@tangle-network/agent-eval 0.94.0 → 0.95.0
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 +15 -0
- package/README.md +44 -30
- package/dist/action-policy.d.ts +24 -0
- package/dist/action-policy.d.ts.map +1 -0
- package/dist/action-policy.test.d.ts +2 -0
- package/dist/action-policy.test.d.ts.map +1 -0
- package/dist/active-learning.d.ts +41 -0
- package/dist/active-learning.d.ts.map +1 -0
- package/dist/adapters/http.d.ts +15 -21
- package/dist/adapters/http.d.ts.map +1 -0
- package/dist/adapters/http.js.map +1 -1
- package/dist/adapters/langchain.d.ts +7 -13
- package/dist/adapters/langchain.d.ts.map +1 -0
- package/dist/adapters/otel.d.ts +13 -24
- package/dist/adapters/otel.d.ts.map +1 -0
- package/dist/agent-profile-cell.d.ts +101 -0
- package/dist/agent-profile-cell.d.ts.map +1 -0
- package/dist/agent-profile.d.ts +27 -0
- package/dist/agent-profile.d.ts.map +1 -0
- package/dist/agent-profile.test.d.ts +2 -0
- package/dist/agent-profile.test.d.ts.map +1 -0
- package/dist/analyst/adapters.d.ts +62 -0
- package/dist/analyst/adapters.d.ts.map +1 -0
- package/dist/analyst/analyst.test.d.ts +2 -0
- package/dist/analyst/analyst.test.d.ts.map +1 -0
- package/dist/analyst/ax-service.d.ts +27 -0
- package/dist/analyst/ax-service.d.ts.map +1 -0
- package/dist/analyst/behavioral-analyst.d.ts +28 -0
- package/dist/analyst/behavioral-analyst.d.ts.map +1 -0
- package/dist/analyst/chat-client.d.ts +91 -0
- package/dist/analyst/chat-client.d.ts.map +1 -0
- package/dist/analyst/default-registry.d.ts +27 -0
- package/dist/analyst/default-registry.d.ts.map +1 -0
- package/dist/analyst/default-registry.test.d.ts +2 -0
- package/dist/analyst/default-registry.test.d.ts.map +1 -0
- package/dist/analyst/finding-signature.d.ts +48 -0
- package/dist/analyst/finding-signature.d.ts.map +1 -0
- package/dist/analyst/finding-subject.d.ts +146 -0
- package/dist/analyst/finding-subject.d.ts.map +1 -0
- package/dist/analyst/finding-subject.test.d.ts +2 -0
- package/dist/analyst/finding-subject.test.d.ts.map +1 -0
- package/dist/analyst/findings-store.d.ts +75 -0
- package/dist/analyst/findings-store.d.ts.map +1 -0
- package/dist/analyst/index.d.ts +28 -256
- package/dist/analyst/index.d.ts.map +1 -0
- package/dist/analyst/index.js +5 -65
- package/dist/analyst/index.js.map +1 -1
- package/dist/{kind-factory-0BhLSI27.d.ts → analyst/kind-factory.d.ts} +11 -63
- package/dist/analyst/kind-factory.d.ts.map +1 -0
- package/dist/analyst/kinds/failure-mode.d.ts +19 -0
- package/dist/analyst/kinds/failure-mode.d.ts.map +1 -0
- package/dist/analyst/kinds/improvement.d.ts +23 -0
- package/dist/analyst/kinds/improvement.d.ts.map +1 -0
- package/dist/analyst/kinds/index.d.ts +22 -0
- package/dist/analyst/kinds/index.d.ts.map +1 -0
- package/dist/analyst/kinds/kinds.test.d.ts +2 -0
- package/dist/analyst/kinds/kinds.test.d.ts.map +1 -0
- package/dist/analyst/kinds/knowledge-gap.d.ts +28 -0
- package/dist/analyst/kinds/knowledge-gap.d.ts.map +1 -0
- package/dist/analyst/kinds/knowledge-poisoning.d.ts +22 -0
- package/dist/analyst/kinds/knowledge-poisoning.d.ts.map +1 -0
- package/dist/analyst/kinds/skill-usage.d.ts +84 -0
- package/dist/analyst/kinds/skill-usage.d.ts.map +1 -0
- package/dist/analyst/kinds/skill-usage.test.d.ts +2 -0
- package/dist/analyst/kinds/skill-usage.test.d.ts.map +1 -0
- package/dist/analyst/parse-tolerant.d.ts +26 -0
- package/dist/analyst/parse-tolerant.d.ts.map +1 -0
- package/dist/analyst/parse-tolerant.test.d.ts +2 -0
- package/dist/analyst/parse-tolerant.test.d.ts.map +1 -0
- package/dist/analyst/registry.budget.test.d.ts +2 -0
- package/dist/analyst/registry.budget.test.d.ts.map +1 -0
- package/dist/{default-registry-6dhErQbs.d.ts → analyst/registry.d.ts} +8 -37
- package/dist/analyst/registry.d.ts.map +1 -0
- package/dist/analyst/steer-firewall.d.ts +35 -0
- package/dist/analyst/steer-firewall.d.ts.map +1 -0
- package/dist/analyst/steer-firewall.test.d.ts +2 -0
- package/dist/analyst/steer-firewall.test.d.ts.map +1 -0
- package/dist/analyst/structure-findings.d.ts +37 -0
- package/dist/analyst/structure-findings.d.ts.map +1 -0
- package/dist/analyst/structure-findings.test.d.ts +2 -0
- package/dist/analyst/structure-findings.test.d.ts.map +1 -0
- package/dist/analyst/tool-groups.d.ts +34 -0
- package/dist/analyst/tool-groups.d.ts.map +1 -0
- package/dist/{types-Ce17tDlG.d.ts → analyst/types.d.ts} +19 -112
- package/dist/analyst/types.d.ts.map +1 -0
- package/dist/anti-slop.d.ts +59 -0
- package/dist/anti-slop.d.ts.map +1 -0
- package/dist/artifact-validator.d.ts +74 -0
- package/dist/artifact-validator.d.ts.map +1 -0
- package/dist/attestation.d.ts +63 -0
- package/dist/attestation.d.ts.map +1 -0
- package/dist/attestation.test.d.ts +2 -0
- package/dist/attestation.test.d.ts.map +1 -0
- package/dist/authenticity/index.d.ts +15 -16
- package/dist/authenticity/index.d.ts.map +1 -0
- package/dist/authenticity/index.test.d.ts +2 -0
- package/dist/authenticity/index.test.d.ts.map +1 -0
- package/dist/auto-pr.d.ts +120 -0
- package/dist/auto-pr.d.ts.map +1 -0
- package/dist/{baseline-Bbid3WoO.d.ts → baseline.d.ts} +8 -44
- package/dist/baseline.d.ts.map +1 -0
- package/dist/behavior-dsl.d.ts +73 -0
- package/dist/behavior-dsl.d.ts.map +1 -0
- package/dist/belief-state/calibration.d.ts +10 -0
- package/dist/belief-state/calibration.d.ts.map +1 -0
- package/dist/belief-state/calibration.test.d.ts +2 -0
- package/dist/belief-state/calibration.test.d.ts.map +1 -0
- package/dist/belief-state/code-agent-corpus.d.ts +66 -0
- package/dist/belief-state/code-agent-corpus.d.ts.map +1 -0
- package/dist/belief-state/code-agent-corpus.test.d.ts +2 -0
- package/dist/belief-state/code-agent-corpus.test.d.ts.map +1 -0
- package/dist/belief-state/code-agent-evidence.d.ts +22 -0
- package/dist/belief-state/code-agent-evidence.d.ts.map +1 -0
- package/dist/belief-state/code-agent-evidence.test.d.ts +2 -0
- package/dist/belief-state/code-agent-evidence.test.d.ts.map +1 -0
- package/dist/belief-state/extract.d.ts +7 -0
- package/dist/belief-state/extract.d.ts.map +1 -0
- package/dist/belief-state/extract.test.d.ts +2 -0
- package/dist/belief-state/extract.test.d.ts.map +1 -0
- package/dist/belief-state/index.d.ts +14 -604
- package/dist/belief-state/index.d.ts.map +1 -0
- package/dist/belief-state/ope.d.ts +17 -0
- package/dist/belief-state/ope.d.ts.map +1 -0
- package/dist/belief-state/ope.test.d.ts +2 -0
- package/dist/belief-state/ope.test.d.ts.map +1 -0
- package/dist/belief-state/phase0-measurement.d.ts +55 -0
- package/dist/belief-state/phase0-measurement.d.ts.map +1 -0
- package/dist/belief-state/report.d.ts +17 -0
- package/dist/belief-state/report.d.ts.map +1 -0
- package/dist/belief-state/report.test.d.ts +2 -0
- package/dist/belief-state/report.test.d.ts.map +1 -0
- package/dist/belief-state/research-evidence.d.ts +23 -0
- package/dist/belief-state/research-evidence.d.ts.map +1 -0
- package/dist/belief-state/research-evidence.test.d.ts +2 -0
- package/dist/belief-state/research-evidence.test.d.ts.map +1 -0
- package/dist/belief-state/runtime-benchmark-corpus.d.ts +32 -0
- package/dist/belief-state/runtime-benchmark-corpus.d.ts.map +1 -0
- package/dist/belief-state/runtime-hooks.d.ts +87 -0
- package/dist/belief-state/runtime-hooks.d.ts.map +1 -0
- package/dist/belief-state/runtime-hooks.test.d.ts +2 -0
- package/dist/belief-state/runtime-hooks.test.d.ts.map +1 -0
- package/dist/belief-state/selective.d.ts +15 -0
- package/dist/belief-state/selective.d.ts.map +1 -0
- package/dist/belief-state/selective.test.d.ts +2 -0
- package/dist/belief-state/selective.test.d.ts.map +1 -0
- package/dist/belief-state/shadow-probe.d.ts +80 -0
- package/dist/belief-state/shadow-probe.d.ts.map +1 -0
- package/dist/belief-state/shadow-probe.test.d.ts +2 -0
- package/dist/belief-state/shadow-probe.test.d.ts.map +1 -0
- package/dist/belief-state/types.d.ts +195 -0
- package/dist/belief-state/types.d.ts.map +1 -0
- package/dist/belief-state/types.test.d.ts +2 -0
- package/dist/belief-state/types.test.d.ts.map +1 -0
- package/dist/benchmark.d.ts +14 -0
- package/dist/benchmark.d.ts.map +1 -0
- package/dist/benchmarks/index.d.ts +23 -4
- package/dist/benchmarks/index.d.ts.map +1 -0
- package/dist/benchmarks/routing/dataset.d.ts +34 -0
- package/dist/benchmarks/routing/dataset.d.ts.map +1 -0
- package/dist/benchmarks/routing/index.d.ts +34 -0
- package/dist/benchmarks/routing/index.d.ts.map +1 -0
- package/dist/benchmarks/types.d.ts +49 -0
- package/dist/benchmarks/types.d.ts.map +1 -0
- package/dist/bisector.d.ts +81 -0
- package/dist/bisector.d.ts.map +1 -0
- package/dist/budget-guard.d.ts +31 -0
- package/dist/budget-guard.d.ts.map +1 -0
- package/dist/builder-eval/builder-session.d.ts +111 -0
- package/dist/builder-eval/builder-session.d.ts.map +1 -0
- package/dist/builder-eval/correlation.d.ts +32 -0
- package/dist/builder-eval/correlation.d.ts.map +1 -0
- package/dist/builder-eval/index.d.ts +5 -250
- package/dist/builder-eval/index.d.ts.map +1 -0
- package/dist/builder-eval/project-registry.d.ts +51 -0
- package/dist/builder-eval/project-registry.d.ts.map +1 -0
- package/dist/builder-eval/three-layer-eval.d.ts +55 -0
- package/dist/builder-eval/three-layer-eval.d.ts.map +1 -0
- package/dist/campaign/analyst-surface.d.ts +108 -0
- package/dist/campaign/analyst-surface.d.ts.map +1 -0
- package/dist/campaign/analyst-surface.test.d.ts +2 -0
- package/dist/campaign/analyst-surface.test.d.ts.map +1 -0
- package/dist/campaign/auto-pr.d.ts +46 -0
- package/dist/campaign/auto-pr.d.ts.map +1 -0
- package/dist/campaign/distillation/agreement-judge.d.ts +69 -0
- package/dist/campaign/distillation/agreement-judge.d.ts.map +1 -0
- package/dist/campaign/distillation/cli.d.ts +35 -0
- package/dist/campaign/distillation/cli.d.ts.map +1 -0
- package/dist/campaign/distillation/distillation.test.d.ts +2 -0
- package/dist/campaign/distillation/distillation.test.d.ts.map +1 -0
- package/dist/campaign/distillation/gold-scenarios.d.ts +54 -0
- package/dist/campaign/distillation/gold-scenarios.d.ts.map +1 -0
- package/dist/campaign/distillation/run-distillation.d.ts +119 -0
- package/dist/campaign/distillation/run-distillation.d.ts.map +1 -0
- package/dist/campaign/gates/compose.d.ts +12 -0
- package/dist/campaign/gates/compose.d.ts.map +1 -0
- package/dist/campaign/gates/default-production-gate.d.ts +58 -0
- package/dist/campaign/gates/default-production-gate.d.ts.map +1 -0
- package/dist/campaign/gates/heldout-gate.d.ts +12 -0
- package/dist/campaign/gates/heldout-gate.d.ts.map +1 -0
- package/dist/campaign/gates/promotion-policy.d.ts +125 -0
- package/dist/campaign/gates/promotion-policy.d.ts.map +1 -0
- package/dist/campaign/gates/promotion-policy.test.d.ts +2 -0
- package/dist/campaign/gates/promotion-policy.test.d.ts.map +1 -0
- package/dist/campaign/gates/sequential.d.ts +146 -0
- package/dist/campaign/gates/sequential.d.ts.map +1 -0
- package/dist/campaign/gates/sequential.test.d.ts +2 -0
- package/dist/campaign/gates/sequential.test.d.ts.map +1 -0
- package/dist/campaign/gates/statistical-heldout.d.ts +99 -0
- package/dist/campaign/gates/statistical-heldout.d.ts.map +1 -0
- package/dist/campaign/gates/statistical-heldout.test.d.ts +2 -0
- package/dist/campaign/gates/statistical-heldout.test.d.ts.map +1 -0
- package/dist/campaign/index.d.ts +38 -1341
- package/dist/campaign/index.d.ts.map +1 -0
- package/dist/campaign/index.js +1863 -1316
- package/dist/campaign/index.js.map +1 -1
- package/dist/campaign/labeled-store/fs-adapter.d.ts +59 -0
- package/dist/campaign/labeled-store/fs-adapter.d.ts.map +1 -0
- package/dist/campaign/presets/compare-proposers.d.ts +146 -0
- package/dist/campaign/presets/compare-proposers.d.ts.map +1 -0
- package/dist/campaign/presets/playback.d.ts +120 -0
- package/dist/campaign/presets/playback.d.ts.map +1 -0
- package/dist/campaign/presets/playback.test.d.ts +2 -0
- package/dist/campaign/presets/playback.test.d.ts.map +1 -0
- package/dist/campaign/presets/run-eval.d.ts +14 -0
- package/dist/campaign/presets/run-eval.d.ts.map +1 -0
- package/dist/campaign/presets/run-improvement-loop.d.ts +63 -0
- package/dist/campaign/presets/run-improvement-loop.d.ts.map +1 -0
- package/dist/campaign/presets/run-improvement-loop.test.d.ts +2 -0
- package/dist/campaign/presets/run-improvement-loop.test.d.ts.map +1 -0
- package/dist/campaign/presets/run-optimization.d.ts +92 -0
- package/dist/campaign/presets/run-optimization.d.ts.map +1 -0
- package/dist/campaign/presets/run-profile-matrix.d.ts +151 -0
- package/dist/campaign/presets/run-profile-matrix.d.ts.map +1 -0
- package/dist/campaign/presets/run-skill-opt.d.ts +96 -0
- package/dist/campaign/presets/run-skill-opt.d.ts.map +1 -0
- package/dist/campaign/proposers/_findings-text.d.ts +22 -0
- package/dist/campaign/proposers/_findings-text.d.ts.map +1 -0
- package/dist/campaign/proposers/ace.d.ts +33 -0
- package/dist/campaign/proposers/ace.d.ts.map +1 -0
- package/dist/campaign/proposers/ace.test.d.ts +2 -0
- package/dist/campaign/proposers/ace.test.d.ts.map +1 -0
- package/dist/campaign/proposers/analysis-edit.d.ts +32 -0
- package/dist/campaign/proposers/analysis-edit.d.ts.map +1 -0
- package/dist/campaign/proposers/evolutionary.d.ts +20 -0
- package/dist/campaign/proposers/evolutionary.d.ts.map +1 -0
- package/dist/campaign/proposers/fapo.d.ts +120 -0
- package/dist/campaign/proposers/fapo.d.ts.map +1 -0
- package/dist/campaign/proposers/gepa.d.ts +86 -0
- package/dist/campaign/proposers/gepa.d.ts.map +1 -0
- package/dist/campaign/proposers/halo.d.ts +44 -0
- package/dist/campaign/proposers/halo.d.ts.map +1 -0
- package/dist/campaign/proposers/halo.test.d.ts +2 -0
- package/dist/campaign/proposers/halo.test.d.ts.map +1 -0
- package/dist/campaign/proposers/memory.d.ts +47 -0
- package/dist/campaign/proposers/memory.d.ts.map +1 -0
- package/dist/campaign/proposers/memory.test.d.ts +2 -0
- package/dist/campaign/proposers/memory.test.d.ts.map +1 -0
- package/dist/campaign/proposers/skill-opt.d.ts +88 -0
- package/dist/campaign/proposers/skill-opt.d.ts.map +1 -0
- package/dist/campaign/proposers/trace-analyst.d.ts +48 -0
- package/dist/campaign/proposers/trace-analyst.d.ts.map +1 -0
- package/dist/campaign/proposers/trace-analyst.test.d.ts +2 -0
- package/dist/campaign/proposers/trace-analyst.test.d.ts.map +1 -0
- package/dist/campaign/provenance.d.ts +185 -0
- package/dist/campaign/provenance.d.ts.map +1 -0
- package/dist/campaign/run-campaign.d.ts +90 -0
- package/dist/campaign/run-campaign.d.ts.map +1 -0
- package/dist/campaign/score-utils.d.ts +26 -0
- package/dist/campaign/score-utils.d.ts.map +1 -0
- package/dist/campaign/skill-patch.d.ts +62 -0
- package/dist/campaign/skill-patch.d.ts.map +1 -0
- package/dist/campaign/storage.d.ts +38 -0
- package/dist/campaign/storage.d.ts.map +1 -0
- package/dist/{types-BU-7W85F.d.ts → campaign/types.d.ts} +98 -99
- package/dist/campaign/types.d.ts.map +1 -0
- package/dist/campaign/worktree/index.d.ts +53 -0
- package/dist/campaign/worktree/index.d.ts.map +1 -0
- package/dist/canary.d.ts +101 -0
- package/dist/canary.d.ts.map +1 -0
- package/dist/causal-attribution.d.ts +45 -0
- package/dist/causal-attribution.d.ts.map +1 -0
- package/dist/{chunk-2K6UUZ7P.js → chunk-2T4EZACH.js} +1 -1
- package/dist/chunk-2T4EZACH.js.map +1 -0
- package/dist/{chunk-CTBHKLEU.js → chunk-77T4STFI.js} +59 -86
- package/dist/chunk-77T4STFI.js.map +1 -0
- package/dist/{chunk-EGPMSBEZ.js → chunk-7QTQKIDD.js} +178 -177
- package/dist/chunk-7QTQKIDD.js.map +1 -0
- package/dist/{chunk-MIFZUPEK.js → chunk-AQ5WQAIV.js} +21 -6
- package/dist/chunk-AQ5WQAIV.js.map +1 -0
- package/dist/chunk-DJWX3GVS.js +81 -0
- package/dist/chunk-DJWX3GVS.js.map +1 -0
- package/dist/{chunk-S6OZEZQK.js → chunk-HMA63UEO.js} +37 -9
- package/dist/{chunk-S6OZEZQK.js.map → chunk-HMA63UEO.js.map} +1 -1
- package/dist/{chunk-TBDR6PAI.js → chunk-IZCEK2HR.js} +2 -2
- package/dist/{chunk-SD2YFWQQ.js → chunk-KKWJD5E6.js} +20 -20
- package/dist/chunk-KKWJD5E6.js.map +1 -0
- package/dist/{chunk-KWRRMR3J.js → chunk-LO6IOIJ2.js} +10 -10
- package/dist/chunk-LO6IOIJ2.js.map +1 -0
- package/dist/{chunk-E4GH6USR.js → chunk-NZEQVRH5.js} +2 -2
- package/dist/chunk-NZEQVRH5.js.map +1 -0
- package/dist/{chunk-MPQWFX6Y.js → chunk-PSWWQXHF.js} +13 -88
- package/dist/chunk-PSWWQXHF.js.map +1 -0
- package/dist/{chunk-Q5LIB7BC.js → chunk-S4SYLDFX.js} +2 -2
- package/dist/chunk-S4SYLDFX.js.map +1 -0
- package/dist/{chunk-KW53MSA5.js → chunk-X74V6ESX.js} +2 -2
- package/dist/{chunk-QMUEXQJS.js → chunk-YBIGNSCZ.js} +81 -4
- package/dist/chunk-YBIGNSCZ.js.map +1 -0
- package/dist/{chunk-2KNZHH3P.js → chunk-Z6L6YSU6.js} +2 -2
- package/dist/ci-gate.d.ts +44 -0
- package/dist/ci-gate.d.ts.map +1 -0
- package/dist/cli.d.ts +2 -0
- package/dist/cli.d.ts.map +1 -0
- package/dist/client.d.ts +77 -0
- package/dist/client.d.ts.map +1 -0
- package/dist/client.test.d.ts +2 -0
- package/dist/client.test.d.ts.map +1 -0
- package/dist/command-runner.d.ts +74 -0
- package/dist/command-runner.d.ts.map +1 -0
- package/dist/command-runner.test.d.ts +2 -0
- package/dist/command-runner.test.d.ts.map +1 -0
- package/dist/completion-verifier.d.ts +147 -0
- package/dist/completion-verifier.d.ts.map +1 -0
- package/dist/completion-verifier.test.d.ts +9 -0
- package/dist/completion-verifier.test.d.ts.map +1 -0
- package/dist/concurrency.d.ts +23 -0
- package/dist/concurrency.d.ts.map +1 -0
- package/dist/contamination-guard.d.ts +81 -0
- package/dist/contamination-guard.d.ts.map +1 -0
- package/dist/{analyze-runs-B6Ljo_dI.d.ts → contract/analyze-runs.d.ts} +9 -10
- package/dist/contract/analyze-runs.d.ts.map +1 -0
- package/dist/contract/define-agent-eval.d.ts +52 -0
- package/dist/contract/define-agent-eval.d.ts.map +1 -0
- package/dist/contract/diff.d.ts +114 -0
- package/dist/contract/diff.d.ts.map +1 -0
- package/dist/contract/index.d.ts +106 -640
- package/dist/contract/index.d.ts.map +1 -0
- package/dist/contract/index.js +127 -17
- package/dist/contract/index.js.map +1 -1
- package/dist/{insight-report-DWl3z9tl.d.ts → contract/insight-report.d.ts} +16 -19
- package/dist/contract/insight-report.d.ts.map +1 -0
- package/dist/contract/insight-types-fwd.d.ts +7 -0
- package/dist/contract/insight-types-fwd.d.ts.map +1 -0
- package/dist/contract/intake/agent-trace.d.ts +97 -0
- package/dist/contract/intake/agent-trace.d.ts.map +1 -0
- package/dist/{code-agent-session-BO8nCnv3.d.ts → contract/intake/code-agent-session.d.ts} +15 -17
- package/dist/contract/intake/code-agent-session.d.ts.map +1 -0
- package/dist/contract/intake/feedback-table.d.ts +87 -0
- package/dist/contract/intake/feedback-table.d.ts.map +1 -0
- package/dist/contract/intake/index.d.ts +22 -0
- package/dist/contract/intake/index.d.ts.map +1 -0
- package/dist/contract/intake/otel-spans.d.ts +33 -0
- package/dist/contract/intake/otel-spans.d.ts.map +1 -0
- package/dist/contract/self-improve.d.ts +284 -0
- package/dist/contract/self-improve.d.ts.map +1 -0
- package/dist/{control-runtime-Acf9CGhw.d.ts → control-runtime.d.ts} +23 -26
- package/dist/control-runtime.d.ts.map +1 -0
- package/dist/control-runtime.test.d.ts +2 -0
- package/dist/control-runtime.test.d.ts.map +1 -0
- package/dist/control.d.ts +11 -9
- package/dist/control.d.ts.map +1 -0
- package/dist/control.js +2 -2
- package/dist/convergence.d.ts +29 -0
- package/dist/convergence.d.ts.map +1 -0
- package/dist/{cost-ledger-DuSqlw5B.d.ts → cost-ledger.d.ts} +10 -11
- package/dist/cost-ledger.d.ts.map +1 -0
- package/dist/cost-ledger.test.d.ts +2 -0
- package/dist/cost-ledger.test.d.ts.map +1 -0
- package/dist/cost-report.d.ts +43 -0
- package/dist/cost-report.d.ts.map +1 -0
- package/dist/cost-report.test.d.ts +2 -0
- package/dist/cost-report.test.d.ts.map +1 -0
- package/dist/cost-tracker.d.ts +76 -0
- package/dist/cost-tracker.d.ts.map +1 -0
- package/dist/{counterfactual-DlOz8PBx.d.ts → counterfactual.d.ts} +12 -13
- package/dist/counterfactual.d.ts.map +1 -0
- package/dist/cross-trace-diff.d.ts +56 -0
- package/dist/cross-trace-diff.d.ts.map +1 -0
- package/dist/{dataset-BbGkaN2I.d.ts → dataset.d.ts} +11 -14
- package/dist/dataset.d.ts.map +1 -0
- package/dist/deploy-gate-layer.d.ts +125 -0
- package/dist/deploy-gate-layer.d.ts.map +1 -0
- package/dist/deploy-gate-layer.test.d.ts +2 -0
- package/dist/deploy-gate-layer.test.d.ts.map +1 -0
- package/dist/description-length-gate.d.ts +119 -0
- package/dist/description-length-gate.d.ts.map +1 -0
- package/dist/detectors/edge.test.d.ts +2 -0
- package/dist/detectors/edge.test.d.ts.map +1 -0
- package/dist/detectors/index.d.ts +81 -0
- package/dist/detectors/index.d.ts.map +1 -0
- package/dist/detectors/index.test.d.ts +2 -0
- package/dist/detectors/index.test.d.ts.map +1 -0
- package/dist/diagnose/causal-sweep.d.ts +100 -0
- package/dist/diagnose/causal-sweep.d.ts.map +1 -0
- package/dist/diagnose/index.d.ts +36 -0
- package/dist/diagnose/index.d.ts.map +1 -0
- package/dist/diagnose/remediation.d.ts +68 -0
- package/dist/diagnose/remediation.d.ts.map +1 -0
- package/dist/diagnose/repair.d.ts +77 -0
- package/dist/diagnose/repair.d.ts.map +1 -0
- package/dist/diagnose.d.ts +1 -251
- package/dist/diagnose.js +1 -1
- package/dist/discover-personas.d.ts +35 -0
- package/dist/discover-personas.d.ts.map +1 -0
- package/dist/driver.d.ts +95 -0
- package/dist/driver.d.ts.map +1 -0
- package/dist/driver.test.d.ts +8 -0
- package/dist/driver.test.d.ts.map +1 -0
- package/dist/dual-agent-bench.d.ts +81 -0
- package/dist/dual-agent-bench.d.ts.map +1 -0
- package/dist/error-count-extractor.d.ts +47 -0
- package/dist/error-count-extractor.d.ts.map +1 -0
- package/dist/error-count-extractor.test.d.ts +2 -0
- package/dist/error-count-extractor.test.d.ts.map +1 -0
- package/dist/{errors-CzMUYo7b.d.ts → errors.d.ts} +10 -11
- package/dist/errors.d.ts.map +1 -0
- package/dist/{researcher-B0C2_fVO.d.ts → eval-campaign.d.ts} +34 -156
- package/dist/eval-campaign.d.ts.map +1 -0
- package/dist/eval-campaign.test.d.ts +2 -0
- package/dist/eval-campaign.test.d.ts.map +1 -0
- package/dist/eval-tools.d.ts +55 -0
- package/dist/eval-tools.d.ts.map +1 -0
- package/dist/eval-trace-store.d.ts +107 -0
- package/dist/eval-trace-store.d.ts.map +1 -0
- package/dist/eval-trace-store.test.d.ts +2 -0
- package/dist/eval-trace-store.test.d.ts.map +1 -0
- package/dist/executor.d.ts +38 -0
- package/dist/executor.d.ts.map +1 -0
- package/dist/executor.test.d.ts +10 -0
- package/dist/executor.test.d.ts.map +1 -0
- package/dist/experiment-tracker.d.ts +178 -0
- package/dist/experiment-tracker.d.ts.map +1 -0
- package/dist/experiment-tracker.test.d.ts +2 -0
- package/dist/experiment-tracker.test.d.ts.map +1 -0
- package/dist/failure-taxonomy.d.ts +38 -0
- package/dist/failure-taxonomy.d.ts.map +1 -0
- package/dist/{feedback-trajectory-BxY0cKfs.d.ts → feedback-trajectory.d.ts} +36 -38
- package/dist/feedback-trajectory.d.ts.map +1 -0
- package/dist/feedback-trajectory.test.d.ts +2 -0
- package/dist/feedback-trajectory.test.d.ts.map +1 -0
- package/dist/flow-layer.d.ts +90 -0
- package/dist/flow-layer.d.ts.map +1 -0
- package/dist/flow-layer.test.d.ts +2 -0
- package/dist/flow-layer.test.d.ts.map +1 -0
- package/dist/fuzz/capsule.d.ts +46 -0
- package/dist/fuzz/capsule.d.ts.map +1 -0
- package/dist/fuzz/cube.d.ts +36 -0
- package/dist/fuzz/cube.d.ts.map +1 -0
- package/dist/fuzz/explorer-cost.test.d.ts +2 -0
- package/dist/fuzz/explorer-cost.test.d.ts.map +1 -0
- package/dist/fuzz/explorer.d.ts +64 -0
- package/dist/fuzz/explorer.d.ts.map +1 -0
- package/dist/fuzz/fuzz-agent.d.ts +16 -0
- package/dist/fuzz/fuzz-agent.d.ts.map +1 -0
- package/dist/fuzz/fuzz-agent.test.d.ts +2 -0
- package/dist/fuzz/fuzz-agent.test.d.ts.map +1 -0
- package/dist/fuzz/gates.d.ts +33 -0
- package/dist/fuzz/gates.d.ts.map +1 -0
- package/dist/fuzz/index.d.ts +26 -0
- package/dist/fuzz/index.d.ts.map +1 -0
- package/dist/fuzz/policies.d.ts +28 -0
- package/dist/fuzz/policies.d.ts.map +1 -0
- package/dist/fuzz/tools.d.ts +20 -0
- package/dist/fuzz/tools.d.ts.map +1 -0
- package/dist/fuzz/types.d.ts +307 -0
- package/dist/fuzz/types.d.ts.map +1 -0
- package/dist/fuzz.d.ts +1 -547
- package/dist/golden-matcher.d.ts +71 -0
- package/dist/golden-matcher.d.ts.map +1 -0
- package/dist/governance/eu-ai-act.d.ts +37 -0
- package/dist/governance/eu-ai-act.d.ts.map +1 -0
- package/dist/governance/index.d.ts +5 -135
- package/dist/governance/index.d.ts.map +1 -0
- package/dist/governance/nist-ai-rmf.d.ts +15 -0
- package/dist/governance/nist-ai-rmf.d.ts.map +1 -0
- package/dist/governance/soc2.d.ts +12 -0
- package/dist/governance/soc2.d.ts.map +1 -0
- package/dist/governance/types.d.ts +66 -0
- package/dist/governance/types.d.ts.map +1 -0
- package/dist/harness-optimizer.d.ts +82 -0
- package/dist/harness-optimizer.d.ts.map +1 -0
- package/dist/held-out-gate.d.ts +135 -0
- package/dist/held-out-gate.d.ts.map +1 -0
- package/dist/hosted/client.d.ts +73 -0
- package/dist/hosted/client.d.ts.map +1 -0
- package/dist/hosted/from-env.test.d.ts +8 -0
- package/dist/hosted/from-env.test.d.ts.map +1 -0
- package/dist/hosted/index.d.ts +10 -238
- package/dist/hosted/index.d.ts.map +1 -0
- package/dist/hosted/types.d.ts +159 -0
- package/dist/hosted/types.d.ts.map +1 -0
- package/dist/index.d.ts +277 -5665
- package/dist/index.d.ts.map +1 -0
- package/dist/index.js +66 -31
- package/dist/index.js.map +1 -1
- package/dist/integrity/backend-integrity.d.ts +71 -0
- package/dist/integrity/backend-integrity.d.ts.map +1 -0
- package/dist/integrity/preflight.d.ts +72 -0
- package/dist/integrity/preflight.d.ts.map +1 -0
- package/dist/integrity/preflight.test.d.ts +2 -0
- package/dist/integrity/preflight.test.d.ts.map +1 -0
- package/dist/integrity/single-backend.d.ts +67 -0
- package/dist/integrity/single-backend.d.ts.map +1 -0
- package/dist/intent-match-judge.d.ts +69 -0
- package/dist/intent-match-judge.d.ts.map +1 -0
- package/dist/intent-match-judge.test.d.ts +2 -0
- package/dist/intent-match-judge.test.d.ts.map +1 -0
- package/dist/{judge-calibration-0p2QcWNE.d.ts → judge-calibration.d.ts} +16 -17
- package/dist/judge-calibration.d.ts.map +1 -0
- package/dist/judge-ensemble.d.ts +66 -0
- package/dist/judge-ensemble.d.ts.map +1 -0
- package/dist/judge-ensemble.test.d.ts +8 -0
- package/dist/judge-ensemble.test.d.ts.map +1 -0
- package/dist/judge-families.d.ts +38 -0
- package/dist/judge-families.d.ts.map +1 -0
- package/dist/judge-panel.d.ts +65 -0
- package/dist/judge-panel.d.ts.map +1 -0
- package/dist/judge-retry.d.ts +70 -0
- package/dist/judge-retry.d.ts.map +1 -0
- package/dist/judge-runner.d.ts +36 -0
- package/dist/judge-runner.d.ts.map +1 -0
- package/dist/judge-runner.test.d.ts +2 -0
- package/dist/judge-runner.test.d.ts.map +1 -0
- package/dist/judges.d.ts +74 -0
- package/dist/judges.d.ts.map +1 -0
- package/dist/keyword-coverage-judge.d.ts +89 -0
- package/dist/keyword-coverage-judge.d.ts.map +1 -0
- package/dist/keyword-coverage-judge.test.d.ts +2 -0
- package/dist/keyword-coverage-judge.test.d.ts.map +1 -0
- package/dist/knowledge/index.d.ts +3 -103
- package/dist/knowledge/index.d.ts.map +1 -0
- package/dist/knowledge/readiness.d.ts +26 -0
- package/dist/knowledge/readiness.d.ts.map +1 -0
- package/dist/knowledge/types.d.ts +75 -0
- package/dist/knowledge/types.d.ts.map +1 -0
- package/dist/live-proof.d.ts +62 -0
- package/dist/live-proof.d.ts.map +1 -0
- package/dist/{llm-client-Bj7g0rqu.d.ts → llm-client.d.ts} +21 -23
- package/dist/llm-client.d.ts.map +1 -0
- package/dist/llm-client.test.d.ts +2 -0
- package/dist/llm-client.test.d.ts.map +1 -0
- package/dist/locked-jsonl-appender.d.ts +19 -0
- package/dist/locked-jsonl-appender.d.ts.map +1 -0
- package/dist/matrix/aggregation.d.ts +16 -0
- package/dist/matrix/aggregation.d.ts.map +1 -0
- package/dist/matrix/index.d.ts +12 -30
- package/dist/matrix/index.d.ts.map +1 -0
- package/dist/matrix/runner.d.ts +15 -0
- package/dist/matrix/runner.d.ts.map +1 -0
- package/dist/{types-mn5Aqk7x.d.ts → matrix/types.d.ts} +11 -15
- package/dist/matrix/types.d.ts.map +1 -0
- package/dist/meta-eval/calibration.d.ts +47 -0
- package/dist/meta-eval/calibration.d.ts.map +1 -0
- package/dist/meta-eval/correlation-study.d.ts +53 -0
- package/dist/meta-eval/correlation-study.d.ts.map +1 -0
- package/dist/meta-eval/index.d.ts +6 -181
- package/dist/meta-eval/index.d.ts.map +1 -0
- package/dist/{outcome-store-rnXLEqSn.d.ts → meta-eval/outcome-store.d.ts} +7 -8
- package/dist/meta-eval/outcome-store.d.ts.map +1 -0
- package/dist/{rubric-predictive-validity-Cy_W-hWZ.d.ts → meta-eval/rubric-predictive-validity.d.ts} +8 -11
- package/dist/meta-eval/rubric-predictive-validity.d.ts.map +1 -0
- package/dist/meta-eval/sentinel.d.ts +169 -0
- package/dist/meta-eval/sentinel.d.ts.map +1 -0
- package/dist/metrics.d.ts +63 -0
- package/dist/metrics.d.ts.map +1 -0
- package/dist/model-seats.d.ts +71 -0
- package/dist/model-seats.d.ts.map +1 -0
- package/dist/model-seats.test.d.ts +2 -0
- package/dist/model-seats.test.d.ts.map +1 -0
- package/dist/muffled-gate-scanner.d.ts +102 -0
- package/dist/muffled-gate-scanner.d.ts.map +1 -0
- package/dist/{multi-layer-verifier-DUZXrPDA.d.ts → multi-layer-verifier.d.ts} +12 -15
- package/dist/multi-layer-verifier.d.ts.map +1 -0
- package/dist/multi-layer-verifier.test.d.ts +2 -0
- package/dist/multi-layer-verifier.test.d.ts.map +1 -0
- package/dist/multi-toolchain-layer.d.ts +80 -0
- package/dist/multi-toolchain-layer.d.ts.map +1 -0
- package/dist/multi-toolchain-layer.test.d.ts +2 -0
- package/dist/multi-toolchain-layer.test.d.ts.map +1 -0
- package/dist/multishot/default-tools.d.ts +34 -0
- package/dist/multishot/default-tools.d.ts.map +1 -0
- package/dist/multishot/index.d.ts +7 -290
- package/dist/multishot/index.d.ts.map +1 -0
- package/dist/multishot/index.js.map +1 -1
- package/dist/multishot/judges.d.ts +32 -0
- package/dist/multishot/judges.d.ts.map +1 -0
- package/dist/multishot/matrix.d.ts +107 -0
- package/dist/multishot/matrix.d.ts.map +1 -0
- package/dist/multishot/multishot.d.ts +23 -0
- package/dist/multishot/multishot.d.ts.map +1 -0
- package/dist/multishot/router.d.ts +37 -0
- package/dist/multishot/router.d.ts.map +1 -0
- package/dist/multishot/types.d.ts +60 -0
- package/dist/multishot/types.d.ts.map +1 -0
- package/dist/observability.d.ts +71 -0
- package/dist/observability.d.ts.map +1 -0
- package/dist/openapi.json +1 -1
- package/dist/oracle.d.ts +55 -0
- package/dist/oracle.d.ts.map +1 -0
- package/dist/orthogonality.d.ts +35 -0
- package/dist/orthogonality.d.ts.map +1 -0
- package/dist/otel-pipeline.d.ts +31 -0
- package/dist/otel-pipeline.d.ts.map +1 -0
- package/dist/paraphrase.d.ts +107 -0
- package/dist/paraphrase.d.ts.map +1 -0
- package/dist/{pareto-E-pembql.d.ts → pareto.d.ts} +9 -10
- package/dist/pareto.d.ts.map +1 -0
- package/dist/partition-held-out.d.ts +70 -0
- package/dist/partition-held-out.d.ts.map +1 -0
- package/dist/partition-held-out.test.d.ts +2 -0
- package/dist/partition-held-out.test.d.ts.map +1 -0
- package/dist/perf/index.d.ts +13 -119
- package/dist/perf/index.d.ts.map +1 -0
- package/dist/perf/integrity.d.ts +30 -0
- package/dist/perf/integrity.d.ts.map +1 -0
- package/dist/perf/journey.d.ts +45 -0
- package/dist/perf/journey.d.ts.map +1 -0
- package/dist/perf/ratchet.d.ts +47 -0
- package/dist/perf/ratchet.d.ts.map +1 -0
- package/dist/pipelines/budget-breach.d.ts +31 -0
- package/dist/pipelines/budget-breach.d.ts.map +1 -0
- package/dist/pipelines/budget-breach.test.d.ts +2 -0
- package/dist/pipelines/budget-breach.test.d.ts.map +1 -0
- package/dist/pipelines/failure-cluster.d.ts +38 -0
- package/dist/pipelines/failure-cluster.d.ts.map +1 -0
- package/dist/pipelines/failure-cluster.test.d.ts +2 -0
- package/dist/pipelines/failure-cluster.test.d.ts.map +1 -0
- package/dist/pipelines/first-divergence.d.ts +26 -0
- package/dist/pipelines/first-divergence.d.ts.map +1 -0
- package/dist/pipelines/first-divergence.test.d.ts +2 -0
- package/dist/pipelines/first-divergence.test.d.ts.map +1 -0
- package/dist/pipelines/index.d.ts +8 -173
- package/dist/pipelines/index.d.ts.map +1 -0
- package/dist/pipelines/judge-agreement.d.ts +26 -0
- package/dist/pipelines/judge-agreement.d.ts.map +1 -0
- package/dist/pipelines/judge-agreement.test.d.ts +2 -0
- package/dist/pipelines/judge-agreement.test.d.ts.map +1 -0
- package/dist/pipelines/regression.d.ts +23 -0
- package/dist/pipelines/regression.d.ts.map +1 -0
- package/dist/pipelines/regression.test.d.ts +2 -0
- package/dist/pipelines/regression.test.d.ts.map +1 -0
- package/dist/pipelines/stuck-loop.d.ts +32 -0
- package/dist/pipelines/stuck-loop.d.ts.map +1 -0
- package/dist/pipelines/stuck-loop.test.d.ts +2 -0
- package/dist/pipelines/stuck-loop.test.d.ts.map +1 -0
- package/dist/pipelines/tool-waste.d.ts +34 -0
- package/dist/pipelines/tool-waste.d.ts.map +1 -0
- package/dist/pipelines/tool-waste.test.d.ts +2 -0
- package/dist/pipelines/tool-waste.test.d.ts.map +1 -0
- package/dist/playbook.d.ts +16 -0
- package/dist/playbook.d.ts.map +1 -0
- package/dist/pr-review-benchmark.d.ts +88 -0
- package/dist/pr-review-benchmark.d.ts.map +1 -0
- package/dist/pr-review-benchmark.test.d.ts +2 -0
- package/dist/pr-review-benchmark.test.d.ts.map +1 -0
- package/dist/pre-registration.d.ts +125 -0
- package/dist/pre-registration.d.ts.map +1 -0
- package/dist/prm/builtin-rubrics.d.ts +33 -0
- package/dist/prm/builtin-rubrics.d.ts.map +1 -0
- package/dist/prm/index.d.ts +5 -104
- package/dist/prm/index.d.ts.map +1 -0
- package/dist/prm/inference.d.ts +29 -0
- package/dist/prm/inference.d.ts.map +1 -0
- package/dist/prm/inference.test.d.ts +2 -0
- package/dist/prm/inference.test.d.ts.map +1 -0
- package/dist/{rubric-Cc6UHvUb.d.ts → prm/rubric.d.ts} +10 -13
- package/dist/prm/rubric.d.ts.map +1 -0
- package/dist/prm/training-export.d.ts +38 -0
- package/dist/prm/training-export.d.ts.map +1 -0
- package/dist/produced-state.d.ts +63 -0
- package/dist/produced-state.d.ts.map +1 -0
- package/dist/produced-state.test.d.ts +8 -0
- package/dist/produced-state.test.d.ts.map +1 -0
- package/dist/profile/baselines.d.ts +37 -0
- package/dist/profile/baselines.d.ts.map +1 -0
- package/dist/profile/index.d.ts +105 -0
- package/dist/profile/index.d.ts.map +1 -0
- package/dist/promotion-gate.d.ts +93 -0
- package/dist/promotion-gate.d.ts.map +1 -0
- package/dist/prompt-registry.d.ts +41 -0
- package/dist/prompt-registry.d.ts.map +1 -0
- package/dist/propose-review-control.d.ts +49 -0
- package/dist/propose-review-control.d.ts.map +1 -0
- package/dist/propose-review-control.test.d.ts +2 -0
- package/dist/propose-review-control.test.d.ts.map +1 -0
- package/dist/propose-review.d.ts +155 -0
- package/dist/propose-review.d.ts.map +1 -0
- package/dist/{red-team-BWdoyleI.d.ts → red-team.d.ts} +14 -16
- package/dist/red-team.d.ts.map +1 -0
- package/dist/reference-replay-steering.d.ts +11 -0
- package/dist/reference-replay-steering.d.ts.map +1 -0
- package/dist/reference-replay.d.ts +176 -0
- package/dist/reference-replay.d.ts.map +1 -0
- package/dist/reflective-mutation.d.ts +79 -0
- package/dist/reflective-mutation.d.ts.map +1 -0
- package/dist/registry.d.ts +31 -0
- package/dist/registry.d.ts.map +1 -0
- package/dist/release-confidence.d.ts +128 -0
- package/dist/release-confidence.d.ts.map +1 -0
- package/dist/release-report.d.ts +11 -0
- package/dist/release-report.d.ts.map +1 -0
- package/dist/replay.d.ts +120 -0
- package/dist/replay.d.ts.map +1 -0
- package/dist/reporter.d.ts +14 -0
- package/dist/reporter.d.ts.map +1 -0
- package/dist/reporting.d.ts +15 -15
- package/dist/reporting.d.ts.map +1 -0
- package/dist/researcher.d.ts +140 -0
- package/dist/researcher.d.ts.map +1 -0
- package/dist/reviewer.d.ts +118 -0
- package/dist/reviewer.d.ts.map +1 -0
- package/dist/reviewer.test.d.ts +2 -0
- package/dist/reviewer.test.d.ts.map +1 -0
- package/dist/reward-model-export.d.ts +60 -0
- package/dist/reward-model-export.d.ts.map +1 -0
- package/dist/rl/active-curriculum.d.ts +110 -0
- package/dist/rl/active-curriculum.d.ts.map +1 -0
- package/dist/rl/adaptation-eval.d.ts +109 -0
- package/dist/rl/adaptation-eval.d.ts.map +1 -0
- package/dist/{adversarial-DIVcDoI_.d.ts → rl/adversarial.d.ts} +6 -7
- package/dist/rl/adversarial.d.ts.map +1 -0
- package/dist/rl/compute-curves.d.ts +127 -0
- package/dist/rl/compute-curves.d.ts.map +1 -0
- package/dist/rl/contamination.d.ts +117 -0
- package/dist/rl/contamination.d.ts.map +1 -0
- package/dist/rl/corpus.d.ts +55 -0
- package/dist/rl/corpus.d.ts.map +1 -0
- package/dist/rl/corpus.test.d.ts +2 -0
- package/dist/rl/corpus.test.d.ts.map +1 -0
- package/dist/rl/dataset.d.ts +102 -0
- package/dist/rl/dataset.d.ts.map +1 -0
- package/dist/rl/dataset.test.d.ts +2 -0
- package/dist/rl/dataset.test.d.ts.map +1 -0
- package/dist/rl/exporters.d.ts +141 -0
- package/dist/rl/exporters.d.ts.map +1 -0
- package/dist/rl/index.d.ts +49 -0
- package/dist/rl/index.d.ts.map +1 -0
- package/dist/{off-policy-DiwuKKg7.d.ts → rl/off-policy.d.ts} +8 -9
- package/dist/rl/off-policy.d.ts.map +1 -0
- package/dist/rl/predictive-validity-researcher.d.ts +69 -0
- package/dist/rl/predictive-validity-researcher.d.ts.map +1 -0
- package/dist/rl/preferences.d.ts +141 -0
- package/dist/rl/preferences.d.ts.map +1 -0
- package/dist/rl/process-reward.d.ts +122 -0
- package/dist/rl/process-reward.d.ts.map +1 -0
- package/dist/rl/reward-hacking.d.ts +104 -0
- package/dist/rl/reward-hacking.d.ts.map +1 -0
- package/dist/rl/rl-campaign.d.ts +85 -0
- package/dist/rl/rl-campaign.d.ts.map +1 -0
- package/dist/rl/run-record-adapters.d.ts +56 -0
- package/dist/rl/run-record-adapters.d.ts.map +1 -0
- package/dist/rl/sim-fidelity.d.ts +166 -0
- package/dist/rl/sim-fidelity.d.ts.map +1 -0
- package/dist/rl/sim-fidelity.test.d.ts +2 -0
- package/dist/rl/sim-fidelity.test.d.ts.map +1 -0
- package/dist/rl/tournament.d.ts +115 -0
- package/dist/rl/tournament.d.ts.map +1 -0
- package/dist/rl/verifiable-reward.d.ts +124 -0
- package/dist/rl/verifiable-reward.d.ts.map +1 -0
- package/dist/rl.d.ts +1 -1192
- package/dist/rl.js +612 -612
- package/dist/rl.js.map +1 -1
- package/dist/{run-campaign-7WNXMDSN.js → run-campaign-WXY7KI67.js} +2 -2
- package/dist/run-critic.d.ts +23 -0
- package/dist/run-critic.d.ts.map +1 -0
- package/dist/run-evidence.d.ts +32 -0
- package/dist/run-evidence.d.ts.map +1 -0
- package/dist/{run-record-e7vj1uZQ.d.ts → run-record.d.ts} +16 -122
- package/dist/run-record.d.ts.map +1 -0
- package/dist/run-record.test.d.ts +2 -0
- package/dist/run-record.test.d.ts.map +1 -0
- package/dist/run-score.d.ts +31 -0
- package/dist/run-score.d.ts.map +1 -0
- package/dist/runtime-trajectory.d.ts +47 -0
- package/dist/runtime-trajectory.d.ts.map +1 -0
- package/dist/{test-graded-scenario-DeODGLra.d.ts → sandbox-harness.d.ts} +15 -60
- package/dist/sandbox-harness.d.ts.map +1 -0
- package/dist/sandbox-harness.test.d.ts +2 -0
- package/dist/sandbox-harness.test.d.ts.map +1 -0
- package/dist/sandbox-pool.d.ts +74 -0
- package/dist/sandbox-pool.d.ts.map +1 -0
- package/dist/sandbox-pool.test.d.ts +2 -0
- package/dist/sandbox-pool.test.d.ts.map +1 -0
- package/dist/scorecard.d.ts +133 -0
- package/dist/scorecard.d.ts.map +1 -0
- package/dist/scorecard.test.d.ts +2 -0
- package/dist/scorecard.test.d.ts.map +1 -0
- package/dist/self-play.d.ts +69 -0
- package/dist/self-play.d.ts.map +1 -0
- package/dist/semantic-concept-judge.d.ts +135 -0
- package/dist/semantic-concept-judge.d.ts.map +1 -0
- package/dist/semantic-concept-judge.test.d.ts +2 -0
- package/dist/semantic-concept-judge.test.d.ts.map +1 -0
- package/dist/{sequential-5iSVfzl2.d.ts → sequential.d.ts} +9 -10
- package/dist/sequential.d.ts.map +1 -0
- package/dist/{series-convergence-D5OWMBg6.d.ts → series-convergence.d.ts} +4 -5
- package/dist/series-convergence.d.ts.map +1 -0
- package/dist/slo.d.ts +48 -0
- package/dist/slo.d.ts.map +1 -0
- package/dist/state-continuity.d.ts +47 -0
- package/dist/state-continuity.d.ts.map +1 -0
- package/dist/{statistics-CCJpTGOS.d.ts → statistics.d.ts} +46 -48
- package/dist/statistics.d.ts.map +1 -0
- package/dist/statistics.test.d.ts +2 -0
- package/dist/statistics.test.d.ts.map +1 -0
- package/dist/steering-optimizer.d.ts +58 -0
- package/dist/steering-optimizer.d.ts.map +1 -0
- package/dist/steering.d.ts +24 -0
- package/dist/steering.d.ts.map +1 -0
- package/dist/storyboard/code-edit.d.ts +64 -0
- package/dist/storyboard/code-edit.d.ts.map +1 -0
- package/dist/storyboard/code-edit.test.d.ts +2 -0
- package/dist/storyboard/code-edit.test.d.ts.map +1 -0
- package/dist/storyboard/index.d.ts +16 -81
- package/dist/storyboard/index.d.ts.map +1 -0
- package/dist/storyboard/index.test.d.ts +2 -0
- package/dist/storyboard/index.test.d.ts.map +1 -0
- package/dist/{summary-report-BDOFevaT.d.ts → summary-report.d.ts} +23 -160
- package/dist/summary-report.d.ts.map +1 -0
- package/dist/telemetry/client.d.ts +35 -0
- package/dist/telemetry/client.d.ts.map +1 -0
- package/dist/telemetry/index.d.ts +17 -35
- package/dist/telemetry/index.d.ts.map +1 -0
- package/dist/telemetry/schema.d.ts +61 -0
- package/dist/telemetry/schema.d.ts.map +1 -0
- package/dist/telemetry/sink-fetch.d.ts +39 -0
- package/dist/telemetry/sink-fetch.d.ts.map +1 -0
- package/dist/telemetry/{file.d.ts → sink-file.d.ts} +5 -7
- package/dist/telemetry/sink-file.d.ts.map +1 -0
- package/dist/test-graded-scenario.d.ts +42 -0
- package/dist/test-graded-scenario.d.ts.map +1 -0
- package/dist/testing.d.ts +5 -0
- package/dist/testing.d.ts.map +1 -0
- package/dist/testing.js +8 -0
- package/dist/testing.js.map +1 -0
- package/dist/tool-use-metrics.d.ts +35 -0
- package/dist/tool-use-metrics.d.ts.map +1 -0
- package/dist/trace/capture-fetch.d.ts +48 -0
- package/dist/trace/capture-fetch.d.ts.map +1 -0
- package/dist/trace/capture-fetch.test.d.ts +2 -0
- package/dist/trace/capture-fetch.test.d.ts.map +1 -0
- package/dist/{emitter-C2rqGH_l.d.ts → trace/emitter.d.ts} +9 -12
- package/dist/trace/emitter.d.ts.map +1 -0
- package/dist/trace/extract-usage.d.ts +46 -0
- package/dist/trace/extract-usage.d.ts.map +1 -0
- package/dist/trace/extract-usage.test.d.ts +2 -0
- package/dist/trace/extract-usage.test.d.ts.map +1 -0
- package/dist/trace/index.d.ts +15 -0
- package/dist/trace/index.d.ts.map +1 -0
- package/dist/{integrity-D2t12mMw.d.ts → trace/integrity.d.ts} +11 -14
- package/dist/trace/integrity.d.ts.map +1 -0
- package/dist/trace/otel-bridge.d.ts +29 -0
- package/dist/trace/otel-bridge.d.ts.map +1 -0
- package/dist/trace/otel-export.d.ts +52 -0
- package/dist/trace/otel-export.d.ts.map +1 -0
- package/dist/trace/otel.d.ts +57 -0
- package/dist/trace/otel.d.ts.map +1 -0
- package/dist/trace/otlp-attributes.d.ts +17 -0
- package/dist/trace/otlp-attributes.d.ts.map +1 -0
- package/dist/trace/query.d.ts +29 -0
- package/dist/trace/query.d.ts.map +1 -0
- package/dist/trace/query.test.d.ts +2 -0
- package/dist/trace/query.test.d.ts.map +1 -0
- package/dist/{raw-provider-sink-C46HDghv.d.ts → trace/raw-provider-sink.d.ts} +13 -14
- package/dist/trace/raw-provider-sink.d.ts.map +1 -0
- package/dist/{redact-B40YG2M_.d.ts → trace/redact.d.ts} +7 -8
- package/dist/trace/redact.d.ts.map +1 -0
- package/dist/{schema-m0gsnbt3.d.ts → trace/schema.d.ts} +29 -30
- package/dist/trace/schema.d.ts.map +1 -0
- package/dist/trace/store-to-otlp.d.ts +72 -0
- package/dist/trace/store-to-otlp.d.ts.map +1 -0
- package/dist/trace/store-to-otlp.test.d.ts +2 -0
- package/dist/trace/store-to-otlp.test.d.ts.map +1 -0
- package/dist/{store-BcFXE6LG.d.ts → trace/store.d.ts} +21 -12
- package/dist/trace/store.d.ts.map +1 -0
- package/dist/trace/store.test.d.ts +2 -0
- package/dist/trace/store.test.d.ts.map +1 -0
- package/dist/{analyst-C8HHvfJp.d.ts → trace-analyst/analyst.d.ts} +12 -14
- package/dist/trace-analyst/analyst.d.ts.map +1 -0
- package/dist/trace-analyst/analyst.test.d.ts +2 -0
- package/dist/trace-analyst/analyst.test.d.ts.map +1 -0
- package/dist/trace-analyst/behavioral-metrics.d.ts +40 -0
- package/dist/trace-analyst/behavioral-metrics.d.ts.map +1 -0
- package/dist/trace-analyst/behavioral-metrics.test.d.ts +2 -0
- package/dist/trace-analyst/behavioral-metrics.test.d.ts.map +1 -0
- package/dist/trace-analyst/hook.d.ts +55 -0
- package/dist/trace-analyst/hook.d.ts.map +1 -0
- package/dist/trace-analyst/index.d.ts +18 -0
- package/dist/trace-analyst/index.d.ts.map +1 -0
- package/dist/trace-analyst/insights.d.ts +71 -0
- package/dist/trace-analyst/insights.d.ts.map +1 -0
- package/dist/trace-analyst/insights.test.d.ts +2 -0
- package/dist/trace-analyst/insights.test.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-flatten.d.ts +42 -0
- package/dist/trace-analyst/otlp-flatten.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-span.d.ts +85 -0
- package/dist/trace-analyst/otlp-span.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-span.test.d.ts +2 -0
- package/dist/trace-analyst/otlp-span.test.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-to-run-records.d.ts +115 -0
- package/dist/trace-analyst/otlp-to-run-records.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-to-run-records.test.d.ts +2 -0
- package/dist/trace-analyst/otlp-to-run-records.test.d.ts.map +1 -0
- package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts +2 -0
- package/dist/trace-analyst/otlp-to-run-records.timestamps.test.d.ts.map +1 -0
- package/dist/trace-analyst/prompts.d.ts +6 -0
- package/dist/trace-analyst/prompts.d.ts.map +1 -0
- package/dist/trace-analyst/store-otlp.d.ts +126 -0
- package/dist/trace-analyst/store-otlp.d.ts.map +1 -0
- package/dist/trace-analyst/store-otlp.test.d.ts +8 -0
- package/dist/trace-analyst/store-otlp.test.d.ts.map +1 -0
- package/dist/trace-analyst/store-otlp.timestamps.test.d.ts +2 -0
- package/dist/trace-analyst/store-otlp.timestamps.test.d.ts.map +1 -0
- package/dist/trace-analyst/store.d.ts +63 -0
- package/dist/trace-analyst/store.d.ts.map +1 -0
- package/dist/trace-analyst/tools.d.ts +44 -0
- package/dist/trace-analyst/tools.d.ts.map +1 -0
- package/dist/trace-analyst/tools.test.d.ts +10 -0
- package/dist/trace-analyst/tools.test.d.ts.map +1 -0
- package/dist/{store-C1YxJDEK.d.ts → trace-analyst/types.d.ts} +18 -74
- package/dist/trace-analyst/types.d.ts.map +1 -0
- package/dist/trace-contracts.d.ts +180 -0
- package/dist/trace-contracts.d.ts.map +1 -0
- package/dist/traced-analyst.d.ts +26 -0
- package/dist/traced-analyst.d.ts.map +1 -0
- package/dist/traced-judges.d.ts +27 -0
- package/dist/traced-judges.d.ts.map +1 -0
- package/dist/traces.d.ts +4 -960
- package/dist/traces.d.ts.map +1 -0
- package/dist/traces.js +41 -11
- package/dist/{trajectory-2TkpSEVh.d.ts → trajectory.d.ts} +8 -9
- package/dist/trajectory.d.ts.map +1 -0
- package/dist/{types-C7DGg5ex.d.ts → types.d.ts} +31 -33
- package/dist/types.d.ts.map +1 -0
- package/dist/ui-finding.d.ts +104 -0
- package/dist/ui-finding.d.ts.map +1 -0
- package/dist/verdict-cache.d.ts +78 -0
- package/dist/verdict-cache.d.ts.map +1 -0
- package/dist/verdict-cache.test.d.ts +2 -0
- package/dist/verdict-cache.test.d.ts.map +1 -0
- package/dist/{verdict-C9MlYujm.d.ts → verdict.d.ts} +2 -3
- package/dist/verdict.d.ts.map +1 -0
- package/dist/visual-diff.d.ts +32 -0
- package/dist/visual-diff.d.ts.map +1 -0
- package/dist/wire/handlers.d.ts +54 -0
- package/dist/wire/handlers.d.ts.map +1 -0
- package/dist/wire/index.d.ts +13 -570
- package/dist/wire/index.d.ts.map +1 -0
- package/dist/wire/openapi.d.ts +3 -0
- package/dist/wire/openapi.d.ts.map +1 -0
- package/dist/wire/rpc.d.ts +21 -0
- package/dist/wire/rpc.d.ts.map +1 -0
- package/dist/wire/rubrics.d.ts +34 -0
- package/dist/wire/rubrics.d.ts.map +1 -0
- package/dist/wire/schemas.d.ts +410 -0
- package/dist/wire/schemas.d.ts.map +1 -0
- package/dist/wire/server.d.ts +60 -0
- package/dist/wire/server.d.ts.map +1 -0
- package/dist/workflow/event-schema.d.ts +5 -0
- package/dist/workflow/event-schema.d.ts.map +1 -0
- package/dist/workflow/feedback-pack.d.ts +99 -0
- package/dist/workflow/feedback-pack.d.ts.map +1 -0
- package/dist/workflow/index.d.ts +22 -495
- package/dist/workflow/index.d.ts.map +1 -0
- package/dist/workflow/index.js +1 -1
- package/dist/workflow/intelligence-export.d.ts +62 -0
- package/dist/workflow/intelligence-export.d.ts.map +1 -0
- package/dist/workflow/partner-report.d.ts +49 -0
- package/dist/workflow/partner-report.d.ts.map +1 -0
- package/dist/workflow/phase-graph.d.ts +43 -0
- package/dist/workflow/phase-graph.d.ts.map +1 -0
- package/dist/workflow/promotion-gate.d.ts +61 -0
- package/dist/workflow/promotion-gate.d.ts.map +1 -0
- package/dist/workflow/run-record.d.ts +12 -0
- package/dist/workflow/run-record.d.ts.map +1 -0
- package/dist/workflow/runtime-adapter.d.ts +20 -0
- package/dist/workflow/runtime-adapter.d.ts.map +1 -0
- package/dist/workflow/sanitize.d.ts +21 -0
- package/dist/workflow/sanitize.d.ts.map +1 -0
- package/dist/workflow/schema.d.ts +5 -0
- package/dist/workflow/schema.d.ts.map +1 -0
- package/dist/workflow/summary.d.ts +43 -0
- package/dist/workflow/summary.d.ts.map +1 -0
- package/dist/workflow/trace-event-fields.d.ts +6 -0
- package/dist/workflow/trace-event-fields.d.ts.map +1 -0
- package/dist/workflow/trajectory.d.ts +15 -0
- package/dist/workflow/trajectory.d.ts.map +1 -0
- package/dist/workflow/types.d.ts +68 -0
- package/dist/workflow/types.d.ts.map +1 -0
- package/dist/workspace-inspector.d.ts +67 -0
- package/dist/workspace-inspector.d.ts.map +1 -0
- package/dist/wrangler-deploy-runner.test.d.ts +2 -0
- package/dist/wrangler-deploy-runner.test.d.ts.map +1 -0
- package/docs/campaign-proposers.md +170 -0
- package/docs/concepts.md +8 -4
- package/docs/customer-journeys.md +15 -13
- package/docs/design/loop-taxonomy.md +34 -66
- package/docs/distributed-driver.md +14 -14
- package/docs/feature-guide.md +1 -1
- package/docs/hosted-ingest-spec.md +2 -3
- package/docs/multi-shot-optimization.md +8 -8
- package/docs/product-eval-adoption.md +1 -1
- package/docs/self-improvement-map.md +33 -29
- package/package.json +9 -15
- package/dist/calibration-BPmzuVPk.d.ts +0 -101
- package/dist/chunk-2K6UUZ7P.js.map +0 -1
- package/dist/chunk-CTBHKLEU.js.map +0 -1
- package/dist/chunk-E4GH6USR.js.map +0 -1
- package/dist/chunk-EGPMSBEZ.js.map +0 -1
- package/dist/chunk-KWRRMR3J.js.map +0 -1
- package/dist/chunk-MIFZUPEK.js.map +0 -1
- package/dist/chunk-MPQWFX6Y.js.map +0 -1
- package/dist/chunk-Q5LIB7BC.js.map +0 -1
- package/dist/chunk-QMUEXQJS.js.map +0 -1
- package/dist/chunk-SD2YFWQQ.js.map +0 -1
- package/dist/control-D6qwHXIR.d.ts +0 -259
- package/dist/corpus-B8A4BDR3.d.ts +0 -560
- package/dist/failure-cluster-DH9Flgcf.d.ts +0 -76
- package/dist/harness-optimizer-mOl9XX_O.d.ts +0 -106
- package/dist/index-Bx3gZ8xl.d.ts +0 -159
- package/dist/pre-registration-mAnCugl9.d.ts +0 -523
- package/dist/provenance-P-bCL2Fo.d.ts +0 -441
- package/dist/query-B7GGjRox.d.ts +0 -32
- package/dist/release-report-BEbWmVYj.d.ts +0 -233
- package/dist/run-critic-CmMf05uV.d.ts +0 -56
- package/dist/run-improvement-loop-DBahB8Ax.d.ts +0 -427
- package/dist/runtime-trajectory-BDgfGZSr.d.ts +0 -49
- package/dist/semantic-concept-judge-B9MgmBnM.d.ts +0 -624
- package/dist/sink-fetch-B1Yg4Til.d.ts +0 -101
- package/docs/design/external-agent-wedge.md +0 -89
- package/docs/design/phase-d-rfc.md +0 -125
- package/docs/design/phase4-consumer-migration.md +0 -70
- package/docs/design/primitives-integration-spec.md +0 -393
- package/docs/design/product-self-improvement-loop.md +0 -146
- package/docs/design/self-improvement-engine.md +0 -140
- package/docs/design/self-improvement-protocol.md +0 -223
- package/docs/design/self-improvement-roadmap.md +0 -106
- package/docs/design/substrate-gaps.md +0 -118
- package/docs/phase-b-pairing-kit.md +0 -188
- package/docs/phase-b-runbook.md +0 -176
- package/docs/pilot/README.md +0 -62
- package/docs/pilot/customer-checklist.md +0 -90
- package/docs/pilot/integration-foreign-stack.md +0 -296
- package/docs/pilot/integration-tangle-stack.md +0 -248
- package/docs/pilot/one-pager.md +0 -161
- package/docs/pilot/sample-insight-report.json +0 -172
- package/docs/quickstart-external.md +0 -229
- package/docs/research/belief-state-agent-eval-roadmap.md +0 -593
- package/docs/research/research-roadmap.md +0 -205
- package/docs/specs/driver-honest-spec.md +0 -251
- package/docs/specs/hermes-self-improvement-audit.md +0 -93
- package/docs/specs/profile-versioning.md +0 -291
- package/docs/three-package-architecture.md +0 -168
- /package/dist/{chunk-TBDR6PAI.js.map → chunk-IZCEK2HR.js.map} +0 -0
- /package/dist/{chunk-KW53MSA5.js.map → chunk-X74V6ESX.js.map} +0 -0
- /package/dist/{chunk-2KNZHH3P.js.map → chunk-Z6L6YSU6.js.map} +0 -0
- /package/dist/{run-campaign-7WNXMDSN.js.map → run-campaign-WXY7KI67.js.map} +0 -0
|
@@ -0,0 +1,55 @@
|
|
|
1
|
+
import type { RunSplitTag } from '../run-record';
|
|
2
|
+
import { type BeliefDecisionResearchEvidencePacket, type BuildBeliefDecisionResearchEvidencePacketOptions } from './research-evidence';
|
|
3
|
+
import { type RuntimeBeliefDecisionPoint, type RuntimeBeliefHookEvent } from './runtime-hooks';
|
|
4
|
+
import type { BeliefDecisionOutcome, BeliefDecisionPoint } from './types';
|
|
5
|
+
export interface RuntimeBeliefPhase0RunRecord {
|
|
6
|
+
runId: string;
|
|
7
|
+
scenarioId?: string;
|
|
8
|
+
splitTag: RunSplitTag;
|
|
9
|
+
}
|
|
10
|
+
export interface RuntimeBeliefDecisionLabel {
|
|
11
|
+
decisionId: string;
|
|
12
|
+
chosenAction: string;
|
|
13
|
+
outcome: BeliefDecisionOutcome;
|
|
14
|
+
confidence?: number;
|
|
15
|
+
behaviorProb?: number;
|
|
16
|
+
targetProb?: number;
|
|
17
|
+
qHat?: number | null;
|
|
18
|
+
costUsd?: number;
|
|
19
|
+
splitTag?: RunSplitTag;
|
|
20
|
+
metadata?: Record<string, unknown>;
|
|
21
|
+
}
|
|
22
|
+
export interface BuildRuntimeBeliefPhase0MeasurementOptions extends Omit<BuildBeliefDecisionResearchEvidencePacketOptions, 'points'> {
|
|
23
|
+
runs: RuntimeBeliefPhase0RunRecord[];
|
|
24
|
+
decisions: RuntimeBeliefDecisionPoint[];
|
|
25
|
+
events?: RuntimeBeliefHookEvent[];
|
|
26
|
+
labels: RuntimeBeliefDecisionLabel[];
|
|
27
|
+
baselinePolicyId?: string;
|
|
28
|
+
}
|
|
29
|
+
export interface RuntimeBeliefPhase0MeasurementSummary {
|
|
30
|
+
runCount: number;
|
|
31
|
+
producerDecisionCount: number;
|
|
32
|
+
lifecycleEventCount: number;
|
|
33
|
+
labelCount: number;
|
|
34
|
+
completedPointCount: number;
|
|
35
|
+
runJoinRate: number;
|
|
36
|
+
labelJoinRate: number;
|
|
37
|
+
missingRunRecordCount: number;
|
|
38
|
+
missingLabelCount: number;
|
|
39
|
+
withEvidence: number;
|
|
40
|
+
withOutcome: number;
|
|
41
|
+
withSplit: number;
|
|
42
|
+
withBehaviorProb: number;
|
|
43
|
+
withTargetProb: number;
|
|
44
|
+
baselinePolicyId: string;
|
|
45
|
+
packetStatus: BeliefDecisionResearchEvidencePacket['status'];
|
|
46
|
+
claimScope: BeliefDecisionResearchEvidencePacket['claimScope'];
|
|
47
|
+
}
|
|
48
|
+
export interface RuntimeBeliefPhase0Measurement {
|
|
49
|
+
points: BeliefDecisionPoint[];
|
|
50
|
+
packet: BeliefDecisionResearchEvidencePacket;
|
|
51
|
+
summary: RuntimeBeliefPhase0MeasurementSummary;
|
|
52
|
+
diagnostics: string[];
|
|
53
|
+
}
|
|
54
|
+
export declare function buildRuntimeBeliefPhase0Measurement(options: BuildRuntimeBeliefPhase0MeasurementOptions): RuntimeBeliefPhase0Measurement;
|
|
55
|
+
//# sourceMappingURL=phase0-measurement.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"phase0-measurement.d.ts","sourceRoot":"","sources":["../../src/belief-state/phase0-measurement.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,eAAe,CAAA;AAChD,OAAO,EACL,KAAK,oCAAoC,EACzC,KAAK,gDAAgD,EAEtD,MAAM,qBAAqB,CAAA;AAC5B,OAAO,EACL,KAAK,0BAA0B,EAC/B,KAAK,sBAAsB,EAE5B,MAAM,iBAAiB,CAAA;AACxB,OAAO,KAAK,EAAE,qBAAqB,EAAE,mBAAmB,EAAE,MAAM,SAAS,CAAA;AAEzE,MAAM,WAAW,4BAA4B;IAC3C,KAAK,EAAE,MAAM,CAAA;IACb,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,QAAQ,EAAE,WAAW,CAAA;CACtB;AAED,MAAM,WAAW,0BAA0B;IACzC,UAAU,EAAE,MAAM,CAAA;IAClB,YAAY,EAAE,MAAM,CAAA;IACpB,OAAO,EAAE,qBAAqB,CAAA;IAC9B,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,IAAI,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;IACpB,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,QAAQ,CAAC,EAAE,WAAW,CAAA;IACtB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,0CACf,SAAQ,IAAI,CAAC,gDAAgD,EAAE,QAAQ,CAAC;IACxE,IAAI,EAAE,4BAA4B,EAAE,CAAA;IACpC,SAAS,EAAE,0BAA0B,EAAE,CAAA;IACvC,MAAM,CAAC,EAAE,sBAAsB,EAAE,CAAA;IACjC,MAAM,EAAE,0BAA0B,EAAE,CAAA;IACpC,gBAAgB,CAAC,EAAE,MAAM,CAAA;CAC1B;AAED,MAAM,WAAW,qCAAqC;IACpD,QAAQ,EAAE,MAAM,CAAA;IAChB,qBAAqB,EAAE,MAAM,CAAA;IAC7B,mBAAmB,EAAE,MAAM,CAAA;IAC3B,UAAU,EAAE,MAAM,CAAA;IAClB,mBAAmB,EAAE,MAAM,CAAA;IAC3B,WAAW,EAAE,MAAM,CAAA;IACnB,aAAa,EAAE,MAAM,CAAA;IACrB,qBAAqB,EAAE,MAAM,CAAA;IAC7B,iBAAiB,EAAE,MAAM,CAAA;IACzB,YAAY,EAAE,MAAM,CAAA;IACpB,WAAW,EAAE,MAAM,CAAA;IACnB,SAAS,EAAE,MAAM,CAAA;IACjB,gBAAgB,EAAE,MAAM,CAAA;IACxB,cAAc,EAAE,MAAM,CAAA;IACtB,gBAAgB,EAAE,MAAM,CAAA;IACxB,YAAY,EAAE,oCAAoC,CAAC,QAAQ,CAAC,CAAA;IAC5D,UAAU,EAAE,oCAAoC,CAAC,YAAY,CAAC,CAAA;CAC/D;AAED,MAAM,WAAW,8BAA8B;IAC7C,MAAM,EAAE,mBAAmB,EAAE,CAAA;IAC7B,MAAM,EAAE,oCAAoC,CAAA;IAC5C,OAAO,EAAE,qCAAqC,CAAA;IAC9C,WAAW,EAAE,MAAM,EAAE,CAAA;CACtB;AAID,wBAAgB,mCAAmC,CACjD,OAAO,EAAE,0CAA0C,GAClD,8BAA8B,CAoEhC"}
|
|
@@ -0,0 +1,17 @@
|
|
|
1
|
+
import { type BeliefCalibrationOptions } from './calibration';
|
|
2
|
+
import { type BeliefOpeOptions } from './ope';
|
|
3
|
+
import { type EvaluateBeliefSelectivePolicyOptions } from './selective';
|
|
4
|
+
import type { BeliefDecisionPoint, BeliefOpeTargetPolicy, BeliefPolicyEvaluationReport, BeliefSelectivePolicy } from './types';
|
|
5
|
+
export interface AnalyzeBeliefPolicyOpeOptions extends BeliefOpeOptions {
|
|
6
|
+
targetPolicy?: BeliefOpeTargetPolicy;
|
|
7
|
+
}
|
|
8
|
+
export interface AnalyzeBeliefPolicyOptions {
|
|
9
|
+
points: BeliefDecisionPoint[];
|
|
10
|
+
policy: BeliefSelectivePolicy;
|
|
11
|
+
selective?: EvaluateBeliefSelectivePolicyOptions;
|
|
12
|
+
calibration?: BeliefCalibrationOptions;
|
|
13
|
+
ope?: AnalyzeBeliefPolicyOpeOptions;
|
|
14
|
+
requireOpe?: boolean;
|
|
15
|
+
}
|
|
16
|
+
export declare function analyzeBeliefPolicy(options: AnalyzeBeliefPolicyOptions): BeliefPolicyEvaluationReport;
|
|
17
|
+
//# sourceMappingURL=report.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"report.d.ts","sourceRoot":"","sources":["../../src/belief-state/report.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,KAAK,wBAAwB,EAA4B,MAAM,eAAe,CAAA;AACvF,OAAO,EAAE,KAAK,gBAAgB,EAA2B,MAAM,OAAO,CAAA;AACtE,OAAO,EACL,KAAK,oCAAoC,EAE1C,MAAM,aAAa,CAAA;AACpB,OAAO,KAAK,EACV,mBAAmB,EAGnB,qBAAqB,EACrB,4BAA4B,EAC5B,qBAAqB,EACtB,MAAM,SAAS,CAAA;AAEhB,MAAM,WAAW,6BAA8B,SAAQ,gBAAgB;IACrE,YAAY,CAAC,EAAE,qBAAqB,CAAA;CACrC;AAED,MAAM,WAAW,0BAA0B;IACzC,MAAM,EAAE,mBAAmB,EAAE,CAAA;IAC7B,MAAM,EAAE,qBAAqB,CAAA;IAC7B,SAAS,CAAC,EAAE,oCAAoC,CAAA;IAChD,WAAW,CAAC,EAAE,wBAAwB,CAAA;IACtC,GAAG,CAAC,EAAE,6BAA6B,CAAA;IACnC,UAAU,CAAC,EAAE,OAAO,CAAA;CACrB;AAED,wBAAgB,mBAAmB,CACjC,OAAO,EAAE,0BAA0B,GAClC,4BAA4B,CA4C9B"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"report.test.d.ts","sourceRoot":"","sources":["../../src/belief-state/report.test.ts"],"names":[],"mappings":""}
|
|
@@ -0,0 +1,23 @@
|
|
|
1
|
+
import { type AnalyzeBeliefDecisionCorpusOptions, type BeliefDecisionCorpusEvaluation } from './code-agent-corpus';
|
|
2
|
+
export type BeliefResearchClaimScope = 'selective' | 'counterfactual';
|
|
3
|
+
export type BeliefResearchEvidenceStatus = 'supported' | 'blocked';
|
|
4
|
+
export type BeliefResearchGateId = 'corpus' | 'selective' | 'calibration' | 'ope';
|
|
5
|
+
export interface BeliefResearchEvidenceGate {
|
|
6
|
+
id: BeliefResearchGateId;
|
|
7
|
+
status: BeliefResearchEvidenceStatus;
|
|
8
|
+
blockers: string[];
|
|
9
|
+
caveats: string[];
|
|
10
|
+
}
|
|
11
|
+
export interface BeliefDecisionResearchEvidencePacket {
|
|
12
|
+
claimScope: BeliefResearchClaimScope;
|
|
13
|
+
status: BeliefResearchEvidenceStatus;
|
|
14
|
+
analysis: BeliefDecisionCorpusEvaluation;
|
|
15
|
+
gates: BeliefResearchEvidenceGate[];
|
|
16
|
+
blockers: string[];
|
|
17
|
+
caveats: string[];
|
|
18
|
+
}
|
|
19
|
+
export interface BuildBeliefDecisionResearchEvidencePacketOptions extends AnalyzeBeliefDecisionCorpusOptions {
|
|
20
|
+
claimScope?: BeliefResearchClaimScope;
|
|
21
|
+
}
|
|
22
|
+
export declare function buildBeliefDecisionResearchEvidencePacket(options: BuildBeliefDecisionResearchEvidencePacketOptions): BeliefDecisionResearchEvidencePacket;
|
|
23
|
+
//# sourceMappingURL=research-evidence.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"research-evidence.d.ts","sourceRoot":"","sources":["../../src/belief-state/research-evidence.ts"],"names":[],"mappings":"AAAA,OAAO,EACL,KAAK,kCAAkC,EAEvC,KAAK,8BAA8B,EACpC,MAAM,qBAAqB,CAAA;AAE5B,MAAM,MAAM,wBAAwB,GAAG,WAAW,GAAG,gBAAgB,CAAA;AACrE,MAAM,MAAM,4BAA4B,GAAG,WAAW,GAAG,SAAS,CAAA;AAClE,MAAM,MAAM,oBAAoB,GAAG,QAAQ,GAAG,WAAW,GAAG,aAAa,GAAG,KAAK,CAAA;AAEjF,MAAM,WAAW,0BAA0B;IACzC,EAAE,EAAE,oBAAoB,CAAA;IACxB,MAAM,EAAE,4BAA4B,CAAA;IACpC,QAAQ,EAAE,MAAM,EAAE,CAAA;IAClB,OAAO,EAAE,MAAM,EAAE,CAAA;CAClB;AAED,MAAM,WAAW,oCAAoC;IACnD,UAAU,EAAE,wBAAwB,CAAA;IACpC,MAAM,EAAE,4BAA4B,CAAA;IACpC,QAAQ,EAAE,8BAA8B,CAAA;IACxC,KAAK,EAAE,0BAA0B,EAAE,CAAA;IACnC,QAAQ,EAAE,MAAM,EAAE,CAAA;IAClB,OAAO,EAAE,MAAM,EAAE,CAAA;CAClB;AAED,MAAM,WAAW,gDACf,SAAQ,kCAAkC;IAC1C,UAAU,CAAC,EAAE,wBAAwB,CAAA;CACtC;AAED,wBAAgB,yCAAyC,CACvD,OAAO,EAAE,gDAAgD,GACxD,oCAAoC,CA4BtC"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"research-evidence.test.d.ts","sourceRoot":"","sources":["../../src/belief-state/research-evidence.test.ts"],"names":[],"mappings":""}
|
|
@@ -0,0 +1,32 @@
|
|
|
1
|
+
import { type ProjectRuntimeTrajectoryEvidenceOptions, type RuntimeTrajectoryEvidenceProjection, type RuntimeTrajectoryRecord } from '../runtime-trajectory';
|
|
2
|
+
import { type BuildRuntimeBeliefPhase0MeasurementOptions, type RuntimeBeliefDecisionLabel, type RuntimeBeliefPhase0Measurement, type RuntimeBeliefPhase0RunRecord } from './phase0-measurement';
|
|
3
|
+
import type { RuntimeBeliefDecisionPoint, RuntimeBeliefHookEvent } from './runtime-hooks';
|
|
4
|
+
type RuntimeBenchmarkTrajectoryRecord = RuntimeTrajectoryRecord & {
|
|
5
|
+
benchmark?: unknown;
|
|
6
|
+
condition?: unknown;
|
|
7
|
+
instanceId?: unknown;
|
|
8
|
+
runtimeDecisionPoints?: unknown;
|
|
9
|
+
};
|
|
10
|
+
export interface BuildRuntimeBenchmarkBeliefPhase0MeasurementOptions extends Omit<BuildRuntimeBeliefPhase0MeasurementOptions, 'runs' | 'events' | 'decisions' | 'labels'> {
|
|
11
|
+
records: RuntimeBenchmarkTrajectoryRecord[];
|
|
12
|
+
decisions?: RuntimeBeliefDecisionPoint[];
|
|
13
|
+
defaultSplitTag?: ProjectRuntimeTrajectoryEvidenceOptions['defaultSplitTag'];
|
|
14
|
+
labels?: RuntimeBeliefDecisionLabel[];
|
|
15
|
+
}
|
|
16
|
+
export interface RuntimeBenchmarkBeliefPhase0Summary {
|
|
17
|
+
decisionCount: number;
|
|
18
|
+
labelCount: number;
|
|
19
|
+
}
|
|
20
|
+
export interface RuntimeBenchmarkBeliefPhase0Measurement {
|
|
21
|
+
runs: RuntimeBeliefPhase0RunRecord[];
|
|
22
|
+
events: RuntimeBeliefHookEvent[];
|
|
23
|
+
decisions: RuntimeBeliefDecisionPoint[];
|
|
24
|
+
labels: RuntimeBeliefDecisionLabel[];
|
|
25
|
+
trajectory: RuntimeTrajectoryEvidenceProjection;
|
|
26
|
+
measurement: RuntimeBeliefPhase0Measurement;
|
|
27
|
+
summary: RuntimeBenchmarkBeliefPhase0Summary;
|
|
28
|
+
diagnostics: string[];
|
|
29
|
+
}
|
|
30
|
+
export declare function buildRuntimeBenchmarkBeliefPhase0Measurement(options: BuildRuntimeBenchmarkBeliefPhase0MeasurementOptions): RuntimeBenchmarkBeliefPhase0Measurement;
|
|
31
|
+
export {};
|
|
32
|
+
//# sourceMappingURL=runtime-benchmark-corpus.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"runtime-benchmark-corpus.d.ts","sourceRoot":"","sources":["../../src/belief-state/runtime-benchmark-corpus.ts"],"names":[],"mappings":"AAAA,OAAO,EACL,KAAK,uCAAuC,EAE5C,KAAK,mCAAmC,EACxC,KAAK,uBAAuB,EAC7B,MAAM,uBAAuB,CAAA;AAC9B,OAAO,EACL,KAAK,0CAA0C,EAE/C,KAAK,0BAA0B,EAC/B,KAAK,8BAA8B,EACnC,KAAK,4BAA4B,EAClC,MAAM,sBAAsB,CAAA;AAC7B,OAAO,KAAK,EAAE,0BAA0B,EAAE,sBAAsB,EAAE,MAAM,iBAAiB,CAAA;AAEzF,KAAK,gCAAgC,GAAG,uBAAuB,GAAG;IAChE,SAAS,CAAC,EAAE,OAAO,CAAA;IACnB,SAAS,CAAC,EAAE,OAAO,CAAA;IACnB,UAAU,CAAC,EAAE,OAAO,CAAA;IACpB,qBAAqB,CAAC,EAAE,OAAO,CAAA;CAChC,CAAA;AAmBD,MAAM,WAAW,mDACf,SAAQ,IAAI,CACV,0CAA0C,EAC1C,MAAM,GAAG,QAAQ,GAAG,WAAW,GAAG,QAAQ,CAC3C;IACD,OAAO,EAAE,gCAAgC,EAAE,CAAA;IAC3C,SAAS,CAAC,EAAE,0BAA0B,EAAE,CAAA;IACxC,eAAe,CAAC,EAAE,uCAAuC,CAAC,iBAAiB,CAAC,CAAA;IAC5E,MAAM,CAAC,EAAE,0BAA0B,EAAE,CAAA;CACtC;AAED,MAAM,WAAW,mCAAmC;IAClD,aAAa,EAAE,MAAM,CAAA;IACrB,UAAU,EAAE,MAAM,CAAA;CACnB;AAED,MAAM,WAAW,uCAAuC;IACtD,IAAI,EAAE,4BAA4B,EAAE,CAAA;IACpC,MAAM,EAAE,sBAAsB,EAAE,CAAA;IAChC,SAAS,EAAE,0BAA0B,EAAE,CAAA;IACvC,MAAM,EAAE,0BAA0B,EAAE,CAAA;IACpC,UAAU,EAAE,mCAAmC,CAAA;IAC/C,WAAW,EAAE,8BAA8B,CAAA;IAC3C,OAAO,EAAE,mCAAmC,CAAA;IAC5C,WAAW,EAAE,MAAM,EAAE,CAAA;CACtB;AAED,wBAAgB,4CAA4C,CAC1D,OAAO,EAAE,mDAAmD,GAC3D,uCAAuC,CA2CzC"}
|
|
@@ -0,0 +1,87 @@
|
|
|
1
|
+
import type { BeliefShadowProbeInput } from './shadow-probe';
|
|
2
|
+
import type { BeliefDecisionKind, BeliefDecisionOutcome, BeliefDecisionPoint, BeliefEvidenceQuality } from './types';
|
|
3
|
+
export interface RuntimeBeliefDecisionEvidenceRef {
|
|
4
|
+
source: string;
|
|
5
|
+
id: string;
|
|
6
|
+
detail?: string;
|
|
7
|
+
quality?: BeliefEvidenceQuality;
|
|
8
|
+
metadata?: Record<string, unknown>;
|
|
9
|
+
}
|
|
10
|
+
export interface RuntimeBeliefDecisionPoint {
|
|
11
|
+
id: string;
|
|
12
|
+
runId: string;
|
|
13
|
+
scenarioId?: string;
|
|
14
|
+
stepIndex: number;
|
|
15
|
+
kind: string;
|
|
16
|
+
candidateActions?: string[];
|
|
17
|
+
context?: string;
|
|
18
|
+
evidence?: RuntimeBeliefDecisionEvidenceRef[];
|
|
19
|
+
metadata?: Record<string, unknown>;
|
|
20
|
+
}
|
|
21
|
+
export interface RuntimeBeliefHookEvent {
|
|
22
|
+
id: string;
|
|
23
|
+
runId: string;
|
|
24
|
+
scenarioId?: string;
|
|
25
|
+
target: string;
|
|
26
|
+
phase: string;
|
|
27
|
+
timestamp: number;
|
|
28
|
+
stepIndex?: number;
|
|
29
|
+
parentId?: string;
|
|
30
|
+
payload?: unknown;
|
|
31
|
+
metadata?: Record<string, unknown>;
|
|
32
|
+
}
|
|
33
|
+
export interface RuntimeBeliefHookContext {
|
|
34
|
+
signal?: AbortSignal;
|
|
35
|
+
}
|
|
36
|
+
export interface RuntimeBeliefHooks {
|
|
37
|
+
onEvent?: (event: RuntimeBeliefHookEvent, context: RuntimeBeliefHookContext) => void | Promise<void>;
|
|
38
|
+
onDecisionPoint?: (point: RuntimeBeliefDecisionPoint, context: RuntimeBeliefHookContext) => void | Promise<void>;
|
|
39
|
+
}
|
|
40
|
+
export interface RuntimeBeliefConversionDiagnostic {
|
|
41
|
+
decisionId: string;
|
|
42
|
+
severity: 'warning' | 'error';
|
|
43
|
+
reason: string;
|
|
44
|
+
}
|
|
45
|
+
export interface RuntimeBeliefShadowProbeInputOptions {
|
|
46
|
+
probeId: string;
|
|
47
|
+
decisionKind?: BeliefDecisionKind;
|
|
48
|
+
includeEvidenceDetail?: boolean;
|
|
49
|
+
includeLifecycleEvidence?: boolean;
|
|
50
|
+
lifecycleEvents?: RuntimeBeliefHookEvent[];
|
|
51
|
+
maxContextChars?: number;
|
|
52
|
+
}
|
|
53
|
+
export interface RuntimeBeliefDecisionPointOptions {
|
|
54
|
+
chosenAction?: string;
|
|
55
|
+
decisionKind?: BeliefDecisionKind;
|
|
56
|
+
confidence?: number;
|
|
57
|
+
behaviorProb?: number;
|
|
58
|
+
targetProb?: number;
|
|
59
|
+
qHat?: number | null;
|
|
60
|
+
costUsd?: number;
|
|
61
|
+
outcome?: BeliefDecisionOutcome;
|
|
62
|
+
metadata?: Record<string, unknown>;
|
|
63
|
+
includeLifecycleEvidence?: boolean;
|
|
64
|
+
lifecycleEvents?: RuntimeBeliefHookEvent[];
|
|
65
|
+
}
|
|
66
|
+
export interface RuntimeBeliefShadowProbeInputReport {
|
|
67
|
+
input?: BeliefShadowProbeInput;
|
|
68
|
+
diagnostics: RuntimeBeliefConversionDiagnostic[];
|
|
69
|
+
}
|
|
70
|
+
export interface RuntimeBeliefDecisionPointReport {
|
|
71
|
+
point?: BeliefDecisionPoint;
|
|
72
|
+
diagnostics: RuntimeBeliefConversionDiagnostic[];
|
|
73
|
+
}
|
|
74
|
+
export interface BeliefRuntimeHookCollector {
|
|
75
|
+
hooks: RuntimeBeliefHooks;
|
|
76
|
+
decisions: RuntimeBeliefDecisionPoint[];
|
|
77
|
+
events: RuntimeBeliefHookEvent[];
|
|
78
|
+
toShadowProbeInputs(options?: Partial<RuntimeBeliefShadowProbeInputOptions>): {
|
|
79
|
+
inputs: BeliefShadowProbeInput[];
|
|
80
|
+
diagnostics: RuntimeBeliefConversionDiagnostic[];
|
|
81
|
+
};
|
|
82
|
+
clear(): void;
|
|
83
|
+
}
|
|
84
|
+
export declare function runtimeDecisionPointToBeliefShadowProbeInput(point: RuntimeBeliefDecisionPoint, options: RuntimeBeliefShadowProbeInputOptions): RuntimeBeliefShadowProbeInputReport;
|
|
85
|
+
export declare function runtimeDecisionPointToBeliefDecisionPoint(point: RuntimeBeliefDecisionPoint, options: RuntimeBeliefDecisionPointOptions): RuntimeBeliefDecisionPointReport;
|
|
86
|
+
export declare function createBeliefRuntimeHookCollector(defaults: RuntimeBeliefShadowProbeInputOptions): BeliefRuntimeHookCollector;
|
|
87
|
+
//# sourceMappingURL=runtime-hooks.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"runtime-hooks.d.ts","sourceRoot":"","sources":["../../src/belief-state/runtime-hooks.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,sBAAsB,EAAE,MAAM,gBAAgB,CAAA;AAC5D,OAAO,KAAK,EACV,kBAAkB,EAClB,qBAAqB,EACrB,mBAAmB,EACnB,qBAAqB,EAEtB,MAAM,SAAS,CAAA;AAGhB,MAAM,WAAW,gCAAgC;IAC/C,MAAM,EAAE,MAAM,CAAA;IACd,EAAE,EAAE,MAAM,CAAA;IACV,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,OAAO,CAAC,EAAE,qBAAqB,CAAA;IAC/B,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,0BAA0B;IACzC,EAAE,EAAE,MAAM,CAAA;IACV,KAAK,EAAE,MAAM,CAAA;IACb,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,SAAS,EAAE,MAAM,CAAA;IACjB,IAAI,EAAE,MAAM,CAAA;IACZ,gBAAgB,CAAC,EAAE,MAAM,EAAE,CAAA;IAC3B,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,QAAQ,CAAC,EAAE,gCAAgC,EAAE,CAAA;IAC7C,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,sBAAsB;IACrC,EAAE,EAAE,MAAM,CAAA;IACV,KAAK,EAAE,MAAM,CAAA;IACb,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,MAAM,EAAE,MAAM,CAAA;IACd,KAAK,EAAE,MAAM,CAAA;IACb,SAAS,EAAE,MAAM,CAAA;IACjB,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,QAAQ,CAAC,EAAE,MAAM,CAAA;IACjB,OAAO,CAAC,EAAE,OAAO,CAAA;IACjB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,wBAAwB;IACvC,MAAM,CAAC,EAAE,WAAW,CAAA;CACrB;AAED,MAAM,WAAW,kBAAkB;IACjC,OAAO,CAAC,EAAE,CACR,KAAK,EAAE,sBAAsB,EAC7B,OAAO,EAAE,wBAAwB,KAC9B,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAA;IACzB,eAAe,CAAC,EAAE,CAChB,KAAK,EAAE,0BAA0B,EACjC,OAAO,EAAE,wBAAwB,KAC9B,IAAI,GAAG,OAAO,CAAC,IAAI,CAAC,CAAA;CAC1B;AAED,MAAM,WAAW,iCAAiC;IAChD,UAAU,EAAE,MAAM,CAAA;IAClB,QAAQ,EAAE,SAAS,GAAG,OAAO,CAAA;IAC7B,MAAM,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,oCAAoC;IACnD,OAAO,EAAE,MAAM,CAAA;IACf,YAAY,CAAC,EAAE,kBAAkB,CAAA;IACjC,qBAAqB,CAAC,EAAE,OAAO,CAAA;IAC/B,wBAAwB,CAAC,EAAE,OAAO,CAAA;IAClC,eAAe,CAAC,EAAE,sBAAsB,EAAE,CAAA;IAC1C,eAAe,CAAC,EAAE,MAAM,CAAA;CACzB;AAED,MAAM,WAAW,iCAAiC;IAChD,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,YAAY,CAAC,EAAE,kBAAkB,CAAA;IACjC,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,IAAI,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;IACpB,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,OAAO,CAAC,EAAE,qBAAqB,CAAA;IAC/B,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;IAClC,wBAAwB,CAAC,EAAE,OAAO,CAAA;IAClC,eAAe,CAAC,EAAE,sBAAsB,EAAE,CAAA;CAC3C;AAED,MAAM,WAAW,mCAAmC;IAClD,KAAK,CAAC,EAAE,sBAAsB,CAAA;IAC9B,WAAW,EAAE,iCAAiC,EAAE,CAAA;CACjD;AAED,MAAM,WAAW,gCAAgC;IAC/C,KAAK,CAAC,EAAE,mBAAmB,CAAA;IAC3B,WAAW,EAAE,iCAAiC,EAAE,CAAA;CACjD;AAED,MAAM,WAAW,0BAA0B;IACzC,KAAK,EAAE,kBAAkB,CAAA;IACzB,SAAS,EAAE,0BAA0B,EAAE,CAAA;IACvC,MAAM,EAAE,sBAAsB,EAAE,CAAA;IAChC,mBAAmB,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC,oCAAoC,CAAC,GAAG;QAC5E,MAAM,EAAE,sBAAsB,EAAE,CAAA;QAChC,WAAW,EAAE,iCAAiC,EAAE,CAAA;KACjD,CAAA;IACD,KAAK,IAAI,IAAI,CAAA;CACd;AAKD,wBAAgB,4CAA4C,CAC1D,KAAK,EAAE,0BAA0B,EACjC,OAAO,EAAE,oCAAoC,GAC5C,mCAAmC,CA2BrC;AAED,wBAAgB,yCAAyC,CACvD,KAAK,EAAE,0BAA0B,EACjC,OAAO,EAAE,iCAAiC,GACzC,gCAAgC,CAiDlC;AAED,wBAAgB,gCAAgC,CAC9C,QAAQ,EAAE,oCAAoC,GAC7C,0BAA0B,CAwC5B"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"runtime-hooks.test.d.ts","sourceRoot":"","sources":["../../src/belief-state/runtime-hooks.test.ts"],"names":[],"mappings":""}
|
|
@@ -0,0 +1,15 @@
|
|
|
1
|
+
import type { BeliefDecisionPoint, BeliefPolicyAction, BeliefSelectivePolicy, BeliefSelectivePolicyMetrics, BeliefUtilityOptions } from './types';
|
|
2
|
+
export interface EvaluateBeliefSelectivePolicyOptions {
|
|
3
|
+
utility?: BeliefUtilityOptions;
|
|
4
|
+
minN?: number;
|
|
5
|
+
minAccepted?: number;
|
|
6
|
+
minUtilityDelta?: number;
|
|
7
|
+
seed?: number;
|
|
8
|
+
}
|
|
9
|
+
export declare function thresholdSelectivePolicy(options: {
|
|
10
|
+
id?: string;
|
|
11
|
+
confidenceThreshold: number;
|
|
12
|
+
belowThresholdAction?: Exclude<BeliefPolicyAction, 'accept'>;
|
|
13
|
+
}): BeliefSelectivePolicy;
|
|
14
|
+
export declare function evaluateBeliefSelectivePolicy(points: BeliefDecisionPoint[], policy: BeliefSelectivePolicy, options?: EvaluateBeliefSelectivePolicyOptions): BeliefSelectivePolicyMetrics;
|
|
15
|
+
//# sourceMappingURL=selective.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"selective.d.ts","sourceRoot":"","sources":["../../src/belief-state/selective.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EACV,mBAAmB,EACnB,kBAAkB,EAClB,qBAAqB,EACrB,4BAA4B,EAC5B,oBAAoB,EACrB,MAAM,SAAS,CAAA;AAEhB,MAAM,WAAW,oCAAoC;IACnD,OAAO,CAAC,EAAE,oBAAoB,CAAA;IAC9B,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,eAAe,CAAC,EAAE,MAAM,CAAA;IACxB,IAAI,CAAC,EAAE,MAAM,CAAA;CACd;AAaD,wBAAgB,wBAAwB,CAAC,OAAO,EAAE;IAChD,EAAE,CAAC,EAAE,MAAM,CAAA;IACX,mBAAmB,EAAE,MAAM,CAAA;IAC3B,oBAAoB,CAAC,EAAE,OAAO,CAAC,kBAAkB,EAAE,QAAQ,CAAC,CAAA;CAC7D,GAAG,qBAAqB,CAsBxB;AAED,wBAAgB,6BAA6B,CAC3C,MAAM,EAAE,mBAAmB,EAAE,EAC7B,MAAM,EAAE,qBAAqB,EAC7B,OAAO,GAAE,oCAAyC,GACjD,4BAA4B,CA+D9B"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"selective.test.d.ts","sourceRoot":"","sources":["../../src/belief-state/selective.test.ts"],"names":[],"mappings":""}
|
|
@@ -0,0 +1,80 @@
|
|
|
1
|
+
import type { BeliefDecisionKind, BeliefDecisionOutcome, BeliefDecisionPoint, BeliefEvidenceQuality } from './types';
|
|
2
|
+
export interface BeliefShadowProbeInput {
|
|
3
|
+
probeId: string;
|
|
4
|
+
decisionId: string;
|
|
5
|
+
runId: string;
|
|
6
|
+
scenarioId?: string;
|
|
7
|
+
stepIndex: number;
|
|
8
|
+
decisionKind: BeliefDecisionKind;
|
|
9
|
+
candidateActions: string[];
|
|
10
|
+
observedAction?: string;
|
|
11
|
+
evidence: BeliefShadowProbeEvidenceRef[];
|
|
12
|
+
context?: string;
|
|
13
|
+
metadata?: Record<string, unknown>;
|
|
14
|
+
}
|
|
15
|
+
export interface BeliefShadowProbeEvidenceRef {
|
|
16
|
+
id: string;
|
|
17
|
+
source: string;
|
|
18
|
+
detail?: string;
|
|
19
|
+
quality?: BeliefEvidenceQuality;
|
|
20
|
+
}
|
|
21
|
+
export interface BeliefShadowProbeResponse {
|
|
22
|
+
predictedAction: string;
|
|
23
|
+
confidence: number;
|
|
24
|
+
beliefSummary?: string;
|
|
25
|
+
uncertainty?: string[];
|
|
26
|
+
evidenceRefs?: string[];
|
|
27
|
+
wouldChangeMindIf?: string[];
|
|
28
|
+
targetProb?: number;
|
|
29
|
+
qHat?: number | null;
|
|
30
|
+
metadata?: Record<string, unknown>;
|
|
31
|
+
}
|
|
32
|
+
export interface BeliefShadowProbeRecord extends BeliefShadowProbeResponse {
|
|
33
|
+
probeId: string;
|
|
34
|
+
decisionId: string;
|
|
35
|
+
runId: string;
|
|
36
|
+
scenarioId?: string;
|
|
37
|
+
stepIndex: number;
|
|
38
|
+
decisionKind: BeliefDecisionKind;
|
|
39
|
+
candidateActions: string[];
|
|
40
|
+
observedAction: string;
|
|
41
|
+
agreesWithObservedAction: boolean;
|
|
42
|
+
outcome?: BeliefDecisionOutcome;
|
|
43
|
+
}
|
|
44
|
+
export interface BeliefShadowProbeDiagnostic {
|
|
45
|
+
decisionId: string;
|
|
46
|
+
severity: 'warning' | 'error';
|
|
47
|
+
reason: string;
|
|
48
|
+
}
|
|
49
|
+
export interface BeliefShadowProbeSummary {
|
|
50
|
+
attempted: number;
|
|
51
|
+
completed: number;
|
|
52
|
+
dropped: number;
|
|
53
|
+
withOutcome: number;
|
|
54
|
+
withTargetProb: number;
|
|
55
|
+
meanConfidence: number | null;
|
|
56
|
+
observedAgreementRate: number | null;
|
|
57
|
+
}
|
|
58
|
+
export interface BeliefShadowProbeRun {
|
|
59
|
+
probeId: string;
|
|
60
|
+
records: BeliefShadowProbeRecord[];
|
|
61
|
+
diagnostics: BeliefShadowProbeDiagnostic[];
|
|
62
|
+
summary: BeliefShadowProbeSummary;
|
|
63
|
+
}
|
|
64
|
+
export interface RunBeliefShadowProbeOptions {
|
|
65
|
+
probeId: string;
|
|
66
|
+
points: BeliefDecisionPoint[];
|
|
67
|
+
probe: (input: BeliefShadowProbeInput) => BeliefShadowProbeResponse | Promise<BeliefShadowProbeResponse>;
|
|
68
|
+
contextOf?: (point: BeliefDecisionPoint) => string | undefined | Promise<string | undefined>;
|
|
69
|
+
metadataOf?: (point: BeliefDecisionPoint) => Record<string, unknown> | undefined | Promise<Record<string, unknown> | undefined>;
|
|
70
|
+
includeObservedAction?: boolean;
|
|
71
|
+
includeEvidenceDetail?: boolean;
|
|
72
|
+
includeOutcomeInRecord?: boolean;
|
|
73
|
+
requireCandidateActions?: boolean;
|
|
74
|
+
allowOutOfSetActions?: boolean;
|
|
75
|
+
concurrency?: number;
|
|
76
|
+
maxContextChars?: number;
|
|
77
|
+
}
|
|
78
|
+
export declare function runBeliefShadowProbe(options: RunBeliefShadowProbeOptions): Promise<BeliefShadowProbeRun>;
|
|
79
|
+
export declare function formatBeliefShadowProbePrompt(input: BeliefShadowProbeInput): string;
|
|
80
|
+
//# sourceMappingURL=shadow-probe.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"shadow-probe.d.ts","sourceRoot":"","sources":["../../src/belief-state/shadow-probe.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,kBAAkB,EAClB,qBAAqB,EACrB,mBAAmB,EACnB,qBAAqB,EACtB,MAAM,SAAS,CAAA;AAEhB,MAAM,WAAW,sBAAsB;IACrC,OAAO,EAAE,MAAM,CAAA;IACf,UAAU,EAAE,MAAM,CAAA;IAClB,KAAK,EAAE,MAAM,CAAA;IACb,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,SAAS,EAAE,MAAM,CAAA;IACjB,YAAY,EAAE,kBAAkB,CAAA;IAChC,gBAAgB,EAAE,MAAM,EAAE,CAAA;IAC1B,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB,QAAQ,EAAE,4BAA4B,EAAE,CAAA;IACxC,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,4BAA4B;IAC3C,EAAE,EAAE,MAAM,CAAA;IACV,MAAM,EAAE,MAAM,CAAA;IACd,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,OAAO,CAAC,EAAE,qBAAqB,CAAA;CAChC;AAED,MAAM,WAAW,yBAAyB;IACxC,eAAe,EAAE,MAAM,CAAA;IACvB,UAAU,EAAE,MAAM,CAAA;IAClB,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,WAAW,CAAC,EAAE,MAAM,EAAE,CAAA;IACtB,YAAY,CAAC,EAAE,MAAM,EAAE,CAAA;IACvB,iBAAiB,CAAC,EAAE,MAAM,EAAE,CAAA;IAC5B,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,IAAI,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;IACpB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,uBAAwB,SAAQ,yBAAyB;IACxE,OAAO,EAAE,MAAM,CAAA;IACf,UAAU,EAAE,MAAM,CAAA;IAClB,KAAK,EAAE,MAAM,CAAA;IACb,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,SAAS,EAAE,MAAM,CAAA;IACjB,YAAY,EAAE,kBAAkB,CAAA;IAChC,gBAAgB,EAAE,MAAM,EAAE,CAAA;IAC1B,cAAc,EAAE,MAAM,CAAA;IACtB,wBAAwB,EAAE,OAAO,CAAA;IACjC,OAAO,CAAC,EAAE,qBAAqB,CAAA;CAChC;AAED,MAAM,WAAW,2BAA2B;IAC1C,UAAU,EAAE,MAAM,CAAA;IAClB,QAAQ,EAAE,SAAS,GAAG,OAAO,CAAA;IAC7B,MAAM,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,wBAAwB;IACvC,SAAS,EAAE,MAAM,CAAA;IACjB,SAAS,EAAE,MAAM,CAAA;IACjB,OAAO,EAAE,MAAM,CAAA;IACf,WAAW,EAAE,MAAM,CAAA;IACnB,cAAc,EAAE,MAAM,CAAA;IACtB,cAAc,EAAE,MAAM,GAAG,IAAI,CAAA;IAC7B,qBAAqB,EAAE,MAAM,GAAG,IAAI,CAAA;CACrC;AAED,MAAM,WAAW,oBAAoB;IACnC,OAAO,EAAE,MAAM,CAAA;IACf,OAAO,EAAE,uBAAuB,EAAE,CAAA;IAClC,WAAW,EAAE,2BAA2B,EAAE,CAAA;IAC1C,OAAO,EAAE,wBAAwB,CAAA;CAClC;AAED,MAAM,WAAW,2BAA2B;IAC1C,OAAO,EAAE,MAAM,CAAA;IACf,MAAM,EAAE,mBAAmB,EAAE,CAAA;IAC7B,KAAK,EAAE,CACL,KAAK,EAAE,sBAAsB,KAC1B,yBAAyB,GAAG,OAAO,CAAC,yBAAyB,CAAC,CAAA;IACnE,SAAS,CAAC,EAAE,CAAC,KAAK,EAAE,mBAAmB,KAAK,MAAM,GAAG,SAAS,GAAG,OAAO,CAAC,MAAM,GAAG,SAAS,CAAC,CAAA;IAC5F,UAAU,CAAC,EAAE,CACX,KAAK,EAAE,mBAAmB,KACvB,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,SAAS,GAAG,OAAO,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,SAAS,CAAC,CAAA;IACvF,qBAAqB,CAAC,EAAE,OAAO,CAAA;IAC/B,qBAAqB,CAAC,EAAE,OAAO,CAAA;IAC/B,sBAAsB,CAAC,EAAE,OAAO,CAAA;IAChC,uBAAuB,CAAC,EAAE,OAAO,CAAA;IACjC,oBAAoB,CAAC,EAAE,OAAO,CAAA;IAC9B,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,eAAe,CAAC,EAAE,MAAM,CAAA;CACzB;AAKD,wBAAsB,oBAAoB,CACxC,OAAO,EAAE,2BAA2B,GACnC,OAAO,CAAC,oBAAoB,CAAC,CA2B/B;AAED,wBAAgB,6BAA6B,CAAC,KAAK,EAAE,sBAAsB,GAAG,MAAM,CAwBnF"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"shadow-probe.test.d.ts","sourceRoot":"","sources":["../../src/belief-state/shadow-probe.test.ts"],"names":[],"mappings":""}
|
|
@@ -0,0 +1,195 @@
|
|
|
1
|
+
import type { CalibrationReport } from '../meta-eval/calibration';
|
|
2
|
+
import type { OffPolicyEstimate } from '../rl/off-policy';
|
|
3
|
+
export declare const BELIEF_DECISION_KINDS: readonly ["continue", "verify", "ask", "retry", "stop", "memory-write", "memory-read", "tool-select", "skill-select", "workflow-select", "surface-promote"];
|
|
4
|
+
export type BeliefDecisionKind = (typeof BELIEF_DECISION_KINDS)[number];
|
|
5
|
+
export declare const BELIEF_EVIDENCE_SOURCES: readonly ["run", "span", "event", "finding", "memory", "knowledge", "policy"];
|
|
6
|
+
export type BeliefEvidenceSource = (typeof BELIEF_EVIDENCE_SOURCES)[number];
|
|
7
|
+
export declare const BELIEF_EVIDENCE_QUALITIES: readonly ["direct", "derived", "self-reported", "unverified", "stale", "contradicted"];
|
|
8
|
+
export type BeliefEvidenceQuality = (typeof BELIEF_EVIDENCE_QUALITIES)[number];
|
|
9
|
+
export declare const BELIEF_EVALUATION_CRITERIA: readonly [{
|
|
10
|
+
readonly id: "capture-integrity";
|
|
11
|
+
readonly label: "Capture integrity";
|
|
12
|
+
readonly reasonCodes: readonly ["trace-missing", "run-record-missing", "backend-integrity-missing"];
|
|
13
|
+
}, {
|
|
14
|
+
readonly id: "decision-completeness";
|
|
15
|
+
readonly label: "Decision completeness";
|
|
16
|
+
readonly reasonCodes: readonly ["candidate-actions-missing", "chosen-action-missing", "decision-evidence-missing"];
|
|
17
|
+
}, {
|
|
18
|
+
readonly id: "evidence-quality";
|
|
19
|
+
readonly label: "Evidence quality";
|
|
20
|
+
readonly reasonCodes: readonly ["evidence-stale", "evidence-contradictory", "evidence-unverified", "evidence-self-reported"];
|
|
21
|
+
}, {
|
|
22
|
+
readonly id: "outcome-quality";
|
|
23
|
+
readonly label: "Outcome quality";
|
|
24
|
+
readonly reasonCodes: readonly ["outcome-missing", "outcome-delayed", "cost-missing"];
|
|
25
|
+
}, {
|
|
26
|
+
readonly id: "calibration";
|
|
27
|
+
readonly label: "Calibration";
|
|
28
|
+
readonly reasonCodes: readonly ["confidence-missing", "calibration-unsupported", "calibration-gap-high"];
|
|
29
|
+
}, {
|
|
30
|
+
readonly id: "accepted-region-risk";
|
|
31
|
+
readonly label: "Accepted-region risk";
|
|
32
|
+
readonly reasonCodes: readonly ["accepted-error-high", "coverage-too-low"];
|
|
33
|
+
}, {
|
|
34
|
+
readonly id: "policy-value";
|
|
35
|
+
readonly label: "Policy value";
|
|
36
|
+
readonly reasonCodes: readonly ["utility-lift-missing", "baseline-dominates", "cost-too-high"];
|
|
37
|
+
}, {
|
|
38
|
+
readonly id: "ope-support";
|
|
39
|
+
readonly label: "OPE support";
|
|
40
|
+
readonly reasonCodes: readonly ["behavior-propensity-missing", "behavior-propensity-invalid", "target-propensity-missing", "target-propensity-invalid", "effective-sample-size-low", "importance-weight-high"];
|
|
41
|
+
}, {
|
|
42
|
+
readonly id: "memory-health";
|
|
43
|
+
readonly label: "Memory health";
|
|
44
|
+
readonly reasonCodes: readonly ["memory-stale", "memory-poisoning-risk", "context-bloat", "memory-write-unverified"];
|
|
45
|
+
}, {
|
|
46
|
+
readonly id: "surface-attribution";
|
|
47
|
+
readonly label: "Surface attribution";
|
|
48
|
+
readonly reasonCodes: readonly ["surface-claim-unsupported", "causal-attribution-missing"];
|
|
49
|
+
}, {
|
|
50
|
+
readonly id: "generalization";
|
|
51
|
+
readonly label: "Generalization";
|
|
52
|
+
readonly reasonCodes: readonly ["split-missing", "holdout-regression", "task-family-coverage-low", "leakage-risk"];
|
|
53
|
+
}, {
|
|
54
|
+
readonly id: "promotion";
|
|
55
|
+
readonly label: "Promotion";
|
|
56
|
+
readonly reasonCodes: readonly ["negative-control-failed", "promotion-gate-failed", "human-review-required"];
|
|
57
|
+
}];
|
|
58
|
+
export type BeliefEvaluationCriterionId = (typeof BELIEF_EVALUATION_CRITERIA)[number]['id'];
|
|
59
|
+
export type BeliefDecisionReasonCode = (typeof BELIEF_EVALUATION_CRITERIA)[number]['reasonCodes'][number];
|
|
60
|
+
export interface BeliefDecisionReason {
|
|
61
|
+
code: BeliefDecisionReasonCode;
|
|
62
|
+
criterion?: BeliefEvaluationCriterionId;
|
|
63
|
+
detail?: string;
|
|
64
|
+
evidenceIds?: string[];
|
|
65
|
+
metadata?: Record<string, unknown>;
|
|
66
|
+
}
|
|
67
|
+
export declare function isBeliefDecisionKind(value: unknown): value is BeliefDecisionKind;
|
|
68
|
+
export declare function isBeliefEvidenceSource(value: unknown): value is BeliefEvidenceSource;
|
|
69
|
+
export interface BeliefEvidenceRef {
|
|
70
|
+
source: BeliefEvidenceSource;
|
|
71
|
+
id: string;
|
|
72
|
+
runId?: string;
|
|
73
|
+
spanId?: string;
|
|
74
|
+
eventId?: string;
|
|
75
|
+
detail?: string;
|
|
76
|
+
quality?: BeliefEvidenceQuality;
|
|
77
|
+
observedAt?: string;
|
|
78
|
+
metadata?: Record<string, unknown>;
|
|
79
|
+
}
|
|
80
|
+
export interface BeliefDecisionOutcome {
|
|
81
|
+
success?: boolean;
|
|
82
|
+
score?: number;
|
|
83
|
+
reward?: number;
|
|
84
|
+
costUsd?: number;
|
|
85
|
+
observedAt?: string;
|
|
86
|
+
metadata?: Record<string, unknown>;
|
|
87
|
+
}
|
|
88
|
+
export interface BeliefDecisionPoint {
|
|
89
|
+
id: string;
|
|
90
|
+
runId: string;
|
|
91
|
+
scenarioId?: string;
|
|
92
|
+
stepIndex: number;
|
|
93
|
+
kind: BeliefDecisionKind;
|
|
94
|
+
chosenAction: string;
|
|
95
|
+
candidateActions?: string[];
|
|
96
|
+
confidence?: number;
|
|
97
|
+
behaviorProb?: number;
|
|
98
|
+
targetProb?: number;
|
|
99
|
+
qHat?: number | null;
|
|
100
|
+
costUsd?: number;
|
|
101
|
+
evidence: BeliefEvidenceRef[];
|
|
102
|
+
outcome?: BeliefDecisionOutcome;
|
|
103
|
+
reasons?: BeliefDecisionReason[];
|
|
104
|
+
metadata?: Record<string, unknown>;
|
|
105
|
+
}
|
|
106
|
+
export interface BeliefDecisionExtractionDiagnostic {
|
|
107
|
+
runId: string;
|
|
108
|
+
eventId?: string;
|
|
109
|
+
severity: 'info' | 'warning' | 'error';
|
|
110
|
+
reason: string;
|
|
111
|
+
}
|
|
112
|
+
export interface BeliefDecisionExtractionReport {
|
|
113
|
+
decisions: BeliefDecisionPoint[];
|
|
114
|
+
diagnostics: BeliefDecisionExtractionDiagnostic[];
|
|
115
|
+
}
|
|
116
|
+
export type BeliefPolicyAction = 'accept' | 'defer' | 'verify' | 'ask' | 'retry' | 'stop';
|
|
117
|
+
export interface BeliefPolicyDecision {
|
|
118
|
+
action: BeliefPolicyAction;
|
|
119
|
+
confidence?: number;
|
|
120
|
+
targetProb?: number;
|
|
121
|
+
qHat?: number | null;
|
|
122
|
+
reason?: string;
|
|
123
|
+
reasons?: BeliefDecisionReason[];
|
|
124
|
+
}
|
|
125
|
+
export interface BeliefSelectivePolicy {
|
|
126
|
+
id: string;
|
|
127
|
+
decide(point: BeliefDecisionPoint): BeliefPolicyDecision;
|
|
128
|
+
}
|
|
129
|
+
export interface BeliefOpeTargetPolicy {
|
|
130
|
+
id: string;
|
|
131
|
+
targetProbOf(point: BeliefDecisionPoint): number | null | undefined;
|
|
132
|
+
qHatOf?(point: BeliefDecisionPoint): number | null | undefined;
|
|
133
|
+
}
|
|
134
|
+
export interface BeliefUtilityOptions {
|
|
135
|
+
successUtility?: number;
|
|
136
|
+
failureUtility?: number;
|
|
137
|
+
deferUtility?: number;
|
|
138
|
+
verifyCost?: number;
|
|
139
|
+
askCost?: number;
|
|
140
|
+
retryCost?: number;
|
|
141
|
+
stopUtility?: number;
|
|
142
|
+
costWeight?: number;
|
|
143
|
+
}
|
|
144
|
+
export interface BeliefSelectivePolicyMetrics {
|
|
145
|
+
policyId: string;
|
|
146
|
+
n: number;
|
|
147
|
+
accepted: number;
|
|
148
|
+
rejected: number;
|
|
149
|
+
coverage: number;
|
|
150
|
+
acceptedErrorRate: number;
|
|
151
|
+
baselineUtility: number;
|
|
152
|
+
policyUtility: number;
|
|
153
|
+
utilityDelta: number;
|
|
154
|
+
utilityCi95: {
|
|
155
|
+
mean: number;
|
|
156
|
+
lower: number;
|
|
157
|
+
upper: number;
|
|
158
|
+
};
|
|
159
|
+
rejectedMeanReward: number | null;
|
|
160
|
+
recommendation: 'ship' | 'hold' | 'need_more_data';
|
|
161
|
+
reasons: string[];
|
|
162
|
+
}
|
|
163
|
+
export interface BeliefOpeSupportDiagnostics {
|
|
164
|
+
supported: boolean;
|
|
165
|
+
n: number;
|
|
166
|
+
dropped: number;
|
|
167
|
+
effectiveSampleSize: number;
|
|
168
|
+
effectiveSampleRatio: number;
|
|
169
|
+
maxImportanceWeight: number;
|
|
170
|
+
reasons: string[];
|
|
171
|
+
}
|
|
172
|
+
export interface BeliefOpeReport {
|
|
173
|
+
targetPolicyId: string;
|
|
174
|
+
ips: OffPolicyEstimate;
|
|
175
|
+
snips: OffPolicyEstimate;
|
|
176
|
+
dr: OffPolicyEstimate;
|
|
177
|
+
support: BeliefOpeSupportDiagnostics;
|
|
178
|
+
}
|
|
179
|
+
export type BeliefEvaluationStatus = 'ship' | 'hold' | 'need_more_data';
|
|
180
|
+
export type BeliefCalibrationStatus = 'supported' | 'unsupported';
|
|
181
|
+
export type BeliefOpeStatus = 'supported' | 'unsupported' | 'not_requested';
|
|
182
|
+
export interface BeliefPolicyEvaluationReport {
|
|
183
|
+
policyId: string;
|
|
184
|
+
n: number;
|
|
185
|
+
status: BeliefEvaluationStatus;
|
|
186
|
+
selectiveStatus: BeliefEvaluationStatus;
|
|
187
|
+
calibrationStatus: BeliefCalibrationStatus;
|
|
188
|
+
opeStatus: BeliefOpeStatus;
|
|
189
|
+
opeTargetPolicyId?: string;
|
|
190
|
+
selective: BeliefSelectivePolicyMetrics;
|
|
191
|
+
calibration?: CalibrationReport;
|
|
192
|
+
ope?: BeliefOpeReport;
|
|
193
|
+
diagnostics: string[];
|
|
194
|
+
}
|
|
195
|
+
//# sourceMappingURL=types.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../../src/belief-state/types.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,0BAA0B,CAAA;AACjE,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,kBAAkB,CAAA;AAEzD,eAAO,MAAM,qBAAqB,6JAYxB,CAAA;AAEV,MAAM,MAAM,kBAAkB,GAAG,CAAC,OAAO,qBAAqB,CAAC,CAAC,MAAM,CAAC,CAAA;AAEvE,eAAO,MAAM,uBAAuB,+EAQ1B,CAAA;AAEV,MAAM,MAAM,oBAAoB,GAAG,CAAC,OAAO,uBAAuB,CAAC,CAAC,MAAM,CAAC,CAAA;AAE3E,eAAO,MAAM,yBAAyB,wFAO5B,CAAA;AAEV,MAAM,MAAM,qBAAqB,GAAG,CAAC,OAAO,yBAAyB,CAAC,CAAC,MAAM,CAAC,CAAA;AAE9E,eAAO,MAAM,0BAA0B;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EAuF7B,CAAA;AAEV,MAAM,MAAM,2BAA2B,GAAG,CAAC,OAAO,0BAA0B,CAAC,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,CAAA;AAC3F,MAAM,MAAM,wBAAwB,GAClC,CAAC,OAAO,0BAA0B,CAAC,CAAC,MAAM,CAAC,CAAC,aAAa,CAAC,CAAC,MAAM,CAAC,CAAA;AAEpE,MAAM,WAAW,oBAAoB;IACnC,IAAI,EAAE,wBAAwB,CAAA;IAC9B,SAAS,CAAC,EAAE,2BAA2B,CAAA;IACvC,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,WAAW,CAAC,EAAE,MAAM,EAAE,CAAA;IACtB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,wBAAgB,oBAAoB,CAAC,KAAK,EAAE,OAAO,GAAG,KAAK,IAAI,kBAAkB,CAEhF;AAED,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,OAAO,GAAG,KAAK,IAAI,oBAAoB,CAIpF;AAED,MAAM,WAAW,iBAAiB;IAChC,MAAM,EAAE,oBAAoB,CAAA;IAC5B,EAAE,EAAE,MAAM,CAAA;IACV,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,OAAO,CAAC,EAAE,qBAAqB,CAAA;IAC/B,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,qBAAqB;IACpC,OAAO,CAAC,EAAE,OAAO,CAAA;IACjB,KAAK,CAAC,EAAE,MAAM,CAAA;IACd,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,mBAAmB;IAClC,EAAE,EAAE,MAAM,CAAA;IACV,KAAK,EAAE,MAAM,CAAA;IACb,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,SAAS,EAAE,MAAM,CAAA;IACjB,IAAI,EAAE,kBAAkB,CAAA;IACxB,YAAY,EAAE,MAAM,CAAA;IACpB,gBAAgB,CAAC,EAAE,MAAM,EAAE,CAAA;IAC3B,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,IAAI,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;IACpB,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,QAAQ,EAAE,iBAAiB,EAAE,CAAA;IAC7B,OAAO,CAAC,EAAE,qBAAqB,CAAA;IAC/B,OAAO,CAAC,EAAE,oBAAoB,EAAE,CAAA;IAChC,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACnC;AAED,MAAM,WAAW,kCAAkC;IACjD,KAAK,EAAE,MAAM,CAAA;IACb,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,QAAQ,EAAE,MAAM,GAAG,SAAS,GAAG,OAAO,CAAA;IACtC,MAAM,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,8BAA8B;IAC7C,SAAS,EAAE,mBAAmB,EAAE,CAAA;IAChC,WAAW,EAAE,kCAAkC,EAAE,CAAA;CAClD;AAED,MAAM,MAAM,kBAAkB,GAAG,QAAQ,GAAG,OAAO,GAAG,QAAQ,GAAG,KAAK,GAAG,OAAO,GAAG,MAAM,CAAA;AAEzF,MAAM,WAAW,oBAAoB;IACnC,MAAM,EAAE,kBAAkB,CAAA;IAC1B,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,IAAI,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;IACpB,MAAM,CAAC,EAAE,MAAM,CAAA;IACf,OAAO,CAAC,EAAE,oBAAoB,EAAE,CAAA;CACjC;AAED,MAAM,WAAW,qBAAqB;IACpC,EAAE,EAAE,MAAM,CAAA;IACV,MAAM,CAAC,KAAK,EAAE,mBAAmB,GAAG,oBAAoB,CAAA;CACzD;AAED,MAAM,WAAW,qBAAqB;IACpC,EAAE,EAAE,MAAM,CAAA;IACV,YAAY,CAAC,KAAK,EAAE,mBAAmB,GAAG,MAAM,GAAG,IAAI,GAAG,SAAS,CAAA;IACnE,MAAM,CAAC,CAAC,KAAK,EAAE,mBAAmB,GAAG,MAAM,GAAG,IAAI,GAAG,SAAS,CAAA;CAC/D;AAED,MAAM,WAAW,oBAAoB;IACnC,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB,cAAc,CAAC,EAAE,MAAM,CAAA;IACvB,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB,UAAU,CAAC,EAAE,MAAM,CAAA;IACnB,OAAO,CAAC,EAAE,MAAM,CAAA;IAChB,SAAS,CAAC,EAAE,MAAM,CAAA;IAClB,WAAW,CAAC,EAAE,MAAM,CAAA;IACpB,UAAU,CAAC,EAAE,MAAM,CAAA;CACpB;AAED,MAAM,WAAW,4BAA4B;IAC3C,QAAQ,EAAE,MAAM,CAAA;IAChB,CAAC,EAAE,MAAM,CAAA;IACT,QAAQ,EAAE,MAAM,CAAA;IAChB,QAAQ,EAAE,MAAM,CAAA;IAChB,QAAQ,EAAE,MAAM,CAAA;IAChB,iBAAiB,EAAE,MAAM,CAAA;IACzB,eAAe,EAAE,MAAM,CAAA;IACvB,aAAa,EAAE,MAAM,CAAA;IACrB,YAAY,EAAE,MAAM,CAAA;IACpB,WAAW,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAA;IAC3D,kBAAkB,EAAE,MAAM,GAAG,IAAI,CAAA;IACjC,cAAc,EAAE,MAAM,GAAG,MAAM,GAAG,gBAAgB,CAAA;IAClD,OAAO,EAAE,MAAM,EAAE,CAAA;CAClB;AAED,MAAM,WAAW,2BAA2B;IAC1C,SAAS,EAAE,OAAO,CAAA;IAClB,CAAC,EAAE,MAAM,CAAA;IACT,OAAO,EAAE,MAAM,CAAA;IACf,mBAAmB,EAAE,MAAM,CAAA;IAC3B,oBAAoB,EAAE,MAAM,CAAA;IAC5B,mBAAmB,EAAE,MAAM,CAAA;IAC3B,OAAO,EAAE,MAAM,EAAE,CAAA;CAClB;AAED,MAAM,WAAW,eAAe;IAC9B,cAAc,EAAE,MAAM,CAAA;IACtB,GAAG,EAAE,iBAAiB,CAAA;IACtB,KAAK,EAAE,iBAAiB,CAAA;IACxB,EAAE,EAAE,iBAAiB,CAAA;IACrB,OAAO,EAAE,2BAA2B,CAAA;CACrC;AAED,MAAM,MAAM,sBAAsB,GAAG,MAAM,GAAG,MAAM,GAAG,gBAAgB,CAAA;AACvE,MAAM,MAAM,uBAAuB,GAAG,WAAW,GAAG,aAAa,CAAA;AACjE,MAAM,MAAM,eAAe,GAAG,WAAW,GAAG,aAAa,GAAG,eAAe,CAAA;AAE3E,MAAM,WAAW,4BAA4B;IAC3C,QAAQ,EAAE,MAAM,CAAA;IAChB,CAAC,EAAE,MAAM,CAAA;IACT,MAAM,EAAE,sBAAsB,CAAA;IAC9B,eAAe,EAAE,sBAAsB,CAAA;IACvC,iBAAiB,EAAE,uBAAuB,CAAA;IAC1C,SAAS,EAAE,eAAe,CAAA;IAC1B,iBAAiB,CAAC,EAAE,MAAM,CAAA;IAC1B,SAAS,EAAE,4BAA4B,CAAA;IACvC,WAAW,CAAC,EAAE,iBAAiB,CAAA;IAC/B,GAAG,CAAC,EAAE,eAAe,CAAA;IACrB,WAAW,EAAE,MAAM,EAAE,CAAA;CACtB"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"types.test.d.ts","sourceRoot":"","sources":["../../src/belief-state/types.test.ts"],"names":[],"mappings":""}
|
|
@@ -0,0 +1,14 @@
|
|
|
1
|
+
import type { TCloud } from '@tangle-network/tcloud';
|
|
2
|
+
import type { BenchmarkReport, BenchmarkRunnerConfig, Scenario } from './types';
|
|
3
|
+
/**
|
|
4
|
+
* BenchmarkRunner — orchestrates scenarios, executor, judges, and scoring.
|
|
5
|
+
*
|
|
6
|
+
* Domain-agnostic. Each agent provides its own scenarios, judges, and system prompt.
|
|
7
|
+
*/
|
|
8
|
+
export declare class BenchmarkRunner {
|
|
9
|
+
private tc;
|
|
10
|
+
private config;
|
|
11
|
+
constructor(tc: TCloud, config: BenchmarkRunnerConfig);
|
|
12
|
+
run(scenarios?: Scenario[]): Promise<BenchmarkReport>;
|
|
13
|
+
}
|
|
14
|
+
//# sourceMappingURL=benchmark.d.ts.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"benchmark.d.ts","sourceRoot":"","sources":["../src/benchmark.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,wBAAwB,CAAA;AAEpD,OAAO,KAAK,EAAE,eAAe,EAAE,qBAAqB,EAAE,QAAQ,EAAkB,MAAM,SAAS,CAAA;AAE/F;;;;GAIG;AACH,qBAAa,eAAe;IAC1B,OAAO,CAAC,EAAE,CAAQ;IAClB,OAAO,CAAC,MAAM,CAAuB;gBAEzB,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,qBAAqB;IAK/C,GAAG,CAAC,SAAS,CAAC,EAAE,QAAQ,EAAE,GAAG,OAAO,CAAC,eAAe,CAAC;CAgK5D"}
|