@tangle-network/agent-eval 0.95.0 → 0.95.1
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/{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/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 +688 -106
- 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-C1NCIZ9o.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/{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-CncDq9qE.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/{matrix/types.d.ts → types-BUxNaJ8c.d.ts} +11 -9
- package/dist/{types.d.ts → types-C7DGg5ex.d.ts} +33 -31
- package/dist/{campaign/types.d.ts → types-DQRY8ZT-.d.ts} +47 -44
- 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/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
|
@@ -1,107 +0,0 @@
|
|
|
1
|
-
import type { AgentProfile } from '@tangle-network/agent-interface';
|
|
2
|
-
import type { MatrixResult } from '../matrix';
|
|
3
|
-
import { type JudgeConfig, type JudgeScore } from './judges';
|
|
4
|
-
import type { MultishotArtifact, MultishotMessage, MultishotPersona, MultishotShape, MultishotToolDefinition, MultishotToolExecutor } from './types';
|
|
5
|
-
export interface ConversationJudgeInput<TPersona extends MultishotPersona> {
|
|
6
|
-
transcript: MultishotMessage[];
|
|
7
|
-
persona: TPersona;
|
|
8
|
-
}
|
|
9
|
-
export interface ArtifactJudgeInput<TPersona extends MultishotPersona> {
|
|
10
|
-
artifact: MultishotArtifact;
|
|
11
|
-
persona: TPersona;
|
|
12
|
-
}
|
|
13
|
-
export interface MultishotJudges<TPersona extends MultishotPersona> {
|
|
14
|
-
/** Scores the full transcript end-to-end (always runs). */
|
|
15
|
-
conversation: JudgeConfig<ConversationJudgeInput<TPersona>>;
|
|
16
|
-
/** Scores each code-type artifact. Optional — omit when domain has no code artifacts. */
|
|
17
|
-
codeReview?: JudgeConfig<ArtifactJudgeInput<TPersona>>;
|
|
18
|
-
/** Scores each non-code (research/content/template) artifact. Optional. */
|
|
19
|
-
contentQuality?: JudgeConfig<ArtifactJudgeInput<TPersona>>;
|
|
20
|
-
/** Which artifact types route to codeReview. Defaults to ['code']. */
|
|
21
|
-
codeArtifactTypes?: string[];
|
|
22
|
-
/** Which artifact types route to contentQuality. Defaults to ['research']. */
|
|
23
|
-
contentArtifactTypes?: string[];
|
|
24
|
-
}
|
|
25
|
-
export interface CellCompositeScore {
|
|
26
|
-
composite: number;
|
|
27
|
-
conversation: JudgeScore;
|
|
28
|
-
codeReview?: {
|
|
29
|
-
perArtifact: Array<JudgeScore & {
|
|
30
|
-
turn: number;
|
|
31
|
-
type: string;
|
|
32
|
-
}>;
|
|
33
|
-
composite: number;
|
|
34
|
-
};
|
|
35
|
-
contentQuality?: {
|
|
36
|
-
perArtifact: Array<JudgeScore & {
|
|
37
|
-
turn: number;
|
|
38
|
-
type: string;
|
|
39
|
-
}>;
|
|
40
|
-
composite: number;
|
|
41
|
-
};
|
|
42
|
-
}
|
|
43
|
-
export interface RunMultishotMatrixOptions<TPersona extends MultishotPersona> {
|
|
44
|
-
/** AgentProfile axis (matrix primary). */
|
|
45
|
-
profiles: Array<{
|
|
46
|
-
id: string;
|
|
47
|
-
value: AgentProfile;
|
|
48
|
-
}>;
|
|
49
|
-
/** Persona axis. */
|
|
50
|
-
personas: TPersona[];
|
|
51
|
-
/** Persona-shaping callbacks. */
|
|
52
|
-
shape: MultishotShape<TPersona>;
|
|
53
|
-
/** Judge configurations. */
|
|
54
|
-
judges: MultishotJudges<TPersona>;
|
|
55
|
-
/** Tool definitions advertised to the agent. Defaults to delegate_research + delegate_code. */
|
|
56
|
-
tools?: MultishotToolDefinition[];
|
|
57
|
-
/** Map from tool name → inline executor. Must align with `tools`. */
|
|
58
|
-
toolExecutors?: Record<string, MultishotToolExecutor>;
|
|
59
|
-
/** Tool name → artifact type label. Defaults to research/code mapping. */
|
|
60
|
-
artifactTypeFor?: (toolName: string) => string | undefined;
|
|
61
|
-
/** Where per-cell artifacts land. Cells write to `<runDir>/<profileId>/<personaId>/rep-N/`. */
|
|
62
|
-
runDir: string;
|
|
63
|
-
/** Replicates per (profile, persona) cell. */
|
|
64
|
-
reps?: number;
|
|
65
|
-
/** Max conversation turns per cell. */
|
|
66
|
-
maxTurns?: number;
|
|
67
|
-
/** Max concurrent cells. */
|
|
68
|
-
maxConcurrency?: number;
|
|
69
|
-
/** Total $ ceiling across the matrix; cells aborted past this. */
|
|
70
|
-
costCeiling?: number;
|
|
71
|
-
/** Agent model. */
|
|
72
|
-
agentModel?: string;
|
|
73
|
-
/** Driver model. */
|
|
74
|
-
driverModel?: string;
|
|
75
|
-
/** Pass-thru fields. */
|
|
76
|
-
apiKey?: string;
|
|
77
|
-
baseUrl?: string;
|
|
78
|
-
}
|
|
79
|
-
interface CellOutput {
|
|
80
|
-
turns: number;
|
|
81
|
-
toolCalls: number;
|
|
82
|
-
artifactCount: number;
|
|
83
|
-
}
|
|
84
|
-
export interface CellCompositeInput {
|
|
85
|
-
conversation: JudgeScore;
|
|
86
|
-
/** Present iff the codeReview judge is configured. */
|
|
87
|
-
codeReviews?: ReadonlyArray<JudgeScore>;
|
|
88
|
-
/** Present iff the contentQuality judge is configured. */
|
|
89
|
-
contentReviews?: ReadonlyArray<JudgeScore>;
|
|
90
|
-
}
|
|
91
|
-
/** Cell composite = mean over configured judge slots, excluding failed
|
|
92
|
-
* scores: a failed conversation judge or an all-failed artifact slot carries
|
|
93
|
-
* no signal and is dropped from the mean. `composite` is 0 only when EVERY
|
|
94
|
-
* configured slot failed (`allJudgesFailed` distinguishes that from a real
|
|
95
|
-
* zero). Pure — exported for deterministic testing. */
|
|
96
|
-
export declare function computeCellComposite(input: CellCompositeInput): {
|
|
97
|
-
composite: number;
|
|
98
|
-
codeComposite: number;
|
|
99
|
-
contentComposite: number;
|
|
100
|
-
allJudgesFailed: boolean;
|
|
101
|
-
};
|
|
102
|
-
export interface RunMultishotMatrixResult {
|
|
103
|
-
matrix: MatrixResult<CellOutput>;
|
|
104
|
-
}
|
|
105
|
-
export declare function runMultishotMatrix<TPersona extends MultishotPersona>(opts: RunMultishotMatrixOptions<TPersona>): Promise<RunMultishotMatrixResult>;
|
|
106
|
-
export {};
|
|
107
|
-
//# sourceMappingURL=matrix.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"matrix.d.ts","sourceRoot":"","sources":["../../src/multishot/matrix.ts"],"names":[],"mappings":"AAUA,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,iCAAiC,CAAA;AACnE,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,WAAW,CAAA;AAE7C,OAAO,EAAE,KAAK,WAAW,EAAE,KAAK,UAAU,EAAY,MAAM,UAAU,CAAA;AAEtE,OAAO,KAAK,EACV,iBAAiB,EACjB,gBAAgB,EAChB,gBAAgB,EAChB,cAAc,EACd,uBAAuB,EACvB,qBAAqB,EACtB,MAAM,SAAS,CAAA;AAEhB,MAAM,WAAW,sBAAsB,CAAC,QAAQ,SAAS,gBAAgB;IACvE,UAAU,EAAE,gBAAgB,EAAE,CAAA;IAC9B,OAAO,EAAE,QAAQ,CAAA;CAClB;AAED,MAAM,WAAW,kBAAkB,CAAC,QAAQ,SAAS,gBAAgB;IACnE,QAAQ,EAAE,iBAAiB,CAAA;IAC3B,OAAO,EAAE,QAAQ,CAAA;CAClB;AAED,MAAM,WAAW,eAAe,CAAC,QAAQ,SAAS,gBAAgB;IAChE,2DAA2D;IAC3D,YAAY,EAAE,WAAW,CAAC,sBAAsB,CAAC,QAAQ,CAAC,CAAC,CAAA;IAC3D,yFAAyF;IACzF,UAAU,CAAC,EAAE,WAAW,CAAC,kBAAkB,CAAC,QAAQ,CAAC,CAAC,CAAA;IACtD,2EAA2E;IAC3E,cAAc,CAAC,EAAE,WAAW,CAAC,kBAAkB,CAAC,QAAQ,CAAC,CAAC,CAAA;IAC1D,sEAAsE;IACtE,iBAAiB,CAAC,EAAE,MAAM,EAAE,CAAA;IAC5B,8EAA8E;IAC9E,oBAAoB,CAAC,EAAE,MAAM,EAAE,CAAA;CAChC;AAED,MAAM,WAAW,kBAAkB;IACjC,SAAS,EAAE,MAAM,CAAA;IACjB,YAAY,EAAE,UAAU,CAAA;IACxB,UAAU,CAAC,EAAE;QACX,WAAW,EAAE,KAAK,CAAC,UAAU,GAAG;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,IAAI,EAAE,MAAM,CAAA;SAAE,CAAC,CAAA;QAC/D,SAAS,EAAE,MAAM,CAAA;KAClB,CAAA;IACD,cAAc,CAAC,EAAE;QACf,WAAW,EAAE,KAAK,CAAC,UAAU,GAAG;YAAE,IAAI,EAAE,MAAM,CAAC;YAAC,IAAI,EAAE,MAAM,CAAA;SAAE,CAAC,CAAA;QAC/D,SAAS,EAAE,MAAM,CAAA;KAClB,CAAA;CACF;AAED,MAAM,WAAW,yBAAyB,CAAC,QAAQ,SAAS,gBAAgB;IAC1E,0CAA0C;IAC1C,QAAQ,EAAE,KAAK,CAAC;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,YAAY,CAAA;KAAE,CAAC,CAAA;IACpD,oBAAoB;IACpB,QAAQ,EAAE,QAAQ,EAAE,CAAA;IACpB,iCAAiC;IACjC,KAAK,EAAE,cAAc,CAAC,QAAQ,CAAC,CAAA;IAC/B,4BAA4B;IAC5B,MAAM,EAAE,eAAe,CAAC,QAAQ,CAAC,CAAA;IACjC,+FAA+F;IAC/F,KAAK,CAAC,EAAE,uBAAuB,EAAE,CAAA;IACjC,qEAAqE;IACrE,aAAa,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,qBAAqB,CAAC,CAAA;IACrD,0EAA0E;IAC1E,eAAe,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,MAAM,GAAG,SAAS,CAAA;IAC1D,+FAA+F;IAC/F,MAAM,EAAE,MAAM,CAAA;IACd,8CAA8C;IAC9C,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,uCAAuC;IACvC,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,4BAA4B;IAC5B,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB,kEAAkE;IAClE,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,mBAAmB;IACnB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,oBAAoB;IACpB,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,wBAAwB;IACxB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,OAAO,CAAC,EAAE,MAAM,CAAA;CACjB;AAED,UAAU,UAAU;IAClB,KAAK,EAAE,MAAM,CAAA;IACb,SAAS,EAAE,MAAM,CAAA;IACjB,aAAa,EAAE,MAAM,CAAA;CACtB;AAcD,MAAM,WAAW,kBAAkB;IACjC,YAAY,EAAE,UAAU,CAAA;IACxB,sDAAsD;IACtD,WAAW,CAAC,EAAE,aAAa,CAAC,UAAU,CAAC,CAAA;IACvC,0DAA0D;IAC1D,cAAc,CAAC,EAAE,aAAa,CAAC,UAAU,CAAC,CAAA;CAC3C;AAED;;;;wDAIwD;AACxD,wBAAgB,oBAAoB,CAAC,KAAK,EAAE,kBAAkB,GAAG;IAC/D,SAAS,EAAE,MAAM,CAAA;IACjB,aAAa,EAAE,MAAM,CAAA;IACrB,gBAAgB,EAAE,MAAM,CAAA;IACxB,eAAe,EAAE,OAAO,CAAA;CACzB,CAoBA;AAED,MAAM,WAAW,wBAAwB;IACvC,MAAM,EAAE,YAAY,CAAC,UAAU,CAAC,CAAA;CACjC;AAED,wBAAsB,kBAAkB,CAAC,QAAQ,SAAS,gBAAgB,EACxE,IAAI,EAAE,yBAAyB,CAAC,QAAQ,CAAC,GACxC,OAAO,CAAC,wBAAwB,CAAC,CA2InC"}
|
|
@@ -1,23 +0,0 @@
|
|
|
1
|
-
import type { AgentProfile } from '@tangle-network/agent-interface';
|
|
2
|
-
import { type MultishotPersona, type MultishotResult, type MultishotShape, type MultishotToolDefinition, type MultishotToolExecutor } from './types';
|
|
3
|
-
export interface RunMultishotOptions<TPersona extends MultishotPersona> {
|
|
4
|
-
profile: AgentProfile;
|
|
5
|
-
persona: TPersona;
|
|
6
|
-
shape: MultishotShape<TPersona>;
|
|
7
|
-
/** Tool definitions advertised to the agent. Defaults to delegate_research + delegate_code. */
|
|
8
|
-
tools?: MultishotToolDefinition[];
|
|
9
|
-
/** Map from tool name → executor invoked inline when the agent emits a tool_call. */
|
|
10
|
-
toolExecutors?: Record<string, MultishotToolExecutor>;
|
|
11
|
-
/** Map from tool name → artifact type label written into MultishotArtifact.type.
|
|
12
|
-
* Tools without a mapping still execute, but their results aren't surfaced as
|
|
13
|
-
* typed artifacts (only as tool messages in the transcript). */
|
|
14
|
-
artifactTypeFor?: (toolName: string) => string | undefined;
|
|
15
|
-
maxTurns?: number;
|
|
16
|
-
agentModel?: string;
|
|
17
|
-
driverModel?: string;
|
|
18
|
-
apiKey?: string;
|
|
19
|
-
baseUrl?: string;
|
|
20
|
-
signal?: AbortSignal;
|
|
21
|
-
}
|
|
22
|
-
export declare function runMultishot<TPersona extends MultishotPersona>(opts: RunMultishotOptions<TPersona>): Promise<MultishotResult>;
|
|
23
|
-
//# sourceMappingURL=multishot.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"multishot.d.ts","sourceRoot":"","sources":["../../src/multishot/multishot.ts"],"names":[],"mappings":"AAQA,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,iCAAiC,CAAA;AAQnE,OAAO,EAIL,KAAK,gBAAgB,EACrB,KAAK,eAAe,EACpB,KAAK,cAAc,EACnB,KAAK,uBAAuB,EAC5B,KAAK,qBAAqB,EAC3B,MAAM,SAAS,CAAA;AAEhB,MAAM,WAAW,mBAAmB,CAAC,QAAQ,SAAS,gBAAgB;IACpE,OAAO,EAAE,YAAY,CAAA;IACrB,OAAO,EAAE,QAAQ,CAAA;IACjB,KAAK,EAAE,cAAc,CAAC,QAAQ,CAAC,CAAA;IAC/B,+FAA+F;IAC/F,KAAK,CAAC,EAAE,uBAAuB,EAAE,CAAA;IACjC,qFAAqF;IACrF,aAAa,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,qBAAqB,CAAC,CAAA;IACrD;;qEAEiE;IACjE,eAAe,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,MAAM,GAAG,SAAS,CAAA;IAC1D,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,MAAM,CAAC,EAAE,WAAW,CAAA;CACrB;AAED,wBAAsB,YAAY,CAAC,QAAQ,SAAS,gBAAgB,EAClE,IAAI,EAAE,mBAAmB,CAAC,QAAQ,CAAC,GAClC,OAAO,CAAC,eAAe,CAAC,CAwI1B"}
|
|
@@ -1,37 +0,0 @@
|
|
|
1
|
-
import type { MultishotToolDefinition } from './types';
|
|
2
|
-
export interface RouterCompletionRequest {
|
|
3
|
-
apiKey: string;
|
|
4
|
-
baseUrl: string;
|
|
5
|
-
model: string;
|
|
6
|
-
messages: Array<Record<string, unknown>>;
|
|
7
|
-
tools?: MultishotToolDefinition[];
|
|
8
|
-
temperature?: number;
|
|
9
|
-
maxTokens?: number;
|
|
10
|
-
signal?: AbortSignal;
|
|
11
|
-
}
|
|
12
|
-
export interface RouterToolCall {
|
|
13
|
-
id: string;
|
|
14
|
-
type: 'function';
|
|
15
|
-
function: {
|
|
16
|
-
name: string;
|
|
17
|
-
arguments: string;
|
|
18
|
-
};
|
|
19
|
-
}
|
|
20
|
-
export interface RouterCompletionResponse {
|
|
21
|
-
message: {
|
|
22
|
-
content?: string | null;
|
|
23
|
-
tool_calls?: RouterToolCall[];
|
|
24
|
-
};
|
|
25
|
-
usage?: {
|
|
26
|
-
prompt_tokens?: number;
|
|
27
|
-
completion_tokens?: number;
|
|
28
|
-
};
|
|
29
|
-
}
|
|
30
|
-
export declare function routerCompletion(req: RouterCompletionRequest): Promise<RouterCompletionResponse>;
|
|
31
|
-
export declare function estimateRouterCost(model: string, usage?: {
|
|
32
|
-
prompt_tokens?: number;
|
|
33
|
-
completion_tokens?: number;
|
|
34
|
-
}): number;
|
|
35
|
-
export declare function defaultRouterBaseUrl(): string;
|
|
36
|
-
export declare function requireRouterApiKey(): string;
|
|
37
|
-
//# sourceMappingURL=router.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"router.d.ts","sourceRoot":"","sources":["../../src/multishot/router.ts"],"names":[],"mappings":"AAIA,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,SAAS,CAAA;AAEtD,MAAM,WAAW,uBAAuB;IACtC,MAAM,EAAE,MAAM,CAAA;IACd,OAAO,EAAE,MAAM,CAAA;IACf,KAAK,EAAE,MAAM,CAAA;IACb,QAAQ,EAAE,KAAK,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,CAAA;IACxC,KAAK,CAAC,EAAE,uBAAuB,EAAE,CAAA;IACjC,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,MAAM,CAAC,EAAE,WAAW,CAAA;CACrB;AAED,MAAM,WAAW,cAAc;IAC7B,EAAE,EAAE,MAAM,CAAA;IACV,IAAI,EAAE,UAAU,CAAA;IAChB,QAAQ,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAA;KAAE,CAAA;CAC9C;AAED,MAAM,WAAW,wBAAwB;IACvC,OAAO,EAAE;QAAE,OAAO,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QAAC,UAAU,CAAC,EAAE,cAAc,EAAE,CAAA;KAAE,CAAA;IACnE,KAAK,CAAC,EAAE;QAAE,aAAa,CAAC,EAAE,MAAM,CAAC;QAAC,iBAAiB,CAAC,EAAE,MAAM,CAAA;KAAE,CAAA;CAC/D;AAED,wBAAsB,gBAAgB,CACpC,GAAG,EAAE,uBAAuB,GAC3B,OAAO,CAAC,wBAAwB,CAAC,CA0BnC;AAID,wBAAgB,kBAAkB,CAChC,KAAK,EAAE,MAAM,EACb,KAAK,CAAC,EAAE;IAAE,aAAa,CAAC,EAAE,MAAM,CAAC;IAAC,iBAAiB,CAAC,EAAE,MAAM,CAAA;CAAE,GAC7D,MAAM,CAiBR;AAED,wBAAgB,oBAAoB,IAAI,MAAM,CAK7C;AAED,wBAAgB,mBAAmB,IAAI,MAAM,CAI5C"}
|
|
@@ -1,60 +0,0 @@
|
|
|
1
|
-
export interface MultishotMessage {
|
|
2
|
-
role: 'user' | 'assistant' | 'tool';
|
|
3
|
-
content: string;
|
|
4
|
-
toolCallId?: string;
|
|
5
|
-
toolCalls?: Array<{
|
|
6
|
-
id: string;
|
|
7
|
-
name: string;
|
|
8
|
-
args: Record<string, unknown>;
|
|
9
|
-
}>;
|
|
10
|
-
}
|
|
11
|
-
export interface MultishotArtifact {
|
|
12
|
-
type: string;
|
|
13
|
-
turn: number;
|
|
14
|
-
invocation: {
|
|
15
|
-
name: string;
|
|
16
|
-
args: Record<string, unknown>;
|
|
17
|
-
};
|
|
18
|
-
content: string;
|
|
19
|
-
}
|
|
20
|
-
export interface MultishotResult {
|
|
21
|
-
transcript: MultishotMessage[];
|
|
22
|
-
artifacts: MultishotArtifact[];
|
|
23
|
-
toolCalls: number;
|
|
24
|
-
durationMs: number;
|
|
25
|
-
costUsd: number;
|
|
26
|
-
}
|
|
27
|
-
export interface MultishotToolDefinition {
|
|
28
|
-
type: 'function';
|
|
29
|
-
function: {
|
|
30
|
-
name: string;
|
|
31
|
-
description: string;
|
|
32
|
-
parameters: Record<string, unknown>;
|
|
33
|
-
};
|
|
34
|
-
}
|
|
35
|
-
export type MultishotToolExecutor = (args: Record<string, unknown>, ctx: {
|
|
36
|
-
apiKey: string;
|
|
37
|
-
baseUrl: string;
|
|
38
|
-
signal?: AbortSignal;
|
|
39
|
-
}) => Promise<{
|
|
40
|
-
content: string;
|
|
41
|
-
costUsd: number;
|
|
42
|
-
}>;
|
|
43
|
-
export interface MultishotPersona {
|
|
44
|
-
/** Stable identifier — used for per-cell artifact paths + matrix axis keys. */
|
|
45
|
-
id: string;
|
|
46
|
-
/** Per-domain payload (income/profile/voice/etc.) shaped by the consumer. */
|
|
47
|
-
[k: string]: unknown;
|
|
48
|
-
}
|
|
49
|
-
export interface MultishotShape<TPersona extends MultishotPersona> {
|
|
50
|
-
/** Opening user message (turn 0) — the persona's first ask. */
|
|
51
|
-
buildOpener: (persona: TPersona) => string;
|
|
52
|
-
/** System prompt the driver LLM uses to roleplay the persona. Should set
|
|
53
|
-
* voice, goals, constraints, time-pressure, and the "never go silent" rule. */
|
|
54
|
-
buildDriverSystemPrompt: (persona: TPersona) => string;
|
|
55
|
-
}
|
|
56
|
-
export declare class MultishotDriverEmptyError extends Error {
|
|
57
|
-
readonly turn: number;
|
|
58
|
-
constructor(turn: number);
|
|
59
|
-
}
|
|
60
|
-
//# sourceMappingURL=types.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/multishot/types.ts"],"names":[],"mappings":"AAEA,MAAM,WAAW,gBAAgB;IAC/B,IAAI,EAAE,MAAM,GAAG,WAAW,GAAG,MAAM,CAAA;IACnC,OAAO,EAAE,MAAM,CAAA;IACf,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,SAAS,CAAC,EAAE,KAAK,CAAC;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KAAE,CAAC,CAAA;CAC/E;AAED,MAAM,WAAW,iBAAiB;IAChC,IAAI,EAAE,MAAM,CAAA;IACZ,IAAI,EAAE,MAAM,CAAA;IACZ,UAAU,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KAAE,CAAA;IAC3D,OAAO,EAAE,MAAM,CAAA;CAChB;AAED,MAAM,WAAW,eAAe;IAC9B,UAAU,EAAE,gBAAgB,EAAE,CAAA;IAC9B,SAAS,EAAE,iBAAiB,EAAE,CAAA;IAC9B,SAAS,EAAE,MAAM,CAAA;IACjB,UAAU,EAAE,MAAM,CAAA;IAClB,OAAO,EAAE,MAAM,CAAA;CAChB;AAED,MAAM,WAAW,uBAAuB;IACtC,IAAI,EAAE,UAAU,CAAA;IAChB,QAAQ,EAAE;QACR,IAAI,EAAE,MAAM,CAAA;QACZ,WAAW,EAAE,MAAM,CAAA;QACnB,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;KACpC,CAAA;CACF;AAED,MAAM,MAAM,qBAAqB,GAAG,CAClC,IAAI,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,EAC7B,GAAG,EAAE;IAAE,MAAM,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,WAAW,CAAA;CAAE,KAC3D,OAAO,CAAC;IAAE,OAAO,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,CAAC,CAAA;AAElD,MAAM,WAAW,gBAAgB;IAC/B,+EAA+E;IAC/E,EAAE,EAAE,MAAM,CAAA;IACV,6EAA6E;IAC7E,CAAC,CAAC,EAAE,MAAM,GAAG,OAAO,CAAA;CACrB;AAED,MAAM,WAAW,cAAc,CAAC,QAAQ,SAAS,gBAAgB;IAC/D,+DAA+D;IAC/D,WAAW,EAAE,CAAC,OAAO,EAAE,QAAQ,KAAK,MAAM,CAAA;IAC1C;oFACgF;IAChF,uBAAuB,EAAE,CAAC,OAAO,EAAE,QAAQ,KAAK,MAAM,CAAA;CACvD;AAED,qBAAa,yBAA0B,SAAQ,KAAK;aACtB,IAAI,EAAE,MAAM;gBAAZ,IAAI,EAAE,MAAM;CAIzC"}
|
package/dist/observability.d.ts
DELETED
|
@@ -1,71 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Observability adapters — bidirectional parity with production backends.
|
|
3
|
-
*
|
|
4
|
-
* `LangfuseAdapter` maps a Run's spans into Langfuse generation/score
|
|
5
|
-
* records (schema-compatible; we don't depend on the SDK — consumers
|
|
6
|
-
* POST the returned JSON to their Langfuse collector).
|
|
7
|
-
*
|
|
8
|
-
* `PrometheusEmitter` converts a TraceStore into a Prometheus text-
|
|
9
|
-
* exposition-format string (counters + gauges for runs, tool calls,
|
|
10
|
-
* errors, cost). Drop into a `/metrics` handler; no SDK needed.
|
|
11
|
-
*
|
|
12
|
-
* `replayTraceThroughJudge` is the canonical "re-score with a new
|
|
13
|
-
* judge" path — takes an existing run, runs a judge function over
|
|
14
|
-
* each LLM span, emits JudgeVerdict spans back into the store.
|
|
15
|
-
*/
|
|
16
|
-
import type { LlmSpan } from './trace/schema';
|
|
17
|
-
import type { TraceStore } from './trace/store';
|
|
18
|
-
export interface LangfuseGeneration {
|
|
19
|
-
id: string;
|
|
20
|
-
traceId: string;
|
|
21
|
-
name: string;
|
|
22
|
-
model: string;
|
|
23
|
-
input: unknown;
|
|
24
|
-
output: unknown;
|
|
25
|
-
startTime: string;
|
|
26
|
-
endTime: string;
|
|
27
|
-
usage: {
|
|
28
|
-
input: number;
|
|
29
|
-
output: number;
|
|
30
|
-
total: number;
|
|
31
|
-
totalCost: number;
|
|
32
|
-
};
|
|
33
|
-
metadata: Record<string, unknown>;
|
|
34
|
-
}
|
|
35
|
-
export interface LangfuseScore {
|
|
36
|
-
id: string;
|
|
37
|
-
traceId: string;
|
|
38
|
-
observationId: string;
|
|
39
|
-
name: string;
|
|
40
|
-
value: number;
|
|
41
|
-
comment?: string;
|
|
42
|
-
}
|
|
43
|
-
export interface LangfuseEnvelope {
|
|
44
|
-
traceId: string;
|
|
45
|
-
generations: LangfuseGeneration[];
|
|
46
|
-
scores: LangfuseScore[];
|
|
47
|
-
}
|
|
48
|
-
export declare function toLangfuseEnvelope(store: TraceStore, runId: string): Promise<LangfuseEnvelope>;
|
|
49
|
-
export declare function toPrometheusText(store: TraceStore): Promise<string>;
|
|
50
|
-
export interface JudgeReplayResult {
|
|
51
|
-
spanId: string;
|
|
52
|
-
targetSpanId: string;
|
|
53
|
-
dimension: string;
|
|
54
|
-
score: number;
|
|
55
|
-
rationale?: string;
|
|
56
|
-
}
|
|
57
|
-
/**
|
|
58
|
-
* Apply a judge function to every LLM span in a run and record the
|
|
59
|
-
* results as JudgeVerdict spans. This is the canonical "no re-execution"
|
|
60
|
-
* re-scoring path — you supply a pure judge `(llmSpan) → verdict`.
|
|
61
|
-
*/
|
|
62
|
-
export declare function replayTraceThroughJudge(store: TraceStore, runId: string, judge: {
|
|
63
|
-
id: string;
|
|
64
|
-
dimension: string;
|
|
65
|
-
score: (span: LlmSpan) => Promise<{
|
|
66
|
-
score: number;
|
|
67
|
-
rationale?: string;
|
|
68
|
-
evidence?: string;
|
|
69
|
-
}>;
|
|
70
|
-
}): Promise<JudgeReplayResult[]>;
|
|
71
|
-
//# sourceMappingURL=observability.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"observability.d.ts","sourceRoot":"","sources":["../src/observability.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAKH,OAAO,KAAK,EAAE,OAAO,EAAQ,MAAM,gBAAgB,CAAA;AACnD,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,eAAe,CAAA;AAI/C,MAAM,WAAW,kBAAkB;IACjC,EAAE,EAAE,MAAM,CAAA;IACV,OAAO,EAAE,MAAM,CAAA;IACf,IAAI,EAAE,MAAM,CAAA;IACZ,KAAK,EAAE,MAAM,CAAA;IACb,KAAK,EAAE,OAAO,CAAA;IACd,MAAM,EAAE,OAAO,CAAA;IACf,SAAS,EAAE,MAAM,CAAA;IACjB,OAAO,EAAE,MAAM,CAAA;IACf,KAAK,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAA;KAAE,CAAA;IAC1E,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CAClC;AAED,MAAM,WAAW,aAAa;IAC5B,EAAE,EAAE,MAAM,CAAA;IACV,OAAO,EAAE,MAAM,CAAA;IACf,aAAa,EAAE,MAAM,CAAA;IACrB,IAAI,EAAE,MAAM,CAAA;IACZ,KAAK,EAAE,MAAM,CAAA;IACb,OAAO,CAAC,EAAE,MAAM,CAAA;CACjB;AAED,MAAM,WAAW,gBAAgB;IAC/B,OAAO,EAAE,MAAM,CAAA;IACf,WAAW,EAAE,kBAAkB,EAAE,CAAA;IACjC,MAAM,EAAE,aAAa,EAAE,CAAA;CACxB;AAED,wBAAsB,kBAAkB,CACtC,KAAK,EAAE,UAAU,EACjB,KAAK,EAAE,MAAM,GACZ,OAAO,CAAC,gBAAgB,CAAC,CAmC3B;AAID,wBAAsB,gBAAgB,CAAC,KAAK,EAAE,UAAU,GAAG,OAAO,CAAC,MAAM,CAAC,CAuDzE;AAQD,MAAM,WAAW,iBAAiB;IAChC,MAAM,EAAE,MAAM,CAAA;IACd,YAAY,EAAE,MAAM,CAAA;IACpB,SAAS,EAAE,MAAM,CAAA;IACjB,KAAK,EAAE,MAAM,CAAA;IACb,SAAS,CAAC,EAAE,MAAM,CAAA;CACnB;AAED;;;;GAIG;AACH,wBAAsB,uBAAuB,CAC3C,KAAK,EAAE,UAAU,EACjB,KAAK,EAAE,MAAM,EACb,KAAK,EAAE;IACL,EAAE,EAAE,MAAM,CAAA;IACV,SAAS,EAAE,MAAM,CAAA;IACjB,KAAK,EAAE,CAAC,IAAI,EAAE,OAAO,KAAK,OAAO,CAAC;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;CAC5F,GACA,OAAO,CAAC,iBAAiB,EAAE,CAAC,CA0B9B"}
|
package/dist/oracle.d.ts
DELETED
|
@@ -1,55 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Declarative oracles — ground-truth assertions without an LLM.
|
|
3
|
-
*
|
|
4
|
-
* Lifted from browser-agent-driver's _oracle.mjs. When you know the
|
|
5
|
-
* expected outcome exactly (a URL, a text fragment, a JSON shape), you
|
|
6
|
-
* don't need an LLM judge — you need a regex. These oracles are
|
|
7
|
-
* composable pass/fail checks over an observation bundle.
|
|
8
|
-
*
|
|
9
|
-
* Each oracle returns { pass, detail, evidence? } and has a short
|
|
10
|
-
* `id` for reporting. `evaluateOracles` runs a batch and aggregates.
|
|
11
|
-
*/
|
|
12
|
-
export interface OracleObservation {
|
|
13
|
-
/** Final observable text output from the agent (response, page snapshot, stdout). */
|
|
14
|
-
text?: string;
|
|
15
|
-
/** Final URL — for browser-style scenarios. */
|
|
16
|
-
url?: string;
|
|
17
|
-
/** Any structured JSON the agent produced. */
|
|
18
|
-
json?: unknown;
|
|
19
|
-
/** Free-form context used by custom oracles. */
|
|
20
|
-
context?: Record<string, unknown>;
|
|
21
|
-
}
|
|
22
|
-
export interface OracleResult {
|
|
23
|
-
id: string;
|
|
24
|
-
pass: boolean;
|
|
25
|
-
detail: string;
|
|
26
|
-
evidence?: string;
|
|
27
|
-
}
|
|
28
|
-
export interface Oracle {
|
|
29
|
-
id: string;
|
|
30
|
-
check(obs: OracleObservation): OracleResult;
|
|
31
|
-
}
|
|
32
|
-
export declare function textInSnapshot(needle: string, opts?: {
|
|
33
|
-
caseSensitive?: boolean;
|
|
34
|
-
}): Oracle;
|
|
35
|
-
export declare function urlContains(fragment: string): Oracle;
|
|
36
|
-
export declare function jsonShape(expected: Record<string, unknown>): Oracle;
|
|
37
|
-
export declare function regexMatches(pattern: RegExp): Oracle;
|
|
38
|
-
/**
|
|
39
|
-
* Anti-bot detector — distinguishes genuine failures from blocked navigation
|
|
40
|
-
* (cloudflare, recaptcha, etc). Returns an Oracle that PASSES when no block
|
|
41
|
-
* marker is present; on block, detail names the blocker so runners can tag
|
|
42
|
-
* results as "blocked" rather than "failed". Lifted from browser-agent-driver.
|
|
43
|
-
*/
|
|
44
|
-
export declare function notBlocked(): Oracle;
|
|
45
|
-
export interface OracleReport {
|
|
46
|
-
results: OracleResult[];
|
|
47
|
-
pass: boolean;
|
|
48
|
-
passCount: number;
|
|
49
|
-
failCount: number;
|
|
50
|
-
/** 0-1 ratio of oracles passed. */
|
|
51
|
-
score: number;
|
|
52
|
-
}
|
|
53
|
-
/** Run all oracles against one observation and aggregate. */
|
|
54
|
-
export declare function evaluateOracles(obs: OracleObservation, oracles: Oracle[]): OracleReport;
|
|
55
|
-
//# sourceMappingURL=oracle.d.ts.map
|
package/dist/oracle.d.ts.map
DELETED
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"oracle.d.ts","sourceRoot":"","sources":["../src/oracle.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,MAAM,WAAW,iBAAiB;IAChC,qFAAqF;IACrF,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,+CAA+C;IAC/C,GAAG,CAAC,EAAE,MAAM,CAAA;IACZ,8CAA8C;IAC9C,IAAI,CAAC,EAAE,OAAO,CAAA;IACd,gDAAgD;IAChD,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CAClC;AAED,MAAM,WAAW,YAAY;IAC3B,EAAE,EAAE,MAAM,CAAA;IACV,IAAI,EAAE,OAAO,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;IACd,QAAQ,CAAC,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,MAAM;IACrB,EAAE,EAAE,MAAM,CAAA;IACV,KAAK,CAAC,GAAG,EAAE,iBAAiB,GAAG,YAAY,CAAA;CAC5C;AAED,wBAAgB,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,IAAI,GAAE;IAAE,aAAa,CAAC,EAAE,OAAO,CAAA;CAAO,GAAG,MAAM,CAiB7F;AAED,wBAAgB,WAAW,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAepD;AAED,wBAAgB,SAAS,CAAC,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,MAAM,CA4BnE;AAED,wBAAgB,YAAY,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAepD;AAED;;;;;GAKG;AACH,wBAAgB,UAAU,IAAI,MAAM,CA4BnC;AAED,MAAM,WAAW,YAAY;IAC3B,OAAO,EAAE,YAAY,EAAE,CAAA;IACvB,IAAI,EAAE,OAAO,CAAA;IACb,SAAS,EAAE,MAAM,CAAA;IACjB,SAAS,EAAE,MAAM,CAAA;IACjB,mCAAmC;IACnC,KAAK,EAAE,MAAM,CAAA;CACd;AAED,6DAA6D;AAC7D,wBAAgB,eAAe,CAAC,GAAG,EAAE,iBAAiB,EAAE,OAAO,EAAE,MAAM,EAAE,GAAG,YAAY,CAWvF"}
|
package/dist/orthogonality.d.ts
DELETED
|
@@ -1,35 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Inter-critic / inter-pass orthogonality.
|
|
3
|
-
*
|
|
4
|
-
* Detects redundant ensembles. When you run N critics (or N audit passes,
|
|
5
|
-
* or N specialized agents) on the same input, you want them to disagree —
|
|
6
|
-
* each contributing distinct signal. If they all converge on the same set
|
|
7
|
-
* of findings, you're paying N× cost for ~1× signal.
|
|
8
|
-
*
|
|
9
|
-
* The metric is `1 − mean pairwise cosine similarity` over bags of words
|
|
10
|
-
* extracted from each pass's outputs. 1.0 = fully orthogonal,
|
|
11
|
-
* 0.0 = fully redundant.
|
|
12
|
-
*
|
|
13
|
-
* Universal primitive: pass anything that produces text (findings, tool
|
|
14
|
-
* calls rendered as JSON, verdict strings) and the matcher derives its own
|
|
15
|
-
* vocabulary.
|
|
16
|
-
*/
|
|
17
|
-
export interface OrthogonalityInput<T> {
|
|
18
|
-
passes: Array<{
|
|
19
|
-
findings: T[];
|
|
20
|
-
}>;
|
|
21
|
-
/** Render one element to text. Default: defaultRender (concatenates string fields). */
|
|
22
|
-
text?: (item: T) => string;
|
|
23
|
-
/** Minimum token length kept in the bag. Default 4 (drops short fillers). */
|
|
24
|
-
minTokenLength?: number;
|
|
25
|
-
}
|
|
26
|
-
export interface OrthogonalityResult {
|
|
27
|
-
/** 1 − mean pairwise cosine similarity across passes. 1=fully orthogonal, 0=fully redundant. */
|
|
28
|
-
orthogonality: number;
|
|
29
|
-
/** Number of passes considered. */
|
|
30
|
-
passCount: number;
|
|
31
|
-
/** Pairwise cosine similarities, in upper-triangular order (for debugging). */
|
|
32
|
-
similarities: number[];
|
|
33
|
-
}
|
|
34
|
-
export declare function passOrthogonality<T>(input: OrthogonalityInput<T>): OrthogonalityResult;
|
|
35
|
-
//# sourceMappingURL=orthogonality.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"orthogonality.d.ts","sourceRoot":"","sources":["../src/orthogonality.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,MAAM,WAAW,kBAAkB,CAAC,CAAC;IACnC,MAAM,EAAE,KAAK,CAAC;QAAE,QAAQ,EAAE,CAAC,EAAE,CAAA;KAAE,CAAC,CAAA;IAChC,uFAAuF;IACvF,IAAI,CAAC,EAAE,CAAC,IAAI,EAAE,CAAC,KAAK,MAAM,CAAA;IAC1B,6EAA6E;IAC7E,cAAc,CAAC,EAAE,MAAM,CAAA;CACxB;AAED,MAAM,WAAW,mBAAmB;IAClC,gGAAgG;IAChG,aAAa,EAAE,MAAM,CAAA;IACrB,mCAAmC;IACnC,SAAS,EAAE,MAAM,CAAA;IACjB,+EAA+E;IAC/E,YAAY,EAAE,MAAM,EAAE,CAAA;CACvB;AAED,wBAAgB,iBAAiB,CAAC,CAAC,EAAE,KAAK,EAAE,kBAAkB,CAAC,CAAC,CAAC,GAAG,mBAAmB,CAoBtF"}
|
package/dist/otel-pipeline.d.ts
DELETED
|
@@ -1,31 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Pipeline-level OTEL integration — auto-attaches an OTEL exporter when
|
|
3
|
-
* OTEL_EXPORTER_OTLP_ENDPOINT is set. Pipelines call `withOtelPipeline()`
|
|
4
|
-
* to get a configured exporter + shutdown handle without manual wiring.
|
|
5
|
-
*/
|
|
6
|
-
import { type OtelExportConfig, type OtelExporter } from './trace/otel-export';
|
|
7
|
-
export interface OtelPipelineHandle {
|
|
8
|
-
/** The active exporter, or undefined if no endpoint is configured. */
|
|
9
|
-
exporter: OtelExporter | undefined;
|
|
10
|
-
/** Call at pipeline end to flush + shutdown. Safe to call even if exporter is undefined. */
|
|
11
|
-
shutdown(): Promise<void>;
|
|
12
|
-
}
|
|
13
|
-
export interface OtelPipelineOptions {
|
|
14
|
-
/** Override OTEL config. */
|
|
15
|
-
otelConfig?: OtelExportConfig;
|
|
16
|
-
/** Pipeline-specific resource attributes. */
|
|
17
|
-
pipelineKind?: string;
|
|
18
|
-
pipelineId?: string;
|
|
19
|
-
}
|
|
20
|
-
/**
|
|
21
|
-
* Create an OTEL exporter scoped to a pipeline run. Auto-reads
|
|
22
|
-
* OTEL_EXPORTER_OTLP_ENDPOINT from env when no explicit config is passed.
|
|
23
|
-
*
|
|
24
|
-
* Returns a handle with `exporter` (possibly undefined) and `shutdown()`.
|
|
25
|
-
*/
|
|
26
|
-
export declare function withOtelPipeline(opts?: OtelPipelineOptions): OtelPipelineHandle;
|
|
27
|
-
/**
|
|
28
|
-
* Check if OTEL export is configured (endpoint is set).
|
|
29
|
-
*/
|
|
30
|
-
export declare function isOtelConfigured(): boolean;
|
|
31
|
-
//# sourceMappingURL=otel-pipeline.d.ts.map
|
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"otel-pipeline.d.ts","sourceRoot":"","sources":["../src/otel-pipeline.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,EAAsB,KAAK,gBAAgB,EAAE,KAAK,YAAY,EAAE,MAAM,qBAAqB,CAAA;AAElG,MAAM,WAAW,kBAAkB;IACjC,sEAAsE;IACtE,QAAQ,EAAE,YAAY,GAAG,SAAS,CAAA;IAClC,4FAA4F;IAC5F,QAAQ,IAAI,OAAO,CAAC,IAAI,CAAC,CAAA;CAC1B;AAED,MAAM,WAAW,mBAAmB;IAClC,4BAA4B;IAC5B,UAAU,CAAC,EAAE,gBAAgB,CAAA;IAC7B,6CAA6C;IAC7C,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,UAAU,CAAC,EAAE,MAAM,CAAA;CACpB;AAED;;;;;GAKG;AACH,wBAAgB,gBAAgB,CAAC,IAAI,CAAC,EAAE,mBAAmB,GAAG,kBAAkB,CAkB/E;AAED;;GAEG;AACH,wBAAgB,gBAAgB,IAAI,OAAO,CAE1C"}
|
package/dist/paraphrase.d.ts
DELETED
|
@@ -1,107 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Paraphrase robustness — mutates a scenario prompt in structure-
|
|
3
|
-
* preserving ways, re-scores, and reports score variance.
|
|
4
|
-
*
|
|
5
|
-
* Mutators are pure functions `(prompt: string) => string`. Ship a
|
|
6
|
-
* default set; consumers add domain-specific ones.
|
|
7
|
-
*
|
|
8
|
-
* Robustness score: 1 - stdDev(scores) / (mean if positive else 1).
|
|
9
|
-
* A perfect agent returns the same answer regardless of typo / case /
|
|
10
|
-
* reordering — any variance signals a brittle prompt.
|
|
11
|
-
*/
|
|
12
|
-
export type Mutator = (prompt: string, seed: number) => string;
|
|
13
|
-
export interface RobustnessResult {
|
|
14
|
-
originalScore: number;
|
|
15
|
-
variantScores: Array<{
|
|
16
|
-
mutator: string;
|
|
17
|
-
score: number;
|
|
18
|
-
mutated: string;
|
|
19
|
-
}>;
|
|
20
|
-
meanScore: number;
|
|
21
|
-
stdDev: number;
|
|
22
|
-
robustness: number;
|
|
23
|
-
}
|
|
24
|
-
export declare function paraphraseRobustness(prompt: string, mutators: Array<{
|
|
25
|
-
id: string;
|
|
26
|
-
fn: Mutator;
|
|
27
|
-
}>, scoreFn: (prompt: string) => Promise<number>, options?: {
|
|
28
|
-
seed?: number;
|
|
29
|
-
}): Promise<RobustnessResult>;
|
|
30
|
-
/** Lowercase the whole prompt. Robust models ignore case. */
|
|
31
|
-
export declare const lowercaseMutator: Mutator;
|
|
32
|
-
/** Reorder sentences. Robust models don't depend on sentence order. */
|
|
33
|
-
export declare const sentenceReorderMutator: Mutator;
|
|
34
|
-
/** Swap adjacent letter pairs (1 per 40 chars, min 1). Robust models tolerate typos. */
|
|
35
|
-
export declare const typoMutator: Mutator;
|
|
36
|
-
/** Add a benign politeness prefix. Robust models ignore flattery. */
|
|
37
|
-
export declare const politenessPrefixMutator: Mutator;
|
|
38
|
-
/** Compact whitespace, strip newlines. Robust models don't depend on formatting. */
|
|
39
|
-
export declare const whitespaceCollapseMutator: Mutator;
|
|
40
|
-
export declare const DEFAULT_MUTATORS: Array<{
|
|
41
|
-
id: string;
|
|
42
|
-
fn: Mutator;
|
|
43
|
-
}>;
|
|
44
|
-
export interface ParaphraseRobustnessScenarioInput {
|
|
45
|
-
scenarios: Array<{
|
|
46
|
-
id: string;
|
|
47
|
-
userTurns: string[];
|
|
48
|
-
}>;
|
|
49
|
-
/**
|
|
50
|
-
* Mutators applied to every user turn in every scenario. Each
|
|
51
|
-
* scenario is paraphrased once per mutator (so `reps` × `scenarios`
|
|
52
|
-
* × `mutators` total paraphrased runs).
|
|
53
|
-
*/
|
|
54
|
-
mutators: Array<{
|
|
55
|
-
name: string;
|
|
56
|
-
mutator: (text: string) => string;
|
|
57
|
-
}>;
|
|
58
|
-
/**
|
|
59
|
-
* Run a (possibly mutated) scenario and return its score in [0,1].
|
|
60
|
-
* Called once for the original turns of each scenario, and once per
|
|
61
|
-
* (scenario × mutator × rep) for the paraphrased variants.
|
|
62
|
-
*/
|
|
63
|
-
runScenario: (args: {
|
|
64
|
-
id: string;
|
|
65
|
-
userTurns: string[];
|
|
66
|
-
}) => Promise<{
|
|
67
|
-
score: number;
|
|
68
|
-
}>;
|
|
69
|
-
/** Times to repeat each (scenario × mutator) pair. Default 1. */
|
|
70
|
-
reps?: number;
|
|
71
|
-
}
|
|
72
|
-
export interface ParaphraseRobustnessScenarioResult {
|
|
73
|
-
/**
|
|
74
|
-
* Aggregate robustness: `mean(paraphrased) / mean(original)`,
|
|
75
|
-
* clipped to `[0, 1]`. `1` = paraphrasing didn't degrade the agent;
|
|
76
|
-
* `0` = paraphrasing destroyed it (or original was 0).
|
|
77
|
-
*/
|
|
78
|
-
score: number;
|
|
79
|
-
perScenario: Array<{
|
|
80
|
-
id: string;
|
|
81
|
-
originalScore: number;
|
|
82
|
-
paraphrasedMean: number;
|
|
83
|
-
/** Per-mutator delta (paraphrased − original); negative = mutator hurt. */
|
|
84
|
-
deltas: Record<string, number>;
|
|
85
|
-
}>;
|
|
86
|
-
mutators: string[];
|
|
87
|
-
}
|
|
88
|
-
/**
|
|
89
|
-
* Multi-turn convenience wrapper around {@link paraphraseRobustness}.
|
|
90
|
-
*
|
|
91
|
-
* Consumers with a list of multi-turn scenarios were hand-wrapping the
|
|
92
|
-
* single-prompt runner per scenario; this iterates for them. Mutators
|
|
93
|
-
* are applied to every user turn (mutator runs once per turn with a
|
|
94
|
-
* stable seed derived from the rep index).
|
|
95
|
-
*
|
|
96
|
-
* Contract:
|
|
97
|
-
* - Calls `runScenario` once with the original `userTurns` to
|
|
98
|
-
* establish the baseline `originalScore`.
|
|
99
|
-
* - For each `(scenario, mutator, rep)` combination, builds a
|
|
100
|
-
* mutated copy of `userTurns` (every turn passed through
|
|
101
|
-
* `mutator.mutator`) and calls `runScenario` again.
|
|
102
|
-
* - Aggregates per-scenario means, then computes the overall
|
|
103
|
-
* `mean(paraphrasedMean) / mean(originalScore)`, clipped to
|
|
104
|
-
* `[0, 1]`. If every original score is 0 the aggregate is 0.
|
|
105
|
-
*/
|
|
106
|
-
export declare function paraphraseRobustnessScenarios(args: ParaphraseRobustnessScenarioInput): Promise<ParaphraseRobustnessScenarioResult>;
|
|
107
|
-
//# sourceMappingURL=paraphrase.d.ts.map
|
package/dist/paraphrase.d.ts.map
DELETED
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"paraphrase.d.ts","sourceRoot":"","sources":["../src/paraphrase.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,MAAM,MAAM,OAAO,GAAG,CAAC,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,KAAK,MAAM,CAAA;AAE9D,MAAM,WAAW,gBAAgB;IAC/B,aAAa,EAAE,MAAM,CAAA;IACrB,aAAa,EAAE,KAAK,CAAC;QAAE,OAAO,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;IACzE,SAAS,EAAE,MAAM,CAAA;IACjB,MAAM,EAAE,MAAM,CAAA;IACd,UAAU,EAAE,MAAM,CAAA;CACnB;AAED,wBAAsB,oBAAoB,CACxC,MAAM,EAAE,MAAM,EACd,QAAQ,EAAE,KAAK,CAAC;IAAE,EAAE,EAAE,MAAM,CAAC;IAAC,EAAE,EAAE,OAAO,CAAA;CAAE,CAAC,EAC5C,OAAO,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,CAAC,EAC5C,OAAO,GAAE;IAAE,IAAI,CAAC,EAAE,MAAM,CAAA;CAAO,GAC9B,OAAO,CAAC,gBAAgB,CAAC,CAiB3B;AAID,6DAA6D;AAC7D,eAAO,MAAM,gBAAgB,EAAE,OAAgC,CAAA;AAE/D,uEAAuE;AACvE,eAAO,MAAM,sBAAsB,EAAE,OAWpC,CAAA;AAED,wFAAwF;AACxF,eAAO,MAAM,WAAW,EAAE,OAmBzB,CAAA;AAED,qEAAqE;AACrE,eAAO,MAAM,uBAAuB,EAAE,OAAwD,CAAA;AAE9F,oFAAoF;AACpF,eAAO,MAAM,yBAAyB,EAAE,OAA8C,CAAA;AAEtF,eAAO,MAAM,gBAAgB,EAAE,KAAK,CAAC;IAAE,EAAE,EAAE,MAAM,CAAC;IAAC,EAAE,EAAE,OAAO,CAAA;CAAE,CAM/D,CAAA;AAID,MAAM,WAAW,iCAAiC;IAChD,SAAS,EAAE,KAAK,CAAC;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,EAAE,CAAA;KAAE,CAAC,CAAA;IACrD;;;;OAIG;IACH,QAAQ,EAAE,KAAK,CAAC;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,MAAM,CAAA;KAAE,CAAC,CAAA;IACpE;;;;OAIG;IACH,WAAW,EAAE,CAAC,IAAI,EAAE;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,EAAE,CAAA;KAAE,KAAK,OAAO,CAAC;QAAE,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;IACtF,iEAAiE;IACjE,IAAI,CAAC,EAAE,MAAM,CAAA;CACd;AAED,MAAM,WAAW,kCAAkC;IACjD;;;;OAIG;IACH,KAAK,EAAE,MAAM,CAAA;IACb,WAAW,EAAE,KAAK,CAAC;QACjB,EAAE,EAAE,MAAM,CAAA;QACV,aAAa,EAAE,MAAM,CAAA;QACrB,eAAe,EAAE,MAAM,CAAA;QACvB,2EAA2E;QAC3E,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;KAC/B,CAAC,CAAA;IACF,QAAQ,EAAE,MAAM,EAAE,CAAA;CACnB;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,wBAAsB,6BAA6B,CACjD,IAAI,EAAE,iCAAiC,GACtC,OAAO,CAAC,kCAAkC,CAAC,CAoD7C"}
|
package/dist/pareto.d.ts.map
DELETED
|
@@ -1 +0,0 @@
|
|
|
1
|
-
{"version":3,"file":"pareto.d.ts","sourceRoot":"","sources":["../src/pareto.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,MAAM,MAAM,SAAS,GAAG,UAAU,GAAG,UAAU,CAAA;AAE/C,MAAM,WAAW,SAAS,CAAC,CAAC;IAC1B,oCAAoC;IACpC,IAAI,EAAE,MAAM,CAAA;IACZ,SAAS,EAAE,SAAS,CAAA;IACpB,KAAK,EAAE,CAAC,SAAS,EAAE,CAAC,KAAK,MAAM,CAAA;CAChC;AAED,MAAM,WAAW,YAAY,CAAC,CAAC;IAC7B,QAAQ,EAAE,CAAC,EAAE,CAAA;IACb,SAAS,EAAE,CAAC,EAAE,CAAA;IACd,+DAA+D;IAC/D,YAAY,EAAE,KAAK,CAAC;QAAE,SAAS,EAAE,CAAC,CAAC;QAAC,SAAS,EAAE,CAAC,EAAE,CAAA;KAAE,CAAC,CAAA;CACtD;AAED,0GAA0G;AAC1G,wBAAgB,SAAS,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,CAAC,EAAE,UAAU,EAAE,SAAS,CAAC,CAAC,CAAC,EAAE,GAAG,OAAO,CAY5E;AAED;;;;GAIG;AACH,wBAAgB,cAAc,CAAC,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,EAAE,UAAU,EAAE,SAAS,CAAC,CAAC,CAAC,EAAE,GAAG,YAAY,CAAC,CAAC,CAAC,CAiB9F;AAED;;;;;;;;GAQG;AACH,wBAAgB,WAAW,CAAC,CAAC,EAC3B,UAAU,EAAE,CAAC,EAAE,EACf,UAAU,EAAE,SAAS,CAAC,CAAC,CAAC,EAAE,EAC1B,OAAO,GAAE;IAAE,OAAO,CAAC,EAAE,OAAO,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CAAA;CAAO,GAC1D,KAAK,CAAC;IAAE,SAAS,EAAE,CAAC,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,CAAC,CA2BxC;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,gBAAgB,CAAC,CAAC,EAChC,UAAU,EAAE,CAAC,EAAE,EACf,UAAU,EAAE,SAAS,CAAC,CAAC,CAAC,EAAE,GACzB,KAAK,CAAC;IAAE,SAAS,EAAE,CAAC,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAA;CAAE,CAAC,CAsB3C;AAED;;;;GAIG;AACH,wBAAgB,0BAA0B,CAAC,CAAC,EAC1C,UAAU,EAAE,CAAC,EAAE,EACf,UAAU,EAAE,SAAS,CAAC,CAAC,CAAC,EAAE,GACzB,KAAK,CAAC;IAAE,SAAS,EAAE,CAAC,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAA;CAAE,CAAC,CAK3C"}
|
|
@@ -1,70 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Deterministic held-out partitioning over stable ids.
|
|
3
|
-
*
|
|
4
|
-
* A promotion gate's paired-delta math needs a fixed partition: the same
|
|
5
|
-
* scenario id lands in the same bucket forever, regardless of when it ran or
|
|
6
|
-
* which loop emitted the record — otherwise the gate has nothing paired to
|
|
7
|
-
* compare against. Picking one global split per run is wrong: every run then
|
|
8
|
-
* stamps either ALL search or ALL holdout, and no id is ever observed on both
|
|
9
|
-
* sides. Per-id deterministic hashing fixes that.
|
|
10
|
-
*
|
|
11
|
-
* Two consumers hand-roll this (agent-builder's `deterministicSplit` and the
|
|
12
|
-
* frontier persona-splitter). The substrate already has a 3-way benchmark
|
|
13
|
-
* `deterministicSplit` in `./benchmarks`; this is the generic, validated batch
|
|
14
|
-
* partitioner: it takes a list of ids, splits them by a stable FNV-1a hash, and
|
|
15
|
-
* fails loud when the inputs can't support a trustworthy held-out comparison
|
|
16
|
-
* (duplicate ids, or a holdout set below the significance floor).
|
|
17
|
-
*/
|
|
18
|
-
/** 32-bit FNV-1a hash. Stable, allocation-free, deterministic across runtimes —
|
|
19
|
-
* the same id+seed maps to the same bucket on every machine and process. */
|
|
20
|
-
export declare function fnv1a32(input: string): number;
|
|
21
|
-
/** Map an id+seed to the half-open unit interval [0, 1) deterministically. */
|
|
22
|
-
export declare function hashToUnit(id: string, seed: string): number;
|
|
23
|
-
export interface PartitionHeldOutOptions {
|
|
24
|
-
/** Partition seed. Bumping it reshuffles every assignment — do that only when
|
|
25
|
-
* the corpus/policy changes meaningfully so old and new records can't pair.
|
|
26
|
-
* Default 'held-out-v1'. */
|
|
27
|
-
seed?: string;
|
|
28
|
-
/** Fraction routed to the held-out bucket, in (0, 1). Default 0.5. */
|
|
29
|
-
holdoutFraction?: number;
|
|
30
|
-
/**
|
|
31
|
-
* Minimum held-out ids required for the split to be admissible. Below this,
|
|
32
|
-
* a paired comparison has too few observations to be significant, so the
|
|
33
|
-
* partition throws rather than hand back an underpowered holdout. Default 1
|
|
34
|
-
* (only the empty-holdout degenerate case is rejected). Raise it to enforce
|
|
35
|
-
* real statistical power. */
|
|
36
|
-
minHoldout?: number;
|
|
37
|
-
/** Same floor for the search side. Default 1. */
|
|
38
|
-
minSearch?: number;
|
|
39
|
-
}
|
|
40
|
-
export interface HeldOutPartition {
|
|
41
|
-
/** Ids the optimizer is allowed to read/optimize against. */
|
|
42
|
-
search: string[];
|
|
43
|
-
/** Ids reserved for gated, held-out evaluation. */
|
|
44
|
-
holdout: string[];
|
|
45
|
-
seed: string;
|
|
46
|
-
holdoutFraction: number;
|
|
47
|
-
}
|
|
48
|
-
/**
|
|
49
|
-
* Assign a single id to `'search'` or `'holdout'` deterministically. Same
|
|
50
|
-
* `(id, seed)` always returns the same tag — the invariant the gate relies on.
|
|
51
|
-
* Use this in a `splitTag:` field literal when stamping records one at a time;
|
|
52
|
-
* use `partitionHeldOut` when you have the whole id list up front and want the
|
|
53
|
-
* validated disjoint split.
|
|
54
|
-
*/
|
|
55
|
-
export declare function assignHeldOutTag(id: string, options?: {
|
|
56
|
-
seed?: string;
|
|
57
|
-
holdoutFraction?: number;
|
|
58
|
-
}): 'search' | 'holdout';
|
|
59
|
-
/**
|
|
60
|
-
* Partition a list of stable ids into disjoint `search` / `holdout` sets by a
|
|
61
|
-
* deterministic hash. Fail-loud:
|
|
62
|
-
* - empty input throws,
|
|
63
|
-
* - duplicate ids throw (a dup would be observed on both sides, double-counting),
|
|
64
|
-
* - a holdout (or search) set below its floor throws (underpowered comparison).
|
|
65
|
-
*
|
|
66
|
-
* Order within each bucket follows the input order, so the result is stable and
|
|
67
|
-
* reproducible.
|
|
68
|
-
*/
|
|
69
|
-
export declare function partitionHeldOut(ids: string[], options?: PartitionHeldOutOptions): HeldOutPartition;
|
|
70
|
-
//# sourceMappingURL=partition-held-out.d.ts.map
|