1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
// Copyright 2024-2026 WritersLogic Contributors
// SPDX-License-Identifier: AGPL-3.0-or-later
//! Counting holographic membership store.
//!
//! A variant of [`super::bloom_memory::HolographicBloomMemory`] that keeps a
//! per-index HIT COUNT instead of a one-bit OR union. The binary union saturates
//! to all-ones as items accumulate, at which point non-members become
//! indistinguishable from members and discrimination collapses to chance. The
//! count preserves the signal the union throws away -- each member adds +1 to each
//! of its k active indices -- and a Poisson z-score readout recovers that shift far
//! past the binary wall.
//!
//! Measured (D=16384, k=64, 20 seeds, member-vs-non-member AUC): the binary
//! readout walls at AUC<0.95 by ~1000 items and is at chance by ~2000; the counting
//! readout holds AUC>0.95 to ~4000 and degrades gracefully (0.84 at 8000). A ~4x
//! usable-capacity gain at matched dimension. See `docs/PREREGISTRATION-binding-
//! readout.md` §19 and `src/bin/bloom-wall.rs`.
//!
//! Cost: one `u32` per index rather than one bit -- this trades a Bloom filter's
//! bit-compactness for capacity. It remains a deterministic integer fold of its
//! inserts, so the store is an exact, replay-verifiable function of what was added.
use crate::core::entangled::EntangledHVec;
/// A z-score threshold of 3.0 bounds the per-query non-member false-positive rate
/// to roughly 0.1% under the Poisson null (the summed statistic is ~N(0,1) for a
/// non-member), independent of load.
pub const DEFAULT_Z_THRESHOLD: f64 = 3.0;
pub struct CountingBloomMemory {
dim: usize,
counts: Vec<u32>,
/// Σ over inserted items of their active-index count -- the Poisson null rate
/// is λ = active_insertions / dim.
active_insertions: u64,
item_count: usize,
}
impl CountingBloomMemory {
pub fn new(dim: usize) -> Self {
Self {
dim,
counts: vec![0; dim],
active_insertions: 0,
item_count: 0,
}
}
/// Add one item, incrementing the count at each of its active indices.
pub fn insert(&mut self, item: &EntangledHVec) {
for &i in item.indices() {
// Indices are < dim by construction; the bounds check guards against a
// mismatched-dimension item rather than trusting the caller.
if let Some(c) = self.counts.get_mut(i as usize) {
*c = c.saturating_add(1);
self.active_insertions += 1;
}
}
self.item_count += 1;
}
pub fn insert_batch(&mut self, items: &[EntangledHVec]) {
for item in items {
self.insert(item);
}
}
/// Poisson z-score membership statistic.
///
/// Under the non-member null each queried index has count ~ Poisson(λ), so the
/// summed excess `Σ(count_i − λ)` over the query's k indices is ~N(0, kλ). A
/// member shifted every one of its indices up by +1, moving the sum by +k. The
/// returned standardized value `Σ(count_i − λ) / √(kλ)` is therefore ~0 for a
/// non-member and ~√(k/λ) for a member. (Ranking is invariant to the √(kλ)
/// scale at fixed k; the normalization is what makes the value an interpretable
/// z-score for thresholding.)
pub fn membership_score(&self, item: &EntangledHVec) -> f64 {
let ix = item.indices();
if ix.is_empty() {
return 0.0;
}
let lambda = self.active_insertions as f64 / self.dim.max(1) as f64;
let excess: f64 = ix
.iter()
.map(|&i| self.counts.get(i as usize).copied().unwrap_or(0) as f64 - lambda)
.sum();
let std = (ix.len() as f64 * lambda).sqrt();
if std <= f64::EPSILON {
// Empty or near-empty store: any present index is unambiguous signal.
return if excess > 0.0 { f64::INFINITY } else { 0.0 };
}
excess / std
}
/// Whether `item` is a member at the given z-score threshold.
pub fn contains(&self, item: &EntangledHVec, z_threshold: f64) -> bool {
self.membership_score(item) >= z_threshold
}
/// Candidate indices whose membership z-score meets `z_threshold`, with scores.
pub fn query_candidates(
&self,
candidates: &[EntangledHVec],
z_threshold: f64,
) -> Vec<(usize, f64)> {
candidates
.iter()
.enumerate()
.map(|(i, item)| (i, self.membership_score(item)))
.filter(|&(_, z)| z >= z_threshold)
.collect()
}
pub fn item_count(&self) -> usize {
self.item_count
}
pub fn dim(&self) -> usize {
self.dim
}
/// Poisson null rate λ = Σk / dim (mean count per index).
pub fn mean_count(&self) -> f64 {
self.active_insertions as f64 / self.dim.max(1) as f64
}
}
#[cfg(test)]
mod tests {
use super::*;
fn store(dim: usize, n: u64) -> CountingBloomMemory {
let mut mem = CountingBloomMemory::new(dim);
let items: Vec<EntangledHVec> = (0..n)
.map(|i| EntangledHVec::new_deterministic(dim, i * 100))
.collect();
mem.insert_batch(&items);
mem
}
fn mean(v: &[f64]) -> f64 {
v.iter().sum::<f64>() / v.len() as f64
}
#[test]
fn single_member_scores_high_nonmember_near_zero() {
let dim = 16384;
let mut mem = CountingBloomMemory::new(dim);
let item = EntangledHVec::new_deterministic(dim, 42);
mem.insert(&item);
let member = mem.membership_score(&item);
assert!(member > 20.0, "lone member must stand out, got {member:.2}");
// At near-empty load the null variance √(kλ) is tiny, so an individual
// non-member sharing an index or two is noisy; the null MEAN over a sample
// still sits near zero, and the member dwarfs any individual non-member.
let nms: Vec<f64> = (0..50)
.map(|j| mem.membership_score(&EntangledHVec::new_deterministic(dim, 900_000 + j)))
.collect();
assert!(
mean(&nms).abs() < 1.0,
"non-member mean should sit near the null"
);
let nm_max = nms.iter().cloned().fold(f64::NEG_INFINITY, f64::max);
assert!(
member > 5.0 * nm_max.max(1.0),
"member must dwarf non-members"
);
}
#[test]
fn empty_query_is_zero() {
let mem = store(16384, 100);
let empty = EntangledHVec::from_indices(Vec::new(), 16384);
assert_eq!(mem.membership_score(&empty), 0.0);
}
#[test]
fn discriminates_at_1000_items() {
// Where the binary readout is already collapsing (AUC~0.87), counting is
// near-perfect. Assert a comfortable, theory-predicted margin.
let dim = 16384;
let mem = store(dim, 1000);
let members: Vec<f64> = (0..1000)
.map(|i| mem.membership_score(&EntangledHVec::new_deterministic(dim, i * 100)))
.collect();
let nonmembers: Vec<f64> = (0..200)
.map(|j| mem.membership_score(&EntangledHVec::new_deterministic(dim, 5_000_000 + j)))
.collect();
let m = mean(&members);
let nm = mean(&nonmembers);
assert!(m > 3.0, "member mean z should clear 3, got {m:.2}");
assert!(
nm.abs() < 1.0,
"non-member mean z should sit near 0, got {nm:.2}"
);
let fpr = nonmembers.iter().filter(|&&z| z >= 3.0).count() as f64 / nonmembers.len() as f64;
assert!(
fpr < 0.05,
"non-member FPR@z>3 should be small, got {fpr:.3}"
);
}
#[test]
fn still_discriminates_where_binary_is_dead() {
// At 2000 items the binary OR-union readout is at chance (AUC~0.51);
// counting still separates the means by a clear margin (~4x capacity).
let dim = 16384;
let mem = store(dim, 2000);
let members: Vec<f64> = (0..400)
.map(|i| mem.membership_score(&EntangledHVec::new_deterministic(dim, i * 100)))
.collect();
let nonmembers: Vec<f64> = (0..400)
.map(|j| mem.membership_score(&EntangledHVec::new_deterministic(dim, 6_000_000 + j)))
.collect();
assert!(
mean(&members) - mean(&nonmembers) > 2.0,
"counting must still separate at 2000 items: members {:.2} vs non {:.2}",
mean(&members),
mean(&nonmembers)
);
}
}