@tangle-network/agent-eval 0.95.0 → 0.96.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +17 -0
- package/dist/adapters/http.d.ts +17 -10
- package/dist/adapters/langchain.d.ts +14 -7
- package/dist/adapters/otel.d.ts +25 -13
- package/dist/{rl/adversarial.d.ts → adversarial-DIVcDoI_.d.ts} +7 -6
- package/dist/analyst/index.d.ts +236 -28
- package/dist/analyst/index.js +1 -1
- package/dist/{trace-analyst/analyst.d.ts → analyst-C8HHvfJp.d.ts} +14 -12
- package/dist/{contract/analyze-runs.d.ts → analyze-runs-DtT6F_6T.d.ts} +10 -9
- package/dist/authenticity/index.d.ts +16 -15
- package/dist/{baseline.d.ts → baseline-Bbid3WoO.d.ts} +44 -8
- package/dist/belief-state/index.d.ts +605 -14
- package/dist/benchmarks/index.d.ts +5 -23
- package/dist/builder-eval/index.d.ts +250 -5
- package/dist/calibration-BPmzuVPk.d.ts +101 -0
- package/dist/campaign/index.d.ts +1452 -38
- package/dist/{chunk-AQ5WQAIV.js → chunk-3NHEO6ZC.js} +2 -2
- package/dist/chunk-3NHEO6ZC.js.map +1 -0
- package/dist/cli.d.ts +0 -2
- package/dist/{contract/intake/code-agent-session.d.ts → code-agent-session-CPHRCb4-.d.ts} +17 -15
- package/dist/contract/index.d.ts +824 -107
- package/dist/contract/index.js +145 -1
- package/dist/contract/index.js.map +1 -1
- package/dist/control-Doncu-B_.d.ts +259 -0
- package/dist/{control-runtime.d.ts → control-runtime-Acf9CGhw.d.ts} +26 -23
- package/dist/control.d.ts +10 -11
- package/dist/corpus-D4YW9UoJ.d.ts +560 -0
- package/dist/{cost-ledger.d.ts → cost-ledger-DuSqlw5B.d.ts} +11 -10
- package/dist/{counterfactual.d.ts → counterfactual-DlOz8PBx.d.ts} +13 -12
- package/dist/{dataset.d.ts → dataset-BbGkaN2I.d.ts} +14 -11
- package/dist/{analyst/registry.d.ts → default-registry-GyE8X5SP.d.ts} +37 -8
- package/dist/diagnose.d.ts +252 -1
- package/dist/{trace/emitter.d.ts → emitter-C2rqGH_l.d.ts} +12 -9
- package/dist/{errors.d.ts → errors-CzMUYo7b.d.ts} +11 -10
- package/dist/failure-cluster-DH9Flgcf.d.ts +76 -0
- package/dist/{feedback-trajectory.d.ts → feedback-trajectory-BxY0cKfs.d.ts} +38 -36
- package/dist/fuzz.d.ts +547 -1
- package/dist/gepa-Dprxvz8r.d.ts +414 -0
- package/dist/governance/index.d.ts +135 -5
- package/dist/harness-optimizer-mOl9XX_O.d.ts +106 -0
- package/dist/hosted/index.d.ts +239 -10
- package/dist/index-_Y4oNOOb.d.ts +159 -0
- package/dist/index.d.ts +5660 -277
- package/dist/index.js +1 -1
- package/dist/{contract/insight-report.d.ts → insight-report-BnRjTibG.d.ts} +19 -16
- package/dist/{trace/integrity.d.ts → integrity-D2t12mMw.d.ts} +14 -11
- package/dist/{judge-calibration.d.ts → judge-calibration-0p2QcWNE.d.ts} +17 -16
- package/dist/{analyst/kind-factory.d.ts → kind-factory-X3eDYbKn.d.ts} +61 -10
- package/dist/knowledge/index.d.ts +103 -3
- package/dist/{llm-client.d.ts → llm-client-Bj7g0rqu.d.ts} +23 -21
- package/dist/matrix/index.d.ts +30 -12
- package/dist/meta-eval/index.d.ts +182 -6
- package/dist/{multi-layer-verifier.d.ts → multi-layer-verifier-DUZXrPDA.d.ts} +15 -12
- package/dist/multishot/index.d.ts +290 -7
- package/dist/{rl/off-policy.d.ts → off-policy-DiwuKKg7.d.ts} +9 -8
- package/dist/openapi.json +1 -1
- package/dist/{meta-eval/outcome-store.d.ts → outcome-store-rnXLEqSn.d.ts} +8 -7
- package/dist/{pareto.d.ts → pareto-E-pembql.d.ts} +10 -9
- package/dist/perf/index.d.ts +119 -13
- package/dist/pipelines/index.d.ts +173 -8
- package/dist/pre-registration-nfUdc9EQ.d.ts +483 -0
- package/dist/prm/index.d.ts +104 -5
- package/dist/provenance-CIxfBnkl.d.ts +426 -0
- package/dist/query-B7GGjRox.d.ts +32 -0
- package/dist/{trace/raw-provider-sink.d.ts → raw-provider-sink-C46HDghv.d.ts} +14 -13
- package/dist/{red-team.d.ts → red-team-BWdoyleI.d.ts} +16 -14
- package/dist/{trace/redact.d.ts → redact-B40YG2M_.d.ts} +8 -7
- package/dist/release-report-pidWUMZ2.d.ts +233 -0
- package/dist/reporting.d.ts +16 -15
- package/dist/{eval-campaign.d.ts → researcher-Jr8ME1dZ.d.ts} +156 -34
- package/dist/rl.d.ts +1193 -1
- package/dist/{prm/rubric.d.ts → rubric-Cc6UHvUb.d.ts} +13 -10
- package/dist/{meta-eval/rubric-predictive-validity.d.ts → rubric-predictive-validity-C2hDKM8Z.d.ts} +11 -8
- package/dist/run-critic-CmMf05uV.d.ts +56 -0
- package/dist/{run-record.d.ts → run-record-CP2ObebC.d.ts} +117 -15
- package/dist/runtime-trajectory-BOUUjI0y.d.ts +49 -0
- package/dist/{trace/schema.d.ts → schema-m0gsnbt3.d.ts} +30 -29
- package/dist/semantic-concept-judge-DSBB2Cfp.d.ts +624 -0
- package/dist/{sequential.d.ts → sequential-5iSVfzl2.d.ts} +10 -9
- package/dist/{series-convergence.d.ts → series-convergence-D5OWMBg6.d.ts} +5 -4
- package/dist/sink-fetch-B1Yg4Til.d.ts +101 -0
- package/dist/{statistics.d.ts → statistics-CCJpTGOS.d.ts} +48 -46
- package/dist/{trace/store.d.ts → store-BcFXE6LG.d.ts} +12 -21
- package/dist/{trace-analyst/types.d.ts → store-C1YxJDEK.d.ts} +74 -18
- package/dist/storyboard/index.d.ts +81 -16
- package/dist/{summary-report.d.ts → summary-report-CInXwsza.d.ts} +160 -23
- package/dist/telemetry/{sink-file.d.ts → file.d.ts} +7 -5
- package/dist/telemetry/index.d.ts +35 -17
- package/dist/{sandbox-harness.d.ts → test-graded-scenario-DeODGLra.d.ts} +60 -15
- package/dist/{locked-jsonl-appender.d.ts → testing-C21CHsq2.d.ts} +4 -3
- package/dist/testing.d.ts +1 -5
- package/dist/traces.d.ts +976 -4
- package/dist/{trajectory.d.ts → trajectory-2TkpSEVh.d.ts} +9 -8
- package/dist/{analyst/types.d.ts → types-B5x54y6n.d.ts} +112 -19
- package/dist/{campaign/types.d.ts → types-BMahhhio.d.ts} +47 -44
- package/dist/{matrix/types.d.ts → types-BUxNaJ8c.d.ts} +11 -9
- package/dist/{types.d.ts → types-C7DGg5ex.d.ts} +33 -31
- package/dist/{verdict.d.ts → verdict-C9MlYujm.d.ts} +3 -2
- package/dist/wire/index.d.ts +570 -13
- package/dist/workflow/index.d.ts +496 -22
- package/package.json +2 -2
- package/dist/action-policy.d.ts +0 -24
- package/dist/action-policy.d.ts.map +0 -1
- package/dist/action-policy.test.d.ts +0 -2
- package/dist/action-policy.test.d.ts.map +0 -1
- package/dist/active-learning.d.ts +0 -41
- package/dist/active-learning.d.ts.map +0 -1
- package/dist/adapters/http.d.ts.map +0 -1
- package/dist/adapters/langchain.d.ts.map +0 -1
- package/dist/adapters/otel.d.ts.map +0 -1
- package/dist/agent-profile-cell.d.ts +0 -101
- package/dist/agent-profile-cell.d.ts.map +0 -1
- package/dist/agent-profile.d.ts +0 -27
- package/dist/agent-profile.d.ts.map +0 -1
- package/dist/agent-profile.test.d.ts +0 -2
- package/dist/agent-profile.test.d.ts.map +0 -1
- package/dist/analyst/adapters.d.ts +0 -62
- package/dist/analyst/adapters.d.ts.map +0 -1
- package/dist/analyst/analyst.test.d.ts +0 -2
- package/dist/analyst/analyst.test.d.ts.map +0 -1
- package/dist/analyst/ax-service.d.ts +0 -27
- package/dist/analyst/ax-service.d.ts.map +0 -1
- package/dist/analyst/behavioral-analyst.d.ts +0 -28
- package/dist/analyst/behavioral-analyst.d.ts.map +0 -1
- package/dist/analyst/chat-client.d.ts +0 -91
- package/dist/analyst/chat-client.d.ts.map +0 -1
- package/dist/analyst/default-registry.d.ts +0 -27
- package/dist/analyst/default-registry.d.ts.map +0 -1
- package/dist/analyst/default-registry.test.d.ts +0 -2
- package/dist/analyst/default-registry.test.d.ts.map +0 -1
- package/dist/analyst/finding-signature.d.ts +0 -48
- package/dist/analyst/finding-signature.d.ts.map +0 -1
- package/dist/analyst/finding-subject.d.ts +0 -146
- package/dist/analyst/finding-subject.d.ts.map +0 -1
- package/dist/analyst/finding-subject.test.d.ts +0 -2
- package/dist/analyst/finding-subject.test.d.ts.map +0 -1
- package/dist/analyst/findings-store.d.ts +0 -75
- package/dist/analyst/findings-store.d.ts.map +0 -1
- package/dist/analyst/index.d.ts.map +0 -1
- package/dist/analyst/kind-factory.d.ts.map +0 -1
- package/dist/analyst/kinds/failure-mode.d.ts +0 -19
- package/dist/analyst/kinds/failure-mode.d.ts.map +0 -1
- package/dist/analyst/kinds/improvement.d.ts +0 -23
- package/dist/analyst/kinds/improvement.d.ts.map +0 -1
- package/dist/analyst/kinds/index.d.ts +0 -22
- package/dist/analyst/kinds/index.d.ts.map +0 -1
- package/dist/analyst/kinds/kinds.test.d.ts +0 -2
- package/dist/analyst/kinds/kinds.test.d.ts.map +0 -1
- package/dist/analyst/kinds/knowledge-gap.d.ts +0 -28
- package/dist/analyst/kinds/knowledge-gap.d.ts.map +0 -1
- package/dist/analyst/kinds/knowledge-poisoning.d.ts +0 -22
- package/dist/analyst/kinds/knowledge-poisoning.d.ts.map +0 -1
- package/dist/analyst/kinds/skill-usage.d.ts +0 -84
- package/dist/analyst/kinds/skill-usage.d.ts.map +0 -1
- package/dist/analyst/kinds/skill-usage.test.d.ts +0 -2
- package/dist/analyst/kinds/skill-usage.test.d.ts.map +0 -1
- package/dist/analyst/parse-tolerant.d.ts +0 -26
- package/dist/analyst/parse-tolerant.d.ts.map +0 -1
- package/dist/analyst/parse-tolerant.test.d.ts +0 -2
- package/dist/analyst/parse-tolerant.test.d.ts.map +0 -1
- package/dist/analyst/registry.budget.test.d.ts +0 -2
- package/dist/analyst/registry.budget.test.d.ts.map +0 -1
- package/dist/analyst/registry.d.ts.map +0 -1
- package/dist/analyst/steer-firewall.d.ts +0 -35
- package/dist/analyst/steer-firewall.d.ts.map +0 -1
- package/dist/analyst/steer-firewall.test.d.ts +0 -2
- package/dist/analyst/steer-firewall.test.d.ts.map +0 -1
- package/dist/analyst/structure-findings.d.ts +0 -37
- package/dist/analyst/structure-findings.d.ts.map +0 -1
- package/dist/analyst/structure-findings.test.d.ts +0 -2
- package/dist/analyst/structure-findings.test.d.ts.map +0 -1
- package/dist/analyst/tool-groups.d.ts +0 -34
- package/dist/analyst/tool-groups.d.ts.map +0 -1
- package/dist/analyst/types.d.ts.map +0 -1
- package/dist/anti-slop.d.ts +0 -59
- package/dist/anti-slop.d.ts.map +0 -1
- package/dist/artifact-validator.d.ts +0 -74
- package/dist/artifact-validator.d.ts.map +0 -1
- package/dist/attestation.d.ts +0 -63
- package/dist/attestation.d.ts.map +0 -1
- package/dist/attestation.test.d.ts +0 -2
- package/dist/attestation.test.d.ts.map +0 -1
- package/dist/authenticity/index.d.ts.map +0 -1
- package/dist/authenticity/index.test.d.ts +0 -2
- package/dist/authenticity/index.test.d.ts.map +0 -1
- package/dist/auto-pr.d.ts +0 -120
- package/dist/auto-pr.d.ts.map +0 -1
- package/dist/baseline.d.ts.map +0 -1
- package/dist/behavior-dsl.d.ts +0 -73
- package/dist/behavior-dsl.d.ts.map +0 -1
- package/dist/belief-state/calibration.d.ts +0 -10
- package/dist/belief-state/calibration.d.ts.map +0 -1
- package/dist/belief-state/calibration.test.d.ts +0 -2
- package/dist/belief-state/calibration.test.d.ts.map +0 -1
- package/dist/belief-state/code-agent-corpus.d.ts +0 -66
- package/dist/belief-state/code-agent-corpus.d.ts.map +0 -1
- package/dist/belief-state/code-agent-corpus.test.d.ts +0 -2
- package/dist/belief-state/code-agent-corpus.test.d.ts.map +0 -1
- package/dist/belief-state/code-agent-evidence.d.ts +0 -22
- package/dist/belief-state/code-agent-evidence.d.ts.map +0 -1
- package/dist/belief-state/code-agent-evidence.test.d.ts +0 -2
- package/dist/belief-state/code-agent-evidence.test.d.ts.map +0 -1
- package/dist/belief-state/extract.d.ts +0 -7
- package/dist/belief-state/extract.d.ts.map +0 -1
- package/dist/belief-state/extract.test.d.ts +0 -2
- package/dist/belief-state/extract.test.d.ts.map +0 -1
- package/dist/belief-state/index.d.ts.map +0 -1
- package/dist/belief-state/ope.d.ts +0 -17
- package/dist/belief-state/ope.d.ts.map +0 -1
- package/dist/belief-state/ope.test.d.ts +0 -2
- package/dist/belief-state/ope.test.d.ts.map +0 -1
- package/dist/belief-state/phase0-measurement.d.ts +0 -55
- package/dist/belief-state/phase0-measurement.d.ts.map +0 -1
- package/dist/belief-state/report.d.ts +0 -17
- package/dist/belief-state/report.d.ts.map +0 -1
- package/dist/belief-state/report.test.d.ts +0 -2
- package/dist/belief-state/report.test.d.ts.map +0 -1
- package/dist/belief-state/research-evidence.d.ts +0 -23
- package/dist/belief-state/research-evidence.d.ts.map +0 -1
- package/dist/belief-state/research-evidence.test.d.ts +0 -2
- package/dist/belief-state/research-evidence.test.d.ts.map +0 -1
- package/dist/belief-state/runtime-benchmark-corpus.d.ts +0 -32
- package/dist/belief-state/runtime-benchmark-corpus.d.ts.map +0 -1
- package/dist/belief-state/runtime-hooks.d.ts +0 -87
- package/dist/belief-state/runtime-hooks.d.ts.map +0 -1
- package/dist/belief-state/runtime-hooks.test.d.ts +0 -2
- package/dist/belief-state/runtime-hooks.test.d.ts.map +0 -1
- package/dist/belief-state/selective.d.ts +0 -15
- package/dist/belief-state/selective.d.ts.map +0 -1
- package/dist/belief-state/selective.test.d.ts +0 -2
- package/dist/belief-state/selective.test.d.ts.map +0 -1
- package/dist/belief-state/shadow-probe.d.ts +0 -80
- package/dist/belief-state/shadow-probe.d.ts.map +0 -1
- package/dist/belief-state/shadow-probe.test.d.ts +0 -2
- package/dist/belief-state/shadow-probe.test.d.ts.map +0 -1
- package/dist/belief-state/types.d.ts +0 -195
- package/dist/belief-state/types.d.ts.map +0 -1
- package/dist/belief-state/types.test.d.ts +0 -2
- package/dist/belief-state/types.test.d.ts.map +0 -1
- package/dist/benchmark.d.ts +0 -14
- package/dist/benchmark.d.ts.map +0 -1
- package/dist/benchmarks/index.d.ts.map +0 -1
- package/dist/benchmarks/routing/dataset.d.ts +0 -34
- package/dist/benchmarks/routing/dataset.d.ts.map +0 -1
- package/dist/benchmarks/routing/index.d.ts +0 -34
- package/dist/benchmarks/routing/index.d.ts.map +0 -1
- package/dist/benchmarks/types.d.ts +0 -49
- package/dist/benchmarks/types.d.ts.map +0 -1
- package/dist/bisector.d.ts +0 -81
- package/dist/bisector.d.ts.map +0 -1
- package/dist/budget-guard.d.ts +0 -31
- package/dist/budget-guard.d.ts.map +0 -1
- package/dist/builder-eval/builder-session.d.ts +0 -111
- package/dist/builder-eval/builder-session.d.ts.map +0 -1
- package/dist/builder-eval/correlation.d.ts +0 -32
- package/dist/builder-eval/correlation.d.ts.map +0 -1
- package/dist/builder-eval/index.d.ts.map +0 -1
- package/dist/builder-eval/project-registry.d.ts +0 -51
- package/dist/builder-eval/project-registry.d.ts.map +0 -1
- package/dist/builder-eval/three-layer-eval.d.ts +0 -55
- package/dist/builder-eval/three-layer-eval.d.ts.map +0 -1
- package/dist/campaign/analyst-surface.d.ts +0 -108
- package/dist/campaign/analyst-surface.d.ts.map +0 -1
- package/dist/campaign/analyst-surface.test.d.ts +0 -2
- package/dist/campaign/analyst-surface.test.d.ts.map +0 -1
- package/dist/campaign/auto-pr.d.ts +0 -46
- package/dist/campaign/auto-pr.d.ts.map +0 -1
- package/dist/campaign/distillation/agreement-judge.d.ts +0 -69
- package/dist/campaign/distillation/agreement-judge.d.ts.map +0 -1
- package/dist/campaign/distillation/cli.d.ts +0 -35
- package/dist/campaign/distillation/cli.d.ts.map +0 -1
- package/dist/campaign/distillation/distillation.test.d.ts +0 -2
- package/dist/campaign/distillation/distillation.test.d.ts.map +0 -1
- package/dist/campaign/distillation/gold-scenarios.d.ts +0 -54
- package/dist/campaign/distillation/gold-scenarios.d.ts.map +0 -1
- package/dist/campaign/distillation/run-distillation.d.ts +0 -119
- package/dist/campaign/distillation/run-distillation.d.ts.map +0 -1
- package/dist/campaign/gates/compose.d.ts +0 -12
- package/dist/campaign/gates/compose.d.ts.map +0 -1
- package/dist/campaign/gates/default-production-gate.d.ts +0 -58
- package/dist/campaign/gates/default-production-gate.d.ts.map +0 -1
- package/dist/campaign/gates/heldout-gate.d.ts +0 -12
- package/dist/campaign/gates/heldout-gate.d.ts.map +0 -1
- package/dist/campaign/gates/promotion-policy.d.ts +0 -125
- package/dist/campaign/gates/promotion-policy.d.ts.map +0 -1
- package/dist/campaign/gates/promotion-policy.test.d.ts +0 -2
- package/dist/campaign/gates/promotion-policy.test.d.ts.map +0 -1
- package/dist/campaign/gates/sequential.d.ts +0 -146
- package/dist/campaign/gates/sequential.d.ts.map +0 -1
- package/dist/campaign/gates/sequential.test.d.ts +0 -2
- package/dist/campaign/gates/sequential.test.d.ts.map +0 -1
- package/dist/campaign/gates/statistical-heldout.d.ts +0 -99
- package/dist/campaign/gates/statistical-heldout.d.ts.map +0 -1
- package/dist/campaign/gates/statistical-heldout.test.d.ts +0 -2
- package/dist/campaign/gates/statistical-heldout.test.d.ts.map +0 -1
- package/dist/campaign/index.d.ts.map +0 -1
- package/dist/campaign/labeled-store/fs-adapter.d.ts +0 -59
- package/dist/campaign/labeled-store/fs-adapter.d.ts.map +0 -1
- package/dist/campaign/presets/compare-proposers.d.ts +0 -146
- package/dist/campaign/presets/compare-proposers.d.ts.map +0 -1
- package/dist/campaign/presets/playback.d.ts +0 -120
- package/dist/campaign/presets/playback.d.ts.map +0 -1
- package/dist/campaign/presets/playback.test.d.ts +0 -2
- package/dist/campaign/presets/playback.test.d.ts.map +0 -1
- package/dist/campaign/presets/run-eval.d.ts +0 -14
- package/dist/campaign/presets/run-eval.d.ts.map +0 -1
- package/dist/campaign/presets/run-improvement-loop.d.ts +0 -63
- package/dist/campaign/presets/run-improvement-loop.d.ts.map +0 -1
- package/dist/campaign/presets/run-improvement-loop.test.d.ts +0 -2
- package/dist/campaign/presets/run-improvement-loop.test.d.ts.map +0 -1
- package/dist/campaign/presets/run-optimization.d.ts +0 -92
- package/dist/campaign/presets/run-optimization.d.ts.map +0 -1
- package/dist/campaign/presets/run-profile-matrix.d.ts +0 -151
- package/dist/campaign/presets/run-profile-matrix.d.ts.map +0 -1
- package/dist/campaign/presets/run-skill-opt.d.ts +0 -96
- package/dist/campaign/presets/run-skill-opt.d.ts.map +0 -1
- package/dist/campaign/proposers/_findings-text.d.ts +0 -22
- package/dist/campaign/proposers/_findings-text.d.ts.map +0 -1
- package/dist/campaign/proposers/ace.d.ts +0 -33
- package/dist/campaign/proposers/ace.d.ts.map +0 -1
- package/dist/campaign/proposers/ace.test.d.ts +0 -2
- package/dist/campaign/proposers/ace.test.d.ts.map +0 -1
- package/dist/campaign/proposers/analysis-edit.d.ts +0 -32
- package/dist/campaign/proposers/analysis-edit.d.ts.map +0 -1
- package/dist/campaign/proposers/evolutionary.d.ts +0 -20
- package/dist/campaign/proposers/evolutionary.d.ts.map +0 -1
- package/dist/campaign/proposers/fapo.d.ts +0 -120
- package/dist/campaign/proposers/fapo.d.ts.map +0 -1
- package/dist/campaign/proposers/gepa.d.ts +0 -86
- package/dist/campaign/proposers/gepa.d.ts.map +0 -1
- package/dist/campaign/proposers/halo.d.ts +0 -44
- package/dist/campaign/proposers/halo.d.ts.map +0 -1
- package/dist/campaign/proposers/halo.test.d.ts +0 -2
- package/dist/campaign/proposers/halo.test.d.ts.map +0 -1
- package/dist/campaign/proposers/memory.d.ts +0 -47
- package/dist/campaign/proposers/memory.d.ts.map +0 -1
- package/dist/campaign/proposers/memory.test.d.ts +0 -2
- package/dist/campaign/proposers/memory.test.d.ts.map +0 -1
- package/dist/campaign/proposers/skill-opt.d.ts +0 -88
- package/dist/campaign/proposers/skill-opt.d.ts.map +0 -1
- package/dist/campaign/proposers/trace-analyst.d.ts +0 -48
- package/dist/campaign/proposers/trace-analyst.d.ts.map +0 -1
- package/dist/campaign/proposers/trace-analyst.test.d.ts +0 -2
- package/dist/campaign/proposers/trace-analyst.test.d.ts.map +0 -1
- package/dist/campaign/provenance.d.ts +0 -185
- package/dist/campaign/provenance.d.ts.map +0 -1
- package/dist/campaign/run-campaign.d.ts +0 -90
- package/dist/campaign/run-campaign.d.ts.map +0 -1
- package/dist/campaign/score-utils.d.ts +0 -26
- package/dist/campaign/score-utils.d.ts.map +0 -1
- package/dist/campaign/skill-patch.d.ts +0 -62
- package/dist/campaign/skill-patch.d.ts.map +0 -1
- package/dist/campaign/storage.d.ts +0 -38
- package/dist/campaign/storage.d.ts.map +0 -1
- package/dist/campaign/types.d.ts.map +0 -1
- package/dist/campaign/worktree/index.d.ts +0 -53
- package/dist/campaign/worktree/index.d.ts.map +0 -1
- package/dist/canary.d.ts +0 -101
- package/dist/canary.d.ts.map +0 -1
- package/dist/causal-attribution.d.ts +0 -45
- package/dist/causal-attribution.d.ts.map +0 -1
- package/dist/chunk-AQ5WQAIV.js.map +0 -1
- package/dist/ci-gate.d.ts +0 -44
- package/dist/ci-gate.d.ts.map +0 -1
- package/dist/cli.d.ts.map +0 -1
- package/dist/client.d.ts +0 -77
- package/dist/client.d.ts.map +0 -1
- package/dist/client.test.d.ts +0 -2
- package/dist/client.test.d.ts.map +0 -1
- package/dist/command-runner.d.ts +0 -74
- package/dist/command-runner.d.ts.map +0 -1
- package/dist/command-runner.test.d.ts +0 -2
- package/dist/command-runner.test.d.ts.map +0 -1
- package/dist/completion-verifier.d.ts +0 -147
- package/dist/completion-verifier.d.ts.map +0 -1
- package/dist/completion-verifier.test.d.ts +0 -9
- package/dist/completion-verifier.test.d.ts.map +0 -1
- package/dist/concurrency.d.ts +0 -23
- package/dist/concurrency.d.ts.map +0 -1
- package/dist/contamination-guard.d.ts +0 -81
- package/dist/contamination-guard.d.ts.map +0 -1
- package/dist/contract/analyze-runs.d.ts.map +0 -1
- package/dist/contract/define-agent-eval.d.ts +0 -52
- package/dist/contract/define-agent-eval.d.ts.map +0 -1
- package/dist/contract/diff.d.ts +0 -114
- package/dist/contract/diff.d.ts.map +0 -1
- package/dist/contract/index.d.ts.map +0 -1
- package/dist/contract/insight-report.d.ts.map +0 -1
- package/dist/contract/insight-types-fwd.d.ts +0 -7
- package/dist/contract/insight-types-fwd.d.ts.map +0 -1
- package/dist/contract/intake/agent-trace.d.ts +0 -97
- package/dist/contract/intake/agent-trace.d.ts.map +0 -1
- package/dist/contract/intake/code-agent-session.d.ts.map +0 -1
- package/dist/contract/intake/feedback-table.d.ts +0 -87
- package/dist/contract/intake/feedback-table.d.ts.map +0 -1
- package/dist/contract/intake/index.d.ts +0 -22
- package/dist/contract/intake/index.d.ts.map +0 -1
- package/dist/contract/intake/otel-spans.d.ts +0 -33
- package/dist/contract/intake/otel-spans.d.ts.map +0 -1
- package/dist/contract/self-improve.d.ts +0 -284
- package/dist/contract/self-improve.d.ts.map +0 -1
- package/dist/control-runtime.d.ts.map +0 -1
- package/dist/control-runtime.test.d.ts +0 -2
- package/dist/control-runtime.test.d.ts.map +0 -1
- package/dist/control.d.ts.map +0 -1
- package/dist/convergence.d.ts +0 -29
- package/dist/convergence.d.ts.map +0 -1
- package/dist/cost-ledger.d.ts.map +0 -1
- package/dist/cost-ledger.test.d.ts +0 -2
- package/dist/cost-ledger.test.d.ts.map +0 -1
- package/dist/cost-report.d.ts +0 -43
- package/dist/cost-report.d.ts.map +0 -1
- package/dist/cost-report.test.d.ts +0 -2
- package/dist/cost-report.test.d.ts.map +0 -1
- package/dist/cost-tracker.d.ts +0 -76
- package/dist/cost-tracker.d.ts.map +0 -1
- package/dist/counterfactual.d.ts.map +0 -1
- package/dist/cross-trace-diff.d.ts +0 -56
- package/dist/cross-trace-diff.d.ts.map +0 -1
- package/dist/dataset.d.ts.map +0 -1
- package/dist/deploy-gate-layer.d.ts +0 -125
- package/dist/deploy-gate-layer.d.ts.map +0 -1
- package/dist/deploy-gate-layer.test.d.ts +0 -2
- package/dist/deploy-gate-layer.test.d.ts.map +0 -1
- package/dist/description-length-gate.d.ts +0 -119
- package/dist/description-length-gate.d.ts.map +0 -1
- package/dist/detectors/edge.test.d.ts +0 -2
- package/dist/detectors/edge.test.d.ts.map +0 -1
- package/dist/detectors/index.d.ts +0 -81
- package/dist/detectors/index.d.ts.map +0 -1
- package/dist/detectors/index.test.d.ts +0 -2
- package/dist/detectors/index.test.d.ts.map +0 -1
- package/dist/diagnose/causal-sweep.d.ts +0 -100
- package/dist/diagnose/causal-sweep.d.ts.map +0 -1
- package/dist/diagnose/index.d.ts +0 -36
- package/dist/diagnose/index.d.ts.map +0 -1
- package/dist/diagnose/remediation.d.ts +0 -68
- package/dist/diagnose/remediation.d.ts.map +0 -1
- package/dist/diagnose/repair.d.ts +0 -77
- package/dist/diagnose/repair.d.ts.map +0 -1
- package/dist/discover-personas.d.ts +0 -35
- package/dist/discover-personas.d.ts.map +0 -1
- package/dist/driver.d.ts +0 -95
- package/dist/driver.d.ts.map +0 -1
- package/dist/driver.test.d.ts +0 -8
- package/dist/driver.test.d.ts.map +0 -1
- package/dist/dual-agent-bench.d.ts +0 -81
- package/dist/dual-agent-bench.d.ts.map +0 -1
- package/dist/error-count-extractor.d.ts +0 -47
- package/dist/error-count-extractor.d.ts.map +0 -1
- package/dist/error-count-extractor.test.d.ts +0 -2
- package/dist/error-count-extractor.test.d.ts.map +0 -1
- package/dist/errors.d.ts.map +0 -1
- package/dist/eval-campaign.d.ts.map +0 -1
- package/dist/eval-campaign.test.d.ts +0 -2
- package/dist/eval-campaign.test.d.ts.map +0 -1
- package/dist/eval-tools.d.ts +0 -55
- package/dist/eval-tools.d.ts.map +0 -1
- package/dist/eval-trace-store.d.ts +0 -107
- package/dist/eval-trace-store.d.ts.map +0 -1
- package/dist/eval-trace-store.test.d.ts +0 -2
- package/dist/eval-trace-store.test.d.ts.map +0 -1
- package/dist/executor.d.ts +0 -38
- package/dist/executor.d.ts.map +0 -1
- package/dist/executor.test.d.ts +0 -10
- package/dist/executor.test.d.ts.map +0 -1
- package/dist/experiment-tracker.d.ts +0 -178
- package/dist/experiment-tracker.d.ts.map +0 -1
- package/dist/experiment-tracker.test.d.ts +0 -2
- package/dist/experiment-tracker.test.d.ts.map +0 -1
- package/dist/failure-taxonomy.d.ts +0 -38
- package/dist/failure-taxonomy.d.ts.map +0 -1
- package/dist/feedback-trajectory.d.ts.map +0 -1
- package/dist/feedback-trajectory.test.d.ts +0 -2
- package/dist/feedback-trajectory.test.d.ts.map +0 -1
- package/dist/flow-layer.d.ts +0 -90
- package/dist/flow-layer.d.ts.map +0 -1
- package/dist/flow-layer.test.d.ts +0 -2
- package/dist/flow-layer.test.d.ts.map +0 -1
- package/dist/fuzz/capsule.d.ts +0 -46
- package/dist/fuzz/capsule.d.ts.map +0 -1
- package/dist/fuzz/cube.d.ts +0 -36
- package/dist/fuzz/cube.d.ts.map +0 -1
- package/dist/fuzz/explorer-cost.test.d.ts +0 -2
- package/dist/fuzz/explorer-cost.test.d.ts.map +0 -1
- package/dist/fuzz/explorer.d.ts +0 -64
- package/dist/fuzz/explorer.d.ts.map +0 -1
- package/dist/fuzz/fuzz-agent.d.ts +0 -16
- package/dist/fuzz/fuzz-agent.d.ts.map +0 -1
- package/dist/fuzz/fuzz-agent.test.d.ts +0 -2
- package/dist/fuzz/fuzz-agent.test.d.ts.map +0 -1
- package/dist/fuzz/gates.d.ts +0 -33
- package/dist/fuzz/gates.d.ts.map +0 -1
- package/dist/fuzz/index.d.ts +0 -26
- package/dist/fuzz/index.d.ts.map +0 -1
- package/dist/fuzz/policies.d.ts +0 -28
- package/dist/fuzz/policies.d.ts.map +0 -1
- package/dist/fuzz/tools.d.ts +0 -20
- package/dist/fuzz/tools.d.ts.map +0 -1
- package/dist/fuzz/types.d.ts +0 -307
- package/dist/fuzz/types.d.ts.map +0 -1
- package/dist/golden-matcher.d.ts +0 -71
- package/dist/golden-matcher.d.ts.map +0 -1
- package/dist/governance/eu-ai-act.d.ts +0 -37
- package/dist/governance/eu-ai-act.d.ts.map +0 -1
- package/dist/governance/index.d.ts.map +0 -1
- package/dist/governance/nist-ai-rmf.d.ts +0 -15
- package/dist/governance/nist-ai-rmf.d.ts.map +0 -1
- package/dist/governance/soc2.d.ts +0 -12
- package/dist/governance/soc2.d.ts.map +0 -1
- package/dist/governance/types.d.ts +0 -66
- package/dist/governance/types.d.ts.map +0 -1
- package/dist/harness-optimizer.d.ts +0 -82
- package/dist/harness-optimizer.d.ts.map +0 -1
- package/dist/held-out-gate.d.ts +0 -135
- package/dist/held-out-gate.d.ts.map +0 -1
- package/dist/hosted/client.d.ts +0 -73
- package/dist/hosted/client.d.ts.map +0 -1
- package/dist/hosted/from-env.test.d.ts +0 -8
- package/dist/hosted/from-env.test.d.ts.map +0 -1
- package/dist/hosted/index.d.ts.map +0 -1
- package/dist/hosted/types.d.ts +0 -159
- package/dist/hosted/types.d.ts.map +0 -1
- package/dist/index.d.ts.map +0 -1
- package/dist/integrity/backend-integrity.d.ts +0 -71
- package/dist/integrity/backend-integrity.d.ts.map +0 -1
- package/dist/integrity/preflight.d.ts +0 -72
- package/dist/integrity/preflight.d.ts.map +0 -1
- package/dist/integrity/preflight.test.d.ts +0 -2
- package/dist/integrity/preflight.test.d.ts.map +0 -1
- package/dist/integrity/single-backend.d.ts +0 -67
- package/dist/integrity/single-backend.d.ts.map +0 -1
- package/dist/intent-match-judge.d.ts +0 -69
- package/dist/intent-match-judge.d.ts.map +0 -1
- package/dist/intent-match-judge.test.d.ts +0 -2
- package/dist/intent-match-judge.test.d.ts.map +0 -1
- package/dist/judge-calibration.d.ts.map +0 -1
- package/dist/judge-ensemble.d.ts +0 -66
- package/dist/judge-ensemble.d.ts.map +0 -1
- package/dist/judge-ensemble.test.d.ts +0 -8
- package/dist/judge-ensemble.test.d.ts.map +0 -1
- package/dist/judge-families.d.ts +0 -38
- package/dist/judge-families.d.ts.map +0 -1
- package/dist/judge-panel.d.ts +0 -65
- package/dist/judge-panel.d.ts.map +0 -1
- package/dist/judge-retry.d.ts +0 -70
- package/dist/judge-retry.d.ts.map +0 -1
- package/dist/judge-runner.d.ts +0 -36
- package/dist/judge-runner.d.ts.map +0 -1
- package/dist/judge-runner.test.d.ts +0 -2
- package/dist/judge-runner.test.d.ts.map +0 -1
- package/dist/judges.d.ts +0 -74
- package/dist/judges.d.ts.map +0 -1
- package/dist/keyword-coverage-judge.d.ts +0 -89
- package/dist/keyword-coverage-judge.d.ts.map +0 -1
- package/dist/keyword-coverage-judge.test.d.ts +0 -2
- package/dist/keyword-coverage-judge.test.d.ts.map +0 -1
- package/dist/knowledge/index.d.ts.map +0 -1
- package/dist/knowledge/readiness.d.ts +0 -26
- package/dist/knowledge/readiness.d.ts.map +0 -1
- package/dist/knowledge/types.d.ts +0 -75
- package/dist/knowledge/types.d.ts.map +0 -1
- package/dist/live-proof.d.ts +0 -62
- package/dist/live-proof.d.ts.map +0 -1
- package/dist/llm-client.d.ts.map +0 -1
- package/dist/llm-client.test.d.ts +0 -2
- package/dist/llm-client.test.d.ts.map +0 -1
- package/dist/locked-jsonl-appender.d.ts.map +0 -1
- package/dist/matrix/aggregation.d.ts +0 -16
- package/dist/matrix/aggregation.d.ts.map +0 -1
- package/dist/matrix/index.d.ts.map +0 -1
- package/dist/matrix/runner.d.ts +0 -15
- package/dist/matrix/runner.d.ts.map +0 -1
- package/dist/matrix/types.d.ts.map +0 -1
- package/dist/meta-eval/calibration.d.ts +0 -47
- package/dist/meta-eval/calibration.d.ts.map +0 -1
- package/dist/meta-eval/correlation-study.d.ts +0 -53
- package/dist/meta-eval/correlation-study.d.ts.map +0 -1
- package/dist/meta-eval/index.d.ts.map +0 -1
- package/dist/meta-eval/outcome-store.d.ts.map +0 -1
- package/dist/meta-eval/rubric-predictive-validity.d.ts.map +0 -1
- package/dist/meta-eval/sentinel.d.ts +0 -169
- package/dist/meta-eval/sentinel.d.ts.map +0 -1
- package/dist/metrics.d.ts +0 -63
- package/dist/metrics.d.ts.map +0 -1
- package/dist/model-seats.d.ts +0 -71
- package/dist/model-seats.d.ts.map +0 -1
- package/dist/model-seats.test.d.ts +0 -2
- package/dist/model-seats.test.d.ts.map +0 -1
- package/dist/muffled-gate-scanner.d.ts +0 -102
- package/dist/muffled-gate-scanner.d.ts.map +0 -1
- package/dist/multi-layer-verifier.d.ts.map +0 -1
- package/dist/multi-layer-verifier.test.d.ts +0 -2
- package/dist/multi-layer-verifier.test.d.ts.map +0 -1
- package/dist/multi-toolchain-layer.d.ts +0 -80
- package/dist/multi-toolchain-layer.d.ts.map +0 -1
- package/dist/multi-toolchain-layer.test.d.ts +0 -2
- package/dist/multi-toolchain-layer.test.d.ts.map +0 -1
- package/dist/multishot/default-tools.d.ts +0 -34
- package/dist/multishot/default-tools.d.ts.map +0 -1
- package/dist/multishot/index.d.ts.map +0 -1
- package/dist/multishot/judges.d.ts +0 -32
- package/dist/multishot/judges.d.ts.map +0 -1
- package/dist/multishot/matrix.d.ts +0 -107
- package/dist/multishot/matrix.d.ts.map +0 -1
- package/dist/multishot/multishot.d.ts +0 -23
- package/dist/multishot/multishot.d.ts.map +0 -1
- package/dist/multishot/router.d.ts +0 -37
- package/dist/multishot/router.d.ts.map +0 -1
- package/dist/multishot/types.d.ts +0 -60
- package/dist/multishot/types.d.ts.map +0 -1
- package/dist/observability.d.ts +0 -71
- package/dist/observability.d.ts.map +0 -1
- package/dist/oracle.d.ts +0 -55
- package/dist/oracle.d.ts.map +0 -1
- package/dist/orthogonality.d.ts +0 -35
- package/dist/orthogonality.d.ts.map +0 -1
- package/dist/otel-pipeline.d.ts +0 -31
- package/dist/otel-pipeline.d.ts.map +0 -1
- package/dist/paraphrase.d.ts +0 -107
- package/dist/paraphrase.d.ts.map +0 -1
- package/dist/pareto.d.ts.map +0 -1
- package/dist/partition-held-out.d.ts +0 -70
- package/dist/partition-held-out.d.ts.map +0 -1
- package/dist/partition-held-out.test.d.ts +0 -2
- package/dist/partition-held-out.test.d.ts.map +0 -1
- package/dist/perf/index.d.ts.map +0 -1
- package/dist/perf/integrity.d.ts +0 -30
- package/dist/perf/integrity.d.ts.map +0 -1
- package/dist/perf/journey.d.ts +0 -45
- package/dist/perf/journey.d.ts.map +0 -1
- package/dist/perf/ratchet.d.ts +0 -47
- package/dist/perf/ratchet.d.ts.map +0 -1
- package/dist/pipelines/budget-breach.d.ts +0 -31
- package/dist/pipelines/budget-breach.d.ts.map +0 -1
- package/dist/pipelines/budget-breach.test.d.ts +0 -2
- package/dist/pipelines/budget-breach.test.d.ts.map +0 -1
- package/dist/pipelines/failure-cluster.d.ts +0 -38
- package/dist/pipelines/failure-cluster.d.ts.map +0 -1
- package/dist/pipelines/failure-cluster.test.d.ts +0 -2
- package/dist/pipelines/failure-cluster.test.d.ts.map +0 -1
- package/dist/pipelines/first-divergence.d.ts +0 -26
- package/dist/pipelines/first-divergence.d.ts.map +0 -1
- package/dist/pipelines/first-divergence.test.d.ts +0 -2
- package/dist/pipelines/first-divergence.test.d.ts.map +0 -1
- package/dist/pipelines/index.d.ts.map +0 -1
- package/dist/pipelines/judge-agreement.d.ts +0 -26
- package/dist/pipelines/judge-agreement.d.ts.map +0 -1
- package/dist/pipelines/judge-agreement.test.d.ts +0 -2
- package/dist/pipelines/judge-agreement.test.d.ts.map +0 -1
- package/dist/pipelines/regression.d.ts +0 -23
- package/dist/pipelines/regression.d.ts.map +0 -1
- package/dist/pipelines/regression.test.d.ts +0 -2
- package/dist/pipelines/regression.test.d.ts.map +0 -1
- package/dist/pipelines/stuck-loop.d.ts +0 -32
- package/dist/pipelines/stuck-loop.d.ts.map +0 -1
- package/dist/pipelines/stuck-loop.test.d.ts +0 -2
- package/dist/pipelines/stuck-loop.test.d.ts.map +0 -1
- package/dist/pipelines/tool-waste.d.ts +0 -34
- package/dist/pipelines/tool-waste.d.ts.map +0 -1
- package/dist/pipelines/tool-waste.test.d.ts +0 -2
- package/dist/pipelines/tool-waste.test.d.ts.map +0 -1
- package/dist/playbook.d.ts +0 -16
- package/dist/playbook.d.ts.map +0 -1
- package/dist/pr-review-benchmark.d.ts +0 -88
- package/dist/pr-review-benchmark.d.ts.map +0 -1
- package/dist/pr-review-benchmark.test.d.ts +0 -2
- package/dist/pr-review-benchmark.test.d.ts.map +0 -1
- package/dist/pre-registration.d.ts +0 -125
- package/dist/pre-registration.d.ts.map +0 -1
- package/dist/prm/builtin-rubrics.d.ts +0 -33
- package/dist/prm/builtin-rubrics.d.ts.map +0 -1
- package/dist/prm/index.d.ts.map +0 -1
- package/dist/prm/inference.d.ts +0 -29
- package/dist/prm/inference.d.ts.map +0 -1
- package/dist/prm/inference.test.d.ts +0 -2
- package/dist/prm/inference.test.d.ts.map +0 -1
- package/dist/prm/rubric.d.ts.map +0 -1
- package/dist/prm/training-export.d.ts +0 -38
- package/dist/prm/training-export.d.ts.map +0 -1
- package/dist/produced-state.d.ts +0 -63
- package/dist/produced-state.d.ts.map +0 -1
- package/dist/produced-state.test.d.ts +0 -8
- package/dist/produced-state.test.d.ts.map +0 -1
- package/dist/profile/baselines.d.ts +0 -37
- package/dist/profile/baselines.d.ts.map +0 -1
- package/dist/profile/index.d.ts +0 -105
- package/dist/profile/index.d.ts.map +0 -1
- package/dist/promotion-gate.d.ts +0 -93
- package/dist/promotion-gate.d.ts.map +0 -1
- package/dist/prompt-registry.d.ts +0 -41
- package/dist/prompt-registry.d.ts.map +0 -1
- package/dist/propose-review-control.d.ts +0 -49
- package/dist/propose-review-control.d.ts.map +0 -1
- package/dist/propose-review-control.test.d.ts +0 -2
- package/dist/propose-review-control.test.d.ts.map +0 -1
- package/dist/propose-review.d.ts +0 -155
- package/dist/propose-review.d.ts.map +0 -1
- package/dist/red-team.d.ts.map +0 -1
- package/dist/reference-replay-steering.d.ts +0 -11
- package/dist/reference-replay-steering.d.ts.map +0 -1
- package/dist/reference-replay.d.ts +0 -176
- package/dist/reference-replay.d.ts.map +0 -1
- package/dist/reflective-mutation.d.ts +0 -79
- package/dist/reflective-mutation.d.ts.map +0 -1
- package/dist/registry.d.ts +0 -31
- package/dist/registry.d.ts.map +0 -1
- package/dist/release-confidence.d.ts +0 -128
- package/dist/release-confidence.d.ts.map +0 -1
- package/dist/release-report.d.ts +0 -11
- package/dist/release-report.d.ts.map +0 -1
- package/dist/replay.d.ts +0 -120
- package/dist/replay.d.ts.map +0 -1
- package/dist/reporter.d.ts +0 -14
- package/dist/reporter.d.ts.map +0 -1
- package/dist/reporting.d.ts.map +0 -1
- package/dist/researcher.d.ts +0 -140
- package/dist/researcher.d.ts.map +0 -1
- package/dist/reviewer.d.ts +0 -118
- package/dist/reviewer.d.ts.map +0 -1
- package/dist/reviewer.test.d.ts +0 -2
- package/dist/reviewer.test.d.ts.map +0 -1
- package/dist/reward-model-export.d.ts +0 -60
- package/dist/reward-model-export.d.ts.map +0 -1
- package/dist/rl/active-curriculum.d.ts +0 -110
- package/dist/rl/active-curriculum.d.ts.map +0 -1
- package/dist/rl/adaptation-eval.d.ts +0 -109
- package/dist/rl/adaptation-eval.d.ts.map +0 -1
- package/dist/rl/adversarial.d.ts.map +0 -1
- package/dist/rl/compute-curves.d.ts +0 -127
- package/dist/rl/compute-curves.d.ts.map +0 -1
- package/dist/rl/contamination.d.ts +0 -117
- package/dist/rl/contamination.d.ts.map +0 -1
- package/dist/rl/corpus.d.ts +0 -55
- package/dist/rl/corpus.d.ts.map +0 -1
- package/dist/rl/corpus.test.d.ts +0 -2
- package/dist/rl/corpus.test.d.ts.map +0 -1
- package/dist/rl/dataset.d.ts +0 -102
- package/dist/rl/dataset.d.ts.map +0 -1
- package/dist/rl/dataset.test.d.ts +0 -2
- package/dist/rl/dataset.test.d.ts.map +0 -1
- package/dist/rl/exporters.d.ts +0 -141
- package/dist/rl/exporters.d.ts.map +0 -1
- package/dist/rl/index.d.ts +0 -49
- package/dist/rl/index.d.ts.map +0 -1
- package/dist/rl/off-policy.d.ts.map +0 -1
- package/dist/rl/predictive-validity-researcher.d.ts +0 -69
- package/dist/rl/predictive-validity-researcher.d.ts.map +0 -1
- package/dist/rl/preferences.d.ts +0 -141
- package/dist/rl/preferences.d.ts.map +0 -1
- package/dist/rl/process-reward.d.ts +0 -122
- package/dist/rl/process-reward.d.ts.map +0 -1
- package/dist/rl/reward-hacking.d.ts +0 -104
- package/dist/rl/reward-hacking.d.ts.map +0 -1
- package/dist/rl/rl-campaign.d.ts +0 -85
- package/dist/rl/rl-campaign.d.ts.map +0 -1
- package/dist/rl/run-record-adapters.d.ts +0 -56
- package/dist/rl/run-record-adapters.d.ts.map +0 -1
- package/dist/rl/sim-fidelity.d.ts +0 -166
- package/dist/rl/sim-fidelity.d.ts.map +0 -1
- package/dist/rl/sim-fidelity.test.d.ts +0 -2
- package/dist/rl/sim-fidelity.test.d.ts.map +0 -1
- package/dist/rl/tournament.d.ts +0 -115
- package/dist/rl/tournament.d.ts.map +0 -1
- package/dist/rl/verifiable-reward.d.ts +0 -124
- package/dist/rl/verifiable-reward.d.ts.map +0 -1
- package/dist/run-critic.d.ts +0 -23
- package/dist/run-critic.d.ts.map +0 -1
- package/dist/run-evidence.d.ts +0 -32
- package/dist/run-evidence.d.ts.map +0 -1
- package/dist/run-record.d.ts.map +0 -1
- package/dist/run-record.test.d.ts +0 -2
- package/dist/run-record.test.d.ts.map +0 -1
- package/dist/run-score.d.ts +0 -31
- package/dist/run-score.d.ts.map +0 -1
- package/dist/runtime-trajectory.d.ts +0 -47
- package/dist/runtime-trajectory.d.ts.map +0 -1
- package/dist/sandbox-harness.d.ts.map +0 -1
- package/dist/sandbox-harness.test.d.ts +0 -2
- package/dist/sandbox-harness.test.d.ts.map +0 -1
- package/dist/sandbox-pool.d.ts +0 -74
- package/dist/sandbox-pool.d.ts.map +0 -1
- package/dist/sandbox-pool.test.d.ts +0 -2
- package/dist/sandbox-pool.test.d.ts.map +0 -1
- package/dist/scorecard.d.ts +0 -133
- package/dist/scorecard.d.ts.map +0 -1
- package/dist/scorecard.test.d.ts +0 -2
- package/dist/scorecard.test.d.ts.map +0 -1
- package/dist/self-play.d.ts +0 -69
- package/dist/self-play.d.ts.map +0 -1
- package/dist/semantic-concept-judge.d.ts +0 -135
- package/dist/semantic-concept-judge.d.ts.map +0 -1
- package/dist/semantic-concept-judge.test.d.ts +0 -2
- package/dist/semantic-concept-judge.test.d.ts.map +0 -1
- package/dist/sequential.d.ts.map +0 -1
- package/dist/series-convergence.d.ts.map +0 -1
- package/dist/slo.d.ts +0 -48
- package/dist/slo.d.ts.map +0 -1
- package/dist/state-continuity.d.ts +0 -47
- package/dist/state-continuity.d.ts.map +0 -1
- package/dist/statistics.d.ts.map +0 -1
- package/dist/statistics.test.d.ts +0 -2
- package/dist/statistics.test.d.ts.map +0 -1
- package/dist/steering-optimizer.d.ts +0 -58
- package/dist/steering-optimizer.d.ts.map +0 -1
- package/dist/steering.d.ts +0 -24
- package/dist/steering.d.ts.map +0 -1
- package/dist/storyboard/code-edit.d.ts +0 -64
- package/dist/storyboard/code-edit.d.ts.map +0 -1
- package/dist/storyboard/code-edit.test.d.ts +0 -2
- package/dist/storyboard/code-edit.test.d.ts.map +0 -1
- package/dist/storyboard/index.d.ts.map +0 -1
- package/dist/storyboard/index.test.d.ts +0 -2
- package/dist/storyboard/index.test.d.ts.map +0 -1
- package/dist/summary-report.d.ts.map +0 -1
- package/dist/telemetry/client.d.ts +0 -35
- package/dist/telemetry/client.d.ts.map +0 -1
- package/dist/telemetry/index.d.ts.map +0 -1
- package/dist/telemetry/schema.d.ts +0 -61
- package/dist/telemetry/schema.d.ts.map +0 -1
- package/dist/telemetry/sink-fetch.d.ts +0 -39
- package/dist/telemetry/sink-fetch.d.ts.map +0 -1
- package/dist/telemetry/sink-file.d.ts.map +0 -1
- package/dist/test-graded-scenario.d.ts +0 -42
- package/dist/test-graded-scenario.d.ts.map +0 -1
- package/dist/testing.d.ts.map +0 -1
- package/dist/tool-use-metrics.d.ts +0 -35
- package/dist/tool-use-metrics.d.ts.map +0 -1
- package/dist/trace/capture-fetch.d.ts +0 -48
- package/dist/trace/capture-fetch.d.ts.map +0 -1
- package/dist/trace/capture-fetch.test.d.ts +0 -2
- package/dist/trace/capture-fetch.test.d.ts.map +0 -1
- package/dist/trace/emitter.d.ts.map +0 -1
- package/dist/trace/extract-usage.d.ts +0 -46
- package/dist/trace/extract-usage.d.ts.map +0 -1
- package/dist/trace/extract-usage.test.d.ts +0 -2
- package/dist/trace/extract-usage.test.d.ts.map +0 -1
- package/dist/trace/index.d.ts +0 -15
- package/dist/trace/index.d.ts.map +0 -1
- package/dist/trace/integrity.d.ts.map +0 -1
- package/dist/trace/otel-bridge.d.ts +0 -29
- package/dist/trace/otel-bridge.d.ts.map +0 -1
- package/dist/trace/otel-export.d.ts +0 -52
- package/dist/trace/otel-export.d.ts.map +0 -1
- package/dist/trace/otel.d.ts +0 -57
- package/dist/trace/otel.d.ts.map +0 -1
- package/dist/trace/otlp-attributes.d.ts +0 -17
- package/dist/trace/otlp-attributes.d.ts.map +0 -1
- package/dist/trace/query.d.ts +0 -29
- package/dist/trace/query.d.ts.map +0 -1
- package/dist/trace/query.test.d.ts +0 -2
- package/dist/trace/query.test.d.ts.map +0 -1
- package/dist/trace/raw-provider-sink.d.ts.map +0 -1
- package/dist/trace/redact.d.ts.map +0 -1
- package/dist/trace/schema.d.ts.map +0 -1
- package/dist/trace/store-to-otlp.d.ts +0 -72
- package/dist/trace/store-to-otlp.d.ts.map +0 -1
- package/dist/trace/store-to-otlp.test.d.ts +0 -2
- package/dist/trace/store-to-otlp.test.d.ts.map +0 -1
- package/dist/trace/store.d.ts.map +0 -1
- package/dist/trace/store.test.d.ts +0 -2
- package/dist/trace/store.test.d.ts.map +0 -1
- package/dist/trace-analyst/analyst.d.ts.map +0 -1
- package/dist/trace-analyst/analyst.test.d.ts +0 -2
- package/dist/trace-analyst/analyst.test.d.ts.map +0 -1
- package/dist/trace-analyst/behavioral-metrics.d.ts +0 -40
- package/dist/trace-analyst/behavioral-metrics.d.ts.map +0 -1
- package/dist/trace-analyst/behavioral-metrics.test.d.ts +0 -2
- package/dist/trace-analyst/behavioral-metrics.test.d.ts.map +0 -1
- package/dist/trace-analyst/hook.d.ts +0 -55
- package/dist/trace-analyst/hook.d.ts.map +0 -1
- package/dist/trace-analyst/index.d.ts +0 -18
- package/dist/trace-analyst/index.d.ts.map +0 -1
- package/dist/trace-analyst/insights.d.ts +0 -71
- package/dist/trace-analyst/insights.d.ts.map +0 -1
- package/dist/trace-analyst/insights.test.d.ts +0 -2
- package/dist/trace-analyst/insights.test.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-flatten.d.ts +0 -42
- package/dist/trace-analyst/otlp-flatten.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-span.d.ts +0 -85
- package/dist/trace-analyst/otlp-span.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-span.test.d.ts +0 -2
- package/dist/trace-analyst/otlp-span.test.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-to-run-records.d.ts +0 -115
- package/dist/trace-analyst/otlp-to-run-records.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-to-run-records.test.d.ts +0 -2
- package/dist/trace-analyst/otlp-to-run-records.test.d.ts.map +0 -1
- package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts +0 -2
- package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts.map +0 -1
- package/dist/trace-analyst/prompts.d.ts +0 -6
- package/dist/trace-analyst/prompts.d.ts.map +0 -1
- package/dist/trace-analyst/store-otlp.d.ts +0 -126
- package/dist/trace-analyst/store-otlp.d.ts.map +0 -1
- package/dist/trace-analyst/store-otlp.test.d.ts +0 -8
- package/dist/trace-analyst/store-otlp.test.d.ts.map +0 -1
- package/dist/trace-analyst/store-otlp.timestamps.test.d.ts +0 -2
- package/dist/trace-analyst/store-otlp.timestamps.test.d.ts.map +0 -1
- package/dist/trace-analyst/store.d.ts +0 -63
- package/dist/trace-analyst/store.d.ts.map +0 -1
- package/dist/trace-analyst/tools.d.ts +0 -44
- package/dist/trace-analyst/tools.d.ts.map +0 -1
- package/dist/trace-analyst/tools.test.d.ts +0 -10
- package/dist/trace-analyst/tools.test.d.ts.map +0 -1
- package/dist/trace-analyst/types.d.ts.map +0 -1
- package/dist/trace-contracts.d.ts +0 -180
- package/dist/trace-contracts.d.ts.map +0 -1
- package/dist/traced-analyst.d.ts +0 -26
- package/dist/traced-analyst.d.ts.map +0 -1
- package/dist/traced-judges.d.ts +0 -27
- package/dist/traced-judges.d.ts.map +0 -1
- package/dist/traces.d.ts.map +0 -1
- package/dist/trajectory.d.ts.map +0 -1
- package/dist/types.d.ts.map +0 -1
- package/dist/ui-finding.d.ts +0 -104
- package/dist/ui-finding.d.ts.map +0 -1
- package/dist/verdict-cache.d.ts +0 -78
- package/dist/verdict-cache.d.ts.map +0 -1
- package/dist/verdict-cache.test.d.ts +0 -2
- package/dist/verdict-cache.test.d.ts.map +0 -1
- package/dist/verdict.d.ts.map +0 -1
- package/dist/visual-diff.d.ts +0 -32
- package/dist/visual-diff.d.ts.map +0 -1
- package/dist/wire/handlers.d.ts +0 -54
- package/dist/wire/handlers.d.ts.map +0 -1
- package/dist/wire/index.d.ts.map +0 -1
- package/dist/wire/openapi.d.ts +0 -3
- package/dist/wire/openapi.d.ts.map +0 -1
- package/dist/wire/rpc.d.ts +0 -21
- package/dist/wire/rpc.d.ts.map +0 -1
- package/dist/wire/rubrics.d.ts +0 -34
- package/dist/wire/rubrics.d.ts.map +0 -1
- package/dist/wire/schemas.d.ts +0 -410
- package/dist/wire/schemas.d.ts.map +0 -1
- package/dist/wire/server.d.ts +0 -60
- package/dist/wire/server.d.ts.map +0 -1
- package/dist/workflow/event-schema.d.ts +0 -5
- package/dist/workflow/event-schema.d.ts.map +0 -1
- package/dist/workflow/feedback-pack.d.ts +0 -99
- package/dist/workflow/feedback-pack.d.ts.map +0 -1
- package/dist/workflow/index.d.ts.map +0 -1
- package/dist/workflow/intelligence-export.d.ts +0 -62
- package/dist/workflow/intelligence-export.d.ts.map +0 -1
- package/dist/workflow/partner-report.d.ts +0 -49
- package/dist/workflow/partner-report.d.ts.map +0 -1
- package/dist/workflow/phase-graph.d.ts +0 -43
- package/dist/workflow/phase-graph.d.ts.map +0 -1
- package/dist/workflow/promotion-gate.d.ts +0 -61
- package/dist/workflow/promotion-gate.d.ts.map +0 -1
- package/dist/workflow/run-record.d.ts +0 -12
- package/dist/workflow/run-record.d.ts.map +0 -1
- package/dist/workflow/runtime-adapter.d.ts +0 -20
- package/dist/workflow/runtime-adapter.d.ts.map +0 -1
- package/dist/workflow/sanitize.d.ts +0 -21
- package/dist/workflow/sanitize.d.ts.map +0 -1
- package/dist/workflow/schema.d.ts +0 -5
- package/dist/workflow/schema.d.ts.map +0 -1
- package/dist/workflow/summary.d.ts +0 -43
- package/dist/workflow/summary.d.ts.map +0 -1
- package/dist/workflow/trace-event-fields.d.ts +0 -6
- package/dist/workflow/trace-event-fields.d.ts.map +0 -1
- package/dist/workflow/trajectory.d.ts +0 -15
- package/dist/workflow/trajectory.d.ts.map +0 -1
- package/dist/workflow/types.d.ts +0 -68
- package/dist/workflow/types.d.ts.map +0 -1
- package/dist/workspace-inspector.d.ts +0 -67
- package/dist/workspace-inspector.d.ts.map +0 -1
- package/dist/wrangler-deploy-runner.test.d.ts +0 -2
- package/dist/wrangler-deploy-runner.test.d.ts.map +0 -1
|
@@ -1,52 +0,0 @@
|
|
|
1
|
-
import type { RunEvalOptions } from '../campaign/presets/run-eval';
|
|
2
|
-
import type { CampaignResult, DispatchContext, JudgeConfig, MutableSurface, Scenario } from '../campaign/types';
|
|
3
|
-
import type { HostedTenant } from '../hosted/client';
|
|
4
|
-
import { type SelfImproveBudget, type SelfImproveLlm, type SelfImproveOptions, type SelfImproveResult } from './self-improve';
|
|
5
|
-
export type AgentEvalAgent<TScenario extends Scenario, TArtifact> = (surface: MutableSurface, scenario: TScenario, ctx: DispatchContext) => Promise<TArtifact>;
|
|
6
|
-
export type DefineAgentEvalOptions<TScenario extends Scenario, TArtifact> = SelfImproveOptions<TScenario, TArtifact>;
|
|
7
|
-
export interface AgentEvalEvaluateOptions<TScenario extends Scenario, TArtifact> extends Omit<RunEvalOptions<TScenario, TArtifact>, 'dispatch' | 'judges' | 'runDir' | 'scenarios'> {
|
|
8
|
-
/** Scenario set to evaluate. Defaults to the scenarios passed to `defineAgentEval`. */
|
|
9
|
-
scenarios?: TScenario[];
|
|
10
|
-
/** Surface to evaluate. Defaults to the baseline surface passed to `defineAgentEval`. */
|
|
11
|
-
surface?: MutableSurface;
|
|
12
|
-
/** Agent to evaluate. Defaults to the agent passed to `defineAgentEval`. */
|
|
13
|
-
agent?: AgentEvalAgent<TScenario, TArtifact>;
|
|
14
|
-
/** Single judge override. Ignored when `judges` is set. */
|
|
15
|
-
judge?: JudgeConfig<TArtifact, TScenario>;
|
|
16
|
-
/** Full judge list override. Defaults to the single judge passed to `defineAgentEval`. */
|
|
17
|
-
judges?: JudgeConfig<TArtifact, TScenario>[];
|
|
18
|
-
/** Logical or filesystem run directory. Defaults to an in-memory run. */
|
|
19
|
-
runDir?: string;
|
|
20
|
-
}
|
|
21
|
-
export type AgentEvalImproveOptions<TScenario extends Scenario, TArtifact> = Omit<Partial<SelfImproveOptions<TScenario, TArtifact>>, 'budget' | 'hostedTenant' | 'llm'> & {
|
|
22
|
-
budget?: Partial<SelfImproveBudget>;
|
|
23
|
-
hostedTenant?: Partial<HostedTenant>;
|
|
24
|
-
llm?: Partial<SelfImproveLlm>;
|
|
25
|
-
};
|
|
26
|
-
export interface DefinedAgentEval<TScenario extends Scenario, TArtifact> {
|
|
27
|
-
/** The default scenarios used by `evaluate()` and `improve()`. */
|
|
28
|
-
readonly scenarios: readonly TScenario[];
|
|
29
|
-
/** The default baseline surface used by `evaluate()` and `improve()`. */
|
|
30
|
-
readonly baselineSurface: MutableSurface;
|
|
31
|
-
/**
|
|
32
|
-
* Run one scored evaluation. Use this for a baseline score or to score one
|
|
33
|
-
* candidate surface without running an improvement loop.
|
|
34
|
-
*/
|
|
35
|
-
evaluate(opts?: AgentEvalEvaluateOptions<TScenario, TArtifact>): Promise<CampaignResult<TArtifact, TScenario>>;
|
|
36
|
-
/**
|
|
37
|
-
* Run the closed improvement loop. Per-call overrides replace the definition
|
|
38
|
-
* except for nested config objects (`budget`, `llm`, `hostedTenant`), which
|
|
39
|
-
* are merged field-by-field so callers can override one knob without
|
|
40
|
-
* repeating secrets or budget defaults.
|
|
41
|
-
*/
|
|
42
|
-
improve(opts?: AgentEvalImproveOptions<TScenario, TArtifact>): Promise<SelfImproveResult<TScenario, TArtifact>>;
|
|
43
|
-
}
|
|
44
|
-
/**
|
|
45
|
-
* Define an agent eval once, then either score a surface with `evaluate()` or
|
|
46
|
-
* run the closed loop with `improve()`.
|
|
47
|
-
*
|
|
48
|
-
* This is a DX wrapper only: it delegates to `runEval()` and `selfImprove()` and
|
|
49
|
-
* returns their native result shapes.
|
|
50
|
-
*/
|
|
51
|
-
export declare function defineAgentEval<TScenario extends Scenario, TArtifact>(defaults: DefineAgentEvalOptions<TScenario, TArtifact>): DefinedAgentEval<TScenario, TArtifact>;
|
|
52
|
-
//# sourceMappingURL=define-agent-eval.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"define-agent-eval.d.ts","sourceRoot":"","sources":["../../src/contract/define-agent-eval.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,8BAA8B,CAAA;AAGlE,OAAO,KAAK,EACV,cAAc,EACd,eAAe,EACf,WAAW,EACX,cAAc,EACd,QAAQ,EACT,MAAM,mBAAmB,CAAA;AAC1B,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,kBAAkB,CAAA;AACpD,OAAO,EACL,KAAK,iBAAiB,EACtB,KAAK,cAAc,EACnB,KAAK,kBAAkB,EACvB,KAAK,iBAAiB,EAEvB,MAAM,gBAAgB,CAAA;AAEvB,MAAM,MAAM,cAAc,CAAC,SAAS,SAAS,QAAQ,EAAE,SAAS,IAAI,CAClE,OAAO,EAAE,cAAc,EACvB,QAAQ,EAAE,SAAS,EACnB,GAAG,EAAE,eAAe,KACjB,OAAO,CAAC,SAAS,CAAC,CAAA;AAEvB,MAAM,MAAM,sBAAsB,CAAC,SAAS,SAAS,QAAQ,EAAE,SAAS,IAAI,kBAAkB,CAC5F,SAAS,EACT,SAAS,CACV,CAAA;AAED,MAAM,WAAW,wBAAwB,CAAC,SAAS,SAAS,QAAQ,EAAE,SAAS,CAC7E,SAAQ,IAAI,CACV,cAAc,CAAC,SAAS,EAAE,SAAS,CAAC,EACpC,UAAU,GAAG,QAAQ,GAAG,QAAQ,GAAG,WAAW,CAC/C;IACD,uFAAuF;IACvF,SAAS,CAAC,EAAE,SAAS,EAAE,CAAA;IACvB,yFAAyF;IACzF,OAAO,CAAC,EAAE,cAAc,CAAA;IACxB,4EAA4E;IAC5E,KAAK,CAAC,EAAE,cAAc,CAAC,SAAS,EAAE,SAAS,CAAC,CAAA;IAC5C,2DAA2D;IAC3D,KAAK,CAAC,EAAE,WAAW,CAAC,SAAS,EAAE,SAAS,CAAC,CAAA;IACzC,0FAA0F;IAC1F,MAAM,CAAC,EAAE,WAAW,CAAC,SAAS,EAAE,SAAS,CAAC,EAAE,CAAA;IAC5C,yEAAyE;IACzE,MAAM,CAAC,EAAE,MAAM,CAAA;CAChB;AAED,MAAM,MAAM,uBAAuB,CAAC,SAAS,SAAS,QAAQ,EAAE,SAAS,IAAI,IAAI,CAC/E,OAAO,CAAC,kBAAkB,CAAC,SAAS,EAAE,SAAS,CAAC,CAAC,EACjD,QAAQ,GAAG,cAAc,GAAG,KAAK,CAClC,GAAG;IACF,MAAM,CAAC,EAAE,OAAO,CAAC,iBAAiB,CAAC,CAAA;IACnC,YAAY,CAAC,EAAE,OAAO,CAAC,YAAY,CAAC,CAAA;IACpC,GAAG,CAAC,EAAE,OAAO,CAAC,cAAc,CAAC,CAAA;CAC9B,CAAA;AAED,MAAM,WAAW,gBAAgB,CAAC,SAAS,SAAS,QAAQ,EAAE,SAAS;IACrE,kEAAkE;IAClE,QAAQ,CAAC,SAAS,EAAE,SAAS,SAAS,EAAE,CAAA;IACxC,yEAAyE;IACzE,QAAQ,CAAC,eAAe,EAAE,cAAc,CAAA;IACxC;;;OAGG;IACH,QAAQ,CACN,IAAI,CAAC,EAAE,wBAAwB,CAAC,SAAS,EAAE,SAAS,CAAC,GACpD,OAAO,CAAC,cAAc,CAAC,SAAS,EAAE,SAAS,CAAC,CAAC,CAAA;IAChD;;;;;OAKG;IACH,OAAO,CACL,IAAI,CAAC,EAAE,uBAAuB,CAAC,SAAS,EAAE,SAAS,CAAC,GACnD,OAAO,CAAC,iBAAiB,CAAC,SAAS,EAAE,SAAS,CAAC,CAAC,CAAA;CACpD;AAED;;;;;;GAMG;AACH,wBAAgB,eAAe,CAAC,SAAS,SAAS,QAAQ,EAAE,SAAS,EACnE,QAAQ,EAAE,sBAAsB,CAAC,SAAS,EAAE,SAAS,CAAC,GACrD,gBAAgB,CAAC,SAAS,EAAE,SAAS,CAAC,CAiDxC"}
|
package/dist/contract/diff.d.ts
DELETED
|
@@ -1,114 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* # `@tangle-network/agent-eval/contract` — eval-run diff primitive.
|
|
3
|
-
*
|
|
4
|
-
* The substrate side of the v-N-versus-v-N+1 dashboard view. Given two
|
|
5
|
-
* `EvalRunEvent`s (or two `EvalRunGenerationSnapshot`s from one run), this
|
|
6
|
-
* returns a normalised diff: per-cell composite + per-judge/per-dimension
|
|
7
|
-
* deltas, surface-hash change, aggregate cost + duration shifts.
|
|
8
|
-
*
|
|
9
|
-
* Consumed by:
|
|
10
|
-
* - The hosted-tier dashboard (intelligence-web) — renders v3 vs v4
|
|
11
|
-
* comparisons of cells × judges × dimensions.
|
|
12
|
-
* - CI reporting — emits a "shipped: composite +0.07, cost +$1.20" line
|
|
13
|
-
* in PR review for autonomous-improvement runs.
|
|
14
|
-
* - Any downstream consumer that needs "what actually changed" without
|
|
15
|
-
* reimplementing the matching + arithmetic.
|
|
16
|
-
*
|
|
17
|
-
* Cells are matched on the natural composite key `(scenarioId, rep)`.
|
|
18
|
-
* Unmatched cells surface as `removed` / `added` so callers can tell
|
|
19
|
-
* "this cell got worse" from "this cell wasn't run."
|
|
20
|
-
*/
|
|
21
|
-
import type { GateDecision } from '../campaign/types';
|
|
22
|
-
import type { EvalRunCellScore, EvalRunEvent, EvalRunGenerationSnapshot } from '../hosted/types';
|
|
23
|
-
/** Per-dimension delta. `before` / `after` are null when the judge did not
|
|
24
|
-
* emit a value for that side. `delta` is `after - before`; null when
|
|
25
|
-
* either side is null. */
|
|
26
|
-
export interface EvalDimensionDelta {
|
|
27
|
-
before: number | null;
|
|
28
|
-
after: number | null;
|
|
29
|
-
delta: number | null;
|
|
30
|
-
}
|
|
31
|
-
/** Per-cell delta, keyed on `(scenarioId, rep)`. */
|
|
32
|
-
export interface EvalCellScoreDelta {
|
|
33
|
-
scenarioId: string;
|
|
34
|
-
rep: number;
|
|
35
|
-
compositeBefore: number;
|
|
36
|
-
compositeAfter: number;
|
|
37
|
-
compositeDelta: number;
|
|
38
|
-
/** Per-judge → per-dimension deltas. Outer key = judge name from
|
|
39
|
-
* `EvalRunCellScore.dimensions`; inner key = dimension name. */
|
|
40
|
-
dimensions: Record<string, Record<string, EvalDimensionDelta>>;
|
|
41
|
-
}
|
|
42
|
-
/** Diff between two generation snapshots — the unit the dashboard renders
|
|
43
|
-
* for a single "v3 vs v4" comparison. */
|
|
44
|
-
export interface EvalGenerationDiff {
|
|
45
|
-
beforeIndex: number;
|
|
46
|
-
afterIndex: number;
|
|
47
|
-
beforeSurfaceHash: string;
|
|
48
|
-
afterSurfaceHash: string;
|
|
49
|
-
surfaceChanged: boolean;
|
|
50
|
-
/** Cells present in both snapshots, matched on `(scenarioId, rep)`. */
|
|
51
|
-
matched: EvalCellScoreDelta[];
|
|
52
|
-
/** Cells present in `before` but missing from `after`. */
|
|
53
|
-
removed: EvalRunCellScore[];
|
|
54
|
-
/** Cells present in `after` but missing from `before`. */
|
|
55
|
-
added: EvalRunCellScore[];
|
|
56
|
-
/** Aggregate composite mean across all cells in the snapshot. */
|
|
57
|
-
compositeBefore: number;
|
|
58
|
-
compositeAfter: number;
|
|
59
|
-
compositeDelta: number;
|
|
60
|
-
costUsdBefore: number;
|
|
61
|
-
costUsdAfter: number;
|
|
62
|
-
costUsdDelta: number;
|
|
63
|
-
durationMsBefore: number;
|
|
64
|
-
durationMsAfter: number;
|
|
65
|
-
durationMsDelta: number;
|
|
66
|
-
}
|
|
67
|
-
/** Diff between two full eval-runs. Includes both baseline-vs-baseline and
|
|
68
|
-
* winner-vs-winner generation diffs when both sides expose them, plus
|
|
69
|
-
* run-level metadata. */
|
|
70
|
-
export interface EvalRunDiff {
|
|
71
|
-
beforeRunId: string;
|
|
72
|
-
afterRunId: string;
|
|
73
|
-
beforeTimestamp: string;
|
|
74
|
-
afterTimestamp: string;
|
|
75
|
-
beforeGateDecision: GateDecision | null;
|
|
76
|
-
afterGateDecision: GateDecision | null;
|
|
77
|
-
beforeHoldoutLift: number | null;
|
|
78
|
-
afterHoldoutLift: number | null;
|
|
79
|
-
holdoutLiftDelta: number | null;
|
|
80
|
-
beforeTotalCostUsd: number;
|
|
81
|
-
afterTotalCostUsd: number;
|
|
82
|
-
totalCostUsdDelta: number;
|
|
83
|
-
beforeTotalDurationMs: number;
|
|
84
|
-
afterTotalDurationMs: number;
|
|
85
|
-
totalDurationMsDelta: number;
|
|
86
|
-
/** Baseline-vs-baseline diff. Null when either run has no baseline. */
|
|
87
|
-
baselineDiff: EvalGenerationDiff | null;
|
|
88
|
-
/** Highest-index-generation comparison. Null when either run has no
|
|
89
|
-
* recorded generations (e.g. baseline-only or errored before any
|
|
90
|
-
* generation completed). */
|
|
91
|
-
winnersDiff: EvalGenerationDiff | null;
|
|
92
|
-
}
|
|
93
|
-
/**
|
|
94
|
-
* Diff two generation snapshots. Cells are matched on `(scenarioId, rep)`;
|
|
95
|
-
* unmatched cells surface in `added` / `removed`. Aggregate fields are
|
|
96
|
-
* recomputed from the snapshot's stored fields, not re-derived from cells —
|
|
97
|
-
* this keeps the diff consistent with whatever aggregation the substrate
|
|
98
|
-
* actually reported.
|
|
99
|
-
*/
|
|
100
|
-
export declare function diffGenerations(before: EvalRunGenerationSnapshot, after: EvalRunGenerationSnapshot): EvalGenerationDiff;
|
|
101
|
-
/**
|
|
102
|
-
* Diff two full eval-runs. Produces baseline-vs-baseline and
|
|
103
|
-
* winner-vs-winner generation diffs when both sides expose them, plus
|
|
104
|
-
* run-level cost / lift / gate-decision deltas.
|
|
105
|
-
*/
|
|
106
|
-
export declare function diffRuns(before: EvalRunEvent, after: EvalRunEvent): EvalRunDiff;
|
|
107
|
-
/**
|
|
108
|
-
* Within-run baseline → winning-generation diff. The natural "what did the
|
|
109
|
-
* improvement loop produce" view for a single run. Returns null when the
|
|
110
|
-
* run never reached a generation past baseline (errored early, or the gate
|
|
111
|
-
* shipped the baseline as-is).
|
|
112
|
-
*/
|
|
113
|
-
export declare function diffRunBaselineToWinner(run: EvalRunEvent): EvalGenerationDiff | null;
|
|
114
|
-
//# sourceMappingURL=diff.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"diff.d.ts","sourceRoot":"","sources":["../../src/contract/diff.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,mBAAmB,CAAA;AACrD,OAAO,KAAK,EAAE,gBAAgB,EAAE,YAAY,EAAE,yBAAyB,EAAE,MAAM,iBAAiB,CAAA;AAIhG;;2BAE2B;AAC3B,MAAM,WAAW,kBAAkB;IACjC,MAAM,EAAE,MAAM,GAAG,IAAI,CAAA;IACrB,KAAK,EAAE,MAAM,GAAG,IAAI,CAAA;IACpB,KAAK,EAAE,MAAM,GAAG,IAAI,CAAA;CACrB;AAED,oDAAoD;AACpD,MAAM,WAAW,kBAAkB;IACjC,UAAU,EAAE,MAAM,CAAA;IAClB,GAAG,EAAE,MAAM,CAAA;IACX,eAAe,EAAE,MAAM,CAAA;IACvB,cAAc,EAAE,MAAM,CAAA;IACtB,cAAc,EAAE,MAAM,CAAA;IACtB;qEACiE;IACjE,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,kBAAkB,CAAC,CAAC,CAAA;CAC/D;AAED;0CAC0C;AAC1C,MAAM,WAAW,kBAAkB;IACjC,WAAW,EAAE,MAAM,CAAA;IACnB,UAAU,EAAE,MAAM,CAAA;IAClB,iBAAiB,EAAE,MAAM,CAAA;IACzB,gBAAgB,EAAE,MAAM,CAAA;IACxB,cAAc,EAAE,OAAO,CAAA;IACvB,uEAAuE;IACvE,OAAO,EAAE,kBAAkB,EAAE,CAAA;IAC7B,0DAA0D;IAC1D,OAAO,EAAE,gBAAgB,EAAE,CAAA;IAC3B,0DAA0D;IAC1D,KAAK,EAAE,gBAAgB,EAAE,CAAA;IACzB,iEAAiE;IACjE,eAAe,EAAE,MAAM,CAAA;IACvB,cAAc,EAAE,MAAM,CAAA;IACtB,cAAc,EAAE,MAAM,CAAA;IACtB,aAAa,EAAE,MAAM,CAAA;IACrB,YAAY,EAAE,MAAM,CAAA;IACpB,YAAY,EAAE,MAAM,CAAA;IACpB,gBAAgB,EAAE,MAAM,CAAA;IACxB,eAAe,EAAE,MAAM,CAAA;IACvB,eAAe,EAAE,MAAM,CAAA;CACxB;AAED;;0BAE0B;AAC1B,MAAM,WAAW,WAAW;IAC1B,WAAW,EAAE,MAAM,CAAA;IACnB,UAAU,EAAE,MAAM,CAAA;IAClB,eAAe,EAAE,MAAM,CAAA;IACvB,cAAc,EAAE,MAAM,CAAA;IACtB,kBAAkB,EAAE,YAAY,GAAG,IAAI,CAAA;IACvC,iBAAiB,EAAE,YAAY,GAAG,IAAI,CAAA;IACtC,iBAAiB,EAAE,MAAM,GAAG,IAAI,CAAA;IAChC,gBAAgB,EAAE,MAAM,GAAG,IAAI,CAAA;IAC/B,gBAAgB,EAAE,MAAM,GAAG,IAAI,CAAA;IAC/B,kBAAkB,EAAE,MAAM,CAAA;IAC1B,iBAAiB,EAAE,MAAM,CAAA;IACzB,iBAAiB,EAAE,MAAM,CAAA;IACzB,qBAAqB,EAAE,MAAM,CAAA;IAC7B,oBAAoB,EAAE,MAAM,CAAA;IAC5B,oBAAoB,EAAE,MAAM,CAAA;IAC5B,uEAAuE;IACvE,YAAY,EAAE,kBAAkB,GAAG,IAAI,CAAA;IACvC;;iCAE6B;IAC7B,WAAW,EAAE,kBAAkB,GAAG,IAAI,CAAA;CACvC;AA4CD;;;;;;GAMG;AACH,wBAAgB,eAAe,CAC7B,MAAM,EAAE,yBAAyB,EACjC,KAAK,EAAE,yBAAyB,GAC/B,kBAAkB,CA8CpB;AAYD;;;;GAIG;AACH,wBAAgB,QAAQ,CAAC,MAAM,EAAE,YAAY,EAAE,KAAK,EAAE,YAAY,GAAG,WAAW,CA8B/E;AAED;;;;;GAKG;AACH,wBAAgB,uBAAuB,CAAC,GAAG,EAAE,YAAY,GAAG,kBAAkB,GAAG,IAAI,CAKpF"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/contract/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAoFG;AAMH,YAAY,EACV,kBAAkB,EAClB,sBAAsB,EACtB,kBAAkB,EAClB,iBAAiB,EACjB,cAAc,EACd,mBAAmB,EACnB,WAAW,EACX,eAAe,EACf,UAAU,IAAI,QAAQ,EACtB,IAAI,EACJ,WAAW,EACX,YAAY,EACZ,UAAU,EACV,mBAAmB,EACnB,gBAAgB,EAChB,WAAW,EACX,cAAc,EACd,UAAU,EACV,cAAc,EACd,OAAO,EACP,oBAAoB,EACpB,eAAe,EACf,QAAQ,EACR,aAAa,EACb,eAAe,GAChB,MAAM,mBAAmB,CAAA;AAI1B,OAAO,EAAE,KAAK,cAAc,EAAE,OAAO,EAAE,MAAM,8BAA8B,CAAA;AAC3E,OAAO,EACL,KAAK,yBAAyB,EAC9B,KAAK,wBAAwB,EAC7B,kBAAkB,GACnB,MAAM,0CAA0C,CAAA;AACjD,OAAO,EAAE,KAAK,kBAAkB,EAAE,WAAW,EAAE,MAAM,0BAA0B,CAAA;AAI/E,OAAO,EACL,KAAK,2BAA2B,EAChC,oBAAoB,GACrB,MAAM,oCAAoC,CAAA;AAC3C,OAAO,EAAE,KAAK,mBAAmB,EAAE,YAAY,EAAE,MAAM,4BAA4B,CAAA;AAInF,OAAO,EAAE,WAAW,EAAE,MAAM,2BAA2B,CAAA;AACvD,OAAO,EACL,KAAK,4BAA4B,EACjC,qBAAqB,GACtB,MAAM,2CAA2C,CAAA;AAClD,OAAO,EAAE,KAAK,kBAAkB,EAAE,WAAW,EAAE,MAAM,gCAAgC,CAAA;AACrF,OAAO,EACL,KAAK,YAAY,EACjB,KAAK,WAAW,EAChB,KAAK,0BAA0B,EAC/B,mBAAmB,EACnB,KAAK,cAAc,EACnB,KAAK,eAAe,EACpB,KAAK,6BAA6B,EAClC,KAAK,kBAAkB,EACvB,KAAK,eAAe,EACpB,YAAY,EACZ,sBAAsB,GACvB,MAAM,oCAAoC,CAAA;AAI3C,OAAO,EACL,KAAK,eAAe,EACpB,iBAAiB,EACjB,uBAAuB,GACxB,MAAM,qBAAqB,CAAA;AAI5B,OAAO,EACL,KAAK,iBAAiB,EACtB,sBAAsB,EACtB,KAAK,6BAA6B,EAClC,oBAAoB,EACpB,KAAK,YAAY,GAClB,MAAM,4BAA4B,CAAA;AAInC,YAAY,EAAE,YAAY,EAAE,MAAM,kBAAkB,CAAA;AACpD,OAAO,EACL,KAAK,cAAc,EACnB,KAAK,wBAAwB,EAC7B,KAAK,uBAAuB,EAC5B,KAAK,sBAAsB,EAC3B,KAAK,gBAAgB,EACrB,eAAe,GAChB,MAAM,qBAAqB,CAAA;AAC5B,OAAO,EACL,KAAK,iBAAiB,EACtB,KAAK,cAAc,EACnB,KAAK,kBAAkB,EACvB,KAAK,wBAAwB,EAC7B,KAAK,iBAAiB,EACtB,WAAW,GACZ,MAAM,gBAAgB,CAAA;AAYvB,OAAO,EACL,2BAA2B,EAC3B,KAAK,6BAA6B,GACnC,MAAM,6BAA6B,CAAA;AACpC,YAAY,EAAE,cAAc,EAAE,MAAM,kBAAkB,CAAA;AACtD,YAAY,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAA;AACxD,OAAO,EAAE,WAAW,EAAE,MAAM,gBAAgB,CAAA;AAC5C,YAAY,EACV,qBAAqB,EACrB,aAAa,EACb,iBAAiB,EACjB,YAAY,EACZ,WAAW,EACX,yBAAyB,EACzB,cAAc,EACd,cAAc,EACd,kBAAkB,GACnB,MAAM,kBAAkB,CAAA;AAIzB,OAAO,EACL,eAAe,EACf,uBAAuB,EACvB,QAAQ,EACR,KAAK,kBAAkB,EACvB,KAAK,kBAAkB,EACvB,KAAK,kBAAkB,EACvB,KAAK,WAAW,GACjB,MAAM,QAAQ,CAAA;AAMf,OAAO,EACL,KAAK,qBAAqB,EAC1B,KAAK,yBAAyB,EAC9B,KAAK,sBAAsB,EAC3B,KAAK,cAAc,EACnB,KAAK,eAAe,EACpB,KAAK,eAAe,EACpB,KAAK,gBAAgB,EACrB,KAAK,mBAAmB,EACxB,KAAK,0BAA0B,EAC/B,KAAK,6BAA6B,EAClC,KAAK,4BAA4B,EACjC,KAAK,uBAAuB,EAC5B,KAAK,sBAAsB,EAC3B,KAAK,iBAAiB,EACtB,KAAK,gBAAgB,EACrB,KAAK,wBAAwB,EAC7B,KAAK,uBAAuB,EAC5B,KAAK,oBAAoB,EACzB,qBAAqB,EACrB,gBAAgB,EAChB,iBAAiB,EACjB,mBAAmB,EACnB,mBAAmB,EACnB,aAAa,EACb,kBAAkB,EAClB,aAAa,EACb,KAAK,oBAAoB,EACzB,KAAK,+BAA+B,EACpC,eAAe,EACf,mBAAmB,EACnB,6BAA6B,GAC9B,MAAM,UAAU,CAAA"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"insight-report.d.ts","sourceRoot":"","sources":["../../src/contract/insight-report.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AAEH,OAAO,KAAK,EAAE,mBAAmB,EAAE,gBAAgB,EAAE,MAAM,mBAAmB,CAAA;AAC9E,OAAO,KAAK,EAAE,mBAAmB,EAAE,MAAM,qBAAqB,CAAA;AAI9D,MAAM,WAAW,aAAa;IAC5B,+BAA+B;IAC/B,CAAC,EAAE,MAAM,CAAA;IAET,oEAAoE;IACpE,SAAS,EAAE,kBAAkB,CAAA;IAE7B;wEACoE;IACpE,YAAY,EAAE,MAAM,CAAC,MAAM,EAAE,kBAAkB,CAAC,CAAA;IAEhD,qDAAqD;IACrD,WAAW,EAAE;QACX,IAAI,EAAE,kBAAkB,CAAA;QACxB,MAAM,EAAE,gBAAgB,CAAA;QACxB;;;6EAGqE;QACrE,QAAQ,CAAC,EAAE;YAAE,IAAI,CAAC,EAAE,MAAM,CAAC;YAAC,MAAM,CAAC,EAAE,MAAM,CAAA;SAAE,CAAA;KAC9C,CAAA;IAED;;8CAE0C;IAC1C,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,YAAY,CAAC,CAAA;IAEpC;;sDAEkD;IAClD,UAAU,CAAC,EAAE,iBAAiB,CAAA;IAE9B;;sEAEkE;IAClE,IAAI,CAAC,EAAE,WAAW,CAAA;IAElB;kDAC8C;IAC9C,eAAe,CAAC,EAAE,qBAAqB,CAAA;IAEvC;mCAC+B;IAC/B,aAAa,CAAC,EAAE,oBAAoB,CAAA;IAEpC;;;4CAGwC;IACxC,kBAAkB,CAAC,EAAE,yBAAyB,CAAA;IAE9C;;;8DAG0D;IAC1D,OAAO,EAAE,cAAc,CAAA;IAEvB;;;;8DAI0D;IAC1D,qBAAqB,CAAC,EAAE,qBAAqB,CAAA;IAE7C;;;2EAGuE;IACvE,YAAY,CAAC,EAAE,gBAAgB,EAAE,CAAA;IAEjC;uEACmE;IACnE,eAAe,EAAE,cAAc,EAAE,CAAA;CAClC;AAID,wDAAwD;AACxD,MAAM,WAAW,kBAAkB;IACjC,qDAAqD;IACrD,CAAC,EAAE,MAAM,CAAA;IACT,IAAI,EAAE,MAAM,CAAA;IACZ,GAAG,EAAE,MAAM,CAAA;IACX,GAAG,EAAE,MAAM,CAAA;IACX,MAAM,EAAE,MAAM,CAAA;IACd,GAAG,EAAE,MAAM,CAAA;IACX,GAAG,EAAE,MAAM,CAAA;IACX,qEAAqE;IACrE,SAAS,EAAE,mBAAmB,EAAE,CAAA;IAChC;;;2DAGuD;IACvD,QAAQ,CAAC,EAAE,KAAK,CAAC;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;CACnD;AAED,MAAM,WAAW,YAAY;IAC3B,+CAA+C;IAC/C,CAAC,EAAE,MAAM,CAAA;IACT,6CAA6C;IAC7C,SAAS,EAAE,MAAM,CAAA;IACjB;8DAC0D;IAC1D,WAAW,CAAC,EAAE,mBAAmB,CAAA;IACjC;gEAC4D;IAC5D,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB;iDAC6C;IAC7C,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB;mBACe;IACf,aAAa,CAAC,EAAE,MAAM,CAAA;CACvB;AAED,MAAM,WAAW,iBAAiB;IAChC,qDAAqD;IACrD,MAAM,EAAE,MAAM,CAAA;IACd,yCAAyC;IACzC,YAAY,EAAE,MAAM,CAAA;IACpB,6CAA6C;IAC7C,KAAK,EAAE,MAAM,CAAA;IACb,4DAA4D;IAC5D,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC/B,2EAA2E;IAC3E,iBAAiB,EAAE,KAAK,CAAC;QACvB,KAAK,EAAE,MAAM,CAAA;QACb,OAAO,EAAE,KAAK,CAAC;YAAE,KAAK,EAAE,MAAM,CAAC;YAAC,KAAK,EAAE,MAAM,CAAA;SAAE,CAAC,CAAA;QAChD,KAAK,EAAE,MAAM,CAAA;KACd,CAAC,CAAA;CACH;AAED,MAAM,WAAW,WAAW;IAC1B,YAAY,EAAE,MAAM,CAAA;IACpB,aAAa,EAAE,MAAM,CAAA;IACrB,4BAA4B;IAC5B,KAAK,EAAE,MAAM,CAAA;IACb,wDAAwD;IACxD,IAAI,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IACtB,6BAA6B;IAC7B,MAAM,EAAE,MAAM,CAAA;IACd,qCAAqC;IACrC,CAAC,EAAE,MAAM,CAAA;IACT,+BAA+B;IAC/B,OAAO,EAAE,MAAM,CAAA;IACf,yDAAyD;IACzD,GAAG,EAAE,MAAM,CAAA;IACX,oEAAoE;IACpE,SAAS,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,qBAAqB;IACpC,2EAA2E;IAC3E,QAAQ,EAAE,KAAK,CAAC;QACd,EAAE,EAAE,MAAM,CAAA;QACV,IAAI,EAAE,MAAM,CAAA;QACZ,qDAAqD;QACrD,KAAK,EAAE,MAAM,CAAA;QACb,2DAA2D;QAC3D,SAAS,EAAE,MAAM,EAAE,CAAA;QACnB,uEAAuE;QACvE,YAAY,CAAC,EAAE,MAAM,CAAA;KACtB,CAAC,CAAA;IACF,aAAa,EAAE,MAAM,CAAA;CACtB;AAED;;;;+BAI+B;AAC/B,MAAM,WAAW,gBAAgB;IAC/B,6BAA6B;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,wCAAwC;IACxC,KAAK,EAAE,MAAM,CAAA;IACb,uCAAuC;IACvC,KAAK,EAAE,MAAM,CAAA;CACd;AAED,MAAM,WAAW,oBAAoB;IACnC,+CAA+C;IAC/C,KAAK,EAAE,MAAM,CAAA;IACb;0CACsC;IACtC,kBAAkB,EAAE,OAAO,CAAA;IAC3B,OAAO,CAAC,EAAE,KAAK,CAAC;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;CACpE;AAED,MAAM,WAAW,yBAAyB;IACxC;wEACoE;IACpE,MAAM,EAAE,MAAM,CAAA;IACd,2CAA2C;IAC3C,CAAC,EAAE,MAAM,CAAA;IACT,+DAA+D;IAC/D,OAAO,EAAE,MAAM,CAAA;IACf,qEAAqE;IACrE,QAAQ,EAAE,MAAM,CAAA;IAChB,oEAAoE;IACpE,WAAW,CAAC,EAAE;QACZ,SAAS,EAAE,MAAM,CAAA;QACjB,KAAK,EAAE,MAAM,CAAA;QACb,EAAE,EAAE,MAAM,CAAA;KACX,CAAA;CACF;AAED,MAAM,WAAW,cAAc;IAC7B;4BACwB;IACxB,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,MAAM,CAAA;IAChC,IAAI,EAAE,KAAK,CAAC;QACV,IAAI,EAAE,cAAc,GAAG,eAAe,GAAG,wBAAwB,CAAA;QACjE,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,MAAM,CAAA;QAChC,MAAM,EAAE,MAAM,CAAA;KACf,CAAC,CAAA;IACF;kDAC8C;IAC9C,MAAM,EAAE,MAAM,EAAE,CAAA;CACjB;AAED,MAAM,WAAW,WAAW;IAC1B,2BAA2B;IAC3B,OAAO,EAAE,MAAM,CAAA;IACf,4BAA4B;IAC5B,QAAQ,EAAE,MAAM,CAAA;IAChB;;4CAEwC;IACxC,KAAK,EAAE,MAAM,CAAA;IACb;0EACsE;IACtE,IAAI,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IACtB,wCAAwC;IACxC,MAAM,EAAE,MAAM,CAAA;IACd,sDAAsD;IACtD,OAAO,EAAE,MAAM,CAAA;IACf,oBAAoB;IACpB,SAAS,EAAE,MAAM,CAAA;IACjB,QAAQ,EAAE,MAAM,CAAA;IAChB;;gDAE4C;IAC5C,WAAW,EAAE,OAAO,CAAA;CACrB;AAED,MAAM,WAAW,qBAAqB;IACpC,qBAAqB;IACrB,SAAS,EAAE,MAAM,CAAA;IACjB,QAAQ,EAAE,MAAM,CAAA;IAChB,0EAA0E;IAC1E,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB;;2DAEuD;IACvD,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,WAAW,CAAC,CAAA;IACpC;sEACkE;IAClE,gBAAgB,EAAE,MAAM,EAAE,CAAA;IAC1B,wEAAwE;IACxE,eAAe,EAAE,MAAM,EAAE,CAAA;CAC1B;AAED,MAAM,WAAW,cAAc;IAC7B,QAAQ,EAAE,UAAU,GAAG,MAAM,GAAG,QAAQ,GAAG,KAAK,CAAA;IAChD,IAAI,EAAE,MAAM,GAAG,MAAM,GAAG,aAAa,GAAG,KAAK,GAAG,aAAa,GAAG,eAAe,CAAA;IAC/E,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;IACd,8DAA8D;IAC9D,YAAY,CAAC,EAAE,MAAM,CAAA;CACtB"}
|
|
@@ -1,7 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Forward type re-export so `insight-report.ts` doesn't pull the entire
|
|
3
|
-
* judge-calibration module's runtime surface into the app-facing contract
|
|
4
|
-
* type graph.
|
|
5
|
-
*/
|
|
6
|
-
export type { ContinuousAgreement } from '../judge-calibration';
|
|
7
|
-
//# sourceMappingURL=insight-types-fwd.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"insight-types-fwd.d.ts","sourceRoot":"","sources":["../../src/contract/insight-types-fwd.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,YAAY,EAAE,mBAAmB,EAAE,MAAM,sBAAsB,CAAA"}
|
|
@@ -1,97 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* `fromAgentTrace` — provenance correlation from Cursor's Agent Trace spec
|
|
3
|
-
* (https://github.com/cursor/agent-trace, RFC v0.1.0).
|
|
4
|
-
*
|
|
5
|
-
* Agent Trace is NOT a run/quality trace — it carries no outcome, score, or
|
|
6
|
-
* cost. It records *code authorship*: for a VCS revision, which AI model /
|
|
7
|
-
* conversation authored which file ranges. It explicitly disclaims quality
|
|
8
|
-
* assessment — which is exactly what `analyzeRuns` adds.
|
|
9
|
-
*
|
|
10
|
-
* The two layers join on a key the substrate already has: a `RunRecord`
|
|
11
|
-
* carries `commitSha`, and an Agent Trace record is keyed by
|
|
12
|
-
* `vcs.revision`. So this adapter does not produce `RunRecord`s — it builds a
|
|
13
|
-
* provenance index by commit and partitions existing runs by their authoring
|
|
14
|
-
* model. Feed each cohort to `analyzeRuns` (or pass one as `baselineRuns`) to
|
|
15
|
-
* answer the question no run-only trace can: *which authoring agent's code
|
|
16
|
-
* fails / regresses / costs more.*
|
|
17
|
-
*
|
|
18
|
-
* Granularity is commit-level (the SHA join). Per-file/per-line correlation
|
|
19
|
-
* would require runs to record which files they exercised — out of scope.
|
|
20
|
-
*/
|
|
21
|
-
import type { RunRecord } from '../../run-record';
|
|
22
|
-
export type AgentTraceContributorType = 'human' | 'ai' | 'mixed' | 'unknown';
|
|
23
|
-
export interface AgentTraceContributor {
|
|
24
|
-
type: AgentTraceContributorType;
|
|
25
|
-
/** models.dev id, e.g. `anthropic/claude-opus-4-5-20251101`. */
|
|
26
|
-
model_id?: string;
|
|
27
|
-
}
|
|
28
|
-
export interface AgentTraceRange {
|
|
29
|
-
start_line: number;
|
|
30
|
-
end_line: number;
|
|
31
|
-
content_hash?: string;
|
|
32
|
-
/** Per-range contributor override (agent handoffs). Wins over the
|
|
33
|
-
* conversation-level contributor for these lines. */
|
|
34
|
-
contributor?: AgentTraceContributor;
|
|
35
|
-
}
|
|
36
|
-
export interface AgentTraceConversation {
|
|
37
|
-
url?: string;
|
|
38
|
-
contributor?: AgentTraceContributor;
|
|
39
|
-
ranges: AgentTraceRange[];
|
|
40
|
-
}
|
|
41
|
-
export interface AgentTraceFile {
|
|
42
|
-
path: string;
|
|
43
|
-
conversations: AgentTraceConversation[];
|
|
44
|
-
}
|
|
45
|
-
export interface AgentTraceRecord {
|
|
46
|
-
version: string;
|
|
47
|
-
id: string;
|
|
48
|
-
timestamp: string;
|
|
49
|
-
vcs?: {
|
|
50
|
-
type: string;
|
|
51
|
-
revision: string;
|
|
52
|
-
};
|
|
53
|
-
tool?: {
|
|
54
|
-
name?: string;
|
|
55
|
-
version?: string;
|
|
56
|
-
};
|
|
57
|
-
files: AgentTraceFile[];
|
|
58
|
-
}
|
|
59
|
-
/** Authorship provenance for one VCS revision, aggregated across the record's
|
|
60
|
-
* files/conversations/ranges. */
|
|
61
|
-
export interface AuthoringProvenance {
|
|
62
|
-
commitSha: string;
|
|
63
|
-
/** Unique AI model ids that authored code in this commit (type ai|mixed). */
|
|
64
|
-
aiModels: string[];
|
|
65
|
-
/** Tools that produced the records (e.g. `cursor`). */
|
|
66
|
-
tools: string[];
|
|
67
|
-
conversationCount: number;
|
|
68
|
-
fileCount: number;
|
|
69
|
-
/** Total attributed lines (sum of range spans). */
|
|
70
|
-
lineCount: number;
|
|
71
|
-
/** True if any range was authored (in whole or part) by a human. */
|
|
72
|
-
humanInvolved: boolean;
|
|
73
|
-
}
|
|
74
|
-
export type AgentTraceIndex = Map<string, AuthoringProvenance>;
|
|
75
|
-
/**
|
|
76
|
-
* Build a commit → provenance index from Agent Trace records. Multiple records
|
|
77
|
-
* for the same revision are merged. Records without `vcs.revision` are skipped
|
|
78
|
-
* (the SHA is the join key — without it there is nothing to correlate against).
|
|
79
|
-
*/
|
|
80
|
-
export declare function parseAgentTrace(records: AgentTraceRecord[]): AgentTraceIndex;
|
|
81
|
-
export interface PartitionByAuthoringModelResult {
|
|
82
|
-
/** Runs grouped by each AI model that authored code in the run's commit. A
|
|
83
|
-
* run whose commit had multiple authoring models appears under EACH — the
|
|
84
|
-
* cohorts overlap by construction at commit granularity. */
|
|
85
|
-
byModel: Map<string, RunRecord[]>;
|
|
86
|
-
/** Runs whose `commitSha` had no Agent Trace provenance (no record, or no
|
|
87
|
-
* AI authorship). Kept separate — never silently folded into a cohort. */
|
|
88
|
-
unattributed: RunRecord[];
|
|
89
|
-
}
|
|
90
|
-
/**
|
|
91
|
-
* Partition runs by the AI model(s) that authored the code at each run's
|
|
92
|
-
* `commitSha`. Feed `byModel.get(modelId)` to `analyzeRuns`, or compare two
|
|
93
|
-
* model cohorts via `analyzeRuns({ runs: a, baselineRuns: b })` for a lift CI
|
|
94
|
-
* on "model A's code vs model B's code".
|
|
95
|
-
*/
|
|
96
|
-
export declare function partitionRunsByAuthoringModel(runs: RunRecord[], index: AgentTraceIndex): PartitionByAuthoringModelResult;
|
|
97
|
-
//# sourceMappingURL=agent-trace.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"agent-trace.d.ts","sourceRoot":"","sources":["../../../src/contract/intake/agent-trace.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;GAmBG;AAEH,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,kBAAkB,CAAA;AAIjD,MAAM,MAAM,yBAAyB,GAAG,OAAO,GAAG,IAAI,GAAG,OAAO,GAAG,SAAS,CAAA;AAE5E,MAAM,WAAW,qBAAqB;IACpC,IAAI,EAAE,yBAAyB,CAAA;IAC/B,gEAAgE;IAChE,QAAQ,CAAC,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,eAAe;IAC9B,UAAU,EAAE,MAAM,CAAA;IAClB,QAAQ,EAAE,MAAM,CAAA;IAChB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB;0DACsD;IACtD,WAAW,CAAC,EAAE,qBAAqB,CAAA;CACpC;AAED,MAAM,WAAW,sBAAsB;IACrC,GAAG,CAAC,EAAE,MAAM,CAAA;IACZ,WAAW,CAAC,EAAE,qBAAqB,CAAA;IACnC,MAAM,EAAE,eAAe,EAAE,CAAA;CAC1B;AAED,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAA;IACZ,aAAa,EAAE,sBAAsB,EAAE,CAAA;CACxC;AAED,MAAM,WAAW,gBAAgB;IAC/B,OAAO,EAAE,MAAM,CAAA;IACf,EAAE,EAAE,MAAM,CAAA;IACV,SAAS,EAAE,MAAM,CAAA;IACjB,GAAG,CAAC,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAA;IACxC,IAAI,CAAC,EAAE;QAAE,IAAI,CAAC,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAA;KAAE,CAAA;IAC1C,KAAK,EAAE,cAAc,EAAE,CAAA;CACxB;AAID;kCACkC;AAClC,MAAM,WAAW,mBAAmB;IAClC,SAAS,EAAE,MAAM,CAAA;IACjB,6EAA6E;IAC7E,QAAQ,EAAE,MAAM,EAAE,CAAA;IAClB,uDAAuD;IACvD,KAAK,EAAE,MAAM,EAAE,CAAA;IACf,iBAAiB,EAAE,MAAM,CAAA;IACzB,SAAS,EAAE,MAAM,CAAA;IACjB,mDAAmD;IACnD,SAAS,EAAE,MAAM,CAAA;IACjB,oEAAoE;IACpE,aAAa,EAAE,OAAO,CAAA;CACvB;AAED,MAAM,MAAM,eAAe,GAAG,GAAG,CAAC,MAAM,EAAE,mBAAmB,CAAC,CAAA;AAM9D;;;;GAIG;AACH,wBAAgB,eAAe,CAAC,OAAO,EAAE,gBAAgB,EAAE,GAAG,eAAe,CA+D5E;AAID,MAAM,WAAW,+BAA+B;IAC9C;;iEAE6D;IAC7D,OAAO,EAAE,GAAG,CAAC,MAAM,EAAE,SAAS,EAAE,CAAC,CAAA;IACjC;+EAC2E;IAC3E,YAAY,EAAE,SAAS,EAAE,CAAA;CAC1B;AAED;;;;;GAKG;AACH,wBAAgB,6BAA6B,CAC3C,IAAI,EAAE,SAAS,EAAE,EACjB,KAAK,EAAE,eAAe,GACrB,+BAA+B,CAkBjC"}
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"code-agent-session.d.ts","sourceRoot":"","sources":["../../../src/contract/intake/code-agent-session.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EAAE,SAAS,EAAE,WAAW,EAAiB,MAAM,kBAAkB,CAAA;AAE7E,MAAM,MAAM,sBAAsB,GAAG,OAAO,GAAG,aAAa,GAAG,UAAU,GAAG,WAAW,GAAG,IAAI,CAAA;AAE9F,MAAM,WAAW,oBAAoB;IACnC,OAAO,EAAE,OAAO,EAAE,CAAA;IAClB,cAAc,EAAE,MAAM,CAAA;CACvB;AAED,MAAM,WAAW,uBAAuB;IACtC,OAAO,EAAE,MAAM,CAAA;IACf,YAAY,EAAE,MAAM,CAAA;IACpB,iBAAiB,EAAE,MAAM,CAAA;IACzB,cAAc,EAAE,MAAM,CAAA;IACtB,SAAS,EAAE,MAAM,CAAA;IACjB,WAAW,EAAE,MAAM,CAAA;IACnB,UAAU,EAAE,MAAM,CAAA;IAClB,aAAa,EAAE,MAAM,CAAA;IACrB,cAAc,EAAE,MAAM,CAAA;IACtB,aAAa,EAAE,MAAM,CAAA;IACrB,YAAY,EAAE,MAAM,CAAA;IACpB,cAAc,EAAE,MAAM,CAAA;IACtB,YAAY,EAAE,MAAM,CAAA;IACpB,kBAAkB,EAAE,MAAM,CAAA;IAC1B,OAAO,EAAE,MAAM,CAAA;IACf,aAAa,EAAE,MAAM,CAAA;IACrB,UAAU,EAAE,MAAM,CAAA;IAClB,UAAU,EAAE,MAAM,CAAA;IAClB,gBAAgB,EAAE,MAAM,CAAA;IACxB,mBAAmB,EAAE,MAAM,CAAA;IAC3B,mBAAmB,EAAE,MAAM,CAAA;IAC3B,eAAe,EAAE,MAAM,CAAA;IACvB,eAAe,EAAE,MAAM,CAAA;IACvB,WAAW,EAAE,MAAM,CAAA;IACnB,YAAY,EAAE,MAAM,CAAA;IACpB,YAAY,EAAE,MAAM,CAAA;IACpB,eAAe,EAAE,MAAM,CAAA;IACvB,MAAM,EAAE,MAAM,CAAA;IACd,YAAY,EAAE,MAAM,CAAA;CACrB;AAED,MAAM,WAAW,0BAA0B;IACzC,MAAM,EAAE,sBAAsB,CAAA;IAC9B,SAAS,EAAE,MAAM,CAAA;IACjB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,OAAO,EAAE,MAAM,CAAA;IACf,cAAc,EAAE,MAAM,CAAA;IACtB,aAAa,EAAE,OAAO,CAAA;IACtB,yBAAyB,EAAE,OAAO,CAAA;IAClC,eAAe,EAAE,OAAO,CAAA;IACxB,aAAa,EAAE,OAAO,CAAA;IACtB,OAAO,EAAE,OAAO,CAAA;IAChB,QAAQ,EAAE,MAAM,EAAE,CAAA;CACnB;AAED,MAAM,WAAW,4BAA4B;IAC3C,IAAI,EAAE,SAAS,EAAE,CAAA;IACjB,WAAW,EAAE,0BAA0B,EAAE,CAAA;IACzC,OAAO,EAAE,uBAAuB,EAAE,CAAA;CACnC;AAED,MAAM,WAAW,6BAA6B;IAC5C,OAAO,EAAE,OAAO,EAAE,CAAA;IAClB,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,QAAQ,CAAC,EAAE,WAAW,CAAA;IACtB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,KAAK,CAAC,EAAE,MAAM,CAAA;CACf;AAED,wBAAgB,mBAAmB,CAAC,KAAK,EAAE,MAAM,GAAG,oBAAoB,CAavE;AAED,wBAAgB,gBAAgB,CAC9B,OAAO,EAAE,6BAA6B,GACrC,4BAA4B,CAE9B;AAED,wBAAgB,qBAAqB,CACnC,OAAO,EAAE,6BAA6B,GACrC,4BAA4B,CAE9B;AAED,wBAAgB,mBAAmB,CACjC,OAAO,EAAE,6BAA6B,GACrC,4BAA4B,CAE9B;AAED,wBAAgB,mBAAmB,CACjC,OAAO,EAAE,6BAA6B,GACrC,4BAA4B,CAE9B;AAED,wBAAgB,aAAa,CAC3B,OAAO,EAAE,6BAA6B,GACrC,4BAA4B,CAE9B;AAED,eAAO,MAAM,kBAAkB,sBAAgB,CAAA"}
|
|
@@ -1,87 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* # `intake/feedback-table` — multi-rater approve/reject corpus → `RunRecord[]`.
|
|
3
|
-
*
|
|
4
|
-
* The generic shape behind Obsidian's `#approved` / `#rejected` tags, a
|
|
5
|
-
* Google Sheet, a Postgres `feedback` table, or any CSV with ratings.
|
|
6
|
-
*
|
|
7
|
-
* Caller supplies one row per (run, rater) tuple plus per-run metadata; the
|
|
8
|
-
* adapter rolls them up into the substrate-canonical `RunRecord` shape so
|
|
9
|
-
* `analyzeRuns({ runs, raterScores })` can produce inter-rater agreement,
|
|
10
|
-
* disagreement triage, and downstream recommendations.
|
|
11
|
-
*
|
|
12
|
-
* Per-run `RunRecord.outcome.searchScore` is the rater-mean rating
|
|
13
|
-
* (normalised to 0..1 when scale is supplied); `outcome.raw` carries the
|
|
14
|
-
* per-rater scores keyed by rater id for downstream attribution.
|
|
15
|
-
*/
|
|
16
|
-
import type { RunRecord, RunSplitTag } from '../../run-record';
|
|
17
|
-
export interface FeedbackTableRow {
|
|
18
|
-
/** Stable id for this run — the unit a rater scored. Drives pairing
|
|
19
|
-
* across analysis primitives. */
|
|
20
|
-
runId: string;
|
|
21
|
-
/** Identifier of the rater that produced this rating. */
|
|
22
|
-
rater: string;
|
|
23
|
-
/** The rating itself. Accepts boolean (approve/reject), 0..1 scalar,
|
|
24
|
-
* or any numeric scale — see `scale`. */
|
|
25
|
-
rating: number | boolean;
|
|
26
|
-
/** Optional metadata carried through to `RunRecord.outcome.raw` and the
|
|
27
|
-
* custom-shape metadata bag. */
|
|
28
|
-
metadata?: Record<string, unknown>;
|
|
29
|
-
}
|
|
30
|
-
export interface FeedbackTableMeta {
|
|
31
|
-
runId: string;
|
|
32
|
-
/** When omitted, defaults to `'feedback-corpus'`. Used to group related
|
|
33
|
-
* runs in `analyzeRuns()` lift analysis. */
|
|
34
|
-
experimentId?: string;
|
|
35
|
-
/** When omitted, defaults to `runId` — each run is its own candidate. */
|
|
36
|
-
candidateId?: string;
|
|
37
|
-
/** Cost in USD, when available. Set to 0 when unknown — the consumer's
|
|
38
|
-
* cost analysis sections will collapse gracefully. */
|
|
39
|
-
costUsd?: number;
|
|
40
|
-
/** Wall-clock ms, when available. Defaults to 0. */
|
|
41
|
-
wallMs?: number;
|
|
42
|
-
/** Model identifier including snapshot. Default `unknown@unknown`. */
|
|
43
|
-
model?: string;
|
|
44
|
-
/** Optional sha256 of the prompt; default `'sha256:unknown'`. */
|
|
45
|
-
promptHash?: string;
|
|
46
|
-
/** Default `'sha256:unknown'`. */
|
|
47
|
-
configHash?: string;
|
|
48
|
-
/** Default `'unknown'`. */
|
|
49
|
-
commitSha?: string;
|
|
50
|
-
/** Default `'holdout'` — feedback corpora are by nature the holdout
|
|
51
|
-
* signal a closed-loop improvement aims at. */
|
|
52
|
-
splitTag?: RunSplitTag;
|
|
53
|
-
/** Free-form metadata available to consumers via the cast-out path on
|
|
54
|
-
* the resulting RunRecord. */
|
|
55
|
-
extras?: Record<string, unknown>;
|
|
56
|
-
}
|
|
57
|
-
export interface FromFeedbackTableOptions {
|
|
58
|
-
/** Per-(run, rater) ratings. */
|
|
59
|
-
ratings: FeedbackTableRow[];
|
|
60
|
-
/** Per-run metadata. When a runId appears in `ratings` but not here, the
|
|
61
|
-
* adapter synthesises minimal metadata with defaults documented above. */
|
|
62
|
-
meta?: FeedbackTableMeta[];
|
|
63
|
-
/** Rating scale. Provide `{ min, max }` for non-0..1 numeric scales.
|
|
64
|
-
* Booleans are normalised: true → 1, false → 0. Default: assumes
|
|
65
|
-
* ratings are already 0..1. */
|
|
66
|
-
scale?: {
|
|
67
|
-
min: number;
|
|
68
|
-
max: number;
|
|
69
|
-
};
|
|
70
|
-
/** When true, the rater scores are emitted into `raterScores` (a sibling
|
|
71
|
-
* array `analyzeRuns()` accepts) instead of being averaged into the
|
|
72
|
-
* run's `outcome.searchScore`. Default `true` — preserves rater-level
|
|
73
|
-
* signal for inter-rater analysis. */
|
|
74
|
-
emitRaterScores?: boolean;
|
|
75
|
-
}
|
|
76
|
-
export interface FromFeedbackTableResult {
|
|
77
|
-
runs: RunRecord[];
|
|
78
|
-
/** Rater-level scores ready to pass into `analyzeRuns({ raterScores })`
|
|
79
|
-
* for inter-rater agreement + disagreement triage. */
|
|
80
|
-
raterScores: Array<{
|
|
81
|
-
runId: string;
|
|
82
|
-
rater: string;
|
|
83
|
-
score: number;
|
|
84
|
-
}>;
|
|
85
|
-
}
|
|
86
|
-
export declare function fromFeedbackTable(opts: FromFeedbackTableOptions): FromFeedbackTableResult;
|
|
87
|
-
//# sourceMappingURL=feedback-table.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"feedback-table.d.ts","sourceRoot":"","sources":["../../../src/contract/intake/feedback-table.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,OAAO,KAAK,EAAiC,SAAS,EAAE,WAAW,EAAE,MAAM,kBAAkB,CAAA;AAE7F,MAAM,WAAW,gBAAgB;IAC/B;sCACkC;IAClC,KAAK,EAAE,MAAM,CAAA;IACb,yDAAyD;IACzD,KAAK,EAAE,MAAM,CAAA;IACb;8CAC0C;IAC1C,MAAM,EAAE,MAAM,GAAG,OAAO,CAAA;IACxB;qCACiC;IACjC,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,iBAAiB;IAChC,KAAK,EAAE,MAAM,CAAA;IACb;iDAC6C;IAC7C,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,yEAAyE;IACzE,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB;2DACuD;IACvD,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,oDAAoD;IACpD,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,sEAAsE;IACtE,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,iEAAiE;IACjE,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,kCAAkC;IAClC,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,2BAA2B;IAC3B,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB;oDACgD;IAChD,QAAQ,CAAC,EAAE,WAAW,CAAA;IACtB;mCAC+B;IAC/B,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACjC;AAED,MAAM,WAAW,wBAAwB;IACvC,gCAAgC;IAChC,OAAO,EAAE,gBAAgB,EAAE,CAAA;IAC3B;+EAC2E;IAC3E,IAAI,CAAC,EAAE,iBAAiB,EAAE,CAAA;IAC1B;;oCAEgC;IAChC,KAAK,CAAC,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,GAAG,EAAE,MAAM,CAAA;KAAE,CAAA;IACpC;;;2CAGuC;IACvC,eAAe,CAAC,EAAE,OAAO,CAAA;CAC1B;AAED,MAAM,WAAW,uBAAuB;IACtC,IAAI,EAAE,SAAS,EAAE,CAAA;IACjB;2DACuD;IACvD,WAAW,EAAE,KAAK,CAAC;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;CACpE;AAED,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,wBAAwB,GAAG,uBAAuB,CAwEzF"}
|
|
@@ -1,22 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* # `@tangle-network/agent-eval/contract/intake` — adapters that turn external
|
|
3
|
-
* data sources into the substrate-canonical `RunRecord` shape.
|
|
4
|
-
*
|
|
5
|
-
* Each adapter is a function `source → RunRecord[]` (or
|
|
6
|
-
* `{ runs, raterScores }` when the adapter exposes multi-rater signal).
|
|
7
|
-
* Consumers pipe the output straight into `analyzeRuns({ runs })`.
|
|
8
|
-
*
|
|
9
|
-
* Adapters:
|
|
10
|
-
* - `fromFeedbackTable` — multi-rater approve/reject corpus
|
|
11
|
-
* (Obsidian, Sheets, CSV, Postgres).
|
|
12
|
-
* - `fromOtelSpans` — OTel `TraceSpanEvent[]` from any OTel-compatible
|
|
13
|
-
* observability stack.
|
|
14
|
-
* - `fromCodexSession` / `fromClaudeCodeSession` / `fromOpenCodeSession` /
|
|
15
|
-
* `fromKimiCodeSession` / `fromPiSession` — local coding-agent and
|
|
16
|
-
* graph-shaped sessions projected into process-scored `RunRecord`s.
|
|
17
|
-
*/
|
|
18
|
-
export { type AgentTraceContributor, type AgentTraceContributorType, type AgentTraceConversation, type AgentTraceFile, type AgentTraceIndex, type AgentTraceRange, type AgentTraceRecord, type AuthoringProvenance, type PartitionByAuthoringModelResult, parseAgentTrace, partitionRunsByAuthoringModel, } from './agent-trace';
|
|
19
|
-
export { type CodeAgentSessionDiagnostic, type CodeAgentSessionIntakeOptions, type CodeAgentSessionIntakeResult, type CodeAgentSessionMetrics, type CodeAgentSessionSource, fromClaudeCodeSession, fromCodexSession, fromKimiCodeSession, fromOpenCodeSession, fromPigraphSession, fromPiSession, type ParsedCodeAgentJsonl, parseCodeAgentJsonl, } from './code-agent-session';
|
|
20
|
-
export { type FeedbackTableMeta, type FeedbackTableRow, type FromFeedbackTableOptions, type FromFeedbackTableResult, fromFeedbackTable, } from './feedback-table';
|
|
21
|
-
export { type FromOtelSpansOptions, fromOtelSpans } from './otel-spans';
|
|
22
|
-
//# sourceMappingURL=index.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../src/contract/intake/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,EACL,KAAK,qBAAqB,EAC1B,KAAK,yBAAyB,EAC9B,KAAK,sBAAsB,EAC3B,KAAK,cAAc,EACnB,KAAK,eAAe,EACpB,KAAK,eAAe,EACpB,KAAK,gBAAgB,EACrB,KAAK,mBAAmB,EACxB,KAAK,+BAA+B,EACpC,eAAe,EACf,6BAA6B,GAC9B,MAAM,eAAe,CAAA;AACtB,OAAO,EACL,KAAK,0BAA0B,EAC/B,KAAK,6BAA6B,EAClC,KAAK,4BAA4B,EACjC,KAAK,uBAAuB,EAC5B,KAAK,sBAAsB,EAC3B,qBAAqB,EACrB,gBAAgB,EAChB,mBAAmB,EACnB,mBAAmB,EACnB,kBAAkB,EAClB,aAAa,EACb,KAAK,oBAAoB,EACzB,mBAAmB,GACpB,MAAM,sBAAsB,CAAA;AAC7B,OAAO,EACL,KAAK,iBAAiB,EACtB,KAAK,gBAAgB,EACrB,KAAK,wBAAwB,EAC7B,KAAK,uBAAuB,EAC5B,iBAAiB,GAClB,MAAM,kBAAkB,CAAA;AACzB,OAAO,EAAE,KAAK,oBAAoB,EAAE,aAAa,EAAE,MAAM,cAAc,CAAA"}
|
|
@@ -1,33 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* # `intake/otel-spans` — OTel `TraceSpanEvent[]` → `RunRecord[]`.
|
|
3
|
-
*
|
|
4
|
-
* Turns an existing observability stream into the substrate-canonical
|
|
5
|
-
* `RunRecord` shape so consumers with logs but no eval discipline can
|
|
6
|
-
* call `analyzeRuns()` against their production traffic immediately.
|
|
7
|
-
*
|
|
8
|
-
* Pivot rule: spans are grouped by `tangle.runId` (the same attribute the
|
|
9
|
-
* hosted-tier wire format uses) or, when absent, by `traceId`. One group
|
|
10
|
-
* becomes one `RunRecord`. The root span (no `parentSpanId`) supplies:
|
|
11
|
-
*
|
|
12
|
-
* - `runId` (the group key)
|
|
13
|
-
* - `wallMs` from `endTimeUnixNano - startTimeUnixNano`
|
|
14
|
-
* - `model` from `gen_ai.request.model` / `llm.model` / `tangle.model`
|
|
15
|
-
* - cost from `cost.usd` / `gen_ai.usage.cost_usd` / `tangle.cost.usd`
|
|
16
|
-
* - token usage from `gen_ai.usage.{input,output}_tokens`
|
|
17
|
-
* - `outcome.searchScore` from `tangle.score` / `eval.score` when
|
|
18
|
-
* present; `outcome.raw` collects every numeric attribute.
|
|
19
|
-
*
|
|
20
|
-
* Spans that ERRORed (`status.code === 'ERROR'`) populate `failureMode`
|
|
21
|
-
* with their `name` so `analyzeRuns()`'s failure clustering sees them.
|
|
22
|
-
*/
|
|
23
|
-
import type { TraceSpanEvent } from '../../hosted/types';
|
|
24
|
-
import type { RunRecord, RunSplitTag } from '../../run-record';
|
|
25
|
-
export interface FromOtelSpansOptions {
|
|
26
|
-
spans: TraceSpanEvent[];
|
|
27
|
-
/** Default split tag for synthesized records. Defaults to `'holdout'`. */
|
|
28
|
-
defaultSplit?: RunSplitTag;
|
|
29
|
-
/** Default `experimentId` when not present on any span. */
|
|
30
|
-
experimentId?: string;
|
|
31
|
-
}
|
|
32
|
-
export declare function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[];
|
|
33
|
-
//# sourceMappingURL=otel-spans.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"otel-spans.d.ts","sourceRoot":"","sources":["../../../src/contract/intake/otel-spans.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;GAqBG;AAEH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,oBAAoB,CAAA;AACxD,OAAO,KAAK,EAGV,SAAS,EACT,WAAW,EAEZ,MAAM,kBAAkB,CAAA;AAezB,MAAM,WAAW,oBAAoB;IACnC,KAAK,EAAE,cAAc,EAAE,CAAA;IACvB,0EAA0E;IAC1E,YAAY,CAAC,EAAE,WAAW,CAAA;IAC1B,2DAA2D;IAC3D,YAAY,CAAC,EAAE,MAAM,CAAA;CACtB;AAED,wBAAgB,aAAa,CAAC,IAAI,EAAE,oBAAoB,GAAG,SAAS,EAAE,CA0DrE"}
|