research: P3 记忆层 + 冲突词隔离（泄漏率 0）+ E5 第三条路径排除 - #338
Open
oratis wants to merge 1 commit into
Open
Conversation
一、P3 记忆层(非参数第一阶段,memory.py)
- (uid, word) 键空间硬分区:没有任何 API 接受跨 uid 检索——跨用户命中
在数据结构层不可表达(不是"过滤掉",是"构不出那个键")
- 三态门是 L2 唯一入口;未毕业条目超龄 expire = 写入前可逆性
(比任何 machine unlearning 都便宜,KNOWLEDGE §5.3 卖点)
- test_memory.py 14 项不变量全过(纯 CPU)
二、冲突词隔离测试(同词两老师教互斥语义,四条件对照)
- ★ 配对方向判别 8/8 = 1.000,冲突词泄漏率 0.000(1.5B 与 3B)
对标 PersistBench cross-domain 53%
- 平均分离度 Δp_yes +0.415/+0.362;键空间不相交断言 True
- 🔴 必报诊断:绝对阈值 acc 仅 0.688 但那不是泄漏——失败全落在
「期望 Yes」侧,无定义时 p(yes)=0.183/0.000(强 No 偏置),同一 probe
下本 uid 与对方 uid 的 p(yes) 每次都分得开 ⟹ 记忆分区完全决定答案
方向。与 P0「门是排序器、绝对阈值另需校准」是同一课
- LEAK 条件 acc 与 ISO 恰好互补 ⟹ 泄漏一旦发生必然可见(测法有效)
三、P0l E5 反事实否定条件化 —— 失败,且自埋诊断项否掉我方假设
- 主候选 neg("means M, not M′")全局 margin 反降 +16.71→+8.00
- G5(E5 高危)五种措辞 margin 全为负
- ★ 诊断判决:emph(纯强调,无对立信息)改善 +18.57 > neg 的 +16.13
⟹ 改善来自「加词」nuisance,非反事实机制
- E5 三条路径全部排除(P0i 探针 / P2-D 言语化对比 / P0l 反事实否定)
- 🟡 旁支 rule(iff 措辞)全局 margin +65%,但系事后从 5 措辞挑最大、
判对仅 16→17 ⟹ 标 💭 待预注册复现,不得写入论文
文档:REPORT §3.5/§3.6 新增、计数同步(13 轮/15 组实证/阶段更新)、
§5 站得住的 +2 条、§6 推进到 P4;DESIGN 两文档同步实测值与已排除路径
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
1. P3 记忆层：隔离是架构保证而非约定
memory.py三条设计承诺，全部有测试（test_memory.py14/14 通过，纯 CPU）：(uid, word)，没有任何 API 接受跨 uid 检索——跨用户命中在数据结构层不可表达2. 冲突词隔离测试：泄漏率 0
🔴 必报诊断（否则会被误读成 31% 泄漏）：绝对阈值 acc 仅 0.688，但失败全部落在「期望 Yes」侧，且无定义时 p(yes)=0.183/0.000（强 No 偏置）；同一 probe 下本 uid 与对方 uid 的 p(yes) 每次都分得开 ⟹ 记忆分区完全决定答案方向，失分来自响应偏置。与 P0「门是排序器、绝对阈值另需校准」是同一课。
3. P0l — E5 第三条解法路径排除
反事实否定条件化（
"{w} means M, not M′"）失败：全局 margin 反降（+16.71 → +8.00），G5 五措辞全负。自埋诊断项给出判决：
emph（纯强调、无对立信息）改善 +18.57 >neg的 +16.13 ⟹ 那点改善来自加词 nuisance，不是反事实机制。E5 三条路径已全部排除（P0i 忠实性探针 / P2-D 言语化对比 / P0l 反事实否定）。
🟡 旁支
rule（iff 措辞）全局 margin +65%，但系事后从 5 措辞挑最大、判对仅 16→17 ⟹ 标 💭 待预注册复现，不得写入论文。🤖 Generated with Claude Code