Versions in this module Expand all Collapse all v0 v0.8.0 Jul 10, 2026 v0.7.0 Jul 7, 2026 Changes in this version type Case + CatalogDir string + ExpectRecall string + Recall *CaseRecall + Workload *CaseWorkload + type CaseRecall struct + MarginGap float64 + MinScore float64 + OutcomeFloor float64 + OutcomePrior float64 + RequireWorkloadMatch bool + SoloFloor float64 + type CaseWorkload struct + Name string + Namespace string type Result + RecallFired bool + RecallShortCircuit bool v0.6.1 Jul 3, 2026 v0.6.0 Jul 3, 2026 v0.5.0 Jul 3, 2026 v0.4.0 Jul 2, 2026 v0.3.0 Jul 2, 2026 Changes in this version type Case + GroundTruth *GroundTruth + type CompareCaseRow struct + Coverage float64 + Flaky bool + Name string + PassRate float64 + Reached bool + Runs int + type CompareSpec struct + Judge *JudgeSpec + Models []ModelEntry + func LoadCompareSpec(path string) (CompareSpec, error) + type ComparedCase struct + Name string + Runs []ComparedRun + type ComparedRun struct + Coverage Coverage + Graded bool + Result Result + Verdict Verdict + type ComparisonReport struct + At string + Judge string + Models []ModelComparison + N int + func NewComparisonReport(at string, n int, judge string, models []ModelComparison) ComparisonReport + func ParseComparisonReport(b []byte) (ComparisonReport, error) + func (r ComparisonReport) JSON() ([]byte, error) + func (r ComparisonReport) Markdown() string + type ComparisonRunner struct + Judge Judge + Log *slog.Logger + Model providers.ModelProvider + func (cr *ComparisonRunner) RunCases(ctx context.Context, cases []Case, n int) []ComparedCase + type CountingModel struct + Inner providers.ModelProvider + func (c *CountingModel) Complete(ctx context.Context, req providers.CompletionRequest) (providers.CompletionResponse, error) + func (c *CountingModel) Total() providers.Usage + type JudgeSpec struct + APIKeyEnv string + BaseURL string + Model string + Provider string + type ModelComparison struct + Cases []CompareCaseRow + ConfidentWrong int + CostUSD *float64 + CoverageMedian float64 + Effort string + GradedRuns int + InputTokens int + Model string + Name string + OutputTokens int + PassRate float64 + Provider string + Reached int + RubricMedian map[string]float64 + Total int + func AggregateModel(entry ModelEntry, cases []ComparedCase, usage providers.Usage) ModelComparison + type ModelEntry struct + APIKeyEnv string + BaseURL string + Effort string + Model string + Name string + Prices *Prices + Provider string + type Prices struct + InputUSD float64 + OutputUSD float64 v0.2.0 Jun 28, 2026 v0.1.0 Jun 26, 2026 Changes in this version + var Rubric = []Dimension + func GateError(c Campaign, failUnder float64) error + func RubricMax() int + func WriteCase(dir string, c Case) error + type Call struct + Args string + Err string + Name string + Output string + type Campaign struct + Aggregates []CaseAggregate + N int + func (c Campaign) FlakyNames() []string + func (c Campaign) JSON() ([]byte, error) + func (c Campaign) PassRate() float64 + func (c Campaign) ReachedCases() int + type Case struct + Expected Expected + Name string + Prompt string + Tools map[string]string + func Load(dir string) ([]Case, error) + func RecordedCase(scn Scenario, calls []Call) Case + type CaseAggregate struct + Confidence float64 + Flaky bool + Missing []string + Name string + OverClaimed []string + PassRate float64 + Reached bool + Runs int + type Coverage struct + Bonus []string + CrossSignal bool + Missing []string + Ratio float64 + ToolErrors []string + Touched []string + func ScoreCoverage(expected, optional []string, calls []Call) Coverage + type Dimension struct + Key string + Max int + type Expected struct + Distractors []string + MinConfidence float64 + MustContain []string + RootCauseEntities []string + type GroundTruth struct + ExpectedAction string + ExpectedSources []string + MustReachRoot bool + OptionalSources []string + RootCause string + type Judge interface + Grade func(ctx context.Context, scn Scenario, inv providers.Investigation) (Verdict, error) + type LiveReport struct + At string + N int + Passed int + Ran int + Results []LiveResult + Skipped int + func NewLiveReport(at string, n int, results []LiveResult) LiveReport + func (rep LiveReport) JSON() []byte + func (rep LiveReport) Markdown() string + func (rep LiveReport) RegressionsVS(prev LiveReport) []string + type LiveResult struct + ConfidentWrong bool + CoverageRatio float64 + DimMedian map[string]int + DimVariance map[string]float64 + Flaky bool + Pass bool + Runs []RunOutcome + Scenario string + SkipReason string + Skipped bool + ToolErrors []string + type LiveRunner struct + BaseTools []investigate.Tool + Judge Judge + Log *slog.Logger + Model providers.ModelProvider + N int + OnRecord func(Scenario, []Call) + Recall *investigate.Recall + Steps StepRunner + func (lr *LiveRunner) RunScenario(ctx context.Context, scn Scenario) LiveResult + type ModelJudge struct + Model providers.ModelProvider + func (j ModelJudge) Grade(ctx context.Context, scn Scenario, inv providers.Investigation) (Verdict, error) + type Recorder struct + func (r *Recorder) Calls() []Call + type Report struct + Results []Result + func (r Report) Passed() int + func (r Report) RCARate() float64 + type Result struct + Confidence float64 + Missing []string + Name string + OverClaimed []string + Pass bool + func Score(name string, inv providers.Investigation, exp Expected) Result + type RunOutcome struct + Coverage Coverage + Investigation providers.Investigation + Verdict Verdict + type Runner struct + Log *slog.Logger + Model providers.ModelProvider + func (r *Runner) Run(ctx context.Context, cases []Case) Report + func (r *Runner) RunN(ctx context.Context, cases []Case, n int) Campaign + type Scenario struct + Category string + Description string + GroundTruth GroundTruth + ID string + Invasive bool + Precheck string + Setup []string + Teardown []string + Trigger Trigger + func LoadScenarios(dir string) ([]Scenario, error) + type StepRunner interface + Run func(ctx context.Context, step string) error + type Trigger struct + Mode string + Namespace string + Symptom string + type Verdict struct + ConfidentWrong bool + Rationale string + Scores map[string]int + func (v Verdict) Total() int
Go to main page