rusty-eval 0.1.5

Agent TestOps for Rusty — versioned evaluation datasets, deterministic assertions over recorded runs, experiment reports, and baseline-vs-candidate comparison.
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
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
//! Deterministic human-feedback operations for evaluation and learning loops.
//!
//! The queue deliberately owns no clock, database, or worker runtime. Callers
//! supply timestamps, persist the serializable queue, and decide who may act as
//! a reviewer or adjudicator. That keeps replay and tests deterministic while
//! still providing the state-machine guarantees needed by a distributed review
//! service: leases expire, reviewers cannot vote twice, disagreements require
//! adjudication, and only resolved feedback can become an evaluation case.

use std::collections::{BTreeMap, BTreeSet};

use serde::{Deserialize, Serialize};
use serde_json::Value;

use crate::dataset::{EvalCase, Expectation, StatePredicate};
use crate::error::{EvalError, Result};

/// The feedback queue format version this build loads and writes.
pub const FEEDBACK_FORMAT_VERSION: u64 = 1;

/// A stable reference to the recorded run that produced an annotation task.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct TraceRef {
    /// Stable run identifier in the Flight Recorder.
    pub run_id: String,
    /// Optional conversation or thread identifier.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub thread_id: Option<String>,
    /// Optional event sequence at which the candidate output was captured.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub event_seq: Option<u64>,
}

impl TraceRef {
    /// Reference a run without a narrower thread or event location.
    pub fn run(run_id: impl Into<String>) -> Self {
        Self {
            run_id: run_id.into(),
            thread_id: None,
            event_seq: None,
        }
    }
}

/// One candidate output presented to a reviewer.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct ReviewCandidate {
    /// Stable identifier within the task.
    pub id: String,
    /// Arbitrary structured output shown to the reviewer.
    pub output: Value,
}

/// One scored dimension in a review rubric.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct RubricCriterion {
    /// Stable criterion identifier used as the score-map key.
    pub id: String,
    /// Human-readable guidance for the reviewer.
    pub description: String,
}

/// A versioned rubric attached to an annotation task.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct ReviewRubric {
    /// Stable rubric name.
    pub name: String,
    /// Caller-managed rubric version.
    pub version: String,
    /// Required criteria. Every submitted review must score every criterion.
    pub criteria: Vec<RubricCriterion>,
}

/// The actionable result of a review.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
#[serde(tag = "kind", rename_all = "snake_case")]
pub enum ReviewDecision {
    /// Prefer one of the task's candidates.
    Select { candidate_id: String },
    /// Supply a corrected output when neither candidate is acceptable.
    Correct { output: Value },
    /// Reject the example so it cannot be promoted into a dataset.
    Reject {
        #[serde(default, skip_serializing_if = "Option::is_none")]
        reason: Option<String>,
    },
}

/// A review submitted while holding a live task lease.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct ReviewSubmission {
    /// Selection, correction, or rejection.
    pub decision: ReviewDecision,
    /// Criterion id to normalized score in the inclusive range `[0, 1]`.
    pub scores: BTreeMap<String, f64>,
    /// Optional reviewer rationale.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub comment: Option<String>,
}

/// An immutable accepted review with provenance.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct StoredReview {
    /// Reviewer identity supplied by the caller.
    pub reviewer: String,
    /// Caller-supplied submission timestamp.
    pub submitted_at_ms: u64,
    /// The validated submission.
    pub submission: ReviewSubmission,
}

/// How a final decision was reached.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
#[serde(tag = "kind", rename_all = "snake_case")]
pub enum ResolutionAuthority {
    /// All required reviewers independently made the same decision.
    Consensus,
    /// A named adjudicator resolved disagreement.
    Adjudicator { id: String },
}

/// The final, promotion-eligible result of an annotation task.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
pub struct ReviewResolution {
    /// Final decision.
    pub decision: ReviewDecision,
    /// Consensus or explicit adjudication.
    pub authority: ResolutionAuthority,
    /// Caller-supplied resolution timestamp.
    pub resolved_at_ms: u64,
}

/// Current task state at a caller-supplied point in time.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum AnnotationStatus {
    /// No accepted reviews or live leases.
    Open,
    /// At least one review or live lease exists; more reviews are required.
    InReview,
    /// The required reviews disagree and an adjudicator must decide.
    NeedsAdjudication,
    /// A consensus or adjudicated resolution exists.
    Resolved,
}

/// Proof that a reviewer currently owns a task reservation.
#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
pub struct ReviewLease {
    /// Claimed annotation task.
    pub task_id: String,
    /// Reviewer holding the lease.
    pub reviewer: String,
    /// Monotonic occurrence id that prevents a stale worker from using a
    /// later lease issued to the same reviewer.
    pub lease_id: u64,
    /// Caller-supplied time at which this occurrence was issued.
    pub issued_at_ms: u64,
    /// Exclusive lease expiry; a lease is expired when `now_ms >=` this value.
    pub expires_at_ms: u64,
}

#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
struct LeaseRecord {
    lease_id: u64,
    issued_at_ms: u64,
    expires_at_ms: u64,
}

/// One pairwise annotation task and its review state.
#[derive(Debug, Clone, PartialEq, Serialize)]
pub struct AnnotationTask {
    id: String,
    source: TraceRef,
    input: Value,
    candidates: Vec<ReviewCandidate>,
    rubric: ReviewRubric,
    priority: i32,
    created_at_ms: u64,
    required_reviews: usize,
    #[serde(default, skip_serializing_if = "BTreeMap::is_empty")]
    leases: BTreeMap<String, LeaseRecord>,
    #[serde(default, skip_serializing_if = "BTreeMap::is_empty")]
    reviews: BTreeMap<String, StoredReview>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    resolution: Option<ReviewResolution>,
}

impl AnnotationTask {
    /// Create a pairwise annotation task.
    #[allow(clippy::too_many_arguments)]
    pub fn new(
        id: impl Into<String>,
        source: TraceRef,
        input: Value,
        candidates: Vec<ReviewCandidate>,
        rubric: ReviewRubric,
        priority: i32,
        created_at_ms: u64,
        required_reviews: usize,
    ) -> Result<Self> {
        let task = Self {
            id: id.into(),
            source,
            input,
            candidates,
            rubric,
            priority,
            created_at_ms,
            required_reviews,
            leases: BTreeMap::new(),
            reviews: BTreeMap::new(),
            resolution: None,
        };
        task.validate_definition()?;
        Ok(task)
    }

    /// Stable task identifier.
    pub fn id(&self) -> &str {
        &self.id
    }

    /// Recorded-run provenance.
    pub fn source(&self) -> &TraceRef {
        &self.source
    }

    /// Input that produced the candidates.
    pub fn input(&self) -> &Value {
        &self.input
    }

    /// The two candidates under comparison.
    pub fn candidates(&self) -> &[ReviewCandidate] {
        &self.candidates
    }

    /// Versioned scoring rubric.
    pub fn rubric(&self) -> &ReviewRubric {
        &self.rubric
    }

    /// Scheduling priority; larger values are claimed first.
    pub fn priority(&self) -> i32 {
        self.priority
    }

    /// Task creation timestamp.
    pub fn created_at_ms(&self) -> u64 {
        self.created_at_ms
    }

    /// Number of independent reviews required before resolution.
    pub fn required_reviews(&self) -> usize {
        self.required_reviews
    }

    /// Accepted reviews, keyed by reviewer identity.
    pub fn reviews(&self) -> &BTreeMap<String, StoredReview> {
        &self.reviews
    }

    /// Final resolution, when one exists.
    pub fn resolution(&self) -> Option<&ReviewResolution> {
        self.resolution.as_ref()
    }

    /// Compute status without mutating or consulting a wall clock.
    pub fn status_at(&self, now_ms: u64) -> AnnotationStatus {
        if self.resolution.is_some() {
            return AnnotationStatus::Resolved;
        }
        if self.reviews.len() >= self.required_reviews {
            return AnnotationStatus::NeedsAdjudication;
        }
        if !self.reviews.is_empty()
            || self
                .leases
                .values()
                .any(|lease| lease.expires_at_ms > now_ms)
        {
            AnnotationStatus::InReview
        } else {
            AnnotationStatus::Open
        }
    }

    /// Convert a resolved selection or correction into a versioned dataset case.
    pub fn promote(
        &self,
        case_id: impl Into<String>,
        output_pointer: impl Into<String>,
    ) -> Result<EvalCase> {
        let case_id = case_id.into();
        require_non_empty("promoted case id", &case_id)?;
        let output_pointer = output_pointer.into();
        validate_json_pointer(&output_pointer)?;
        let resolution = self
            .resolution
            .as_ref()
            .ok_or_else(|| EvalError::Feedback(format!("task `{}` is not resolved", self.id)))?;
        let expected = match &resolution.decision {
            ReviewDecision::Select { candidate_id } => self
                .candidates
                .iter()
                .find(|candidate| candidate.id == *candidate_id)
                .map(|candidate| candidate.output.clone())
                .ok_or_else(|| {
                    EvalError::Feedback(format!(
                        "task `{}` resolution selects unknown candidate `{candidate_id}`",
                        self.id
                    ))
                })?,
            ReviewDecision::Correct { output } => output.clone(),
            ReviewDecision::Reject { .. } => {
                return Err(EvalError::Feedback(format!(
                    "task `{}` was rejected and cannot be promoted",
                    self.id
                )));
            }
        };
        let authority = match &resolution.authority {
            ResolutionAuthority::Consensus => "consensus".to_owned(),
            ResolutionAuthority::Adjudicator { id } => format!("adjudicated:{id}"),
        };

        Ok(EvalCase {
            id: case_id,
            input: self.input.clone(),
            expect: Expectation {
                state: vec![StatePredicate {
                    pointer: output_pointer,
                    expected,
                }],
                ..Expectation::default()
            },
            tags: vec![
                "human-feedback".to_owned(),
                format!("feedback-task:{}", self.id),
                format!("source-run:{}", self.source.run_id),
                authority,
            ],
        })
    }

    fn validate_definition(&self) -> Result<()> {
        require_non_empty("task id", &self.id)?;
        require_non_empty("source run id", &self.source.run_id)?;
        if let Some(thread_id) = &self.source.thread_id {
            require_non_empty("source thread id", thread_id)?;
        }
        if self.candidates.len() != 2 {
            return feedback_error(format!(
                "task `{}` must contain exactly two candidates",
                self.id
            ));
        }
        let mut candidate_ids = BTreeSet::new();
        for candidate in &self.candidates {
            require_non_empty("candidate id", &candidate.id)?;
            if !candidate_ids.insert(candidate.id.as_str()) {
                return feedback_error(format!(
                    "task `{}` has duplicate candidate id `{}`",
                    self.id, candidate.id
                ));
            }
        }
        require_non_empty("rubric name", &self.rubric.name)?;
        require_non_empty("rubric version", &self.rubric.version)?;
        if self.rubric.criteria.is_empty() {
            return feedback_error(format!(
                "task `{}` rubric must contain at least one criterion",
                self.id
            ));
        }
        let mut criterion_ids = BTreeSet::new();
        for criterion in &self.rubric.criteria {
            require_non_empty("rubric criterion id", &criterion.id)?;
            require_non_empty("rubric criterion description", &criterion.description)?;
            if !criterion_ids.insert(criterion.id.as_str()) {
                return feedback_error(format!(
                    "task `{}` has duplicate rubric criterion `{}`",
                    self.id, criterion.id
                ));
            }
        }
        if self.required_reviews == 0 {
            return feedback_error(format!(
                "task `{}` must require at least one review",
                self.id
            ));
        }
        Ok(())
    }

    fn validate_submission(&self, submission: &ReviewSubmission) -> Result<()> {
        self.validate_decision(&submission.decision)?;
        let expected: BTreeSet<_> = self
            .rubric
            .criteria
            .iter()
            .map(|criterion| criterion.id.as_str())
            .collect();
        let actual: BTreeSet<_> = submission.scores.keys().map(String::as_str).collect();
        if actual != expected {
            return feedback_error(format!(
                "task `{}` review scores must exactly match rubric criteria",
                self.id
            ));
        }
        for (criterion, score) in &submission.scores {
            if !score.is_finite() || !(0.0..=1.0).contains(score) {
                return feedback_error(format!(
                    "task `{}` score for `{criterion}` must be finite and between 0 and 1",
                    self.id
                ));
            }
        }
        Ok(())
    }

    fn validate_decision(&self, decision: &ReviewDecision) -> Result<()> {
        if let ReviewDecision::Select { candidate_id } = decision {
            if !self
                .candidates
                .iter()
                .any(|candidate| candidate.id == *candidate_id)
            {
                return feedback_error(format!(
                    "task `{}` has no candidate `{candidate_id}`",
                    self.id
                ));
            }
        }
        Ok(())
    }

    fn validate_state(&self) -> Result<()> {
        self.validate_definition()?;
        if self.reviews.len() > self.required_reviews {
            return feedback_error(format!(
                "task `{}` contains more reviews than required",
                self.id
            ));
        }
        for (reviewer, stored) in &self.reviews {
            require_non_empty("reviewer id", reviewer)?;
            if stored.reviewer != *reviewer {
                return feedback_error(format!(
                    "task `{}` review key does not match reviewer",
                    self.id
                ));
            }
            if self.leases.contains_key(reviewer) {
                return feedback_error(format!(
                    "task `{}` reviewer `{reviewer}` has both a lease and review",
                    self.id
                ));
            }
            if stored.submitted_at_ms < self.created_at_ms {
                return feedback_error(format!(
                    "task `{}` review by `{reviewer}` predates task creation",
                    self.id
                ));
            }
            self.validate_submission(&stored.submission)?;
        }
        for (reviewer, lease) in &self.leases {
            require_non_empty("reviewer id", reviewer)?;
            if lease.issued_at_ms < self.created_at_ms {
                return feedback_error(format!(
                    "task `{}` lease for `{reviewer}` predates task creation",
                    self.id
                ));
            }
            if lease.expires_at_ms <= lease.issued_at_ms {
                return feedback_error(format!(
                    "task `{}` lease for `{reviewer}` has an invalid time range",
                    self.id
                ));
            }
        }
        if let Some(resolution) = &self.resolution {
            if self.reviews.len() < self.required_reviews {
                return feedback_error(format!(
                    "task `{}` resolved before collecting required reviews",
                    self.id
                ));
            }
            if self
                .reviews
                .values()
                .any(|review| review.submitted_at_ms > resolution.resolved_at_ms)
            {
                return feedback_error(format!(
                    "task `{}` resolution predates a submitted review",
                    self.id
                ));
            }
            if !self.leases.is_empty() {
                return feedback_error(format!(
                    "task `{}` is resolved but still contains reviewer leases",
                    self.id
                ));
            }
            self.validate_decision(&resolution.decision)?;
            let mut review_decisions = self
                .reviews
                .values()
                .map(|review| &review.submission.decision);
            let first_review_decision = review_decisions.next().expect("required reviews exist");
            let unanimous = review_decisions.all(|decision| decision == first_review_decision);
            match &resolution.authority {
                ResolutionAuthority::Consensus => {
                    if !unanimous || *first_review_decision != resolution.decision {
                        return feedback_error(format!(
                            "task `{}` consensus resolution does not match every review",
                            self.id
                        ));
                    }
                }
                ResolutionAuthority::Adjudicator { id } => {
                    require_non_empty("adjudicator id", id)?;
                    if unanimous {
                        return feedback_error(format!(
                            "task `{}` uses adjudication despite unanimous reviews",
                            self.id
                        ));
                    }
                }
            }
        } else {
            if self.reviews.len() + self.leases.len() > self.required_reviews {
                return feedback_error(format!(
                    "task `{}` has more reviews and leases than required",
                    self.id
                ));
            }
            if self.reviews.len() == self.required_reviews {
                let mut decisions = self
                    .reviews
                    .values()
                    .map(|review| &review.submission.decision);
                if let Some(first) = decisions.next() {
                    if decisions.all(|decision| decision == first) {
                        return feedback_error(format!(
                            "task `{}` has review consensus but no resolution",
                            self.id
                        ));
                    }
                }
            }
        }
        Ok(())
    }
}

/// A named, versioned collection of annotation tasks.
#[derive(Debug, Clone, PartialEq, Serialize)]
pub struct AnnotationQueue {
    format_version: u64,
    name: String,
    tasks: BTreeMap<String, AnnotationTask>,
    next_lease_id: u64,
}

impl AnnotationQueue {
    /// Create an empty queue.
    pub fn new(name: impl Into<String>) -> Result<Self> {
        let name = name.into();
        require_non_empty("queue name", &name)?;
        Ok(Self {
            format_version: FEEDBACK_FORMAT_VERSION,
            name,
            tasks: BTreeMap::new(),
            next_lease_id: 1,
        })
    }

    /// Queue name.
    pub fn name(&self) -> &str {
        &self.name
    }

    /// All tasks in stable id order.
    pub fn tasks(&self) -> &BTreeMap<String, AnnotationTask> {
        &self.tasks
    }

    /// Find a task by id.
    pub fn task(&self, task_id: &str) -> Option<&AnnotationTask> {
        self.tasks.get(task_id)
    }

    /// Add a validated task without overwriting an existing id.
    pub fn enqueue(&mut self, task: AnnotationTask) -> Result<()> {
        task.validate_state()?;
        if self.tasks.contains_key(task.id()) {
            return feedback_error(format!("duplicate annotation task `{}`", task.id()));
        }
        self.tasks.insert(task.id.clone(), task);
        Ok(())
    }

    /// Claim the highest-priority eligible task.
    ///
    /// Ordering is deterministic: priority descending, creation time
    /// ascending, then task id ascending. Calling claim again while the same
    /// reviewer owns a live lease returns that lease idempotently.
    pub fn claim(
        &mut self,
        reviewer: impl Into<String>,
        now_ms: u64,
        lease_ms: u64,
    ) -> Result<Option<ReviewLease>> {
        let reviewer = reviewer.into();
        require_non_empty("reviewer id", &reviewer)?;
        if lease_ms == 0 {
            return feedback_error("lease duration must be greater than zero");
        }
        self.prune_expired_leases(now_ms);
        if let Some((task_id, lease)) = self.tasks.iter().find_map(|(task_id, task)| {
            task.leases
                .get(&reviewer)
                .map(|lease| (task_id.clone(), lease.clone()))
        }) {
            return Ok(Some(ReviewLease {
                task_id,
                reviewer,
                lease_id: lease.lease_id,
                issued_at_ms: lease.issued_at_ms,
                expires_at_ms: lease.expires_at_ms,
            }));
        }

        let task_id = self
            .tasks
            .values()
            .filter(|task| {
                task.created_at_ms <= now_ms
                    && task.resolution.is_none()
                    && !task.reviews.contains_key(&reviewer)
                    && task.reviews.len() < task.required_reviews
                    && task.reviews.len() + task.leases.len() < task.required_reviews
            })
            .min_by(|left, right| {
                right
                    .priority
                    .cmp(&left.priority)
                    .then_with(|| left.created_at_ms.cmp(&right.created_at_ms))
                    .then_with(|| left.id.cmp(&right.id))
            })
            .map(|task| task.id.clone());

        let Some(task_id) = task_id else {
            return Ok(None);
        };
        let expires_at_ms = now_ms.checked_add(lease_ms).ok_or_else(|| {
            EvalError::Feedback("lease expiry overflows u64 milliseconds".to_owned())
        })?;
        let lease_id = self.next_lease_id;
        self.next_lease_id = self.next_lease_id.checked_add(1).ok_or_else(|| {
            EvalError::Feedback("lease occurrence id space is exhausted".to_owned())
        })?;
        self.tasks
            .get_mut(&task_id)
            .expect("selected task remains present")
            .leases
            .insert(
                reviewer.clone(),
                LeaseRecord {
                    lease_id,
                    issued_at_ms: now_ms,
                    expires_at_ms,
                },
            );
        Ok(Some(ReviewLease {
            task_id,
            reviewer,
            lease_id,
            issued_at_ms: now_ms,
            expires_at_ms,
        }))
    }

    /// Submit a review under a live lease.
    pub fn submit(
        &mut self,
        lease: &ReviewLease,
        now_ms: u64,
        submission: ReviewSubmission,
    ) -> Result<AnnotationStatus> {
        require_non_empty("reviewer id", &lease.reviewer)?;
        let task_id = &lease.task_id;
        let reviewer = &lease.reviewer;
        let task = self
            .tasks
            .get_mut(task_id)
            .ok_or_else(|| EvalError::Feedback(format!("unknown annotation task `{task_id}`")))?;
        if task.resolution.is_some() {
            return feedback_error(format!("task `{task_id}` is already resolved"));
        }
        if task.reviews.contains_key(reviewer) {
            return feedback_error(format!(
                "reviewer `{reviewer}` already reviewed task `{task_id}`"
            ));
        }
        if task.reviews.len() >= task.required_reviews {
            return feedback_error(format!(
                "task `{task_id}` already collected all required reviews"
            ));
        }
        let active = task.leases.get(reviewer).cloned().ok_or_else(|| {
            EvalError::Feedback(format!(
                "reviewer `{reviewer}` does not hold a lease for task `{task_id}`"
            ))
        })?;
        if active.lease_id != lease.lease_id
            || active.issued_at_ms != lease.issued_at_ms
            || active.expires_at_ms != lease.expires_at_ms
        {
            return feedback_error(format!(
                "lease {} for task `{task_id}` is stale",
                lease.lease_id
            ));
        }
        if now_ms >= active.expires_at_ms {
            task.leases.remove(reviewer);
            return feedback_error(format!(
                "reviewer `{reviewer}` lease for task `{task_id}` expired"
            ));
        }
        if now_ms < active.issued_at_ms || now_ms < task.created_at_ms {
            return feedback_error(format!(
                "review timestamp for task `{task_id}` predates its lease or task creation"
            ));
        }
        if task
            .reviews
            .values()
            .any(|review| review.submitted_at_ms > now_ms)
        {
            return feedback_error(format!(
                "review timestamp for task `{task_id}` predates an accepted review"
            ));
        }
        task.validate_submission(&submission)?;
        task.leases.remove(reviewer);
        task.reviews.insert(
            reviewer.to_owned(),
            StoredReview {
                reviewer: reviewer.to_owned(),
                submitted_at_ms: now_ms,
                submission,
            },
        );

        if task.reviews.len() == task.required_reviews {
            let mut decisions = task
                .reviews
                .values()
                .map(|review| &review.submission.decision);
            let first = decisions.next().expect("at least one review").clone();
            if decisions.all(|decision| *decision == first) {
                task.resolution = Some(ReviewResolution {
                    decision: first,
                    authority: ResolutionAuthority::Consensus,
                    resolved_at_ms: now_ms,
                });
                task.leases.clear();
            }
        }
        Ok(task.status_at(now_ms))
    }

    /// Resolve a disagreement after all required reviews have arrived.
    pub fn adjudicate(
        &mut self,
        task_id: &str,
        adjudicator: impl Into<String>,
        now_ms: u64,
        decision: ReviewDecision,
    ) -> Result<()> {
        let adjudicator = adjudicator.into();
        require_non_empty("adjudicator id", &adjudicator)?;
        let task = self
            .tasks
            .get_mut(task_id)
            .ok_or_else(|| EvalError::Feedback(format!("unknown annotation task `{task_id}`")))?;
        if task.resolution.is_some() {
            return feedback_error(format!("task `{task_id}` is already resolved"));
        }
        if task.reviews.len() < task.required_reviews {
            return feedback_error(format!(
                "task `{task_id}` has not collected all required reviews"
            ));
        }
        if task
            .reviews
            .values()
            .any(|review| review.submitted_at_ms > now_ms)
        {
            return feedback_error(format!(
                "adjudication for task `{task_id}` predates a submitted review"
            ));
        }
        task.validate_decision(&decision)?;
        task.resolution = Some(ReviewResolution {
            decision,
            authority: ResolutionAuthority::Adjudicator { id: adjudicator },
            resolved_at_ms: now_ms,
        });
        task.leases.clear();
        Ok(())
    }

    /// Serialize the queue with stable map ordering.
    pub fn to_json(&self) -> Result<String> {
        Ok(serde_json::to_string_pretty(self)?)
    }

    /// Load a queue, rejecting unsupported versions and invalid state.
    pub fn from_json(text: &str) -> Result<Self> {
        let header: FeedbackVersionHeader = serde_json::from_str(text)?;
        if header.format_version != FEEDBACK_FORMAT_VERSION {
            return Err(EvalError::UnsupportedFeedbackVersion {
                found: header.format_version,
                supported: FEEDBACK_FORMAT_VERSION,
            });
        }
        let wire: AnnotationQueueWire = serde_json::from_str(text)?;
        let queue = wire.into_queue()?;
        require_non_empty("queue name", &queue.name)?;
        let mut lease_ids = BTreeSet::new();
        let mut leased_reviewers = BTreeSet::new();
        for (task_id, task) in &queue.tasks {
            if task_id != task.id() {
                return feedback_error(format!(
                    "annotation task map key `{task_id}` does not match task id `{}`",
                    task.id()
                ));
            }
            task.validate_state()?;
            for lease in task.leases.values() {
                if lease.lease_id == 0 || lease.lease_id >= queue.next_lease_id {
                    return feedback_error(format!(
                        "task `{task_id}` contains invalid lease occurrence {}",
                        lease.lease_id
                    ));
                }
                if !lease_ids.insert(lease.lease_id) {
                    return feedback_error(format!(
                        "duplicate lease occurrence {}",
                        lease.lease_id
                    ));
                }
            }
            for reviewer in task.leases.keys() {
                if !leased_reviewers.insert(reviewer.as_str()) {
                    return feedback_error(format!(
                        "reviewer `{reviewer}` holds leases for multiple tasks"
                    ));
                }
            }
        }
        Ok(queue)
    }

    fn prune_expired_leases(&mut self, now_ms: u64) {
        for task in self.tasks.values_mut() {
            task.leases.retain(|_, lease| lease.expires_at_ms > now_ms);
        }
    }
}

#[derive(Deserialize)]
struct FeedbackVersionHeader {
    format_version: u64,
}

#[derive(Deserialize)]
struct AnnotationTaskWire {
    id: String,
    source: TraceRef,
    input: Value,
    candidates: Vec<ReviewCandidate>,
    rubric: ReviewRubric,
    priority: i32,
    created_at_ms: u64,
    required_reviews: usize,
    #[serde(default)]
    leases: BTreeMap<String, LeaseRecord>,
    #[serde(default)]
    reviews: BTreeMap<String, StoredReview>,
    #[serde(default)]
    resolution: Option<ReviewResolution>,
}

impl AnnotationTaskWire {
    fn into_task(self) -> AnnotationTask {
        AnnotationTask {
            id: self.id,
            source: self.source,
            input: self.input,
            candidates: self.candidates,
            rubric: self.rubric,
            priority: self.priority,
            created_at_ms: self.created_at_ms,
            required_reviews: self.required_reviews,
            leases: self.leases,
            reviews: self.reviews,
            resolution: self.resolution,
        }
    }
}

#[derive(Deserialize)]
struct AnnotationQueueWire {
    format_version: u64,
    name: String,
    tasks: BTreeMap<String, AnnotationTaskWire>,
    next_lease_id: u64,
}

impl AnnotationQueueWire {
    fn into_queue(self) -> Result<AnnotationQueue> {
        if self.next_lease_id == 0 {
            return feedback_error("next lease occurrence id must be greater than zero");
        }
        Ok(AnnotationQueue {
            format_version: self.format_version,
            name: self.name,
            tasks: self
                .tasks
                .into_iter()
                .map(|(id, task)| (id, task.into_task()))
                .collect(),
            next_lease_id: self.next_lease_id,
        })
    }
}

fn require_non_empty(label: &str, value: &str) -> Result<()> {
    if value.trim().is_empty() {
        feedback_error(format!("{label} must not be empty"))
    } else {
        Ok(())
    }
}

fn validate_json_pointer(pointer: &str) -> Result<()> {
    if !pointer.is_empty() && !pointer.starts_with('/') {
        return feedback_error("output pointer must be an RFC 6901 JSON pointer");
    }
    for segment in pointer.split('/').skip(1) {
        let mut chars = segment.chars();
        while let Some(character) = chars.next() {
            if character == '~' && !matches!(chars.next(), Some('0' | '1')) {
                return feedback_error("output pointer contains an invalid RFC 6901 escape");
            }
        }
    }
    Ok(())
}

fn feedback_error<T>(message: impl Into<String>) -> Result<T> {
    Err(EvalError::Feedback(message.into()))
}