noos 0.4.1

Reliability layer for Rust LLM agents: scope drift, cost circuit breaks, and procedural correction memory as event-driven Decisions.
Documentation
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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
//! Eval: Does Non-Cortical Processing Matter?
//!
//! Run: `cargo run --features candle --example eval_intervention`
//!
//! Tests whether Noos's subcortical modulation (Tầng 1 + 2) creates
//! measurable, meaningful differences in model behavior.
//!
//! For each scenario:
//! 1. CognitiveSession processes message → TurnResult (arousal, gain, budget, etc.)
//! 2. Model forward pass in 2 modes: baseline vs cognitive (Tầng 2 delta modulation)
//! 3. Metrics: KL divergence, entropy change, top-k overlap
//!
//! Neuroscience predictions:
//! - Phasic gain → entropy should DECREASE (focused)
//! - Tonic gain → entropy should INCREASE (exploratory)
//! - High arousal → delta gain > 1.0 (attend to current input)
//! - Depleted budget → delta gain < 1.0 (conservation)

#[cfg(feature = "candle")]
use noos::cognition::delta_modulation::compute_delta_modulation;
#[cfg(feature = "candle")]
use noos::inference::cognitive_model::CognitiveModel;
#[cfg(feature = "candle")]
use noos::inference::cognitive_gate::CognitiveGateConfig;
#[cfg(feature = "candle")]
use noos::inference::mamba::{CognitiveMambaModel, CognitiveMambaWithGate, HfTokenizer, MambaConfig};
#[cfg(feature = "candle")]
use noos::inference::model::LocalModel;
#[cfg(feature = "candle")]
use noos::inference::tokenizer::NoosTokenizer;
#[cfg(feature = "candle")]
use noos::math::softmax::softmax_f32;
#[cfg(feature = "candle")]
use noos::session::{CognitiveSession, TurnResult};
#[cfg(feature = "candle")]
use noos::types::world::GainMode;

fn main() {
    #[cfg(not(feature = "candle"))]
    eprintln!("Requires `candle` feature: cargo run --features candle --example eval_intervention");

    #[cfg(feature = "candle")]
    run();
}

#[cfg(feature = "candle")]
struct Scenario {
    name: &'static str,
    /// Messages to process (last message is the test message).
    messages: Vec<&'static str>,
    /// What we expect to see.
    prediction: &'static str,
}

#[cfg(feature = "candle")]
struct ScenarioResult {
    name: String,
    // Cognitive state from session
    arousal: f64,
    gain_mode: String,
    body_budget: f64,
    sensory_pe: f64,
    // Tầng 1 params
    temperature: f64,
    top_p: f64,
    freq_penalty: f64,
    // Tầng 2 params
    delta_gain: f64,
    modulated_layers: usize,
    // Metrics
    kl_divergence: f64,
    entropy_baseline: f64,
    entropy_tang2: f64,
    entropy_tang1_baseline: f64,
    entropy_tang1_tang2: f64,
    top10_overlap: usize,
}

#[cfg(feature = "candle")]
fn run() {
    println!("=== Noos Eval: Does Non-Cortical Processing Matter? ===\n");

    // ── Load model ──
    let model_id = "state-spaces/mamba-130m-hf";
    let config = MambaConfig::mamba_130m();

    println!("Loading tokenizer...");
    let tokenizer = match HfTokenizer::from_pretrained(model_id) {
        Ok(t) => t,
        Err(e) => { eprintln!("Tokenizer failed: {e}"); return; }
    };

    println!("Loading model (~500MB, first run downloads)...");
    let mut model = match CognitiveMambaModel::from_pretrained(model_id, config) {
        Ok(m) => {
            println!("  {} layers, ready.\n", m.num_layers());
            m
        }
        Err(e) => { eprintln!("Model failed: {e}"); return; }
    };

    let num_layers = model.num_layers();

    // ── Define scenarios ──
    let scenarios = vec![
        Scenario {
            name: "1. Routine greeting (control)",
            messages: vec!["Hello, how are you?"],
            prediction: "Minimal modulation — gain ~1.0, low arousal",
        },
        Scenario {
            name: "2. Emotional stress",
            messages: vec!["I'm really stressed about this deadline and everything is going wrong!!!"],
            prediction: "High arousal, negative → delta gain > 1.0 (focus on current threat)",
        },
        Scenario {
            name: "3. Creative exploration",
            messages: vec!["What if we could redesign cities completely from scratch, reimagining how people live and interact?"],
            prediction: "Exploratory → if tonic, entropy should increase (broader distribution)",
        },
        Scenario {
            name: "4. Technical question",
            messages: vec!["How does the TCP protocol handle packet loss and retransmission?"],
            prediction: "Focused → if phasic, entropy should decrease (narrower distribution)",
        },
        Scenario {
            name: "5. Vietnamese philosophical",
            messages: vec!["Khám phá ý nghĩa của cuộc sống và sự tồn tại con người"],
            prediction: "Novel + Vietnamese detection → high PE, possible NE burst",
        },
        Scenario {
            name: "6. Multi-turn stress (depleted budget)",
            messages: vec![
                "This is terrible, I lost all my work!",
                "And now the server is down too!",
                "Everything keeps getting worse and I don't know what to do!!!",
            ],
            prediction: "Body budget depleted → delta gain < 1.0 (conservation mode)",
        },
    ];

    let prompt = "The meaning of life is";
    let prompt_tokens = tokenizer.encode(prompt, false).unwrap_or_default();
    if prompt_tokens.is_empty() {
        eprintln!("Failed to tokenize prompt");
        return;
    }

    println!("Probe prompt: \"{prompt}\" ({} tokens)", prompt_tokens.len());
    println!("Comparing: baseline forward() vs cognitive forward_cognitive(delta)\n");
    println!("{}", "=".repeat(72));

    let mut results: Vec<ScenarioResult> = Vec::new();

    for scenario in &scenarios {
        // Fresh session for each scenario.
        let mut session = CognitiveSession::with_model_layers(num_layers);

        // Process all messages through CognitiveSession.
        let mut turn_opt: Option<TurnResult> = None;
        for (i, msg) in scenario.messages.iter().enumerate() {
            let t = session.process_message(msg);
            // Simulate response for multi-turn (so consolidate runs).
            if i < scenario.messages.len() - 1 {
                session.process_response("I understand, that's difficult.", 0.5);
            }
            turn_opt = Some(t);
        }
        let turn = turn_opt.expect("Scenario must have at least one message");

        // ── Baseline: forward() with no cognitive modulation ──
        model.reset_cache();
        let logits_baseline = match model.forward(&prompt_tokens, 0) {
            Ok(l) => l,
            Err(e) => { eprintln!("Baseline forward failed: {e}"); continue; }
        };

        // ── Tầng 2: forward_cognitive() with delta modulation ──
        model.reset_cache();
        let delta_mod = compute_delta_modulation(&turn.cognitive_state, num_layers);
        let result_tang2 = match model.forward_cognitive(&prompt_tokens, 0, &delta_mod) {
            Ok(r) => r,
            Err(e) => { eprintln!("Cognitive forward failed: {e}"); continue; }
        };
        let logits_tang2 = &result_tang2.logits;

        // ── Compute metrics ──

        // Raw logit comparison (Tầng 2 effect on model processing).
        let probs_baseline = softmax_f32(&logits_baseline);
        let probs_tang2 = softmax_f32(logits_tang2);

        let kl = kl_divergence(&probs_baseline, &probs_tang2);
        let ent_baseline = entropy(&probs_baseline);
        let ent_tang2 = entropy(&probs_tang2);
        let overlap = top_k_overlap(&logits_baseline, logits_tang2, 10);

        // Tầng 1 effect: apply temperature to raw logits → compare entropy.
        let temp = turn.sampling.temperature as f32;
        let scaled_baseline: Vec<f32> = logits_baseline.iter().map(|l| l / temp).collect();
        let scaled_tang2: Vec<f32> = logits_tang2.iter().map(|l| l / temp).collect();
        let probs_t1_baseline = softmax_f32(&scaled_baseline);
        let probs_t1_tang2 = softmax_f32(&scaled_tang2);
        let ent_t1_baseline = entropy(&probs_t1_baseline);
        let ent_t1_tang2 = entropy(&probs_t1_tang2);

        // ── Print ──
        println!("\n{}", scenario.name);
        println!("{}", "-".repeat(72));
        if scenario.messages.len() > 1 {
            for msg in &scenario.messages {
                println!("  msg: \"{}\"", truncate(msg, 60));
            }
        } else {
            println!("  msg: \"{}\"", truncate(scenario.messages[0], 60));
        }
        println!("  prediction: {}", scenario.prediction);
        println!();

        println!("  Cognitive state:");
        println!("    arousal={:.3}  valence={:?}  gain={:?}  budget={:.3}",
            turn.arousal, turn.valence, turn.gain_mode, turn.body_budget);
        println!("    sensory_pe={:.3}  pe_volatility={:.3}  gate={:?}",
            turn.sensory_pe, turn.cognitive_state.pe_volatility, turn.gate_type);
        println!();

        println!("  Tầng 1 (sampling): temp={:.3}  top_p={:.3}  freq_penalty={:.3}  presence_penalty={:.3}",
            turn.sampling.temperature, turn.sampling.top_p,
            turn.sampling.frequency_penalty, turn.sampling.presence_penalty);
        println!("  Tầng 2 (delta):    gain={:.4}  layers={}/{} (40-60% depth)",
            delta_mod.gain_factor,
            result_tang2.modulated_layers.len(), num_layers);
        println!();

        println!("  Metrics (baseline vs Tầng 2):");
        println!("    KL divergence:    {:.6}", kl);
        println!("    Entropy baseline: {:.4}", ent_baseline);
        println!("    Entropy Tầng 2:   {:.4} ({:+.1}%)",
            ent_tang2, pct_change(ent_baseline, ent_tang2));
        println!("    Top-10 overlap:   {}/10", overlap);
        println!();

        println!("  Tầng 1 temperature effect (temp={:.3}):", temp);
        println!("    Entropy baseline+temp: {:.4}", ent_t1_baseline);
        println!("    Entropy Tầng1+2:       {:.4} ({:+.1}% from baseline+temp)",
            ent_t1_tang2, pct_change(ent_t1_baseline, ent_t1_tang2));

        // Top-5 tokens comparison.
        println!();
        print_top_tokens("    Baseline", &logits_baseline, &tokenizer, 5);
        print_top_tokens("    Tầng 1+2", logits_tang2, &tokenizer, 5);

        results.push(ScenarioResult {
            name: scenario.name.to_string(),
            arousal: turn.arousal,
            gain_mode: format!("{:?}", turn.gain_mode),
            body_budget: turn.body_budget,
            sensory_pe: turn.sensory_pe,
            temperature: turn.sampling.temperature,
            top_p: turn.sampling.top_p,
            freq_penalty: turn.sampling.frequency_penalty,
            delta_gain: delta_mod.gain_factor,
            modulated_layers: result_tang2.modulated_layers.len(),
            kl_divergence: kl,
            entropy_baseline: ent_baseline,
            entropy_tang2: ent_tang2,
            entropy_tang1_baseline: ent_t1_baseline,
            entropy_tang1_tang2: ent_t1_tang2,
            top10_overlap: overlap,
        });
    }

    // ══════════════════════════════════════════════════════════════════
    // PART B: Manual cognitive states (prove Tầng 2 works independently)
    // ══════════════════════════════════════════════════════════════════
    println!("\n{}", "=".repeat(72));
    println!("PART B: Manual Cognitive States (bypass text processing)");
    println!("{}", "=".repeat(72));
    println!();
    println!("Testing Tầng 2 with hand-crafted states to isolate the");
    println!("delta modulation effect from CognitiveSession signal generation.\n");

    use noos::types::belief::AffectValence;
    use noos::types::intervention::CognitiveState;

    let manual_states: Vec<(&str, CognitiveState)> = vec![
        ("Neutral (control)", CognitiveState::default()),
        ("Phasic (focused)", CognitiveState {
            gain_mode: GainMode::Phasic,
            arousal: 0.3,
            certainty: 0.8,
            ..CognitiveState::default()
        }),
        ("Tonic (exploratory)", CognitiveState {
            gain_mode: GainMode::Tonic,
            arousal: 0.2,
            certainty: 0.3,
            ..CognitiveState::default()
        }),
        ("High arousal threat", CognitiveState {
            arousal: 0.9,
            valence: AffectValence::Negative,
            body_budget: 0.6,
            sensory_pe: 0.8,
            ..CognitiveState::default()
        }),
        ("Depleted budget", CognitiveState {
            body_budget: 0.2,
            arousal: 0.4,
            valence: AffectValence::Negative,
            ..CognitiveState::default()
        }),
        ("Volatile environment", CognitiveState {
            pe_volatility: 0.8,
            sensory_pe: 0.7,
            arousal: 0.5,
            ..CognitiveState::default()
        }),
    ];

    // Baseline logits (neutral state, already computed — recompute for clarity).
    model.reset_cache();
    let logits_neutral = model.forward(&prompt_tokens, 0)
        .expect("Baseline forward failed");
    let probs_neutral = softmax_f32(&logits_neutral);
    let ent_neutral = entropy(&probs_neutral);

    println!("{:<25} {:>7} {:>8} {:>8} {:>7} {:>7}",
        "State", "Gain", "KL div", "Ent %", "Top10", "Layers");
    println!("{}", "-".repeat(72));

    for (name, state) in &manual_states {
        model.reset_cache();
        let dm = compute_delta_modulation(state, num_layers);
        let result = model.forward_cognitive(&prompt_tokens, 0, &dm)
            .expect("Cognitive forward failed");

        let probs = softmax_f32(&result.logits);
        let kl = kl_divergence(&probs_neutral, &probs);
        let ent = entropy(&probs);
        let overlap = top_k_overlap(&logits_neutral, &result.logits, 10);

        println!("{:<25} {:>7.4} {:>8.6} {:>+7.1}% {:>5}/10 {:>3}/{}",
            name, dm.gain_factor, kl, pct_change(ent_neutral, ent),
            overlap, result.modulated_layers.len(), num_layers);
    }

    // ══════════════════════════════════════════════════════════════════
    // PART C: Closed-loop with CognitiveGate (thalamocortical feedback)
    // ══════════════════════════════════════════════════════════════════
    println!("\n{}", "=".repeat(72));
    println!("PART C: Closed-Loop with CognitiveGate (thalamocortical feedback)");
    println!("{}", "=".repeat(72));
    println!();
    println!("Loading CognitiveMambaWithGate (fresh gate, untrained)...");

    use noos::cognition::intervention::build_cognitive_state;
    use noos::types::intervention::ForwardResult;

    let gate_config = CognitiveGateConfig::from_mamba_config(&MambaConfig::mamba_130m());
    let gate_model = CognitiveMambaWithGate::from_pretrained_with_gate(
        model_id, MambaConfig::mamba_130m(), gate_config,
    );

    match gate_model {
        Ok((mut gmodel, _varmap)) => {
            let gate_layers = gmodel.num_layers();
            println!("  Gate model loaded. {} layers, gate at mid-depth.\n", gate_layers);

            let test_msg = "I'm really stressed about this deadline and everything is going wrong!!!";
            let gen_tokens = 15;

            println!("Comparing open-loop (base model) vs closed-loop (gate model + feedback)");
            println!("Message: \"{}\"\n", truncate(test_msg, 60));

            // ── Open-loop: base model, fixed cognitive state ──
            let mut session_open = CognitiveSession::with_model_layers(gate_layers);
            let turn_open = session_open.process_message(test_msg);
            let state_open = turn_open.cognitive_state.clone();

            model.reset_cache();
            let _ = model.forward(&prompt_tokens, 0);
            let mut open_alphas: Vec<f64> = Vec::new();
            let mut open_arousals: Vec<f64> = Vec::new();

            for _ in 0..gen_tokens {
                let dm = compute_delta_modulation(&state_open, num_layers);
                let result = model.forward_cognitive(&[0], 0, &dm)
                    .unwrap_or_else(|_| ForwardResult::from_logits(vec![0.0; 50280]));
                open_alphas.push(result.gate_alpha.unwrap_or(0.0));
                open_arousals.push(state_open.arousal);
            }

            // ── Closed-loop: gate model, feedback per token ──
            let mut session_closed = CognitiveSession::with_model_layers(gate_layers);
            let _turn_closed = session_closed.process_message(test_msg);

            gmodel.reset_cache();
            let _ = <CognitiveMambaWithGate as LocalModel>::forward(&mut gmodel, &prompt_tokens, 0);
            let mut closed_alphas: Vec<f64> = Vec::new();
            let mut closed_arousals: Vec<f64> = Vec::new();
            let mut closed_deltas: Vec<f64> = Vec::new();

            for _ in 0..gen_tokens {
                let current_state = build_cognitive_state(
                    session_closed.world_model(),
                    session_closed.world_model().learned.gain_mode,
                );
                let dm = compute_delta_modulation(&current_state, gate_layers);
                let result = gmodel.forward_cognitive(&[0], 0, &dm)
                    .unwrap_or_else(|_| ForwardResult::from_logits(vec![0.0; 50280]));

                let ga = result.gate_alpha.unwrap_or(0.0);
                let gd = result.gate_delta_gain.unwrap_or(1.0);
                closed_alphas.push(ga);
                closed_arousals.push(current_state.arousal);
                closed_deltas.push(gd);

                session_closed.inject_gate_feedback(ga, gd);
            }

            // ── Print comparison ──
            println!("{:<6} {:>10} {:>10} {:>10} {:>10} {:>10}",
                "Token", "Base α", "Gate α", "Base arou", "Gate arou", "Gate δ");
            println!("{}", "-".repeat(62));
            for i in 0..gen_tokens {
                println!("{:<6} {:>10.4} {:>10.4} {:>10.4} {:>10.4} {:>10.4}",
                    i + 1,
                    open_alphas[i], closed_alphas[i],
                    open_arousals[i], closed_arousals[i],
                    closed_deltas[i]);
            }

            let open_arousal_change = open_arousals.last().unwrap_or(&0.0)
                - open_arousals.first().unwrap_or(&0.0);
            let closed_arousal_change = closed_arousals.last().unwrap_or(&0.0)
                - closed_arousals.first().unwrap_or(&0.0);

            println!();
            println!("  Open-loop arousal:   {:+.4} (fixed)", open_arousal_change);
            println!("  Closed-loop arousal: {:+.4} (gate feedback)", closed_arousal_change);

            let avg_gate_alpha: f64 = closed_alphas.iter().sum::<f64>() / gen_tokens as f64;
            let avg_gate_delta: f64 = closed_deltas.iter().sum::<f64>() / gen_tokens as f64;
            println!("  Avg gate α: {:.4}, avg gate δ: {:.4}", avg_gate_alpha, avg_gate_delta);

            if avg_gate_alpha > 0.01 {
                println!("  GATE ACTIVE: thalamocortical feedback loop producing real signals.");
            }
            if (closed_arousal_change - open_arousal_change).abs() > 1e-6 {
                println!("  FEEDBACK LOOP: arousal evolves differently with gate feedback.");
            }
        }
        Err(e) => {
            println!("  Failed to load gate model: {e}");
            println!("  Skipping Part C (gate model required).");
        }
    }

    // ── Summary ──
    println!("\n{}", "=".repeat(72));
    println!("SUMMARY");
    println!("{}", "=".repeat(72));
    println!();
    println!("{:<35} {:>6} {:>6} {:>8} {:>8} {:>7}",
        "Scenario", "Arousal", "Gain", "KL div", "Ent %", "Top10");
    println!("{}", "-".repeat(72));

    for r in &results {
        println!("{:<35} {:>6.3} {:>6} {:>8.6} {:>+7.1}% {:>5}/10",
            truncate(&r.name, 35),
            r.arousal,
            &r.gain_mode[..r.gain_mode.len().min(6)],
            r.kl_divergence,
            pct_change(r.entropy_baseline, r.entropy_tang2),
            r.top10_overlap,
        );
    }

    // ── Verdict ──
    println!("\n{}", "-".repeat(72));
    println!("VERDICT\n");

    let any_kl = results.iter().any(|r| r.kl_divergence > 1e-6);
    let routine_minimal = results.first()
        .map(|r| r.kl_divergence < results.iter().skip(1)
            .map(|r2| r2.kl_divergence)
            .fold(f64::MAX, f64::min))
        .unwrap_or(false);

    if any_kl {
        println!("  Tầng 2 (delta modulation) produces MEASURABLY DIFFERENT logit distributions.");
        if routine_minimal {
            println!("  Control (routine) shows LESS modulation than aroused scenarios.");
            println!("  Non-cortical processing IS context-dependent — not random noise.");
        }
    } else {
        println!("  No measurable difference. Investigate: gain range, layer targeting.");
    }

    // Check neuroscience predictions.
    let stress = results.iter().find(|r| r.name.contains("Emotional"));
    let routine = results.first();
    if let (Some(s), Some(r)) = (stress, routine) {
        if s.delta_gain > r.delta_gain {
            println!("  Stress → higher delta gain than routine (attention to threat).");
        }
    }

    let depleted = results.iter().find(|r| r.name.contains("Multi-turn"));
    if let Some(d) = depleted {
        if d.body_budget < 0.95 {
            println!("  Multi-turn stress depletes body budget to {:.3} (allostasis active).", d.body_budget);
        }
        if d.delta_gain < 1.0 {
            println!("  Depleted budget → delta gain < 1.0 (conservation mode confirmed).");
        }
    }

    println!();
}

// ── Metrics ──────────────────────────────────────────────────────────────

#[cfg(feature = "candle")]
fn kl_divergence(p: &[f32], q: &[f32]) -> f64 {
    let mut kl = 0.0f64;
    for (&pi, &qi) in p.iter().zip(q.iter()) {
        let pi = pi as f64;
        let qi = qi as f64;
        if pi > 1e-10 && qi > 1e-10 {
            kl += pi * (pi / qi).ln();
        }
    }
    kl.max(0.0)
}

#[cfg(feature = "candle")]
fn entropy(probs: &[f32]) -> f64 {
    let mut h = 0.0f64;
    for &p in probs {
        let p = p as f64;
        if p > 1e-10 {
            h -= p * p.ln();
        }
    }
    h
}

#[cfg(feature = "candle")]
fn top_k_overlap(logits_a: &[f32], logits_b: &[f32], k: usize) -> usize {
    let top_a = top_k_indices(logits_a, k);
    let top_b = top_k_indices(logits_b, k);
    top_a.iter().filter(|idx| top_b.contains(idx)).count()
}

#[cfg(feature = "candle")]
fn top_k_indices(logits: &[f32], k: usize) -> Vec<usize> {
    let mut indexed: Vec<(usize, f32)> = logits.iter().copied().enumerate().collect();
    indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
    indexed.iter().take(k).map(|(i, _)| *i).collect()
}

#[cfg(feature = "candle")]
fn pct_change(from: f64, to: f64) -> f64 {
    if from.abs() < 1e-10 { 0.0 } else { (to - from) / from * 100.0 }
}

// ── Display helpers ─────────────────────────────────────────────────────

#[cfg(feature = "candle")]
fn truncate(s: &str, max: usize) -> String {
    if s.len() <= max { s.to_string() }
    else { format!("{}...", &s[..max.saturating_sub(3)]) }
}

#[cfg(feature = "candle")]
fn print_top_tokens(label: &str, logits: &[f32], tokenizer: &HfTokenizer, k: usize) {
    let indices = top_k_indices(logits, k);
    print!("{}: ", label);
    for (i, &idx) in indices.iter().enumerate() {
        let tok = tokenizer.decode_token(idx as u32)
            .unwrap_or_else(|_| format!("[{}]", idx));
        let sep = if i < k - 1 { " | " } else { "" };
        print!("{:.1}:{}{sep}", logits[idx], tok.trim());
    }
    println!();
}