1use std::path::Path;
29
30use firstpass_core::conformal::{self, ConformalResult};
31use firstpass_core::eprocess;
32use firstpass_core::ltt::{self, LttResult};
33use firstpass_core::{Attempt, DeferredVerdict, GateResult, Score, Trace, Verdict};
34
35use crate::store::{self, StoreError};
36
37#[derive(Debug, Clone)]
39pub struct CalibrationReport {
40 pub n_pairs: usize,
43 pub conformal: ConformalResult,
45 pub empirical_served_failure: f64,
49 pub n_served: usize,
51}
52
53impl CalibrationReport {
54 #[must_use]
56 pub fn render(&self) -> String {
57 format!(
58 "pairs: {n_pairs} ({n_served} served at threshold)\n\
59 threshold: {threshold:.4}\n\
60 feasible: {feasible}\n\
61 target alpha: {alpha:.4} (delta {delta:.4})\n\
62 calibration risk: {calib_risk:.4}\n\
63 empirical served-failure: {empirical:.4}\n",
64 n_pairs = self.n_pairs,
65 n_served = self.n_served,
66 threshold = self.conformal.threshold,
67 feasible = self.conformal.feasible,
68 alpha = self.conformal.alpha,
69 delta = self.conformal.delta,
70 calib_risk = self.conformal.calib_risk,
71 empirical = self.empirical_served_failure,
72 )
73 }
74}
75
76#[must_use]
80pub fn calibrate_pairs(
81 pairs: &[(f64, bool)],
82 alpha: f64,
83 delta: f64,
84 min_n: usize,
85) -> CalibrationReport {
86 let result = conformal::calibrate(pairs, alpha, delta, min_n);
87 let (empirical_served_failure, n_served) =
88 conformal::served_failure_rate(pairs, result.threshold);
89 CalibrationReport {
90 n_pairs: pairs.len(),
91 conformal: result,
92 empirical_served_failure,
93 n_served,
94 }
95}
96
97pub(crate) fn gate_score(gates: &[GateResult], verdict: Verdict) -> f64 {
106 let numeric: Vec<f64> = gates
107 .iter()
108 .filter_map(|g| g.score.map(Score::value))
109 .collect();
110 if numeric.is_empty() {
111 f64::from(verdict == Verdict::Pass)
112 } else {
113 numeric.iter().sum::<f64>() / numeric.len() as f64
114 }
115}
116
117fn attempt_score(attempt: &Attempt) -> f64 {
119 gate_score(&attempt.gates, attempt.verdict)
120}
121
122fn trace_pair(trace: &Trace, deferred: &[DeferredVerdict]) -> Option<(f64, bool)> {
126 let last = deferred.last()?;
127 let served_rung = trace.final_.served_rung?;
128 let attempt = trace.attempts.iter().find(|a| a.rung == served_rung)?;
129 Some((attempt_score(attempt), last.verdict == Verdict::Pass))
130}
131
132#[derive(Debug, Clone)]
134pub struct LttReport {
135 pub n_pairs: usize,
137 pub ltt: LttResult,
139}
140
141impl LttReport {
142 #[must_use]
145 pub fn render(&self) -> String {
146 let far = match self.ltt.false_accept_rate {
147 Some(r) => format!("{r:.4}"),
148 None => "N/A (no incorrect items in calibration set)".to_owned(),
149 };
150 format!(
151 "method: ltt\n\
152 pairs: {n_pairs} ({n_served} served at threshold)\n\
153 threshold: {threshold:.4}\n\
154 feasible: {feasible}\n\
155 target alpha: {alpha:.4} (delta {delta:.4})\n\
156 empirical risk: {risk:.4}\n\
157 false-accept rate: {far} (P(score >= lambda | incorrect); verifier ROC point)\n",
158 n_pairs = self.n_pairs,
159 n_served = self.ltt.n_served,
160 threshold = self.ltt.threshold,
161 feasible = self.ltt.feasible,
162 alpha = self.ltt.alpha,
163 delta = self.ltt.delta,
164 risk = self.ltt.empirical_risk,
165 )
166 }
167}
168
169#[must_use]
172pub fn calibrate_pairs_ltt(
173 pairs: &[(f64, bool)],
174 alpha: f64,
175 delta: f64,
176 min_n: usize,
177) -> LttReport {
178 LttReport {
179 n_pairs: pairs.len(),
180 ltt: ltt::calibrate(pairs, alpha, delta, min_n),
181 }
182}
183
184#[derive(Debug, Clone)]
193pub struct EProcessReport {
194 pub n_pairs: usize,
196 pub certification: Option<firstpass_core::eprocess::Certification>,
198 pub alpha: f64,
200 pub delta: f64,
202 pub crossing_level: f64,
204 pub realized_served_failure: f64,
206}
207
208impl EProcessReport {
209 #[must_use]
211 pub fn render(&self) -> String {
212 let cert = match &self.certification {
213 Some(c) => format!(
214 "{:.4} (e-value {:.1} >= {:.1}, first certified at round {})",
215 c.threshold, c.e_value, self.crossing_level, c.certified_at_round
216 ),
217 None => "NONE — no threshold has earned a guarantee on this stream yet".to_owned(),
221 };
222 format!(
223 "method: eprocess (anytime-valid)\n\
224 pairs: {n_pairs}\n\
225 certified threshold: {cert}\n\
226 target alpha: {alpha:.4} (delta {delta:.4}, family-wise across grid AND rounds)\n\
227 realized served-failure: {realized:.4}\n\
228 guarantee: holds at EVERY round, no exchangeability assumed (Ville / e-process)\n",
229 n_pairs = self.n_pairs,
230 alpha = self.alpha,
231 delta = self.delta,
232 realized = self.realized_served_failure,
233 )
234 }
235}
236
237#[must_use]
243pub fn calibrate_pairs_eprocess(pairs: &[(f64, bool)], alpha: f64, delta: f64) -> EProcessReport {
244 const GRID_QUANTUM: f64 = 100.0;
262 let mut grid: Vec<f64> = pairs
263 .iter()
264 .map(|&(s, _)| (s * GRID_QUANTUM).round() / GRID_QUANTUM)
265 .collect();
266 grid.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
267 grid.dedup();
268
269 let mut ctrl = eprocess::EProcessRiskControl::new(alpha, delta, eprocess::DEFAULT_BET, &grid);
270 for &(score, correct) in pairs {
271 ctrl.observe_served(score, correct);
272 }
273 EProcessReport {
274 n_pairs: pairs.len(),
275 certification: ctrl.certified_threshold(),
276 alpha,
277 delta,
278 crossing_level: ctrl.crossing_level(),
279 realized_served_failure: ctrl.realized_served_failure(),
280 }
281}
282
283pub fn calibrate_from_store_eprocess(
290 db_path: impl AsRef<Path>,
291 tenant: &str,
292 alpha: f64,
293 delta: f64,
294) -> Result<EProcessReport, StoreError> {
295 let db_path = db_path.as_ref();
296 let traces = store::load_tenant_traces(db_path, tenant).unwrap_or_default();
297 let mut pairs = Vec::with_capacity(traces.len());
298 for trace in &traces {
299 let deferred = store::load_deferred(db_path, &trace.trace_id.to_string())?;
300 if let Some(pair) = trace_pair(trace, &deferred) {
301 pairs.push(pair);
302 }
303 }
304 Ok(calibrate_pairs_eprocess(&pairs, alpha, delta))
305}
306
307pub fn calibrate_from_store_ltt(
314 db_path: impl AsRef<Path>,
315 tenant: &str,
316 alpha: f64,
317 delta: f64,
318 min_n: usize,
319) -> Result<LttReport, StoreError> {
320 let traces = store::load_tenant_traces(&db_path, tenant).unwrap_or_default();
321 let mut pairs = Vec::with_capacity(traces.len());
322 for trace in &traces {
323 let deferred = store::load_deferred(&db_path, &trace.trace_id.to_string())?;
324 if let Some(pair) = trace_pair(trace, &deferred) {
325 pairs.push(pair);
326 }
327 }
328 Ok(calibrate_pairs_ltt(&pairs, alpha, delta, min_n))
329}
330
331pub fn calibrate_from_store(
340 db_path: impl AsRef<Path>,
341 tenant: &str,
342 alpha: f64,
343 delta: f64,
344 min_n: usize,
345) -> Result<CalibrationReport, StoreError> {
346 let traces = store::load_tenant_traces(&db_path, tenant).unwrap_or_default();
350 let mut pairs = Vec::with_capacity(traces.len());
351 for trace in &traces {
352 let deferred = store::load_deferred(&db_path, &trace.trace_id.to_string())?;
353 if let Some(pair) = trace_pair(trace, &deferred) {
354 pairs.push(pair);
355 }
356 }
357 Ok(calibrate_pairs(&pairs, alpha, delta, min_n))
358}
359
360#[cfg(test)]
361mod tests {
362 use firstpass_core::{
363 Features, FinalOutcome, GENESIS_HASH, GateResult, Mode, PolicyRef, RequestInfo, ServedFrom,
364 TaskKind,
365 };
366
367 use super::*;
368 use crate::store;
369
370 fn trace_with_score(score: f64) -> Trace {
373 let verdict = if score >= 0.5 {
374 Verdict::Pass
375 } else {
376 Verdict::Fail
377 };
378 let attempt = Attempt {
379 rung: 0,
380 model: "claude-haiku-4-5".to_owned(),
381 provider: "anthropic".to_owned(),
382 in_tokens: 10,
383 cache_write_tokens: 0,
384 cache_read_tokens: 0,
385 out_tokens: 5,
386 cost_usd: 0.001,
387 latency_ms: 12,
388 gates: vec![GateResult {
389 gate_id: "gate@v1".to_owned(),
390 verdict,
391 score: Some(Score::clamped(score)),
392 cost_usd: 0.0,
393 ms: 10,
394 reason: None,
395 evidence_ref: None,
396 }],
397 verdict,
398 };
399 let mut trace = Trace {
400 trace_id: uuid::Uuid::now_v7(),
401 prev_hash: GENESIS_HASH.to_owned(),
402 tenant_id: "tenant-a".to_owned(),
403 session_id: "session-1".to_owned(),
404 ts: jiff::Timestamp::now(),
405 mode: Mode::Enforce,
406 policy: PolicyRef {
407 id: "test@v0".to_owned(),
408 explore: false,
409 propensity: None,
410 mode_profile: None,
411 },
412 request: RequestInfo {
413 api: "anthropic.messages".to_owned(),
414 prompt_hash: "deadbeef".to_owned(),
415 features: Features::new(TaskKind::Other),
416 },
417 attempts: vec![attempt],
418 deferred: Vec::new(),
419 final_: FinalOutcome {
420 served_rung: Some(0),
421 served_from: ServedFrom::Attempt,
422 total_cost_usd: 0.001,
423 gate_cost_usd: 0.0,
424 total_latency_ms: 12,
425 escalations: 0,
426 counterfactual_baseline_usd: 0.001,
427 savings_usd: 0.0,
428 cache_source: None,
429 },
430 probe: None,
431 rollout: None,
432 shadow: None,
433 route_ix: None,
434 predicted_pass: None,
435 elastic: None,
436 };
437 trace.recompute_savings();
438 trace
439 }
440
441 #[test]
442 fn calibrate_pairs_finds_a_feasible_threshold_on_clean_pairs() {
443 let mut pairs = Vec::new();
449 for i in 0..60u32 {
450 pairs.push((0.7 + f64::from(i % 10) * 0.01, true));
451 }
452 for i in 0..60u32 {
453 pairs.push((0.2 + f64::from(i % 10) * 0.01, false));
454 }
455 let report = calibrate_pairs(&pairs, 0.2, 0.1, 30);
456 assert!(
457 report.conformal.feasible,
458 "clean separation must be feasible"
459 );
460 assert!(
461 report.conformal.threshold >= 0.2 && report.conformal.threshold <= 0.79,
462 "threshold {} must land inside the observed score range",
463 report.conformal.threshold
464 );
465 assert_eq!(report.n_pairs, 120);
466 assert!(
467 report.empirical_served_failure <= 0.2 + 1e-9,
468 "empirical served-failure {} must respect alpha — the conformal guarantee",
469 report.empirical_served_failure
470 );
471 }
472
473 #[test]
474 fn calibrate_pairs_infeasible_below_min_n() {
475 let pairs = vec![(0.9, true), (0.9, true), (0.1, false)];
476 let report = calibrate_pairs(&pairs, 0.1, 0.1, 30);
477 assert!(
478 !report.conformal.feasible,
479 "too few pairs must be infeasible"
480 );
481 }
482
483 #[tokio::test]
484 async fn calibrate_from_store_pairs_only_traces_with_deferred_feedback() {
485 let db_path = std::env::temp_dir().join(format!(
486 "firstpass-calibrate-test-{}.db",
487 uuid::Uuid::now_v7()
488 ));
489 let (tx, handle) = store::open(&db_path).unwrap();
490
491 let mut correct_ids = Vec::new();
494 let mut incorrect_ids = Vec::new();
495 for i in 0..40u32 {
496 let t = trace_with_score(0.7 + f64::from(i % 10) * 0.01);
497 correct_ids.push(t.trace_id.to_string());
498 tx.try_send(t).unwrap();
499 }
500 for i in 0..40u32 {
501 let t = trace_with_score(0.2 + f64::from(i % 10) * 0.01);
502 incorrect_ids.push(t.trace_id.to_string());
503 tx.try_send(t).unwrap();
504 }
505 for i in 0..5u32 {
506 tx.try_send(trace_with_score(0.5 + f64::from(i) * 0.01))
507 .unwrap();
508 }
509 drop(tx);
510 handle.await.unwrap();
511
512 for trace_id in &correct_ids {
513 let dv = DeferredVerdict {
514 gate_id: "outcome".to_owned(),
515 verdict: Verdict::Pass,
516 score: None,
517 reported_at: jiff::Timestamp::now(),
518 reporter: "unit-test".to_owned(),
519 };
520 store::append_deferred(&db_path, trace_id, &dv).unwrap();
521 }
522 for trace_id in &incorrect_ids {
523 let dv = DeferredVerdict {
524 gate_id: "outcome".to_owned(),
525 verdict: Verdict::Fail,
526 score: None,
527 reported_at: jiff::Timestamp::now(),
528 reporter: "unit-test".to_owned(),
529 };
530 store::append_deferred(&db_path, trace_id, &dv).unwrap();
531 }
532
533 let report = calibrate_from_store(&db_path, "tenant-a", 0.2, 0.1, 30).unwrap();
535 assert_eq!(
536 report.n_pairs, 80,
537 "only the 80 traces with deferred feedback pair up"
538 );
539 assert!(report.conformal.feasible);
540 assert!(
541 report.empirical_served_failure <= 0.2 + 1e-9,
542 "empirical served-failure {} must respect alpha on clean synthetic data",
543 report.empirical_served_failure
544 );
545
546 let other = calibrate_from_store(&db_path, "tenant-b", 0.2, 0.1, 30).unwrap();
548 assert_eq!(
549 other.n_pairs, 0,
550 "tenant-b must not see tenant-a's feedback"
551 );
552
553 let _ = std::fs::remove_file(&db_path);
554 }
555
556 #[test]
559 fn calibrate_pairs_ltt_feasible_on_clean_pairs() {
560 let mut pairs = Vec::new();
562 for i in 0..60u32 {
563 pairs.push((0.7 + f64::from(i % 10) * 0.01, true));
564 }
565 for i in 0..60u32 {
566 pairs.push((0.2 + f64::from(i % 10) * 0.01, false));
567 }
568 let report = calibrate_pairs_ltt(&pairs, 0.2, 0.1, 30);
569 assert!(
570 report.ltt.feasible,
571 "clean separation must be feasible with LTT"
572 );
573 assert!(
574 report.ltt.threshold >= 0.2 && report.ltt.threshold <= 0.79,
575 "threshold {} must land inside the observed score range",
576 report.ltt.threshold
577 );
578 assert_eq!(report.n_pairs, 120);
579 assert!(
580 report.ltt.empirical_risk <= 0.2 + 1e-9,
581 "empirical risk {} must respect alpha",
582 report.ltt.empirical_risk
583 );
584 }
585
586 #[test]
587 fn calibrate_pairs_ltt_infeasible_below_min_n() {
588 let pairs = vec![(0.9, true), (0.9, true), (0.1, false)];
589 let report = calibrate_pairs_ltt(&pairs, 0.1, 0.05, 30);
590 assert!(
591 !report.ltt.feasible,
592 "too few pairs must be infeasible with LTT"
593 );
594 }
595
596 #[test]
607 fn a_high_precision_score_distribution_does_not_collapse_statistical_power() {
608 let pairs: Vec<(f64, bool)> = (0..4000)
611 .map(|i| {
612 let score = 0.90 + f64::from(i) * 1e-7;
613 (score, i % 20 != 0) })
615 .collect();
616 let report = calibrate_pairs_eprocess(&pairs, 0.20, 0.05);
617
618 assert!(
621 report.crossing_level <= 101.0 / 0.05,
622 "the grid must be bounded by quantisation, not by the data: crossing level {} implies \
623 {} grid points",
624 report.crossing_level,
625 report.crossing_level * 0.05
626 );
627 assert!(
628 report.certification.is_some(),
629 "4000 clean pairs must still certify — bounding the grid must not cost the power it \
630 was meant to protect"
631 );
632 }
633
634 #[test]
640 fn eprocess_reports_no_threshold_when_nothing_is_certified() {
641 let pairs: Vec<(f64, bool)> = vec![(0.9, true), (0.8, true)];
642 let report = calibrate_pairs_eprocess(&pairs, 0.10, 0.05);
643 assert!(
644 report.certification.is_none(),
645 "two pairs cannot certify anything at delta=0.05"
646 );
647 let rendered = report.render();
648 assert!(
649 rendered.contains("NONE"),
650 "an uncertified stream must say so plainly, got: {rendered}"
651 );
652 assert!(
653 rendered.contains("anytime-valid"),
654 "the report must name the guarantee it provides, got: {rendered}"
655 );
656 }
657
658 #[test]
661 fn eprocess_certifies_and_renders_the_threshold_on_a_clean_stream() {
662 let mut pairs: Vec<(f64, bool)> = Vec::new();
663 for i in 0..2000 {
664 pairs.push((0.9, i % 20 != 0));
666 }
667 let report = calibrate_pairs_eprocess(&pairs, 0.20, 0.05);
668 let cert = report
669 .certification
670 .as_ref()
671 .expect("2000 clean pairs must certify");
672 assert!(
673 cert.e_value >= report.crossing_level,
674 "a certified threshold must meet the Bonferroni crossing level: {} < {}",
675 cert.e_value,
676 report.crossing_level
677 );
678 assert!(
679 report.render().contains(&format!("{:.4}", cert.threshold)),
680 "the rendered report must show the certified threshold"
681 );
682 }
683
684 #[test]
685 fn ltt_report_render_includes_method_and_far() {
686 let mut pairs: Vec<(f64, bool)> = Vec::new();
688 for _ in 0..200 {
689 pairs.push((0.9, true));
690 }
691 for _ in 0..5 {
692 pairs.push((0.9, false));
693 }
694 for _ in 0..15 {
695 pairs.push((0.2, false));
696 }
697 let report = calibrate_pairs_ltt(&pairs, 0.10, 0.05, 30);
698 let rendered = report.render();
699 assert!(
700 rendered.contains("method: ltt"),
701 "render must tag the method"
702 );
703 assert!(
704 rendered.contains("false-accept rate:"),
705 "render must include verifier ROC note"
706 );
707 assert!(
708 rendered.contains("feasible:"),
709 "render must include feasibility"
710 );
711 }
712}