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
224
225
226
227
228
//! Split out of `core/mod.rs` by AGENTS.md rule 19 (a MOVE: no behaviour, no public surface, no
//! renamed method).
use super::*;
impl CoreEngine {
// ── Smoothed fitness ────────────────────────────────────────────────────
/// Per-net fitness over the live population, in `live_hashes()` order — shared input for
/// ranking and `RaceSnapshot` (the window's family CENTRE, uncharged).
pub(crate) fn smoothed_fitness_values(&self) -> Vec<f32> {
// Only nets with a non-empty rolling buffer get a value — a caught-up child
// qualifies at once, and phantom 0.0 values never drag the mean down.
self.state
.live_hashes()
.into_iter()
.filter_map(|h| self.ranking_score_of(&h))
.collect()
}
/// The net's FITNESS value: the window score (the family CENTRE over `smoothing_window` step fitnesses).
pub fn smoothed_fitness_of(&self, hash: &str) -> Option<f32> {
self.race_reading_of(hash).map(|s| s.score())
}
/// One informative metric's smoothed centre for a net, read under the SAME mean family and
/// policy-scoped rule the fitness centre uses. `None` when the metric has no reading yet.
pub(crate) fn metric_center_of(&self, hash: &str, index: usize) -> Option<f32> {
let buf = self.rolling_metrics.get(hash)?.get(index)?;
let take = self.policy_scoped_take(hash, buf.len());
rolling_stats_tail(buf, take, self.config.mean_type).map(|s| s.score())
}
/// The net's RANKING value: the fitness window centre plus every `selection_metrics` term (`weight × that metric's smoothed centre`).
pub fn ranking_score_of(&self, hash: &str) -> Option<f32> {
let mut score = self.smoothed_fitness_of(hash)?;
for &(index, weight) in &self.selection_terms {
// A metric with no reading yet contributes nothing — the same contract the fitness
// side applies to a net that has not been measured (`0.0`).
if let Some(center) = self.metric_center_of(hash, index) {
score += weight * center;
}
}
Some(score)
}
/// The net's RACE READING with its own error bar: the smoothed mean plus how many
/// step readings are behind it and how far apart they were ([`WindowStats`]).
pub(crate) fn race_reading_of(&self, hash: &str) -> Option<WindowStats> {
self.rolling_fitness.get(hash).and_then(|buf| {
let take = self.policy_scoped_take(hash, buf.len());
rolling_stats_tail(buf, take, self.config.mean_type)
})
}
/// Readings taken on the CURRENT weights: the epoch count, or the whole buffer while no
/// update has landed yet (every reading is still on the current weights).
fn policy_scoped_take(&self, hash: &str, len: usize) -> usize {
if !self.config.policy_scoped_fitness {
return len;
}
match self.rolling_epoch.get(hash).copied() {
Some(n) if n > 0 => n.min(len),
_ => len,
}
}
/// A live net's parameter count (`NetMeta::params`) — the ranking
/// tie-break key.
pub(crate) fn net_params(&self, hash: &str) -> Option<i64> {
self.state.net(hash).and_then(|s| s.meta.params)
}
/// THE ranking primitive: live nets paired with smoothed fitness, ordered
/// best-first. `require_measured` is the eligibility contract (true = ranked_live).
pub(crate) fn ranked_live_impl(&self, require_measured: bool) -> Vec<(String, f32)> {
let direction = self.fitness.direction();
// `(hash, score, params)`: the size rides along so the comparator below is a pure function
// of the tuple — ONE map lookup per net instead of two per comparison, and no borrow
let mut ranked: Vec<(String, f32, i64)> = self
.state
.live_hashes()
.into_iter()
.filter_map(|h| {
let measured = self.race_reading_of(&h).filter(|s| s.n > 0);
match (measured, require_measured) {
(Some(reading), _) => Some((
h.clone(),
self.ranking_score_of(&h).unwrap_or_else(|| reading.score()),
)),
(None, true) => None,
// Never measured: the parent-selection contract counts it as 0.0.
(None, false) => Some((h, 0.0)),
}
})
.map(|(h, score)| {
let params = self.net_params(&h).unwrap_or(i64::MAX);
(h, score, params)
})
.collect();
// A TOTAL order, so no card's numbering depends on hash-map or insertion order:
// risk-adjusted score first, then the SMALLER net (equal score ⇒ prefer the cheaper one
ranked.sort_by(|a, b| {
direction
.cmp(b.1, a.1)
.then_with(|| a.2.cmp(&b.2))
.then_with(|| a.0.cmp(&b.0))
});
ranked.into_iter().map(|(h, score, _)| (h, score)).collect()
}
/// Live nets that have a measured (non-empty) rolling buffer, best-first.
pub(crate) fn ranked_live(&self) -> Vec<(String, f32)> {
self.ranked_live_impl(true)
}
/// Every live net, best-first; a net with no measurement yet counts
/// 0.0 — the parent-selection contract.
pub(crate) fn ranked_live_all(&self) -> Vec<(String, f32)> {
self.ranked_live_impl(false)
}
/// Whether this clock prints the `lottery │ who` line: the `Summ` reporting shape
/// lottery actually ARMED (`lottery_frame` returned a field).
pub(crate) fn lottery_line_on(&self, armed: bool) -> bool {
armed && self.log_level == crate::engine::config::LogLevel::Summ
}
/// Split one clock's nets into `(trained, withheld)` from the LEDGER snapshot taken BEFORE the clock ran: a net whose idle clocks grew got no ticket, and under…
pub(crate) fn lottery_clock_report(
&self,
group: &[String],
ledger_before: &[(usize, usize)],
) -> (Vec<String>, Vec<String>) {
let mut trained: Vec<String> = Vec::new();
let mut withheld: Vec<String> = Vec::new();
for (i, h) in group.iter().enumerate() {
// `(idle_clocks, frozen_clocks)` as they stood BEFORE this clock's group loop.
let Some(&(idle_before, frozen_before)) = ledger_before.get(i) else {
continue;
};
let Some(state) = self.state.net(h) else {
continue;
};
// TWO ledgers answer "did the coin withhold this clock": the idle one under the default contract (nothing ran), and the act-only one under learn-only (the net ran and was…
let denied = state.idle_clocks() > idle_before
|| (self.config.train_merit_learn_only && state.frozen_clocks() > frozen_before);
if denied {
withheld.push(h[..8.min(h.len())].to_string());
} else {
trained.push(h[..8.min(h.len())].to_string());
}
}
(trained, withheld)
}
/// Group nets whose IDLE clock grew since the ledger snapshot — the clocks that did not run at all.
pub(crate) fn idled_since(&self, group: &[String], ledger_before: &[(usize, usize)]) -> usize {
group
.iter()
.enumerate()
.filter(|(i, h)| {
ledger_before.get(*i).is_some_and(|&(idle_before, _)| {
self.state
.net(h)
.is_some_and(|s| s.idle_clocks() > idle_before)
})
})
.count()
}
/// Every count the per-step rollup needs, from ONE walk of the live population:
/// `newborn = live − group.len()` (a net born this clock sits the clock out).
pub(crate) fn step_vitals(
&self,
group: &[String],
lottery_skips: usize,
idled_skips: usize,
) -> StepVitals {
let live = self.state.live_count();
let would_train = group.len();
StepVitals {
// The one honest count in BOTH contracts: subtract the clocks that actually IDLED.
stepped: group.len().saturating_sub(idled_skips.min(lottery_skips)),
trained: would_train.saturating_sub(lottery_skips),
// Clamped: a stray idle delta can never make the denied split exceed the denied total.
idled_skips: idled_skips.min(lottery_skips),
learn_only: self.config.train_merit_learn_only,
live,
would_train,
lottery_skips,
newborn: live.saturating_sub(group.len()),
}
}
// ── Selection helpers ───────────────────────────────────────────────────
/// The live net with the best smoothed fitness (direction-aware). Test-only
/// since the random-topology fallback replaced the clone-fittest path.
#[cfg(test)]
pub(crate) fn fittest_net_hash(
&self,
) -> std::result::Result<String, crate::utils::error::EngineError> {
match self.ranked_live_all().into_iter().next() {
Some((h, _)) => Ok(h),
None => Err(crate::utils::error::EngineError::InvalidOptions(
"race: no live nets to select fittest from".into(),
)),
}
}
/// The worst live net hashes by smoothed fitness (opposite of fittest).
/// Uses the live smoothed fitness; the caller filters out probation.
pub(crate) fn worst_nets_by_smoothed_fitness(&self, count: usize) -> Result<Vec<String>> {
// `ranked_live()` is best-first, so the worst is its reverse.
Ok(self
.ranked_live()
.into_iter()
.rev()
.map(|(h, _)| h)
.take(count)
.collect())
}
// ── Shared batch materialization ────────────────────────────────────────
// ── Instrumentals ───────────────────────────────────────────────────────
}