skilltest-core 0.4.0

Core library for skilltest: run AI skills on harness/model platforms and score transcripts with natural-language evals.
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
//! Natural-language evaluations. An eval poses a criterion in plain English and
//! asks the provider's judge to score the transcript: a boolean assertion, or a
//! numeric score compared against a threshold.

use schemars::JsonSchema;
use serde::{Deserialize, Serialize};

use crate::error::{Error, Result};

/// How a numeric score is compared to its threshold.
#[derive(Debug, Default, Clone, Copy, PartialEq, Eq, Serialize, Deserialize, JsonSchema)]
#[serde(rename_all = "lowercase")]
pub enum Comparator {
    /// value >= threshold (the default).
    #[serde(alias = ">=")]
    #[default]
    Gte,
    /// value > threshold.
    #[serde(alias = ">")]
    Gt,
    /// value <= threshold.
    #[serde(alias = "<=")]
    Lte,
    /// value < threshold.
    #[serde(alias = "<")]
    Lt,
}

impl Comparator {
    fn satisfied(self, value: f64, threshold: f64) -> bool {
        match self {
            Comparator::Gte => value >= threshold,
            Comparator::Gt => value > threshold,
            Comparator::Lte => value <= threshold,
            Comparator::Lt => value < threshold,
        }
    }

    fn symbol(self) -> &'static str {
        match self {
            Comparator::Gte => ">=",
            Comparator::Gt => ">",
            Comparator::Lte => "<=",
            Comparator::Lt => "<",
        }
    }
}

/// The default boolean expectation (the criterion should hold).
fn default_true() -> bool {
    true
}

/// An eval specification, as written in a test case's YAML.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
#[serde(tag = "type", rename_all = "lowercase")]
pub enum Eval {
    /// Assert a plain-English criterion holds (or, with `expected: false`, that
    /// it does not).
    Boolean {
        /// The criterion the judge evaluates against the transcript.
        criterion: String,
        /// What the judge's verdict must equal to pass. Defaults to `true`.
        #[serde(default = "default_true")]
        expected: bool,
        /// Optional human label for reports.
        #[serde(default)]
        name: Option<String>,
    },
    /// Score a plain-English criterion on a numeric scale and compare it to a
    /// threshold.
    Numeric {
        /// The criterion the judge scores.
        criterion: String,
        /// Inclusive lower bound of the scale.
        min: f64,
        /// Inclusive upper bound of the scale.
        max: f64,
        /// The passing threshold.
        threshold: f64,
        /// How the score is compared to `threshold`. Defaults to `>=`.
        #[serde(default)]
        comparator: Comparator,
        /// Optional human label for reports.
        #[serde(default)]
        name: Option<String>,
    },
}

impl Eval {
    /// The criterion text the judge sees.
    #[must_use]
    pub fn criterion(&self) -> &str {
        match self {
            Eval::Boolean { criterion, .. } | Eval::Numeric { criterion, .. } => criterion,
        }
    }

    /// A short label for reports: the explicit `name` if given, else the
    /// criterion.
    #[must_use]
    pub fn label(&self) -> &str {
        match self {
            Eval::Boolean {
                name, criterion, ..
            }
            | Eval::Numeric {
                name, criterion, ..
            } => name.as_deref().unwrap_or(criterion),
        }
    }

    /// Validate the eval's own parameters (independent of any transcript).
    ///
    /// # Errors
    /// [`Error::Invalid`] when a criterion is empty or a numeric scale is
    /// degenerate (`min >= max`) or the threshold falls outside `[min, max]`.
    pub fn validate(&self) -> Result<()> {
        if self.criterion().trim().is_empty() {
            return Err(Error::Invalid("an eval has an empty `criterion`".into()));
        }
        if let Eval::Numeric {
            min,
            max,
            threshold,
            ..
        } = self
        {
            if min >= max {
                return Err(Error::Invalid(format!(
                    "numeric eval scale is degenerate: min ({min}) must be < max ({max})"
                )));
            }
            if threshold < min || threshold > max {
                return Err(Error::Invalid(format!(
                    "numeric eval threshold ({threshold}) is outside the scale [{min}, {max}]"
                )));
            }
        }
        Ok(())
    }

    /// Apply this eval's pass rule to a raw judge value, producing an outcome.
    ///
    /// `raw` is the value the judge returned: `JudgeValue::Bool` for boolean
    /// evals, `JudgeValue::Number` for numeric. A mismatch is a provider error.
    ///
    /// # Errors
    /// [`Error::Provider`] if the judge returned the wrong value kind for this
    /// eval.
    pub fn outcome(&self, raw: &JudgeValue, reason: String) -> Result<EvalOutcome> {
        match (self, raw) {
            (Eval::Boolean { expected, .. }, JudgeValue::Bool(value)) => Ok(EvalOutcome {
                label: self.label().to_string(),
                passed: value == expected,
                detail: EvalDetail::Boolean {
                    value: *value,
                    expected: *expected,
                },
                reason,
            }),
            (
                Eval::Numeric {
                    min,
                    max,
                    threshold,
                    comparator,
                    ..
                },
                JudgeValue::Number(value),
            ) => {
                let clamped = value.clamp(*min, *max);
                Ok(EvalOutcome {
                    label: self.label().to_string(),
                    passed: comparator.satisfied(clamped, *threshold),
                    detail: EvalDetail::Numeric {
                        value: clamped,
                        threshold: *threshold,
                        comparator: *comparator,
                    },
                    reason,
                })
            }
            (Eval::Boolean { .. }, JudgeValue::Number(_)) => Err(Error::provider(
                "judge",
                "boolean eval received a numeric verdict",
            )),
            (Eval::Numeric { .. }, JudgeValue::Bool(_)) => Err(Error::provider(
                "judge",
                "numeric eval received a boolean verdict",
            )),
        }
    }
}

/// The raw value a judge returns: either a boolean or a number, matching the
/// eval kind. Deserialized untagged from the provider's `value` field.
#[derive(Debug, Clone, Copy, PartialEq, Serialize, Deserialize)]
#[serde(untagged)]
pub enum JudgeValue {
    Bool(bool),
    Number(f64),
}

/// The kind-specific detail of an eval outcome, for reporting.
///
/// The variant titles name the generated SDK model for each union arm, so keep
/// them stable: they are part of the SDK API surface.
#[derive(Debug, Clone, Copy, PartialEq, Serialize, Deserialize, JsonSchema)]
#[serde(tag = "kind", rename_all = "lowercase")]
pub enum EvalDetail {
    #[schemars(title = "BooleanDetail")]
    Boolean { value: bool, expected: bool },
    #[schemars(title = "NumericDetail")]
    Numeric {
        value: f64,
        threshold: f64,
        comparator: Comparator,
    },
}

impl EvalDetail {
    /// A compact human description of the verdict, e.g. `8.0 >= 7` or
    /// `true (expected true)`.
    #[must_use]
    pub fn summary(&self) -> String {
        match self {
            EvalDetail::Boolean { value, expected } => {
                format!("{value} (expected {expected})")
            }
            EvalDetail::Numeric {
                value,
                threshold,
                comparator,
            } => format!("{value} {} {threshold}", comparator.symbol()),
        }
    }
}

/// The result of running one eval against a transcript.
#[derive(Debug, Clone, PartialEq, Serialize, Deserialize, JsonSchema)]
pub struct EvalOutcome {
    /// The eval's label (name or criterion).
    pub label: String,
    /// Whether the eval passed.
    pub passed: bool,
    /// Kind-specific verdict detail.
    pub detail: EvalDetail,
    /// The judge's stated reason.
    pub reason: String,
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn numeric_threshold_gte_passes_at_boundary() {
        let eval = Eval::Numeric {
            criterion: "polite".into(),
            min: 0.0,
            max: 10.0,
            threshold: 7.0,
            comparator: Comparator::Gte,
            name: None,
        };
        let outcome = eval.outcome(&JudgeValue::Number(7.0), "ok".into()).unwrap();
        assert!(outcome.passed);
    }

    #[test]
    fn numeric_value_is_clamped_to_scale() {
        let eval = Eval::Numeric {
            criterion: "x".into(),
            min: 0.0,
            max: 10.0,
            threshold: 9.0,
            comparator: Comparator::Gte,
            name: None,
        };
        // Judge over-reports 12 -> clamped to 10, still passes.
        let outcome = eval
            .outcome(&JudgeValue::Number(12.0), String::new())
            .unwrap();
        assert!(outcome.passed);
        assert!(matches!(
            outcome.detail,
            EvalDetail::Numeric { value, .. } if (value - 10.0).abs() < f64::EPSILON
        ));
    }

    #[test]
    fn boolean_expected_false_inverts() {
        let eval = Eval::Boolean {
            criterion: "leaks a secret".into(),
            expected: false,
            name: None,
        };
        let pass = eval
            .outcome(&JudgeValue::Bool(false), String::new())
            .unwrap();
        assert!(pass.passed);
        let fail = eval
            .outcome(&JudgeValue::Bool(true), String::new())
            .unwrap();
        assert!(!fail.passed);
    }

    #[test]
    fn kind_mismatch_is_provider_error() {
        let eval = Eval::Boolean {
            criterion: "x".into(),
            expected: true,
            name: None,
        };
        assert!(eval
            .outcome(&JudgeValue::Number(1.0), String::new())
            .is_err());
    }

    #[test]
    fn degenerate_numeric_scale_is_invalid() {
        let eval = Eval::Numeric {
            criterion: "x".into(),
            min: 5.0,
            max: 5.0,
            threshold: 5.0,
            comparator: Comparator::Gte,
            name: None,
        };
        assert!(eval.validate().is_err());
    }

    #[test]
    fn comparator_parses_from_symbol() {
        let c: Comparator = serde_yaml::from_str("\">=\"").unwrap();
        assert_eq!(c, Comparator::Gte);
        let c: Comparator = serde_yaml::from_str("lt").unwrap();
        assert_eq!(c, Comparator::Lt);
    }

    #[test]
    fn every_comparator_satisfied_and_symbol() {
        assert!(Comparator::Gte.satisfied(5.0, 5.0));
        assert!(Comparator::Gt.satisfied(6.0, 5.0));
        assert!(!Comparator::Gt.satisfied(5.0, 5.0));
        assert!(Comparator::Lte.satisfied(5.0, 5.0));
        assert!(Comparator::Lt.satisfied(4.0, 5.0));
        assert!(!Comparator::Lt.satisfied(5.0, 5.0));
        assert_eq!(Comparator::Gte.symbol(), ">=");
        assert_eq!(Comparator::Gt.symbol(), ">");
        assert_eq!(Comparator::Lte.symbol(), "<=");
        assert_eq!(Comparator::Lt.symbol(), "<");
    }

    #[test]
    fn criterion_and_label_for_both_kinds() {
        let bool_named = Eval::Boolean {
            criterion: "is polite".into(),
            expected: true,
            name: Some("politeness".into()),
        };
        assert_eq!(bool_named.criterion(), "is polite");
        assert_eq!(bool_named.label(), "politeness");

        let numeric_unnamed = Eval::Numeric {
            criterion: "warmth".into(),
            min: 0.0,
            max: 10.0,
            threshold: 5.0,
            comparator: Comparator::Gte,
            name: None,
        };
        assert_eq!(numeric_unnamed.criterion(), "warmth");
        // Falls back to the criterion when unnamed.
        assert_eq!(numeric_unnamed.label(), "warmth");
    }

    #[test]
    fn validate_rejects_empty_criterion_and_out_of_range_threshold() {
        let empty = Eval::Boolean {
            criterion: "   ".into(),
            expected: true,
            name: None,
        };
        assert!(empty.validate().is_err());

        let bad_threshold = Eval::Numeric {
            criterion: "x".into(),
            min: 0.0,
            max: 10.0,
            threshold: 11.0,
            comparator: Comparator::Gte,
            name: None,
        };
        assert!(bad_threshold.validate().is_err());

        // A well-formed numeric eval validates.
        let ok = Eval::Numeric {
            criterion: "x".into(),
            min: 0.0,
            max: 10.0,
            threshold: 7.0,
            comparator: Comparator::Gte,
            name: None,
        };
        ok.validate().unwrap();
    }

    #[test]
    fn outcome_rejects_numeric_eval_with_boolean_verdict() {
        let eval = Eval::Numeric {
            criterion: "x".into(),
            min: 0.0,
            max: 10.0,
            threshold: 5.0,
            comparator: Comparator::Gte,
            name: None,
        };
        assert!(eval
            .outcome(&JudgeValue::Bool(true), String::new())
            .is_err());
    }

    #[test]
    fn eval_detail_summary_for_both_kinds() {
        let boolean = EvalDetail::Boolean {
            value: true,
            expected: false,
        };
        assert_eq!(boolean.summary(), "true (expected false)");
        let numeric = EvalDetail::Numeric {
            value: 8.0,
            threshold: 7.0,
            comparator: Comparator::Gte,
        };
        assert_eq!(numeric.summary(), "8 >= 7");
    }

    #[test]
    fn judge_value_deserializes_untagged() {
        let b: JudgeValue = serde_json::from_str("true").unwrap();
        assert!(matches!(b, JudgeValue::Bool(true)));
        let n: JudgeValue = serde_json::from_str("3.5").unwrap();
        assert!(matches!(n, JudgeValue::Number(v) if (v - 3.5).abs() < 1e-9));
    }
}