是同一人”联系。
- 提到但不在场的人(如“我妈”)作为从属于说话人的人物主体保存,例如“小林的妈妈”。

Expand Down
10 changes: 10 additions & 0 deletions evals/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -253,6 +253,16 @@ M2(记忆质量)的范围、跨阶段划分和门槛口径见 DECISIONS.md 2
- 公开学习语料的学习请求逐字一致(过滤默认关闭)。
- 人物线改变回复准备的返回时,按召回守则比较,并证明学习请求不变。

## 全自动阶段 I 的人物复核评测

默认使用冻结的 [people_merge_v1](people_merge_v1_notes.md),含 30 例、70 条待复核联系,其中 24 条应合并。每例隔离数据库并固定时钟。入口、主体、联系和原始消息按案例写入,再通过每日维护执行真实判断。模型材料只来自产品字段,标注在写回后用于计分。

```bash
uv run python evals/people_merge_eval.py --out /absolute/people-merge-dev --workers 4
```

允许使用 `--corpus` 替换输入。评测按最终的合并记录计分。误合并包括合并了应保持可能的联系以及保留主体错误,识别率按应合并联系中正确合并的比例计算。预先登记的公开验证目标为误合并 0、识别率至少 70%。复核失败、超时、限流或 SQLite 外键检查失败的运行作废,完整结果及指纹保存在仓库外。公开结果用于候选验证,默认开启仍须通过独立保管的 people_merge_hidden_v1。

## GLM 阶段的起点(2026-10-05,历史)

2026-10-05 起,开发和评测的对话模型改为 glm-5.3-flash,embedding 仍为 doubao-embedding-vision,开发设备改为 macOS(见 DECISIONS.md 同日记录)。MiniMax-M3 时期的评测报告和执行记录已从仓库删除,原文保留在 git 历史中(main `2ee81d8`)。
Expand Down
184 changes: 184 additions & 0 deletions evals/people_merge_eval.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,184 @@
"""Run daily identity review on the frozen public corpus and score actual merges."""
from __future__ import annotations

import argparse
import hashlib
import json
import re
from concurrent.futures import ThreadPoolExecutor
from datetime import datetime
from pathlib import Path
from tempfile import TemporaryDirectory

from iris.db import Store
from iris.goals import AUTO_REVIEW_KEY, DUPLICATE_AUTO_KEY
from iris.maintenance import Maintenance, PHASES
from iris.model_health import ModelHealth
from iris.models import Gateway, load_test_models
from iris.people_review import PROMPT, PURPOSE, SETTING

ROOT = Path(__file__).resolve().parents[1]
PRODUCT_FIELDS = ('now', 'entries', 'subjects', 'links', 'memories', 'messages')


def seed(store, material, *, zone, role_name):
store.set_setting('timezone', zone)
store.set_setting('role_name', role_name)
store.set_setting('lifecycle', {SETTING: True})
store.set_setting('goals', {AUTO_REVIEW_KEY: False, DUPLICATE_AUTO_KEY: False})
messages, links = {}, {}
current = datetime.fromisoformat(material['now'])
with store.write() as conn:
for entry in material['entries']:
conn.execute('INSERT INTO entries(id,name,platform,kind) VALUES(?,?,?,?)',
(entry['id'], entry['name'], 'eval', entry['kind']))
for person in material['subjects']:
conn.execute("INSERT INTO subjects(id,kind,name,created_at) VALUES(?,'person',?,?)",
(person['id'], person['name'], person['created_at']))
conn.executemany('INSERT INTO subject_aliases(subject_id,alias) VALUES(?,?)',
[(person['id'], alias) for alias in person['aliases']])
if person['account']:
conn.execute("INSERT INTO platform_identities(subject_id,platform,account_id,display_name) VALUES(?,'eval',?,?)",
(person['id'], person['account'], person['name']))
for person in material['subjects']:
if person['parent']:
conn.execute('UPDATE subjects SET parent_id=? WHERE id=?', (person['parent'], person['id']))
for message in material['messages']:
if datetime.fromisoformat(message['at']) > current:
continue
messages[message['key']] = conn.execute('''INSERT INTO messages(entry_id,kind,sender_subject_id,
content,quote_author_subject_id,quote_content,occurred_at,received_at,dedupe_key,learning_state)
VALUES(?,?,?,?,?,?,?,?,?,'learned')''', (message['entry_id'], message['kind'], message['sender'],
message['content'], message['quote_author'], message['quote_content'], message['at'], message['at'],
message['key'])).lastrowid
for link in material['links']:
a, b = (sorted((link['a'], link['b'])) if link['kind'] == 'same_as' else (link['a'], link['b']))
links[link['key']] = conn.execute('''INSERT INTO subject_links(subject_a,subject_b,kind,status,belief,
source_message_id,created_at) VALUES(?,?,?,?,?,?,?)''',
(a, b, link['kind'], link['status'], link['belief'], messages.get(link['source']), link['created_at'])).lastrowid
for memory in material['memories']:
stamp = memory['created_at']
mid = conn.execute('''INSERT INTO memories(content,kind,speaker_subject_id,stance,belief,importance,
retention,created_at,updated_at,first_confirmed_at,last_confirmed_at)
VALUES(?,'关系',?,?,60,50,50,?,?,?,?)''',
(memory['content'], memory['speaker'], memory['stance'], stamp, stamp, stamp, stamp)).lastrowid
conn.executemany('INSERT INTO memory_subjects(memory_id,subject_id) VALUES(?,?)',
[(mid, sid) for sid in memory['about']])
conn.executemany("INSERT INTO sources(memory_id,kind,message_id,created_at) VALUES(?,'message',?,?)",
[(mid, messages[key], stamp) for key in memory['sources'] if key in messages])
# These are historical fixture states. Install them after their original
# evidence and from branches to roots, while each destination is live.
merged = {person['id']: person['merged_into'] for person in material['subjects'] if person['merged_into']}
while merged:
leaves = sorted(merged.keys() - set(merged.values()))
if not leaves:
raise ValueError('cyclic fixture merge')
for sid in leaves:
conn.execute('UPDATE subjects SET merged_into=?,merged_at=? WHERE id=?',
(merged.pop(sid), material['now'], sid))
return links, {mid: key for key, mid in messages.items()}


def run_case(case_id, material, configs, *, zone, role_name):
current = datetime.fromisoformat(material['now'])

def clock():
return current

with TemporaryDirectory(prefix='iris-people-eval-') as directory:
store = Store(Path(directory) / 'iris.db')
gateway = Gateway(configs, store, health=ModelHealth(store, configs, clock=clock), clock=clock)
try:
links, message_keys = seed(store, material, zone=zone, role_name=role_name)
maintenance = Maintenance(store, gateway=gateway, clock=clock)
run_id = maintenance.request()
with store.write() as conn:
conn.execute('UPDATE maintenance_runs SET phase=? WHERE id=?', (PHASES.index('goals'), run_id))
maintenance.run(run_id)
with store.read() as conn:
items = {row['object_id']: dict(row) for row in conn.execute(
"SELECT object_id,outcome,reason,details_json FROM maintenance_items WHERE run_id=? AND phase='people'",
(run_id,))}
merges = {row['link_id']: dict(row) for row in conn.execute(
"SELECT link_id,source_id,target_id,evidence_ids_json FROM subject_merges WHERE actor='maintenance'")}
calls = [dict(row) for row in conn.execute('''SELECT purpose,status_code,result_category,
timed_out,finish_reason,prompt_tokens,completion_tokens,duration_ms FROM model_calls''')]
foreign_key_errors = bool(conn.execute('PRAGMA foreign_key_check').fetchone())
results = {}
for key, link_id in links.items():
item, merge = items.get(link_id, {}), merges.get(link_id)
results[key] = {'target': merge['target_id'] if merge else None,
'outcome': item.get('outcome'), 'reason': item.get('reason'),
'evidence': [message_keys[mid] for mid in json.loads(merge['evidence_ids_json'])] if merge else []}
return {'id': case_id, 'links': results, 'calls': calls, 'foreign_key_errors': foreign_key_errors}
finally:
gateway.close()
store.close()


def score(corpus, results):
links = []
for case, result in zip(corpus['cases'], results, strict=True):
for expected in case['expected']['links']:
actual = result['links'][expected['key']]
should_merge = expected['expected'] == 'merge'
links.append({'case': case['id'], 'link': expected['key'], 'should_merge': should_merge, **actual,
'expected_target': expected['target'],
'correct': actual['target'] == (expected['target'] if should_merge else None)})
positives = sum(row['should_merge'] for row in links)
identified = sum(row['should_merge'] and row['correct'] for row in links)
wrong_merges = sum(row['target'] is not None and not row['correct'] for row in links)
failures = sum(row['outcome'] in (None, 'failed') for row in links)
calls = [call for result in results for call in result['calls']]
valid = not (failures or any(result['foreign_key_errors'] for result in results)
or any(call['timed_out'] or call['status_code'] == 429 for call in calls))
return {'valid': valid, 'cases': len(results), 'links': len(links), 'correct': sum(row['correct'] for row in links),
'expected_merges': positives, 'correct_merges': identified, 'wrong_merges': wrong_merges,
'identification': identified / positives if positives else None, 'model_calls': len(calls),
'failed_reviews': failures, 'meets_targets': bool(valid and positives and not wrong_merges and identified / positives >= .70),
'link_results': links}


def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--corpus', type=Path, default=ROOT / 'evals/people_merge_v1.json')
parser.add_argument('--out', type=Path, required=True)
parser.add_argument('--workers', type=int, choices=range(1, 7), default=1)
args = parser.parse_args()
corpus_bytes = args.corpus.read_bytes()
corpus = json.loads(corpus_bytes)
if corpus['format_version'] != 1 or not corpus['cases']:
parser.error('unsupported corpus format')
case_ids = [case['id'] for case in corpus['cases']]
if any(not re.fullmatch(r'[A-Za-z0-9_-]{1,48}', key) for key in case_ids) or len(case_ids) != len(set(case_ids)):
parser.error('case IDs must be unique and safe as filenames')
out = args.out.resolve()
if out.is_relative_to(ROOT):
parser.error('--out must be outside the repository')
configs = {'chat': load_test_models()['chat']}
out.mkdir(parents=True, exist_ok=False)
source_files = sorted(path for path in (ROOT / 'src/iris').rglob('*') if path.suffix in ('.py', '.sql', '.md'))
metadata = {'corpus': corpus['name'], 'corpus_sha256': hashlib.sha256(corpus_bytes).hexdigest(),
'source_sha256': hashlib.sha256(b''.join(
path.relative_to(ROOT).as_posix().encode() + path.read_bytes() for path in source_files)).hexdigest(),
'runner_sha256': hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),
'prompt_sha256': hashlib.sha256(PROMPT.encode()).hexdigest(),
'purpose': PURPOSE, 'model': configs['chat'].model, 'reasoning_effort': configs['chat'].reasoning_effort}
(out / 'metadata.json').write_text(json.dumps(metadata, indent=2) + '\n', encoding='utf-8')
results = []
with ThreadPoolExecutor(max_workers=args.workers) as pool:
jobs = [pool.submit(run_case, case['id'], {field: case[field] for field in PRODUCT_FIELDS}, configs,
zone=corpus['timezone'], role_name=corpus['role_name']) for case in corpus['cases']]
for job in jobs:
result = job.result()
results.append(result)
(out / (result['id'] + '.json')).write_text(json.dumps(result, indent=2) + '\n', encoding='utf-8')
print(result['id'], 'calls', len(result['calls']), flush=True)
report = {**metadata, **score(corpus, results)}
(out / 'report.json').write_text(json.dumps(report, ensure_ascii=False, indent=2) + '\n', encoding='utf-8')
print(json.dumps({key: value for key, value in report.items() if key != 'link_results'}, ensure_ascii=False), flush=True)
return 0 if report['meets_targets'] else 1


if __name__ == '__main__':
raise SystemExit(main())
13 changes: 13 additions & 0 deletions frontend/src/LifecycleSettings.tsx
Original file line number Diff line number Diff line change
Expand Up @@ -201,6 +201,19 @@ export function LifecycleSettings({
/>
<small>放弃后第几天重试,逗号分隔</small>
</label>
<label className="check-label">
<input
type="checkbox"
checked={draft.people_auto_merge_enabled ?? false}
onChange={(e) =>
change("people_auto_merge_enabled", e.target.checked)
}
/>
每日维护自动合并同一人
</label>
<p className="lifecycle-help">
默认关闭。开启后,维护会复核“可能是同一人”的联系。合并须有跨日期的独立证据,可以在人物页查看证据并拆分。
</p>
<p className="lifecycle-help">
维护按实际执行次数衰减,遗忘和置顶记忆不衰减。依据失效时每个依据只扣一次,置顶期间暂缓扣减。已放弃的批次按上面的日期(角色时区)各自动重试一次,每次使用完整尝试额度;停机错过的日期恢复后最多补一次。用完后缺口只作记录,仍可手动重新学习,手动重新学习不占自动重试次数。内容拒绝不自动重试。
</p>
Expand Down
Loading