spawnfile 0.1.25 → 0.1.27
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/README.md +1 -1
- package/dist/audit/readAudit.d.ts +2 -2
- package/dist/auth/trainingAuth.js +1 -1
- package/dist/cli/runCli.d.ts +2 -2
- package/dist/cli/runCli.js +2 -2
- package/dist/cli/trainCommand.js +4 -5
- package/dist/cli/{paideiaDelegation.d.ts → trainingExecution.d.ts} +3 -4
- package/dist/cli/trainingExecution.js +151 -0
- package/dist/compiler/containerDaimonGrokWorkerRender.d.ts +2 -2
- package/dist/compiler/containerDaimonGrokWorkerRender.js +3 -2
- package/dist/compiler/containerDaimonUidEntrypointRender.js +4 -3
- package/dist/compiler/containerPackageOverrides.d.ts +1 -1
- package/dist/compiler/training/broker/declaration.d.ts +2 -2
- package/dist/compiler/training/broker/declaration.js +3 -2
- package/dist/compiler/training/broker/entrypoint.d.ts +3 -3
- package/dist/compiler/training/broker/entrypoint.js +7 -6
- package/dist/compiler/training/broker/evaluatorSecrets.js +4 -2
- package/dist/compiler/training/broker/paths.d.ts +15 -15
- package/dist/compiler/training/broker/paths.js +23 -23
- package/dist/compiler/training/broker/projection.d.ts +4 -4
- package/dist/compiler/training/broker/projection.js +9 -7
- package/dist/compiler/training/broker/provisioning.d.ts +1 -1
- package/dist/compiler/training/broker/provisioning.js +3 -3
- package/dist/compiler/training/broker/registration.d.ts +1 -1
- package/dist/compiler/training/broker/registration.js +1 -1
- package/dist/compiler/training/broker/supervisor.d.ts +1 -1
- package/dist/compiler/training/container/contract.js +2 -1
- package/dist/compiler/training/container/evaluators.d.ts +3 -3
- package/dist/compiler/training/container/evaluators.js +6 -6
- package/dist/compiler/training/container/launch.js +15 -7
- package/dist/compiler/training/container/security.js +3 -3
- package/dist/compiler/training/container/subjectIsolation.js +2 -1
- package/dist/compiler/training/execution.d.ts +8 -10
- package/dist/compiler/training/execution.js +1 -1
- package/dist/compiler/training/moments.js +2 -2
- package/dist/compiler/training/plan/contract.d.ts +7 -7
- package/dist/compiler/training/plan/load.js +2 -0
- package/dist/compiler/training/preparation/assets/Dockerfile +8 -17
- package/dist/compiler/training/preparation/assets/package-lock.json +2012 -316
- package/dist/compiler/training/preparation/broker.js +2 -1
- package/dist/compiler/training/preparation/contract.d.ts +5 -9
- package/dist/compiler/training/preparation/contract.js +4 -3
- package/dist/compiler/training/preparation/image.js +2 -4
- package/dist/compiler/training/preparation/prepare.js +3 -3
- package/dist/compiler/training/repair/compatibility.js +17 -6
- package/dist/compiler/training/repair/contract.d.ts +14 -22
- package/dist/compiler/training/repair/contract.js +3 -2
- package/dist/compiler/training/repair/prepare.d.ts +5 -9
- package/dist/compiler/training/repair/prepare.js +10 -4
- package/dist/compiler/training/repair/witness.d.ts +4 -8
- package/dist/compiler/training/repair/witness.js +16 -2
- package/dist/compiler/training/rewind/contract.d.ts +9 -9
- package/dist/compiler/training/subjectDeclaration.d.ts +2 -1
- package/dist/compiler/training/subjectDeclaration.js +4 -2
- package/dist/compiler/training/subjectLease.d.ts +9 -8
- package/dist/compiler/training/subjectLease.js +17 -13
- package/dist/distribution/imageRuntimeAuth.js +33 -24
- package/dist/manifest/executionSchemas.d.ts +9 -9
- package/dist/manifest/mcpSchemas.d.ts +3 -3
- package/dist/manifest/schemas.d.ts +71 -71
- package/dist/manifest/teamNetworkAccessSchemas.d.ts +1 -1
- package/dist/manifest/teamNetworkAuthSchemas.d.ts +1 -1
- package/dist/manifest/teamNetworkSchemas.d.ts +2 -2
- package/dist/manifest/teamNetworkServerSchemas.d.ts +2 -2
- package/dist/manifest/trainingDataSchemas.d.ts +2 -2
- package/dist/runtime/codex-exec/training/adapter.d.ts +13 -0
- package/dist/runtime/codex-exec/training/adapter.js +101 -0
- package/dist/runtime/codex-exec/training/adapter.js.map +1 -0
- package/dist/runtime/codex-exec/training/environment.d.ts +3 -0
- package/dist/runtime/codex-exec/training/environment.js +26 -0
- package/dist/runtime/codex-exec/training/environment.js.map +1 -0
- package/dist/runtime/codex-exec/training/index.d.ts +2 -0
- package/dist/runtime/codex-exec/training/index.js +3 -0
- package/dist/runtime/codex-exec/training/index.js.map +1 -0
- package/dist/runtime/codex-exec/training/protocol.d.ts +15 -0
- package/dist/runtime/codex-exec/training/protocol.js +102 -0
- package/dist/runtime/codex-exec/training/protocol.js.map +1 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/index.d.ts +1 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/index.js +2 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/index.js.map +1 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/protocol.d.ts +17 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/protocol.js +80 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/protocol.js.map +1 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/provider.d.ts +12 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/provider.js +125 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/provider.js.map +1 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/strict-schema.d.ts +6 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/strict-schema.js +219 -0
- package/dist/runtime/codex-exec/training/providers/codex-subscription/strict-schema.js.map +1 -0
- package/dist/runtime/container.js +1 -1
- package/dist/runtime/containerPackageOverrides.d.ts +2 -2
- package/dist/runtime/containerPackageOverrides.js +1 -1
- package/dist/runtime/daimon/adapter.js +9 -9
- package/dist/runtime/daimon/config.d.ts +2 -1
- package/dist/runtime/daimon/config.js +25 -25
- package/dist/runtime/daimon/contract-manifest.json +1 -1
- package/dist/runtime/daimon/contract-manifest.sha256 +1 -1
- package/dist/runtime/daimon/contractManifest.d.ts +206 -202
- package/dist/runtime/daimon/contractManifest.js +48 -145
- package/dist/runtime/daimon/contractManifest.js.map +1 -0
- package/dist/runtime/daimon/contractManifestSchema.d.ts +1455 -0
- package/dist/runtime/daimon/contractManifestSchema.js +19 -0
- package/dist/runtime/daimon/contractManifestSchema.js.map +1 -0
- package/dist/runtime/daimon/engineDescriptorSchema.d.ts +1826 -0
- package/dist/runtime/daimon/engineDescriptorSchema.js +107 -0
- package/dist/runtime/daimon/engineDescriptorSchema.js.map +1 -0
- package/dist/runtime/daimon/engineDescriptors.generated.d.ts +477 -0
- package/dist/runtime/daimon/engineDescriptors.generated.js +5 -0
- package/dist/runtime/daimon/engineDescriptors.generated.js.map +1 -0
- package/dist/runtime/daimon/engines.d.ts +664 -0
- package/dist/runtime/daimon/engines.js +25 -0
- package/dist/runtime/daimon/engines.js.map +1 -0
- package/dist/runtime/daimon/grokModel.d.ts +2 -1
- package/dist/runtime/daimon/grokModel.js +10 -7
- package/dist/runtime/daimon/grokSkillsSupport.js +2 -1
- package/dist/runtime/daimon/grokTurnLimits.js +3 -3
- package/dist/runtime/daimon/grokWorkerContract.d.ts +2 -2
- package/dist/runtime/daimon/grokWorkerContract.js +10 -9
- package/dist/runtime/daimon/index.d.ts +4 -0
- package/dist/runtime/daimon/index.js +5 -0
- package/dist/runtime/daimon/index.js.map +1 -0
- package/dist/runtime/daimon/packageName.d.ts +2 -0
- package/dist/runtime/daimon/packageName.js +3 -0
- package/dist/runtime/daimon/packageName.js.map +1 -0
- package/dist/runtime/daimon/runAuth.d.ts +4 -4
- package/dist/runtime/daimon/runAuth.js +71 -53
- package/dist/runtime/daimon/training/index.d.ts +2 -0
- package/dist/runtime/daimon/training/index.js +7 -0
- package/dist/runtime/daimon/training/index.js.map +1 -0
- package/dist/runtime/daimon/training/legacy/adapter.d.ts +11 -0
- package/dist/runtime/daimon/training/legacy/adapter.js +175 -0
- package/dist/runtime/daimon/training/legacy/adapter.js.map +1 -0
- package/dist/runtime/daimon/training/legacy/index.d.ts +2 -0
- package/dist/runtime/daimon/training/legacy/index.js +3 -0
- package/dist/runtime/daimon/training/legacy/index.js.map +1 -0
- package/dist/runtime/daimon/training/legacy/public.d.ts +49 -0
- package/dist/runtime/daimon/training/legacy/public.js +16 -0
- package/dist/runtime/daimon/training/legacy/public.js.map +1 -0
- package/dist/runtime/daimon/training/native/adapter.d.ts +18 -0
- package/dist/runtime/daimon/training/native/adapter.js +231 -0
- package/dist/runtime/daimon/training/native/adapter.js.map +1 -0
- package/dist/runtime/daimon/training/native/admission.d.ts +7 -0
- package/dist/runtime/daimon/training/native/admission.js +52 -0
- package/dist/runtime/daimon/training/native/admission.js.map +1 -0
- package/dist/runtime/daimon/training/native/child/arguments.d.ts +16 -0
- package/dist/runtime/daimon/training/native/child/arguments.js +69 -0
- package/dist/runtime/daimon/training/native/child/arguments.js.map +1 -0
- package/dist/runtime/daimon/training/native/child/deferred-trial.d.ts +24 -0
- package/dist/runtime/daimon/training/native/child/deferred-trial.js +63 -0
- package/dist/runtime/daimon/training/native/child/deferred-trial.js.map +1 -0
- package/dist/runtime/daimon/training/native/child/diagnostics.d.ts +58 -0
- package/dist/runtime/daimon/training/native/child/diagnostics.js +215 -0
- package/dist/runtime/daimon/training/native/child/diagnostics.js.map +1 -0
- package/dist/runtime/daimon/training/native/child/fixture.d.ts +16 -0
- package/dist/runtime/daimon/training/native/child/fixture.js +157 -0
- package/dist/runtime/daimon/training/native/child/fixture.js.map +1 -0
- package/dist/runtime/daimon/training/native/child/grok.d.ts +14 -0
- package/dist/runtime/daimon/training/native/child/grok.js +47 -0
- package/dist/runtime/daimon/training/native/child/grok.js.map +1 -0
- package/dist/runtime/daimon/training/native/child/index.d.ts +7 -0
- package/dist/runtime/daimon/training/native/child/index.js +5 -0
- package/dist/runtime/daimon/training/native/child/index.js.map +1 -0
- package/dist/runtime/daimon/training/native/child/lifecycle.d.ts +40 -0
- package/dist/runtime/daimon/training/native/child/lifecycle.js +391 -0
- package/dist/runtime/daimon/training/native/child/lifecycle.js.map +1 -0
- package/dist/runtime/daimon/training/native/child/probe.d.ts +2 -0
- package/dist/runtime/daimon/training/native/child/probe.js +11 -0
- package/dist/runtime/daimon/training/native/child/probe.js.map +1 -0
- package/dist/runtime/daimon/training/native/child/sandbox.d.ts +4 -0
- package/dist/runtime/daimon/training/native/child/sandbox.js +76 -0
- package/dist/runtime/daimon/training/native/child/sandbox.js.map +1 -0
- package/dist/runtime/daimon/training/native/child/types.d.ts +15 -0
- package/dist/runtime/daimon/training/native/child/types.js +2 -0
- package/dist/runtime/daimon/training/native/child/types.js.map +1 -0
- package/dist/runtime/daimon/training/native/container/docker.d.ts +29 -0
- package/dist/runtime/daimon/training/native/container/docker.js +105 -0
- package/dist/runtime/daimon/training/native/container/docker.js.map +1 -0
- package/dist/runtime/daimon/training/native/container/grok-slot.d.ts +22 -0
- package/dist/runtime/daimon/training/native/container/grok-slot.js +38 -0
- package/dist/runtime/daimon/training/native/container/grok-slot.js.map +1 -0
- package/dist/runtime/daimon/training/native/container/index.d.ts +6 -0
- package/dist/runtime/daimon/training/native/container/index.js +6 -0
- package/dist/runtime/daimon/training/native/container/index.js.map +1 -0
- package/dist/runtime/daimon/training/native/container/lifecycle.d.ts +8 -0
- package/dist/runtime/daimon/training/native/container/lifecycle.js +110 -0
- package/dist/runtime/daimon/training/native/container/lifecycle.js.map +1 -0
- package/dist/runtime/daimon/training/native/container/manifest.d.ts +17 -0
- package/dist/runtime/daimon/training/native/container/manifest.js +144 -0
- package/dist/runtime/daimon/training/native/container/manifest.js.map +1 -0
- package/dist/runtime/daimon/training/native/container/native-preflight.d.ts +8 -0
- package/dist/runtime/daimon/training/native/container/native-preflight.js +40 -0
- package/dist/runtime/daimon/training/native/container/native-preflight.js.map +1 -0
- package/dist/runtime/daimon/training/native/container/preflight.d.ts +15 -0
- package/dist/runtime/daimon/training/native/container/preflight.js +79 -0
- package/dist/runtime/daimon/training/native/container/preflight.js.map +1 -0
- package/dist/runtime/daimon/training/native/container/slot.d.ts +4 -0
- package/dist/runtime/daimon/training/native/container/slot.js +58 -0
- package/dist/runtime/daimon/training/native/container/slot.js.map +1 -0
- package/dist/runtime/daimon/training/native/external.d.ts +37 -0
- package/dist/runtime/daimon/training/native/external.js +75 -0
- package/dist/runtime/daimon/training/native/external.js.map +1 -0
- package/dist/runtime/daimon/training/native/grants/authority.d.ts +36 -0
- package/dist/runtime/daimon/training/native/grants/authority.js +49 -0
- package/dist/runtime/daimon/training/native/grants/authority.js.map +1 -0
- package/dist/runtime/daimon/training/native/grants/index.d.ts +2 -0
- package/dist/runtime/daimon/training/native/grants/index.js +3 -0
- package/dist/runtime/daimon/training/native/grants/index.js.map +1 -0
- package/dist/runtime/daimon/training/native/grants/pins.d.ts +30 -0
- package/dist/runtime/daimon/training/native/grants/pins.js +20 -0
- package/dist/runtime/daimon/training/native/grants/pins.js.map +1 -0
- package/dist/runtime/daimon/training/native/grants/test-socket.d.ts +6 -0
- package/dist/runtime/daimon/training/native/grants/test-socket.js +35 -0
- package/dist/runtime/daimon/training/native/grants/test-socket.js.map +1 -0
- package/dist/runtime/daimon/training/native/grok-fixtures.d.ts +34 -0
- package/dist/runtime/daimon/training/native/grok-fixtures.js +179 -0
- package/dist/runtime/daimon/training/native/grok-fixtures.js.map +1 -0
- package/dist/runtime/daimon/training/native/index.d.ts +38 -0
- package/dist/runtime/daimon/training/native/index.js +27 -0
- package/dist/runtime/daimon/training/native/index.js.map +1 -0
- package/dist/runtime/daimon/training/native/live-checkout.d.ts +2 -0
- package/dist/runtime/daimon/training/native/live-checkout.js +15 -0
- package/dist/runtime/daimon/training/native/live-checkout.js.map +1 -0
- package/dist/runtime/daimon/training/native/observations.d.ts +51 -0
- package/dist/runtime/daimon/training/native/observations.js +217 -0
- package/dist/runtime/daimon/training/native/observations.js.map +1 -0
- package/dist/runtime/daimon/training/native/outcome.d.ts +13 -0
- package/dist/runtime/daimon/training/native/outcome.js +37 -0
- package/dist/runtime/daimon/training/native/outcome.js.map +1 -0
- package/dist/runtime/daimon/training/native/preparation/engine-config.d.ts +21 -0
- package/dist/runtime/daimon/training/native/preparation/engine-config.js +30 -0
- package/dist/runtime/daimon/training/native/preparation/engine-config.js.map +1 -0
- package/dist/runtime/daimon/training/native/preparation/fixtures.d.ts +17 -0
- package/dist/runtime/daimon/training/native/preparation/fixtures.js +50 -0
- package/dist/runtime/daimon/training/native/preparation/fixtures.js.map +1 -0
- package/dist/runtime/daimon/training/native/preparation/index.d.ts +7 -0
- package/dist/runtime/daimon/training/native/preparation/index.js +5 -0
- package/dist/runtime/daimon/training/native/preparation/index.js.map +1 -0
- package/dist/runtime/daimon/training/native/preparation/layout.d.ts +66 -0
- package/dist/runtime/daimon/training/native/preparation/layout.js +87 -0
- package/dist/runtime/daimon/training/native/preparation/layout.js.map +1 -0
- package/dist/runtime/daimon/training/native/preparation/subject.d.ts +14 -0
- package/dist/runtime/daimon/training/native/preparation/subject.js +67 -0
- package/dist/runtime/daimon/training/native/preparation/subject.js.map +1 -0
- package/dist/runtime/daimon/training/native/preparation/types.d.ts +69 -0
- package/dist/runtime/daimon/training/native/preparation/types.js +2 -0
- package/dist/runtime/daimon/training/native/preparation/types.js.map +1 -0
- package/dist/runtime/daimon/training/native/receipt/artifacts.d.ts +25 -0
- package/dist/runtime/daimon/training/native/receipt/artifacts.js +32 -0
- package/dist/runtime/daimon/training/native/receipt/artifacts.js.map +1 -0
- package/dist/runtime/daimon/training/native/receipt/grok-clocked-pins.d.ts +129 -0
- package/dist/runtime/daimon/training/native/receipt/grok-clocked-pins.js +45 -0
- package/dist/runtime/daimon/training/native/receipt/grok-clocked-pins.js.map +1 -0
- package/dist/runtime/daimon/training/native/receipt/grok.d.ts +246 -0
- package/dist/runtime/daimon/training/native/receipt/grok.js +195 -0
- package/dist/runtime/daimon/training/native/receipt/grok.js.map +1 -0
- package/dist/runtime/daimon/training/native/receipt/index.d.ts +10 -0
- package/dist/runtime/daimon/training/native/receipt/index.js +7 -0
- package/dist/runtime/daimon/training/native/receipt/index.js.map +1 -0
- package/dist/runtime/daimon/training/native/receipt/paths.d.ts +3 -0
- package/dist/runtime/daimon/training/native/receipt/paths.js +18 -0
- package/dist/runtime/daimon/training/native/receipt/paths.js.map +1 -0
- package/dist/runtime/daimon/training/native/receipt/sandbox.d.ts +42 -0
- package/dist/runtime/daimon/training/native/receipt/sandbox.js +35 -0
- package/dist/runtime/daimon/training/native/receipt/sandbox.js.map +1 -0
- package/dist/runtime/daimon/training/native/receipt/schema.d.ts +486 -0
- package/dist/runtime/daimon/training/native/receipt/schema.js +138 -0
- package/dist/runtime/daimon/training/native/receipt/schema.js.map +1 -0
- package/dist/runtime/daimon/training/native/receipt/task-clock.d.ts +19 -0
- package/dist/runtime/daimon/training/native/receipt/task-clock.js +49 -0
- package/dist/runtime/daimon/training/native/receipt/task-clock.js.map +1 -0
- package/dist/runtime/daimon/training/native/sealed-usage.d.ts +218 -0
- package/dist/runtime/daimon/training/native/sealed-usage.js +261 -0
- package/dist/runtime/daimon/training/native/sealed-usage.js.map +1 -0
- package/dist/runtime/daimon/training/native/test-fixtures.d.ts +27 -0
- package/dist/runtime/daimon/training/native/test-fixtures.js +158 -0
- package/dist/runtime/daimon/training/native/test-fixtures.js.map +1 -0
- package/dist/runtime/daimon/training/native/transcript.d.ts +96 -0
- package/dist/runtime/daimon/training/native/transcript.js +239 -0
- package/dist/runtime/daimon/training/native/transcript.js.map +1 -0
- package/dist/runtime/daimon/training/native/transport/delivery.d.ts +27 -0
- package/dist/runtime/daimon/training/native/transport/delivery.js +98 -0
- package/dist/runtime/daimon/training/native/transport/delivery.js.map +1 -0
- package/dist/runtime/daimon/training/native/transport/http.d.ts +3 -0
- package/dist/runtime/daimon/training/native/transport/http.js +81 -0
- package/dist/runtime/daimon/training/native/transport/http.js.map +1 -0
- package/dist/runtime/daimon/training/native/transport/index.d.ts +4 -0
- package/dist/runtime/daimon/training/native/transport/index.js +3 -0
- package/dist/runtime/daimon/training/native/transport/index.js.map +1 -0
- package/dist/runtime/daimon/training/native/transport/types.d.ts +15 -0
- package/dist/runtime/daimon/training/native/transport/types.js +2 -0
- package/dist/runtime/daimon/training/native/transport/types.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/cli.d.ts +1 -0
- package/dist/runtime/daimon/training/native/worker/cli.js +17 -0
- package/dist/runtime/daimon/training/native/worker/cli.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/control.d.ts +20 -0
- package/dist/runtime/daimon/training/native/worker/control.js +266 -0
- package/dist/runtime/daimon/training/native/worker/control.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/deliveries.d.ts +38 -0
- package/dist/runtime/daimon/training/native/worker/deliveries.js +91 -0
- package/dist/runtime/daimon/training/native/worker/deliveries.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/evidence/collect.d.ts +54 -0
- package/dist/runtime/daimon/training/native/worker/evidence/collect.js +278 -0
- package/dist/runtime/daimon/training/native/worker/evidence/collect.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/evidence/execution.d.ts +102 -0
- package/dist/runtime/daimon/training/native/worker/evidence/execution.js +191 -0
- package/dist/runtime/daimon/training/native/worker/evidence/execution.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/evidence/index.d.ts +3 -0
- package/dist/runtime/daimon/training/native/worker/evidence/index.js +4 -0
- package/dist/runtime/daimon/training/native/worker/evidence/index.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/evidence/io.d.ts +20 -0
- package/dist/runtime/daimon/training/native/worker/evidence/io.js +130 -0
- package/dist/runtime/daimon/training/native/worker/evidence/io.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/evidence/seal.d.ts +3 -0
- package/dist/runtime/daimon/training/native/worker/evidence/seal.js +67 -0
- package/dist/runtime/daimon/training/native/worker/evidence/seal.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/index.d.ts +11 -0
- package/dist/runtime/daimon/training/native/worker/index.js +8 -0
- package/dist/runtime/daimon/training/native/worker/index.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/profile.d.ts +23 -0
- package/dist/runtime/daimon/training/native/worker/profile.js +149 -0
- package/dist/runtime/daimon/training/native/worker/profile.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/public.d.ts +41 -0
- package/dist/runtime/daimon/training/native/worker/public.js +22 -0
- package/dist/runtime/daimon/training/native/worker/public.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/sandbox.d.ts +4 -0
- package/dist/runtime/daimon/training/native/worker/sandbox.js +52 -0
- package/dist/runtime/daimon/training/native/worker/sandbox.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/server.d.ts +4 -0
- package/dist/runtime/daimon/training/native/worker/server.js +102 -0
- package/dist/runtime/daimon/training/native/worker/server.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/types.d.ts +318 -0
- package/dist/runtime/daimon/training/native/worker/types.js +2 -0
- package/dist/runtime/daimon/training/native/worker/types.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/index.d.ts +4 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/index.js +4 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/index.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/source.d.ts +3 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/source.js +16 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/source.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/terminal.d.ts +4 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/terminal.js +52 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/terminal.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/transport.d.ts +21 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/transport.js +198 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/transport.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/types.d.ts +91 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/types.js +2 -0
- package/dist/runtime/daimon/training/native/worker/wake-delivery/types.js.map +1 -0
- package/dist/runtime/daimon/training/native/worker/worker.d.ts +1 -0
- package/dist/runtime/daimon/training/native/worker/worker.js +103 -0
- package/dist/runtime/daimon/training/native/worker/worker.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-broker/broker.d.ts +20 -0
- package/dist/runtime/daimon/training/providers/grok-broker/broker.js +55 -0
- package/dist/runtime/daimon/training/providers/grok-broker/broker.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-fixtures.d.ts +15 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-fixtures.js +29 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-fixtures.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-home.d.ts +16 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-home.js +60 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-home.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-pool.d.ts +81 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-pool.js +221 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-pool.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-releases.d.ts +31 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-releases.js +78 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grant-releases.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grants.d.ts +63 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grants.js +20 -0
- package/dist/runtime/daimon/training/providers/grok-broker/grants.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-broker/index.d.ts +3 -0
- package/dist/runtime/daimon/training/providers/grok-broker/index.js +4 -0
- package/dist/runtime/daimon/training/providers/grok-broker/index.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/cli.d.ts +21 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/cli.js +81 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/cli.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/index.d.ts +4 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/index.js +5 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/index.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/protocol.d.ts +26 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/protocol.js +102 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/protocol.js.map +1 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/provider.d.ts +9 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/provider.js +38 -0
- package/dist/runtime/daimon/training/providers/grok-subscription/provider.js.map +1 -0
- package/dist/runtime/daimon/training/services/freshness.d.ts +15 -0
- package/dist/runtime/daimon/training/services/freshness.js +19 -0
- package/dist/runtime/daimon/training/services/freshness.js.map +1 -0
- package/dist/runtime/daimon/training/services/index.d.ts +3 -0
- package/dist/runtime/daimon/training/services/index.js +3 -0
- package/dist/runtime/daimon/training/services/index.js.map +1 -0
- package/dist/runtime/daimon/training/services/slot-freshness.d.ts +34 -0
- package/dist/runtime/daimon/training/services/slot-freshness.js +178 -0
- package/dist/runtime/daimon/training/services/slot-freshness.js.map +1 -0
- package/dist/runtime/daimon/wakeMoments.js +1 -0
- package/dist/runtime/daimon/wakeMoments.js.map +1 -0
- package/dist/runtime/pi/appPreludeSource.js +3 -2
- package/dist/runtime/pi/appTemplate.d.ts +1 -1
- package/dist/runtime/pi/appTemplate.js +2 -1
- package/dist/runtime/registry.d.ts +6 -0
- package/dist/runtime/registry.js +2 -0
- package/dist/runtime/trainingCli.d.ts +5 -0
- package/dist/runtime/trainingCli.js +23 -0
- package/dist/runtime/trainingGraderProviders.d.ts +4 -0
- package/dist/runtime/trainingGraderProviders.js +55 -0
- package/dist/runtime/trainingProviders.d.ts +4 -0
- package/dist/runtime/trainingProviders.js +89 -0
- package/dist/runtime/trainingSubjects.d.ts +129 -0
- package/dist/runtime/trainingSubjects.js +32 -0
- package/dist/shared/daimonCodexDocker.js +9 -10
- package/dist/shared/errors.js +1 -0
- package/dist/shared/errors.js.map +1 -0
- package/dist/shared/index.d.ts +1 -0
- package/dist/shared/index.js +1 -0
- package/dist/shared/trainingLegacy/index.d.ts +5 -0
- package/dist/shared/trainingLegacy/index.js +6 -0
- package/dist/shared/trainingLegacy/index.js.map +1 -0
- package/dist/shared/trainingLegacy/trainingLegacyEnvironment.d.ts +6 -0
- package/dist/shared/trainingLegacy/trainingLegacyEnvironment.js +40 -0
- package/dist/shared/trainingLegacy/trainingLegacyEnvironment.js.map +1 -0
- package/dist/shared/trainingLegacy/trainingLegacyFiles.d.ts +4 -0
- package/dist/shared/trainingLegacy/trainingLegacyFiles.js +34 -0
- package/dist/shared/trainingLegacy/trainingLegacyFiles.js.map +1 -0
- package/dist/shared/trainingLegacy/trainingLegacyGraderProviders.d.ts +5 -0
- package/dist/shared/trainingLegacy/trainingLegacyGraderProviders.js +6 -0
- package/dist/shared/trainingLegacy/trainingLegacyGraderProviders.js.map +1 -0
- package/dist/shared/trainingLegacy/trainingLegacyHashDomains.d.ts +39 -0
- package/dist/shared/trainingLegacy/trainingLegacyHashDomains.js +41 -0
- package/dist/shared/trainingLegacy/trainingLegacyHashDomains.js.map +1 -0
- package/dist/shared/trainingLegacy/trainingLegacyWire.d.ts +15 -0
- package/dist/shared/trainingLegacy/trainingLegacyWire.js +21 -0
- package/dist/shared/trainingLegacy/trainingLegacyWire.js.map +1 -0
- package/dist/shared/trainingLegacy/trainingLegacyWireAliases.json +90 -0
- package/dist/target/contracts.d.ts +2 -2
- package/dist/training/candidates/authored.d.ts +3 -0
- package/dist/training/candidates/authored.js +27 -0
- package/dist/training/candidates/authored.js.map +1 -0
- package/dist/training/candidates/derive.d.ts +3 -0
- package/dist/training/candidates/derive.js +66 -0
- package/dist/training/candidates/derive.js.map +1 -0
- package/dist/training/candidates/filesystem.d.ts +7 -0
- package/dist/training/candidates/filesystem.js +41 -0
- package/dist/training/candidates/filesystem.js.map +1 -0
- package/dist/training/candidates/headings.d.ts +9 -0
- package/dist/training/candidates/headings.js +47 -0
- package/dist/training/candidates/headings.js.map +1 -0
- package/dist/training/candidates/identity.d.ts +6 -0
- package/dist/training/candidates/identity.js +29 -0
- package/dist/training/candidates/identity.js.map +1 -0
- package/dist/training/candidates/index.d.ts +6 -0
- package/dist/training/candidates/index.js +6 -0
- package/dist/training/candidates/index.js.map +1 -0
- package/dist/training/candidates/materialize.d.ts +2 -0
- package/dist/training/candidates/materialize.js +39 -0
- package/dist/training/candidates/materialize.js.map +1 -0
- package/dist/training/candidates/snapshot.d.ts +5 -0
- package/dist/training/candidates/snapshot.js +22 -0
- package/dist/training/candidates/snapshot.js.map +1 -0
- package/dist/training/candidates/types.d.ts +41 -0
- package/dist/training/candidates/types.js +2 -0
- package/dist/training/candidates/types.js.map +1 -0
- package/dist/training/candidates/validation.d.ts +4 -0
- package/dist/training/candidates/validation.js +89 -0
- package/dist/training/candidates/validation.js.map +1 -0
- package/dist/training/capture/index.d.ts +2 -0
- package/dist/training/capture/index.js +2 -0
- package/dist/training/capture/index.js.map +1 -0
- package/dist/training/capture/trace.d.ts +16 -0
- package/dist/training/capture/trace.js +74 -0
- package/dist/training/capture/trace.js.map +1 -0
- package/dist/training/cases/builtin-evaluators.d.ts +2 -0
- package/dist/training/cases/builtin-evaluators.js +120 -0
- package/dist/training/cases/builtin-evaluators.js.map +1 -0
- package/dist/training/cases/check-schema.d.ts +2 -0
- package/dist/training/cases/check-schema.js +60 -0
- package/dist/training/cases/check-schema.js.map +1 -0
- package/dist/training/cases/cost-projection.d.ts +27 -0
- package/dist/training/cases/cost-projection.js +29 -0
- package/dist/training/cases/cost-projection.js.map +1 -0
- package/dist/training/cases/fixtures.d.ts +5 -0
- package/dist/training/cases/fixtures.js +134 -0
- package/dist/training/cases/fixtures.js.map +1 -0
- package/dist/training/cases/grader.d.ts +14 -0
- package/dist/training/cases/grader.js +122 -0
- package/dist/training/cases/grader.js.map +1 -0
- package/dist/training/cases/index.d.ts +12 -0
- package/dist/training/cases/index.js +11 -0
- package/dist/training/cases/index.js.map +1 -0
- package/dist/training/cases/json-schema.d.ts +3 -0
- package/dist/training/cases/json-schema.js +6 -0
- package/dist/training/cases/json-schema.js.map +1 -0
- package/dist/training/cases/parser.d.ts +6 -0
- package/dist/training/cases/parser.js +45 -0
- package/dist/training/cases/parser.js.map +1 -0
- package/dist/training/cases/plans.d.ts +14 -0
- package/dist/training/cases/plans.js +186 -0
- package/dist/training/cases/plans.js.map +1 -0
- package/dist/training/cases/resolver.d.ts +5 -0
- package/dist/training/cases/resolver.js +90 -0
- package/dist/training/cases/resolver.js.map +1 -0
- package/dist/training/cases/schema.d.ts +5 -0
- package/dist/training/cases/schema.js +154 -0
- package/dist/training/cases/schema.js.map +1 -0
- package/dist/training/cases/splits.d.ts +16 -0
- package/dist/training/cases/splits.js +71 -0
- package/dist/training/cases/splits.js.map +1 -0
- package/dist/training/cases/text-checks.d.ts +3 -0
- package/dist/training/cases/text-checks.js +50 -0
- package/dist/training/cases/text-checks.js.map +1 -0
- package/dist/training/cases/training/identity.d.ts +6 -0
- package/dist/training/cases/training/identity.js +26 -0
- package/dist/training/cases/training/identity.js.map +1 -0
- package/dist/training/cases/training/index.d.ts +2 -0
- package/dist/training/cases/training/index.js +3 -0
- package/dist/training/cases/training/index.js.map +1 -0
- package/dist/training/cases/training/schema.d.ts +4 -0
- package/dist/training/cases/training/schema.js +207 -0
- package/dist/training/cases/training/schema.js.map +1 -0
- package/dist/training/cases/turn-bindings.d.ts +3 -0
- package/dist/training/cases/turn-bindings.js +61 -0
- package/dist/training/cases/turn-bindings.js.map +1 -0
- package/dist/training/cases/validation.d.ts +17 -0
- package/dist/training/cases/validation.js +92 -0
- package/dist/training/cases/validation.js.map +1 -0
- package/dist/training/cli/arguments.d.ts +18 -0
- package/dist/training/cli/arguments.js +165 -0
- package/dist/training/cli/arguments.js.map +1 -0
- package/dist/training/cli/commands.d.ts +3 -0
- package/dist/training/cli/commands.js +255 -0
- package/dist/training/cli/commands.js.map +1 -0
- package/dist/training/cli/datasets.d.ts +30 -0
- package/dist/training/cli/datasets.js +116 -0
- package/dist/training/cli/datasets.js.map +1 -0
- package/dist/training/cli/dry-run.d.ts +7 -0
- package/dist/training/cli/dry-run.js +37 -0
- package/dist/training/cli/dry-run.js.map +1 -0
- package/dist/training/cli/grader/atomic/decompose.d.ts +43 -0
- package/dist/training/cli/grader/atomic/decompose.js +208 -0
- package/dist/training/cli/grader/atomic/decompose.js.map +1 -0
- package/dist/training/cli/grader/atomic/files.d.ts +159 -0
- package/dist/training/cli/grader/atomic/files.js +97 -0
- package/dist/training/cli/grader/atomic/files.js.map +1 -0
- package/dist/training/cli/grader/atomic/index.d.ts +8 -0
- package/dist/training/cli/grader/atomic/index.js +34 -0
- package/dist/training/cli/grader/atomic/index.js.map +1 -0
- package/dist/training/cli/grader/atomic/reference.d.ts +131 -0
- package/dist/training/cli/grader/atomic/reference.js +133 -0
- package/dist/training/cli/grader/atomic/reference.js.map +1 -0
- package/dist/training/cli/grader/atomic/runtime.d.ts +3 -0
- package/dist/training/cli/grader/atomic/runtime.js +91 -0
- package/dist/training/cli/grader/atomic/runtime.js.map +1 -0
- package/dist/training/cli/grader/atomic/traps.d.ts +15 -0
- package/dist/training/cli/grader/atomic/traps.js +38 -0
- package/dist/training/cli/grader/atomic/traps.js.map +1 -0
- package/dist/training/cli/grader/backend.d.ts +19 -0
- package/dist/training/cli/grader/backend.js +39 -0
- package/dist/training/cli/grader/backend.js.map +1 -0
- package/dist/training/cli/grader/cache.d.ts +2 -0
- package/dist/training/cli/grader/cache.js +75 -0
- package/dist/training/cli/grader/cache.js.map +1 -0
- package/dist/training/cli/grader/calibrate.d.ts +19 -0
- package/dist/training/cli/grader/calibrate.js +237 -0
- package/dist/training/cli/grader/calibrate.js.map +1 -0
- package/dist/training/cli/grader/cost.d.ts +24 -0
- package/dist/training/cli/grader/cost.js +54 -0
- package/dist/training/cli/grader/cost.js.map +1 -0
- package/dist/training/cli/grader/exclusions.d.ts +12 -0
- package/dist/training/cli/grader/exclusions.js +49 -0
- package/dist/training/cli/grader/exclusions.js.map +1 -0
- package/dist/training/cli/grader/grade.d.ts +26 -0
- package/dist/training/cli/grader/grade.js +33 -0
- package/dist/training/cli/grader/grade.js.map +1 -0
- package/dist/training/cli/grader/index.d.ts +10 -0
- package/dist/training/cli/grader/index.js +73 -0
- package/dist/training/cli/grader/index.js.map +1 -0
- package/dist/training/cli/grader/label-call.d.ts +13 -0
- package/dist/training/cli/grader/label-call.js +67 -0
- package/dist/training/cli/grader/label-call.js.map +1 -0
- package/dist/training/cli/grader/label-import.d.ts +16 -0
- package/dist/training/cli/grader/label-import.js +84 -0
- package/dist/training/cli/grader/label-import.js.map +1 -0
- package/dist/training/cli/grader/label-inputs.d.ts +9 -0
- package/dist/training/cli/grader/label-inputs.js +56 -0
- package/dist/training/cli/grader/label-inputs.js.map +1 -0
- package/dist/training/cli/grader/label.d.ts +25 -0
- package/dist/training/cli/grader/label.js +172 -0
- package/dist/training/cli/grader/label.js.map +1 -0
- package/dist/training/cli/grader/records.d.ts +20 -0
- package/dist/training/cli/grader/records.js +65 -0
- package/dist/training/cli/grader/records.js.map +1 -0
- package/dist/training/cli/grader/report.d.ts +26 -0
- package/dist/training/cli/grader/report.js +42 -0
- package/dist/training/cli/grader/report.js.map +1 -0
- package/dist/training/cli/grader/runtime.d.ts +5 -0
- package/dist/training/cli/grader/runtime.js +64 -0
- package/dist/training/cli/grader/runtime.js.map +1 -0
- package/dist/training/cli/grader/validate.d.ts +5 -0
- package/dist/training/cli/grader/validate.js +42 -0
- package/dist/training/cli/grader/validate.js.map +1 -0
- package/dist/training/cli/index.d.ts +6 -0
- package/dist/training/cli/index.js +4 -0
- package/dist/training/cli/index.js.map +1 -0
- package/dist/training/cli/launch.d.ts +3 -0
- package/dist/training/cli/launch.js +17 -0
- package/dist/training/cli/launch.js.map +1 -0
- package/dist/training/cli/native.d.ts +235 -0
- package/dist/training/cli/native.js +83 -0
- package/dist/training/cli/native.js.map +1 -0
- package/dist/training/cli/output.d.ts +21 -0
- package/dist/training/cli/output.js +95 -0
- package/dist/training/cli/output.js.map +1 -0
- package/dist/training/cli/progress/activity.d.ts +11 -0
- package/dist/training/cli/progress/activity.js +45 -0
- package/dist/training/cli/progress/activity.js.map +1 -0
- package/dist/training/cli/progress/failure.d.ts +37 -0
- package/dist/training/cli/progress/failure.js +48 -0
- package/dist/training/cli/progress/failure.js.map +1 -0
- package/dist/training/cli/progress/generations.d.ts +11 -0
- package/dist/training/cli/progress/generations.js +50 -0
- package/dist/training/cli/progress/generations.js.map +1 -0
- package/dist/training/cli/progress/index.d.ts +6 -0
- package/dist/training/cli/progress/index.js +5 -0
- package/dist/training/cli/progress/index.js.map +1 -0
- package/dist/training/cli/progress/journal.d.ts +38 -0
- package/dist/training/cli/progress/journal.js +101 -0
- package/dist/training/cli/progress/journal.js.map +1 -0
- package/dist/training/cli/report/diff.d.ts +2 -0
- package/dist/training/cli/report/diff.js +24 -0
- package/dist/training/cli/report/diff.js.map +1 -0
- package/dist/training/cli/report/export.d.ts +9 -0
- package/dist/training/cli/report/export.js +81 -0
- package/dist/training/cli/report/export.js.map +1 -0
- package/dist/training/cli/report/fixture.d.ts +9 -0
- package/dist/training/cli/report/fixture.js +65 -0
- package/dist/training/cli/report/fixture.js.map +1 -0
- package/dist/training/cli/report/index.d.ts +1 -0
- package/dist/training/cli/report/index.js +2 -0
- package/dist/training/cli/report/index.js.map +1 -0
- package/dist/training/cli/report/markdown.d.ts +2 -0
- package/dist/training/cli/report/markdown.js +60 -0
- package/dist/training/cli/report/markdown.js.map +1 -0
- package/dist/training/cli/report/summary.d.ts +130 -0
- package/dist/training/cli/report/summary.js +165 -0
- package/dist/training/cli/report/summary.js.map +1 -0
- package/dist/training/cli/spawnfile/context.d.ts +75 -0
- package/dist/training/cli/spawnfile/context.js +78 -0
- package/dist/training/cli/spawnfile/context.js.map +1 -0
- package/dist/training/cli/spawnfile/index.d.ts +5 -0
- package/dist/training/cli/spawnfile/index.js +4 -0
- package/dist/training/cli/spawnfile/index.js.map +1 -0
- package/dist/training/cli/spawnfile/moments.d.ts +44 -0
- package/dist/training/cli/spawnfile/moments.js +71 -0
- package/dist/training/cli/spawnfile/moments.js.map +1 -0
- package/dist/training/cli/spawnfile/selection.d.ts +83 -0
- package/dist/training/cli/spawnfile/selection.js +54 -0
- package/dist/training/cli/spawnfile/selection.js.map +1 -0
- package/dist/training/cli/training/cache.d.ts +4 -0
- package/dist/training/cli/training/cache.js +43 -0
- package/dist/training/cli/training/cache.js.map +1 -0
- package/dist/training/cli/training/checkpoint.d.ts +19 -0
- package/dist/training/cli/training/checkpoint.js +62 -0
- package/dist/training/cli/training/checkpoint.js.map +1 -0
- package/dist/training/cli/training/comparison.d.ts +8 -0
- package/dist/training/cli/training/comparison.js +22 -0
- package/dist/training/cli/training/comparison.js.map +1 -0
- package/dist/training/cli/training/configuration.d.ts +10 -0
- package/dist/training/cli/training/configuration.js +81 -0
- package/dist/training/cli/training/configuration.js.map +1 -0
- package/dist/training/cli/training/deadlines.d.ts +7 -0
- package/dist/training/cli/training/deadlines.js +14 -0
- package/dist/training/cli/training/deadlines.js.map +1 -0
- package/dist/training/cli/training/index.d.ts +8 -0
- package/dist/training/cli/training/index.js +9 -0
- package/dist/training/cli/training/index.js.map +1 -0
- package/dist/training/cli/training/inference.d.ts +48 -0
- package/dist/training/cli/training/inference.js +96 -0
- package/dist/training/cli/training/inference.js.map +1 -0
- package/dist/training/cli/training/optimizer.d.ts +33 -0
- package/dist/training/cli/training/optimizer.js +92 -0
- package/dist/training/cli/training/optimizer.js.map +1 -0
- package/dist/training/cli/training/output-lock.d.ts +2 -0
- package/dist/training/cli/training/output-lock.js +30 -0
- package/dist/training/cli/training/output-lock.js.map +1 -0
- package/dist/training/cli/types.d.ts +101 -0
- package/dist/training/cli/types.js +2 -0
- package/dist/training/cli/types.js.map +1 -0
- package/dist/training/cli/viewer.d.ts +40 -0
- package/dist/training/cli/viewer.js +210 -0
- package/dist/training/cli/viewer.js.map +1 -0
- package/dist/training/contracts/atomic-feedback.d.ts +2 -0
- package/dist/training/contracts/atomic-feedback.js +2 -0
- package/dist/training/contracts/atomic-feedback.js.map +1 -0
- package/dist/training/contracts/dataset.d.ts +174 -0
- package/dist/training/contracts/dataset.js +2 -0
- package/dist/training/contracts/dataset.js.map +1 -0
- package/dist/training/contracts/evaluation.d.ts +256 -0
- package/dist/training/contracts/evaluation.js +41 -0
- package/dist/training/contracts/evaluation.js.map +1 -0
- package/dist/training/contracts/execution.d.ts +363 -0
- package/dist/training/contracts/execution.js +15 -0
- package/dist/training/contracts/execution.js.map +1 -0
- package/dist/training/contracts/grader-providers.d.ts +40 -0
- package/dist/training/contracts/grader-providers.js +27 -0
- package/dist/training/contracts/grader-providers.js.map +1 -0
- package/dist/training/contracts/grading.d.ts +74 -0
- package/dist/training/contracts/grading.js +2 -0
- package/dist/training/contracts/grading.js.map +1 -0
- package/dist/training/contracts/index.d.ts +16 -0
- package/dist/training/contracts/index.js +8 -0
- package/dist/training/contracts/index.js.map +1 -0
- package/dist/training/contracts/inference-grants.d.ts +34 -0
- package/dist/training/contracts/inference-grants.js +2 -0
- package/dist/training/contracts/inference-grants.js.map +1 -0
- package/dist/training/contracts/inference-registry.d.ts +38 -0
- package/dist/training/contracts/inference-registry.js +7 -0
- package/dist/training/contracts/inference-registry.js.map +1 -0
- package/dist/training/contracts/training-config.d.ts +95 -0
- package/dist/training/contracts/training-config.js +15 -0
- package/dist/training/contracts/training-config.js.map +1 -0
- package/dist/training/contracts/trial-slot.d.ts +12 -0
- package/dist/training/contracts/trial-slot.js +2 -0
- package/dist/training/contracts/trial-slot.js.map +1 -0
- package/dist/training/dspy/AGENTS.md +29 -0
- package/dist/training/dspy/README.md +206 -0
- package/dist/training/dspy/protocol.md +311 -0
- package/dist/training/dspy/pyproject.toml +36 -0
- package/dist/training/dspy/requirements.lock +69 -0
- package/dist/training/dspy/spawnfile_training_dspy/AGENTS.md +63 -0
- package/dist/training/dspy/spawnfile_training_dspy/__init__.py +6 -0
- package/dist/training/dspy/spawnfile_training_dspy/checkpoint.py +146 -0
- package/dist/training/dspy/spawnfile_training_dspy/legacy_wire_aliases.json +90 -0
- package/dist/training/dspy/spawnfile_training_dspy/main.py +36 -0
- package/dist/training/dspy/spawnfile_training_dspy/native.py +89 -0
- package/dist/training/dspy/spawnfile_training_dspy/optimizer.py +74 -0
- package/dist/training/dspy/spawnfile_training_dspy/proposal_history.py +275 -0
- package/dist/training/dspy/spawnfile_training_dspy/protocol.py +149 -0
- package/dist/training/dspy/spawnfile_training_dspy/reflection.py +199 -0
- package/dist/training/dspy/spawnfile_training_dspy/search/AGENTS.md +41 -0
- package/dist/training/dspy/spawnfile_training_dspy/search/__init__.py +1 -0
- package/dist/training/dspy/spawnfile_training_dspy/search/behavior_change.py +152 -0
- package/dist/training/dspy/spawnfile_training_dspy/search/boundaries.py +38 -0
- package/dist/training/dspy/spawnfile_training_dspy/search/instruction_text.py +47 -0
- package/dist/training/dspy/spawnfile_training_dspy/search/population.py +213 -0
- package/dist/training/dspy/spawnfile_training_dspy/search/selection.py +24 -0
- package/dist/training/dspy/spawnfile_training_dspy/search/variants.py +83 -0
- package/dist/training/dspy/spawnfile_training_dspy/session.py +116 -0
- package/dist/training/dspy/spawnfile_training_dspy/transport.py +129 -0
- package/dist/training/evaluation/atomic-grader.d.ts +5 -0
- package/dist/training/evaluation/atomic-grader.js +55 -0
- package/dist/training/evaluation/atomic-grader.js.map +1 -0
- package/dist/training/evaluation/embedding/index.d.ts +1 -0
- package/dist/training/evaluation/embedding/index.js +2 -0
- package/dist/training/evaluation/embedding/index.js.map +1 -0
- package/dist/training/evaluation/embedding/similarity.d.ts +4 -0
- package/dist/training/evaluation/embedding/similarity.js +38 -0
- package/dist/training/evaluation/embedding/similarity.js.map +1 -0
- package/dist/training/evaluation/engine.d.ts +6 -0
- package/dist/training/evaluation/engine.js +191 -0
- package/dist/training/evaluation/engine.js.map +1 -0
- package/dist/training/evaluation/grader.d.ts +4 -0
- package/dist/training/evaluation/grader.js +67 -0
- package/dist/training/evaluation/grader.js.map +1 -0
- package/dist/training/evaluation/index.d.ts +10 -0
- package/dist/training/evaluation/index.js +9 -0
- package/dist/training/evaluation/index.js.map +1 -0
- package/dist/training/evaluation/logic/index.d.ts +1 -0
- package/dist/training/evaluation/logic/index.js +2 -0
- package/dist/training/evaluation/logic/index.js.map +1 -0
- package/dist/training/evaluation/logic/results.d.ts +7 -0
- package/dist/training/evaluation/logic/results.js +70 -0
- package/dist/training/evaluation/logic/results.js.map +1 -0
- package/dist/training/evaluation/mechanical/index.d.ts +6 -0
- package/dist/training/evaluation/mechanical/index.js +7 -0
- package/dist/training/evaluation/mechanical/index.js.map +1 -0
- package/dist/training/evaluation/mechanical/predicates.d.ts +2 -0
- package/dist/training/evaluation/mechanical/predicates.js +27 -0
- package/dist/training/evaluation/mechanical/predicates.js.map +1 -0
- package/dist/training/evaluation/mechanical/sequence.d.ts +3 -0
- package/dist/training/evaluation/mechanical/sequence.js +75 -0
- package/dist/training/evaluation/mechanical/sequence.js.map +1 -0
- package/dist/training/evaluation/mechanical/text-checks.d.ts +3 -0
- package/dist/training/evaluation/mechanical/text-checks.js +65 -0
- package/dist/training/evaluation/mechanical/text-checks.js.map +1 -0
- package/dist/training/evaluation/mechanical/text-values.d.ts +7 -0
- package/dist/training/evaluation/mechanical/text-values.js +28 -0
- package/dist/training/evaluation/mechanical/text-values.js.map +1 -0
- package/dist/training/evaluation/mechanical/text.d.ts +3 -0
- package/dist/training/evaluation/mechanical/text.js +40 -0
- package/dist/training/evaluation/mechanical/text.js.map +1 -0
- package/dist/training/evaluation/mechanical/token-efficiency.d.ts +4 -0
- package/dist/training/evaluation/mechanical/token-efficiency.js +28 -0
- package/dist/training/evaluation/mechanical/token-efficiency.js.map +1 -0
- package/dist/training/evaluation/mechanical/tools.d.ts +3 -0
- package/dist/training/evaluation/mechanical/tools.js +39 -0
- package/dist/training/evaluation/mechanical/tools.js.map +1 -0
- package/dist/training/evaluation/registry.d.ts +4 -0
- package/dist/training/evaluation/registry.js +64 -0
- package/dist/training/evaluation/registry.js.map +1 -0
- package/dist/training/evaluation/repair.d.ts +14 -0
- package/dist/training/evaluation/repair.js +73 -0
- package/dist/training/evaluation/repair.js.map +1 -0
- package/dist/training/evaluation/schedule.d.ts +21 -0
- package/dist/training/evaluation/schedule.js +98 -0
- package/dist/training/evaluation/schedule.js.map +1 -0
- package/dist/training/evaluation/selectors.d.ts +22 -0
- package/dist/training/evaluation/selectors.js +61 -0
- package/dist/training/evaluation/selectors.js.map +1 -0
- package/dist/training/evaluation/semantic/attempts.d.ts +12 -0
- package/dist/training/evaluation/semantic/attempts.js +71 -0
- package/dist/training/evaluation/semantic/attempts.js.map +1 -0
- package/dist/training/evaluation/semantic/cache.d.ts +9 -0
- package/dist/training/evaluation/semantic/cache.js +40 -0
- package/dist/training/evaluation/semantic/cache.js.map +1 -0
- package/dist/training/evaluation/semantic/evidence-units.d.ts +14 -0
- package/dist/training/evaluation/semantic/evidence-units.js +58 -0
- package/dist/training/evaluation/semantic/evidence-units.js.map +1 -0
- package/dist/training/evaluation/semantic/failure.d.ts +30 -0
- package/dist/training/evaluation/semantic/failure.js +79 -0
- package/dist/training/evaluation/semantic/failure.js.map +1 -0
- package/dist/training/evaluation/semantic/index.d.ts +5 -0
- package/dist/training/evaluation/semantic/index.js +5 -0
- package/dist/training/evaluation/semantic/index.js.map +1 -0
- package/dist/training/evaluation/semantic/judgment.d.ts +21 -0
- package/dist/training/evaluation/semantic/judgment.js +50 -0
- package/dist/training/evaluation/semantic/judgment.js.map +1 -0
- package/dist/training/evaluation/semantic/protocol.d.ts +118 -0
- package/dist/training/evaluation/semantic/protocol.js +48 -0
- package/dist/training/evaluation/semantic/protocol.js.map +1 -0
- package/dist/training/evaluation/semantic/rubric.d.ts +3 -0
- package/dist/training/evaluation/semantic/rubric.js +202 -0
- package/dist/training/evaluation/semantic/rubric.js.map +1 -0
- package/dist/training/evaluation/semantic/templates.d.ts +7 -0
- package/dist/training/evaluation/semantic/templates.js +75 -0
- package/dist/training/evaluation/semantic/templates.js.map +1 -0
- package/dist/training/experiments/comparison/index.d.ts +2 -0
- package/dist/training/experiments/comparison/index.js +2 -0
- package/dist/training/experiments/comparison/index.js.map +1 -0
- package/dist/training/experiments/comparison/run.d.ts +3 -0
- package/dist/training/experiments/comparison/run.js +92 -0
- package/dist/training/experiments/comparison/run.js.map +1 -0
- package/dist/training/experiments/comparison/types.d.ts +39 -0
- package/dist/training/experiments/comparison/types.js +2 -0
- package/dist/training/experiments/comparison/types.js.map +1 -0
- package/dist/training/experiments/concurrency/index.d.ts +2 -0
- package/dist/training/experiments/concurrency/index.js +2 -0
- package/dist/training/experiments/concurrency/index.js.map +1 -0
- package/dist/training/experiments/concurrency/slots.d.ts +50 -0
- package/dist/training/experiments/concurrency/slots.js +120 -0
- package/dist/training/experiments/concurrency/slots.js.map +1 -0
- package/dist/training/experiments/costs/assumptions.d.ts +4 -0
- package/dist/training/experiments/costs/assumptions.js +18 -0
- package/dist/training/experiments/costs/assumptions.js.map +1 -0
- package/dist/training/experiments/costs/catalog.d.ts +3 -0
- package/dist/training/experiments/costs/catalog.js +24 -0
- package/dist/training/experiments/costs/catalog.js.map +1 -0
- package/dist/training/experiments/costs/config.d.ts +6 -0
- package/dist/training/experiments/costs/config.js +22 -0
- package/dist/training/experiments/costs/config.js.map +1 -0
- package/dist/training/experiments/costs/decisions.d.ts +5 -0
- package/dist/training/experiments/costs/decisions.js +48 -0
- package/dist/training/experiments/costs/decisions.js.map +1 -0
- package/dist/training/experiments/costs/index.d.ts +9 -0
- package/dist/training/experiments/costs/index.js +7 -0
- package/dist/training/experiments/costs/index.js.map +1 -0
- package/dist/training/experiments/costs/planning.d.ts +71 -0
- package/dist/training/experiments/costs/planning.js +140 -0
- package/dist/training/experiments/costs/planning.js.map +1 -0
- package/dist/training/experiments/costs/report.d.ts +71 -0
- package/dist/training/experiments/costs/report.js +269 -0
- package/dist/training/experiments/costs/report.js.map +1 -0
- package/dist/training/experiments/costs/types.d.ts +53 -0
- package/dist/training/experiments/costs/types.js +2 -0
- package/dist/training/experiments/costs/types.js.map +1 -0
- package/dist/training/experiments/costs/valuation.d.ts +3 -0
- package/dist/training/experiments/costs/valuation.js +124 -0
- package/dist/training/experiments/costs/valuation.js.map +1 -0
- package/dist/training/experiments/index.d.ts +12 -0
- package/dist/training/experiments/index.js +12 -0
- package/dist/training/experiments/index.js.map +1 -0
- package/dist/training/experiments/repair/import.d.ts +21 -0
- package/dist/training/experiments/repair/import.js +104 -0
- package/dist/training/experiments/repair/import.js.map +1 -0
- package/dist/training/experiments/repair/index.d.ts +6 -0
- package/dist/training/experiments/repair/index.js +5 -0
- package/dist/training/experiments/repair/index.js.map +1 -0
- package/dist/training/experiments/repair/measurement.d.ts +14 -0
- package/dist/training/experiments/repair/measurement.js +101 -0
- package/dist/training/experiments/repair/measurement.js.map +1 -0
- package/dist/training/experiments/repair/projection.d.ts +3 -0
- package/dist/training/experiments/repair/projection.js +63 -0
- package/dist/training/experiments/repair/projection.js.map +1 -0
- package/dist/training/experiments/repair/receipt.d.ts +43 -0
- package/dist/training/experiments/repair/receipt.js +52 -0
- package/dist/training/experiments/repair/receipt.js.map +1 -0
- package/dist/training/experiments/repair/record.d.ts +4 -0
- package/dist/training/experiments/repair/record.js +24 -0
- package/dist/training/experiments/repair/record.js.map +1 -0
- package/dist/training/experiments/replay/fixture.d.ts +10 -0
- package/dist/training/experiments/replay/fixture.js +53 -0
- package/dist/training/experiments/replay/fixture.js.map +1 -0
- package/dist/training/experiments/replay/index.d.ts +3 -0
- package/dist/training/experiments/replay/index.js +3 -0
- package/dist/training/experiments/replay/index.js.map +1 -0
- package/dist/training/experiments/replay/rescore.d.ts +3 -0
- package/dist/training/experiments/replay/rescore.js +63 -0
- package/dist/training/experiments/replay/rescore.js.map +1 -0
- package/dist/training/experiments/replay/schema.d.ts +408 -0
- package/dist/training/experiments/replay/schema.js +126 -0
- package/dist/training/experiments/replay/schema.js.map +1 -0
- package/dist/training/experiments/replay/source.d.ts +6 -0
- package/dist/training/experiments/replay/source.js +35 -0
- package/dist/training/experiments/replay/source.js.map +1 -0
- package/dist/training/experiments/replay/subject-usage.d.ts +3 -0
- package/dist/training/experiments/replay/subject-usage.js +18 -0
- package/dist/training/experiments/replay/subject-usage.js.map +1 -0
- package/dist/training/experiments/replay/types.d.ts +63 -0
- package/dist/training/experiments/replay/types.js +2 -0
- package/dist/training/experiments/replay/types.js.map +1 -0
- package/dist/training/experiments/training/checkpoint.d.ts +51 -0
- package/dist/training/experiments/training/checkpoint.js +200 -0
- package/dist/training/experiments/training/checkpoint.js.map +1 -0
- package/dist/training/experiments/training/criteria.d.ts +3 -0
- package/dist/training/experiments/training/criteria.js +23 -0
- package/dist/training/experiments/training/criteria.js.map +1 -0
- package/dist/training/experiments/training/feedback.d.ts +15 -0
- package/dist/training/experiments/training/feedback.js +133 -0
- package/dist/training/experiments/training/feedback.js.map +1 -0
- package/dist/training/experiments/training/held-out.d.ts +24 -0
- package/dist/training/experiments/training/held-out.js +41 -0
- package/dist/training/experiments/training/held-out.js.map +1 -0
- package/dist/training/experiments/training/index.d.ts +7 -0
- package/dist/training/experiments/training/index.js +7 -0
- package/dist/training/experiments/training/index.js.map +1 -0
- package/dist/training/experiments/training/partition.d.ts +20 -0
- package/dist/training/experiments/training/partition.js +24 -0
- package/dist/training/experiments/training/partition.js.map +1 -0
- package/dist/training/experiments/training/recovery.d.ts +17 -0
- package/dist/training/experiments/training/recovery.js +54 -0
- package/dist/training/experiments/training/recovery.js.map +1 -0
- package/dist/training/experiments/training/retry.d.ts +14 -0
- package/dist/training/experiments/training/retry.js +16 -0
- package/dist/training/experiments/training/retry.js.map +1 -0
- package/dist/training/experiments/training/scoring.d.ts +4 -0
- package/dist/training/experiments/training/scoring.js +42 -0
- package/dist/training/experiments/training/scoring.js.map +1 -0
- package/dist/training/experiments/training/training.d.ts +85 -0
- package/dist/training/experiments/training/training.js +208 -0
- package/dist/training/experiments/training/training.js.map +1 -0
- package/dist/training/experiments/trials/bindings.d.ts +4 -0
- package/dist/training/experiments/trials/bindings.js +46 -0
- package/dist/training/experiments/trials/bindings.js.map +1 -0
- package/dist/training/experiments/trials/closure.d.ts +2 -0
- package/dist/training/experiments/trials/closure.js +26 -0
- package/dist/training/experiments/trials/closure.js.map +1 -0
- package/dist/training/experiments/trials/control.d.ts +13 -0
- package/dist/training/experiments/trials/control.js +74 -0
- package/dist/training/experiments/trials/control.js.map +1 -0
- package/dist/training/experiments/trials/diagnostics.d.ts +25 -0
- package/dist/training/experiments/trials/diagnostics.js +99 -0
- package/dist/training/experiments/trials/diagnostics.js.map +1 -0
- package/dist/training/experiments/trials/files.d.ts +8 -0
- package/dist/training/experiments/trials/files.js +113 -0
- package/dist/training/experiments/trials/files.js.map +1 -0
- package/dist/training/experiments/trials/fixture.d.ts +13 -0
- package/dist/training/experiments/trials/fixture.js +40 -0
- package/dist/training/experiments/trials/fixture.js.map +1 -0
- package/dist/training/experiments/trials/index.d.ts +5 -0
- package/dist/training/experiments/trials/index.js +5 -0
- package/dist/training/experiments/trials/index.js.map +1 -0
- package/dist/training/experiments/trials/live.d.ts +20 -0
- package/dist/training/experiments/trials/live.js +68 -0
- package/dist/training/experiments/trials/live.js.map +1 -0
- package/dist/training/experiments/trials/paths.d.ts +5 -0
- package/dist/training/experiments/trials/paths.js +90 -0
- package/dist/training/experiments/trials/paths.js.map +1 -0
- package/dist/training/experiments/trials/preflight.d.ts +3 -0
- package/dist/training/experiments/trials/preflight.js +23 -0
- package/dist/training/experiments/trials/preflight.js.map +1 -0
- package/dist/training/experiments/trials/stored.d.ts +23 -0
- package/dist/training/experiments/trials/stored.js +46 -0
- package/dist/training/experiments/trials/stored.js.map +1 -0
- package/dist/training/experiments/trials/subject-usage.d.ts +7 -0
- package/dist/training/experiments/trials/subject-usage.js +66 -0
- package/dist/training/experiments/trials/subject-usage.js.map +1 -0
- package/dist/training/experiments/trials/trial.d.ts +11 -0
- package/dist/training/experiments/trials/trial.js +361 -0
- package/dist/training/experiments/trials/trial.js.map +1 -0
- package/dist/training/experiments/trials/types.d.ts +124 -0
- package/dist/training/experiments/trials/types.js +2 -0
- package/dist/training/experiments/trials/types.js.map +1 -0
- package/dist/training/graders/admission.d.ts +11 -0
- package/dist/training/graders/admission.js +22 -0
- package/dist/training/graders/admission.js.map +1 -0
- package/dist/training/graders/aggregation.d.ts +30 -0
- package/dist/training/graders/aggregation.js +71 -0
- package/dist/training/graders/aggregation.js.map +1 -0
- package/dist/training/graders/atomic/aggregate.d.ts +10 -0
- package/dist/training/graders/atomic/aggregate.js +31 -0
- package/dist/training/graders/atomic/aggregate.js.map +1 -0
- package/dist/training/graders/atomic/feedback.d.ts +3 -0
- package/dist/training/graders/atomic/feedback.js +16 -0
- package/dist/training/graders/atomic/feedback.js.map +1 -0
- package/dist/training/graders/atomic/format.d.ts +61 -0
- package/dist/training/graders/atomic/format.js +53 -0
- package/dist/training/graders/atomic/format.js.map +1 -0
- package/dist/training/graders/atomic/index.d.ts +7 -0
- package/dist/training/graders/atomic/index.js +6 -0
- package/dist/training/graders/atomic/index.js.map +1 -0
- package/dist/training/graders/atomic/runtime.d.ts +4 -0
- package/dist/training/graders/atomic/runtime.js +94 -0
- package/dist/training/graders/atomic/runtime.js.map +1 -0
- package/dist/training/graders/atomic/schema.d.ts +5 -0
- package/dist/training/graders/atomic/schema.js +37 -0
- package/dist/training/graders/atomic/schema.js.map +1 -0
- package/dist/training/graders/atomic/types.d.ts +68 -0
- package/dist/training/graders/atomic/types.js +2 -0
- package/dist/training/graders/atomic/types.js.map +1 -0
- package/dist/training/graders/calibration/escalation.d.ts +50 -0
- package/dist/training/graders/calibration/escalation.js +145 -0
- package/dist/training/graders/calibration/escalation.js.map +1 -0
- package/dist/training/graders/calibration/fitting.d.ts +28 -0
- package/dist/training/graders/calibration/fitting.js +121 -0
- package/dist/training/graders/calibration/fitting.js.map +1 -0
- package/dist/training/graders/calibration/index.d.ts +7 -0
- package/dist/training/graders/calibration/index.js +5 -0
- package/dist/training/graders/calibration/index.js.map +1 -0
- package/dist/training/graders/calibration/metrics.d.ts +31 -0
- package/dist/training/graders/calibration/metrics.js +145 -0
- package/dist/training/graders/calibration/metrics.js.map +1 -0
- package/dist/training/graders/calibration/selection.d.ts +13 -0
- package/dist/training/graders/calibration/selection.js +30 -0
- package/dist/training/graders/calibration/selection.js.map +1 -0
- package/dist/training/graders/calibration/types.d.ts +68 -0
- package/dist/training/graders/calibration/types.js +2 -0
- package/dist/training/graders/calibration/types.js.map +1 -0
- package/dist/training/graders/consensus.d.ts +11 -0
- package/dist/training/graders/consensus.js +29 -0
- package/dist/training/graders/consensus.js.map +1 -0
- package/dist/training/graders/decomposition/evidence.d.ts +21 -0
- package/dist/training/graders/decomposition/evidence.js +107 -0
- package/dist/training/graders/decomposition/evidence.js.map +1 -0
- package/dist/training/graders/decomposition/format.d.ts +265 -0
- package/dist/training/graders/decomposition/format.js +98 -0
- package/dist/training/graders/decomposition/format.js.map +1 -0
- package/dist/training/graders/decomposition/index.d.ts +10 -0
- package/dist/training/graders/decomposition/index.js +6 -0
- package/dist/training/graders/decomposition/index.js.map +1 -0
- package/dist/training/graders/decomposition/instantiate.d.ts +26 -0
- package/dist/training/graders/decomposition/instantiate.js +129 -0
- package/dist/training/graders/decomposition/instantiate.js.map +1 -0
- package/dist/training/graders/decomposition/location.d.ts +20 -0
- package/dist/training/graders/decomposition/location.js +75 -0
- package/dist/training/graders/decomposition/location.js.map +1 -0
- package/dist/training/graders/decomposition/metadata.d.ts +8 -0
- package/dist/training/graders/decomposition/metadata.js +14 -0
- package/dist/training/graders/decomposition/metadata.js.map +1 -0
- package/dist/training/graders/decomposition/spans.d.ts +6 -0
- package/dist/training/graders/decomposition/spans.js +18 -0
- package/dist/training/graders/decomposition/spans.js.map +1 -0
- package/dist/training/graders/decomposition/transport.d.ts +76 -0
- package/dist/training/graders/decomposition/transport.js +14 -0
- package/dist/training/graders/decomposition/transport.js.map +1 -0
- package/dist/training/graders/format/binding.d.ts +3 -0
- package/dist/training/graders/format/binding.js +37 -0
- package/dist/training/graders/format/binding.js.map +1 -0
- package/dist/training/graders/format/common.d.ts +13 -0
- package/dist/training/graders/format/common.js +77 -0
- package/dist/training/graders/format/common.js.map +1 -0
- package/dist/training/graders/format/grader.d.ts +236 -0
- package/dist/training/graders/format/grader.js +161 -0
- package/dist/training/graders/format/grader.js.map +1 -0
- package/dist/training/graders/format/index.d.ts +11 -0
- package/dist/training/graders/format/index.js +8 -0
- package/dist/training/graders/format/index.js.map +1 -0
- package/dist/training/graders/format/label-import.d.ts +42 -0
- package/dist/training/graders/format/label-import.js +44 -0
- package/dist/training/graders/format/label-import.js.map +1 -0
- package/dist/training/graders/format/labels.d.ts +45 -0
- package/dist/training/graders/format/labels.js +109 -0
- package/dist/training/graders/format/labels.js.map +1 -0
- package/dist/training/graders/format/questions.d.ts +67 -0
- package/dist/training/graders/format/questions.js +104 -0
- package/dist/training/graders/format/questions.js.map +1 -0
- package/dist/training/graders/format/row-regex.d.ts +2 -0
- package/dist/training/graders/format/row-regex.js +56 -0
- package/dist/training/graders/format/row-regex.js.map +1 -0
- package/dist/training/graders/format/rows.d.ts +15 -0
- package/dist/training/graders/format/rows.js +31 -0
- package/dist/training/graders/format/rows.js.map +1 -0
- package/dist/training/graders/format/sentences.d.ts +2 -0
- package/dist/training/graders/format/sentences.js +30 -0
- package/dist/training/graders/format/sentences.js.map +1 -0
- package/dist/training/graders/identity-calibration.d.ts +4 -0
- package/dist/training/graders/identity-calibration.js +15 -0
- package/dist/training/graders/identity-calibration.js.map +1 -0
- package/dist/training/graders/identity.d.ts +119 -0
- package/dist/training/graders/identity.js +119 -0
- package/dist/training/graders/identity.js.map +1 -0
- package/dist/training/graders/index.d.ts +23 -0
- package/dist/training/graders/index.js +14 -0
- package/dist/training/graders/index.js.map +1 -0
- package/dist/training/graders/instances.d.ts +21 -0
- package/dist/training/graders/instances.js +121 -0
- package/dist/training/graders/instances.js.map +1 -0
- package/dist/training/graders/labeling/import.d.ts +8 -0
- package/dist/training/graders/labeling/import.js +51 -0
- package/dist/training/graders/labeling/import.js.map +1 -0
- package/dist/training/graders/labeling/index.d.ts +3 -0
- package/dist/training/graders/labeling/index.js +3 -0
- package/dist/training/graders/labeling/index.js.map +1 -0
- package/dist/training/graders/labeling/planning.d.ts +39 -0
- package/dist/training/graders/labeling/planning.js +117 -0
- package/dist/training/graders/labeling/planning.js.map +1 -0
- package/dist/training/graders/references/council.d.ts +2 -0
- package/dist/training/graders/references/council.js +182 -0
- package/dist/training/graders/references/council.js.map +1 -0
- package/dist/training/graders/references/index.d.ts +5 -0
- package/dist/training/graders/references/index.js +5 -0
- package/dist/training/graders/references/index.js.map +1 -0
- package/dist/training/graders/references/metrics.d.ts +133 -0
- package/dist/training/graders/references/metrics.js +113 -0
- package/dist/training/graders/references/metrics.js.map +1 -0
- package/dist/training/graders/references/policy.d.ts +10 -0
- package/dist/training/graders/references/policy.js +16 -0
- package/dist/training/graders/references/policy.js.map +1 -0
- package/dist/training/graders/references/protocol.d.ts +11 -0
- package/dist/training/graders/references/protocol.js +63 -0
- package/dist/training/graders/references/protocol.js.map +1 -0
- package/dist/training/graders/references/types.d.ts +92 -0
- package/dist/training/graders/references/types.js +2 -0
- package/dist/training/graders/references/types.js.map +1 -0
- package/dist/training/graders/rows.d.ts +14 -0
- package/dist/training/graders/rows.js +39 -0
- package/dist/training/graders/rows.js.map +1 -0
- package/dist/training/graders/runtime/accounting.d.ts +10 -0
- package/dist/training/graders/runtime/accounting.js +43 -0
- package/dist/training/graders/runtime/accounting.js.map +1 -0
- package/dist/training/graders/runtime/grade.d.ts +3 -0
- package/dist/training/graders/runtime/grade.js +161 -0
- package/dist/training/graders/runtime/grade.js.map +1 -0
- package/dist/training/graders/runtime/identity.d.ts +6 -0
- package/dist/training/graders/runtime/identity.js +24 -0
- package/dist/training/graders/runtime/identity.js.map +1 -0
- package/dist/training/graders/runtime/index.d.ts +5 -0
- package/dist/training/graders/runtime/index.js +5 -0
- package/dist/training/graders/runtime/index.js.map +1 -0
- package/dist/training/graders/runtime/schema.d.ts +29 -0
- package/dist/training/graders/runtime/schema.js +27 -0
- package/dist/training/graders/runtime/schema.js.map +1 -0
- package/dist/training/graders/runtime/teacher.d.ts +8 -0
- package/dist/training/graders/runtime/teacher.js +43 -0
- package/dist/training/graders/runtime/teacher.js.map +1 -0
- package/dist/training/graders/runtime/types.d.ts +48 -0
- package/dist/training/graders/runtime/types.js +2 -0
- package/dist/training/graders/runtime/types.js.map +1 -0
- package/dist/training/graders/traps/generate.d.ts +6 -0
- package/dist/training/graders/traps/generate.js +94 -0
- package/dist/training/graders/traps/generate.js.map +1 -0
- package/dist/training/graders/traps/index.d.ts +3 -0
- package/dist/training/graders/traps/index.js +3 -0
- package/dist/training/graders/traps/index.js.map +1 -0
- package/dist/training/graders/traps/mutations.d.ts +10 -0
- package/dist/training/graders/traps/mutations.js +48 -0
- package/dist/training/graders/traps/mutations.js.map +1 -0
- package/dist/training/graders/traps/types.d.ts +37 -0
- package/dist/training/graders/traps/types.js +2 -0
- package/dist/training/graders/traps/types.js.map +1 -0
- package/dist/training/graders/traps/validation.d.ts +4 -0
- package/dist/training/graders/traps/validation.js +118 -0
- package/dist/training/graders/traps/validation.js.map +1 -0
- package/dist/training/index.d.ts +17 -0
- package/dist/training/index.js +16 -0
- package/dist/training/index.js.map +1 -0
- package/dist/training/inference/claude/index.d.ts +1 -0
- package/dist/training/inference/claude/index.js +2 -0
- package/dist/training/inference/claude/index.js.map +1 -0
- package/dist/training/inference/claude/protocol.d.ts +16 -0
- package/dist/training/inference/claude/protocol.js +72 -0
- package/dist/training/inference/claude/protocol.js.map +1 -0
- package/dist/training/inference/claude/provider.d.ts +9 -0
- package/dist/training/inference/claude/provider.js +153 -0
- package/dist/training/inference/claude/provider.js.map +1 -0
- package/dist/training/inference/claude/refusal.d.ts +2 -0
- package/dist/training/inference/claude/refusal.js +37 -0
- package/dist/training/inference/claude/refusal.js.map +1 -0
- package/dist/training/inference/decision/http.d.ts +14 -0
- package/dist/training/inference/decision/http.js +57 -0
- package/dist/training/inference/decision/http.js.map +1 -0
- package/dist/training/inference/decision/index.d.ts +6 -0
- package/dist/training/inference/decision/index.js +6 -0
- package/dist/training/inference/decision/index.js.map +1 -0
- package/dist/training/inference/decision/jev.d.ts +4 -0
- package/dist/training/inference/decision/jev.js +163 -0
- package/dist/training/inference/decision/jev.js.map +1 -0
- package/dist/training/inference/decision/ollama.d.ts +9 -0
- package/dist/training/inference/decision/ollama.js +101 -0
- package/dist/training/inference/decision/ollama.js.map +1 -0
- package/dist/training/inference/decision/prompt.d.ts +6 -0
- package/dist/training/inference/decision/prompt.js +19 -0
- package/dist/training/inference/decision/prompt.js.map +1 -0
- package/dist/training/inference/decision/teacher.d.ts +7 -0
- package/dist/training/inference/decision/teacher.js +61 -0
- package/dist/training/inference/decision/teacher.js.map +1 -0
- package/dist/training/inference/decision/typesafe/index.d.ts +9 -0
- package/dist/training/inference/decision/typesafe/index.js +82 -0
- package/dist/training/inference/decision/typesafe/index.js.map +1 -0
- package/dist/training/inference/decision/typesafe/transport.d.ts +44 -0
- package/dist/training/inference/decision/typesafe/transport.js +151 -0
- package/dist/training/inference/decision/typesafe/transport.js.map +1 -0
- package/dist/training/inference/decision/typesafe/wire.d.ts +25 -0
- package/dist/training/inference/decision/typesafe/wire.js +49 -0
- package/dist/training/inference/decision/typesafe/wire.js.map +1 -0
- package/dist/training/inference/diagnostic.d.ts +3 -0
- package/dist/training/inference/diagnostic.js +14 -0
- package/dist/training/inference/diagnostic.js.map +1 -0
- package/dist/training/inference/evaluator/index.d.ts +3 -0
- package/dist/training/inference/evaluator/index.js +4 -0
- package/dist/training/inference/evaluator/index.js.map +1 -0
- package/dist/training/inference/evaluator/ledger.d.ts +29 -0
- package/dist/training/inference/evaluator/ledger.js +70 -0
- package/dist/training/inference/evaluator/ledger.js.map +1 -0
- package/dist/training/inference/evaluator/policy.d.ts +39 -0
- package/dist/training/inference/evaluator/policy.js +67 -0
- package/dist/training/inference/evaluator/policy.js.map +1 -0
- package/dist/training/inference/evaluator/provider.d.ts +6 -0
- package/dist/training/inference/evaluator/provider.js +45 -0
- package/dist/training/inference/evaluator/provider.js.map +1 -0
- package/dist/training/inference/failure.d.ts +27 -0
- package/dist/training/inference/failure.js +36 -0
- package/dist/training/inference/failure.js.map +1 -0
- package/dist/training/inference/index.d.ts +8 -0
- package/dist/training/inference/index.js +9 -0
- package/dist/training/inference/index.js.map +1 -0
- package/dist/training/inference/ollama/context.d.ts +2 -0
- package/dist/training/inference/ollama/context.js +28 -0
- package/dist/training/inference/ollama/context.js.map +1 -0
- package/dist/training/inference/ollama/index.d.ts +2 -0
- package/dist/training/inference/ollama/index.js +3 -0
- package/dist/training/inference/ollama/index.js.map +1 -0
- package/dist/training/inference/ollama/provider.d.ts +12 -0
- package/dist/training/inference/ollama/provider.js +100 -0
- package/dist/training/inference/ollama/provider.js.map +1 -0
- package/dist/training/inference/request-cap.d.ts +7 -0
- package/dist/training/inference/request-cap.js +22 -0
- package/dist/training/inference/request-cap.js.map +1 -0
- package/dist/training/inference/response-schema.d.ts +9 -0
- package/dist/training/inference/response-schema.js +54 -0
- package/dist/training/inference/response-schema.js.map +1 -0
- package/dist/training/inference/routing.d.ts +17 -0
- package/dist/training/inference/routing.js +152 -0
- package/dist/training/inference/routing.js.map +1 -0
- package/dist/training/optimization/boundaries.d.ts +8 -0
- package/dist/training/optimization/boundaries.js +31 -0
- package/dist/training/optimization/boundaries.js.map +1 -0
- package/dist/training/optimization/bridge-fixtures.d.ts +13 -0
- package/dist/training/optimization/bridge-fixtures.js +40 -0
- package/dist/training/optimization/bridge-fixtures.js.map +1 -0
- package/dist/training/optimization/bridge-identity.d.ts +8 -0
- package/dist/training/optimization/bridge-identity.js +110 -0
- package/dist/training/optimization/bridge-identity.js.map +1 -0
- package/dist/training/optimization/checkpoint.d.ts +35 -0
- package/dist/training/optimization/checkpoint.js +187 -0
- package/dist/training/optimization/checkpoint.js.map +1 -0
- package/dist/training/optimization/evidence.d.ts +4 -0
- package/dist/training/optimization/evidence.js +65 -0
- package/dist/training/optimization/evidence.js.map +1 -0
- package/dist/training/optimization/host.d.ts +27 -0
- package/dist/training/optimization/host.js +144 -0
- package/dist/training/optimization/host.js.map +1 -0
- package/dist/training/optimization/index.d.ts +8 -0
- package/dist/training/optimization/index.js +6 -0
- package/dist/training/optimization/index.js.map +1 -0
- package/dist/training/optimization/inference.d.ts +4 -0
- package/dist/training/optimization/inference.js +33 -0
- package/dist/training/optimization/inference.js.map +1 -0
- package/dist/training/optimization/population-evidence.d.ts +4 -0
- package/dist/training/optimization/population-evidence.js +152 -0
- package/dist/training/optimization/population-evidence.js.map +1 -0
- package/dist/training/optimization/protocol.d.ts +532 -0
- package/dist/training/optimization/protocol.js +146 -0
- package/dist/training/optimization/protocol.js.map +1 -0
- package/dist/training/optimization/runner.d.ts +5 -0
- package/dist/training/optimization/runner.js +129 -0
- package/dist/training/optimization/runner.js.map +1 -0
- package/dist/training/optimization/testSupport.d.ts +9 -0
- package/dist/training/optimization/testSupport.js +29 -0
- package/dist/training/optimization/testSupport.js.map +1 -0
- package/dist/training/optimization/types.d.ts +73 -0
- package/dist/training/optimization/types.js +11 -0
- package/dist/training/optimization/types.js.map +1 -0
- package/dist/training/platform/index.d.ts +1 -0
- package/dist/training/platform/index.js +2 -0
- package/dist/training/platform/index.js.map +1 -0
- package/dist/training/platform/mutations/index.d.ts +5 -0
- package/dist/training/platform/mutations/index.js +10 -0
- package/dist/training/platform/mutations/index.js.map +1 -0
- package/dist/training/platform/process/index.d.ts +1 -0
- package/dist/training/platform/process/index.js +2 -0
- package/dist/training/platform/process/index.js.map +1 -0
- package/dist/training/platform/process/runner.d.ts +32 -0
- package/dist/training/platform/process/runner.js +190 -0
- package/dist/training/platform/process/runner.js.map +1 -0
- package/dist/training/providers/index.d.ts +2 -0
- package/dist/training/providers/index.js +3 -0
- package/dist/training/providers/index.js.map +1 -0
- package/dist/training/providers/policy.d.ts +23 -0
- package/dist/training/providers/policy.js +27 -0
- package/dist/training/providers/policy.js.map +1 -0
- package/dist/training/providers/registry.d.ts +2 -0
- package/dist/training/providers/registry.js +62 -0
- package/dist/training/providers/registry.js.map +1 -0
- package/dist/training/storage/cache/index.d.ts +2 -0
- package/dist/training/storage/cache/index.js +2 -0
- package/dist/training/storage/cache/index.js.map +1 -0
- package/dist/training/storage/cache/persistent.d.ts +12 -0
- package/dist/training/storage/cache/persistent.js +124 -0
- package/dist/training/storage/cache/persistent.js.map +1 -0
- package/dist/training/storage/cache/schema.d.ts +2 -0
- package/dist/training/storage/cache/schema.js +62 -0
- package/dist/training/storage/cache/schema.js.map +1 -0
- package/dist/training/storage/content.d.ts +13 -0
- package/dist/training/storage/content.js +44 -0
- package/dist/training/storage/content.js.map +1 -0
- package/dist/training/storage/evaluation-json.d.ts +12 -0
- package/dist/training/storage/evaluation-json.js +46 -0
- package/dist/training/storage/evaluation-json.js.map +1 -0
- package/dist/training/storage/filesystem.d.ts +6 -0
- package/dist/training/storage/filesystem.js +95 -0
- package/dist/training/storage/filesystem.js.map +1 -0
- package/dist/training/storage/index.d.ts +6 -0
- package/dist/training/storage/index.js +5 -0
- package/dist/training/storage/index.js.map +1 -0
- package/dist/training/storage/journal.d.ts +11 -0
- package/dist/training/storage/journal.js +81 -0
- package/dist/training/storage/journal.js.map +1 -0
- package/dist/training/storage/permissions.d.ts +5 -0
- package/dist/training/storage/permissions.js +40 -0
- package/dist/training/storage/permissions.js.map +1 -0
- package/dist/training/storage/store.d.ts +2 -0
- package/dist/training/storage/store.js +162 -0
- package/dist/training/storage/store.js.map +1 -0
- package/dist/training/storage/types.d.ts +67 -0
- package/dist/training/storage/types.js +2 -0
- package/dist/training/storage/types.js.map +1 -0
- package/dist/training/subjects/environment.d.ts +2 -0
- package/dist/training/subjects/environment.js +14 -0
- package/dist/training/subjects/environment.js.map +1 -0
- package/dist/training/subjects/index.d.ts +3 -0
- package/dist/training/subjects/index.js +3 -0
- package/dist/training/subjects/index.js.map +1 -0
- package/dist/training/subjects/registry.d.ts +15 -0
- package/dist/training/subjects/registry.js +13 -0
- package/dist/training/subjects/registry.js.map +1 -0
- package/dist/training/testing/evidence.d.ts +6 -0
- package/dist/training/testing/evidence.js +21 -0
- package/dist/training/testing/evidence.js.map +1 -0
- package/dist/training/testing/index.d.ts +1 -0
- package/dist/training/testing/index.js +2 -0
- package/dist/training/testing/index.js.map +1 -0
- package/dist/training/viewer/cockpit/activity.d.ts +2 -0
- package/dist/training/viewer/cockpit/activity.js +14 -0
- package/dist/training/viewer/cockpit/activity.js.map +1 -0
- package/dist/training/viewer/cockpit/evidence.d.ts +4 -0
- package/dist/training/viewer/cockpit/evidence.js +58 -0
- package/dist/training/viewer/cockpit/evidence.js.map +1 -0
- package/dist/training/viewer/cockpit/format.d.ts +11 -0
- package/dist/training/viewer/cockpit/format.js +27 -0
- package/dist/training/viewer/cockpit/format.js.map +1 -0
- package/dist/training/viewer/cockpit/index.d.ts +4 -0
- package/dist/training/viewer/cockpit/index.js +5 -0
- package/dist/training/viewer/cockpit/index.js.map +1 -0
- package/dist/training/viewer/cockpit/interaction.d.ts +2 -0
- package/dist/training/viewer/cockpit/interaction.js +103 -0
- package/dist/training/viewer/cockpit/interaction.js.map +1 -0
- package/dist/training/viewer/cockpit/overview.d.ts +4 -0
- package/dist/training/viewer/cockpit/overview.js +62 -0
- package/dist/training/viewer/cockpit/overview.js.map +1 -0
- package/dist/training/viewer/cockpit/page.d.ts +2 -0
- package/dist/training/viewer/cockpit/page.js +39 -0
- package/dist/training/viewer/cockpit/page.js.map +1 -0
- package/dist/training/viewer/cockpit/plots.d.ts +6 -0
- package/dist/training/viewer/cockpit/plots.js +67 -0
- package/dist/training/viewer/cockpit/plots.js.map +1 -0
- package/dist/training/viewer/cockpit/responsive.d.ts +2 -0
- package/dist/training/viewer/cockpit/responsive.js +86 -0
- package/dist/training/viewer/cockpit/responsive.js.map +1 -0
- package/dist/training/viewer/cockpit/styles.d.ts +2 -0
- package/dist/training/viewer/cockpit/styles.js +349 -0
- package/dist/training/viewer/cockpit/styles.js.map +1 -0
- package/dist/training/viewer/costs.d.ts +4 -0
- package/dist/training/viewer/costs.js +81 -0
- package/dist/training/viewer/costs.js.map +1 -0
- package/dist/training/viewer/from-trials.d.ts +33 -0
- package/dist/training/viewer/from-trials.js +96 -0
- package/dist/training/viewer/from-trials.js.map +1 -0
- package/dist/training/viewer/index.d.ts +4 -0
- package/dist/training/viewer/index.js +4 -0
- package/dist/training/viewer/index.js.map +1 -0
- package/dist/training/viewer/metrics/candidates.d.ts +10 -0
- package/dist/training/viewer/metrics/candidates.js +56 -0
- package/dist/training/viewer/metrics/candidates.js.map +1 -0
- package/dist/training/viewer/metrics/index.d.ts +2 -0
- package/dist/training/viewer/metrics/index.js +2 -0
- package/dist/training/viewer/metrics/index.js.map +1 -0
- package/dist/training/viewer/metrics/model.d.ts +4 -0
- package/dist/training/viewer/metrics/model.js +11 -0
- package/dist/training/viewer/metrics/model.js.map +1 -0
- package/dist/training/viewer/metrics/recorded-costs.d.ts +4 -0
- package/dist/training/viewer/metrics/recorded-costs.js +55 -0
- package/dist/training/viewer/metrics/recorded-costs.js.map +1 -0
- package/dist/training/viewer/metrics/scores.d.ts +11 -0
- package/dist/training/viewer/metrics/scores.js +139 -0
- package/dist/training/viewer/metrics/scores.js.map +1 -0
- package/dist/training/viewer/metrics/types.d.ts +95 -0
- package/dist/training/viewer/metrics/types.js +2 -0
- package/dist/training/viewer/metrics/types.js.map +1 -0
- package/dist/training/viewer/page.d.ts +4 -0
- package/dist/training/viewer/page.js +7 -0
- package/dist/training/viewer/page.js.map +1 -0
- package/dist/training/viewer/projection.d.ts +3 -0
- package/dist/training/viewer/projection.js +98 -0
- package/dist/training/viewer/projection.js.map +1 -0
- package/dist/training/viewer/server.d.ts +3 -0
- package/dist/training/viewer/server.js +80 -0
- package/dist/training/viewer/server.js.map +1 -0
- package/dist/training/viewer/stream.d.ts +7 -0
- package/dist/training/viewer/stream.js +70 -0
- package/dist/training/viewer/stream.js.map +1 -0
- package/dist/training/viewer/types.d.ts +172 -0
- package/dist/training/viewer/types.js +2 -0
- package/dist/training/viewer/types.js.map +1 -0
- package/examples/training/AGENTS.md +19 -0
- package/examples/training/fixed-questions/AGENTS.md +7 -0
- package/examples/training/fixed-questions/README.md +19 -0
- package/examples/training/fixed-questions/evals/AGENTS.md +7 -0
- package/examples/training/fixed-questions/evals/delivery.spawnfile-training.yaml +33 -0
- package/examples/training/fixed-questions/evals/item.grader.yaml +41 -0
- package/examples/training/native-formatting/AGENTS.md +10 -0
- package/examples/training/native-formatting/PROMPT.md +4 -0
- package/examples/training/native-formatting/README.md +63 -0
- package/examples/training/native-formatting/SOUL.md +4 -0
- package/examples/training/native-formatting/evals/AGENTS.md +6 -0
- package/examples/training/native-formatting/evals/test.spawnfile-training.yaml +22 -0
- package/examples/training/native-formatting/evals/train.spawnfile-training.yaml +32 -0
- package/examples/training/newsroom-reporting/.spawnignore +4 -0
- package/examples/training/newsroom-reporting/AGENTS.md +24 -0
- package/examples/training/newsroom-reporting/README.md +255 -0
- package/examples/training/newsroom-reporting/Spawnfile +248 -0
- package/examples/training/newsroom-reporting/environment/AGENTS.md +14 -0
- package/examples/training/newsroom-reporting/environment/scribe/AGENTS.md +28 -0
- package/examples/training/newsroom-reporting/environment/scribe/RUNBOOK.md +50 -0
- package/examples/training/newsroom-reporting/environment/scribe/SOUL.md +6 -0
- package/examples/training/newsroom-reporting/evals/AGENTS.md +14 -0
- package/examples/training/newsroom-reporting/evals/test.spawnfile-training.yaml +165 -0
- package/examples/training/newsroom-reporting/evals/train.spawnfile-training.yaml +195 -0
- package/package.json +19 -6
- package/runtimes.yaml +6 -6
- package/dist/cli/paideiaDelegation.js +0 -211
- package/dist/cli/paideiaSupervisor.d.ts +0 -17
- package/dist/cli/paideiaSupervisor.js +0 -48
|
@@ -0,0 +1,67 @@
|
|
|
1
|
+
import { evaluateAtomicGrader } from './atomic-grader.js';
|
|
2
|
+
import { allQuestions, DecisionAccountingError, gradeOutput } from '../graders/index.js';
|
|
3
|
+
import { collectJudgeCalls, result } from './logic/index.js';
|
|
4
|
+
import { selectEvidence } from './selectors.js';
|
|
5
|
+
/** All provider construction belongs to the host; this evaluator only consumes injected routes. */
|
|
6
|
+
export async function evaluateGrader(check, context) {
|
|
7
|
+
const runtime = context.graders?.[check.grader];
|
|
8
|
+
if (!runtime)
|
|
9
|
+
throw new Error(`Unresolved grader runtime ${String(check.grader)}`);
|
|
10
|
+
if ('kind' in runtime && runtime.kind === 'atomic')
|
|
11
|
+
return evaluateAtomicGrader(check, context, runtime);
|
|
12
|
+
const legacy = runtime;
|
|
13
|
+
const binding = check.grader_binding;
|
|
14
|
+
if (!binding || legacy.definitionDigest !== binding.definitionDigest || legacy.calibrationDigest !== binding.calibrationDigest) {
|
|
15
|
+
throw new Error('Grader runtime identity does not match resolved grader/calibration files');
|
|
16
|
+
}
|
|
17
|
+
const selected = selectEvidence(check.input, context.snapshot);
|
|
18
|
+
if (!selected.found)
|
|
19
|
+
return { ...result('unknown', 'Graded artifact was not observed', context.snapshot.closed), evidence: selected.evidence };
|
|
20
|
+
const bound = {};
|
|
21
|
+
const evidence = [...selected.evidence];
|
|
22
|
+
const selectors = (check.context ?? {});
|
|
23
|
+
for (const name of Object.keys(selectors)) {
|
|
24
|
+
if (!Object.hasOwn(legacy.grader.context, name))
|
|
25
|
+
throw new Error(`Unknown grader context ${name}`);
|
|
26
|
+
const selection = selectEvidence(selectors[name], context.snapshot);
|
|
27
|
+
if (selection.found)
|
|
28
|
+
bound[name] = selection.value;
|
|
29
|
+
evidence.push(...selection.evidence);
|
|
30
|
+
}
|
|
31
|
+
// Direct engine clients must not bypass the resolution guarantee; absent trial data
|
|
32
|
+
// stays absent so instance expansion records missing coverage instead of empty evidence.
|
|
33
|
+
const unbound = allQuestions(legacy.grader).flatMap(question => [...question.reads, ...(typeof question.when === 'object' ? [question.when.field] : [])])
|
|
34
|
+
.filter(read => read.startsWith('context.') && !Object.hasOwn(selectors, read.slice(8)));
|
|
35
|
+
if (unbound.length)
|
|
36
|
+
return { ...result('unknown', `Unbound grader context: ${unbound.join(', ')}`, context.snapshot.closed), evidence };
|
|
37
|
+
const children = [];
|
|
38
|
+
const gateOutcomes = {};
|
|
39
|
+
const artifact = context.snapshot.artifacts[check.input.slice(9).split('#')[0]];
|
|
40
|
+
for (const gate of binding.gates) {
|
|
41
|
+
if (!context.evaluateCheck)
|
|
42
|
+
throw new Error('Grader gates require the evaluation engine');
|
|
43
|
+
const evaluated = await context.evaluateCheck(gate.check, artifact);
|
|
44
|
+
children.push({ ...evaluated, id: gate.id });
|
|
45
|
+
evidence.push(...evaluated.evidence);
|
|
46
|
+
gateOutcomes[gate.id] = evaluated.status === 'skipped' ? 'unknown' : evaluated.status;
|
|
47
|
+
if (evaluated.status === 'fail')
|
|
48
|
+
break;
|
|
49
|
+
}
|
|
50
|
+
let graded;
|
|
51
|
+
try {
|
|
52
|
+
graded = await gradeOutput({ ...legacy, output: selected.value, context: bound, gateOutcomes, signal: context.signal });
|
|
53
|
+
}
|
|
54
|
+
catch (error) {
|
|
55
|
+
if (!(error instanceof DecisionAccountingError))
|
|
56
|
+
throw error;
|
|
57
|
+
const judgeCalls = collectJudgeCalls(children);
|
|
58
|
+
return { ...result('error', error.message, context.snapshot.closed), evidence, children, judgeCalls,
|
|
59
|
+
decisionAccounting: error.decisionAccounting,
|
|
60
|
+
usage: judgeCalls.flatMap(call => call.usage ? [call.usage] : []) };
|
|
61
|
+
}
|
|
62
|
+
const judgeCalls = collectJudgeCalls([...children, { ...result(graded.status, 'Grader'), judgeCalls: graded.judgeCalls }]);
|
|
63
|
+
return { ...result(graded.status, `Grader ${legacy.grader.id}; coverage ${graded.coverage}`, context.snapshot.closed),
|
|
64
|
+
score: graded.score, grading: graded, evidence, children, judgeCalls,
|
|
65
|
+
usage: judgeCalls.flatMap(call => call.usage ? [call.usage] : []) };
|
|
66
|
+
}
|
|
67
|
+
//# sourceMappingURL=grader.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"grader.js","sourceRoot":"","sources":["../../../src/training/evaluation/grader.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,oBAAoB,EAAE,MAAM,oBAAoB,CAAC;AAI1D,OAAO,EAAE,YAAY,EAAE,uBAAuB,EAAE,WAAW,EAAoB,MAAM,qBAAqB,CAAC;AAC3G,OAAO,EAAE,iBAAiB,EAAE,MAAM,EAAE,MAAM,kBAAkB,CAAC;AAC7D,OAAO,EAAE,cAAc,EAAE,MAAM,gBAAgB,CAAC;AAEhD,mGAAmG;AACnG,MAAM,CAAC,KAAK,UAAU,cAAc,CAAC,KAAgB,EAAE,OAA0B;IAC/E,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,EAAE,CAAC,KAAK,CAAC,MAAgB,CAAC,CAAC;IAC1D,IAAI,CAAC,OAAO;QAAE,MAAM,IAAI,KAAK,CAAC,6BAA6B,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC;IACnF,IAAI,MAAM,IAAI,OAAO,IAAI,OAAO,CAAC,IAAI,KAAK,QAAQ;QAAE,OAAO,oBAAoB,CAAC,KAAK,EAAE,OAAO,EAAE,OAAO,CAAC,CAAC;IACzG,MAAM,MAAM,GAAG,OAAuD,CAAC;IACvE,MAAM,OAAO,GAAG,KAAK,CAAC,cAA2C,CAAC;IAClE,IAAI,CAAC,OAAO,IAAI,MAAM,CAAC,gBAAgB,KAAK,OAAO,CAAC,gBAAgB,IAAI,MAAM,CAAC,iBAAiB,KAAK,OAAO,CAAC,iBAAiB,EAAE,CAAC;QAC/H,MAAM,IAAI,KAAK,CAAC,0EAA0E,CAAC,CAAC;IAC9F,CAAC;IACD,MAAM,QAAQ,GAAG,cAAc,CAAC,KAAK,CAAC,KAAe,EAAE,OAAO,CAAC,QAAQ,CAAC,CAAC;IACzE,IAAI,CAAC,QAAQ,CAAC,KAAK;QAAE,OAAO,EAAE,GAAG,MAAM,CAAC,SAAS,EAAE,kCAAkC,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,QAAQ,EAAE,QAAQ,CAAC,QAAQ,EAAE,CAAC;IAC/I,MAAM,KAAK,GAA4B,EAAE,CAAC;IAC1C,MAAM,QAAQ,GAAG,CAAC,GAAG,QAAQ,CAAC,QAAQ,CAAC,CAAC;IACxC,MAAM,SAAS,GAAG,CAAC,KAAK,CAAC,OAAO,IAAI,EAAE,CAA2B,CAAC;IAClE,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,IAAI,CAAC,SAAS,CAAC,EAAE,CAAC;QAC1C,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,OAAO,EAAE,IAAI,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,0BAA0B,IAAI,EAAE,CAAC,CAAC;QACnG,MAAM,SAAS,GAAG,cAAc,CAAC,SAAS,CAAC,IAAI,CAAE,EAAE,OAAO,CAAC,QAAQ,CAAC,CAAC;QACrE,IAAI,SAAS,CAAC,KAAK;YAAE,KAAK,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC,KAAK,CAAC;QACnD,QAAQ,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC,QAAQ,CAAC,CAAC;IACvC,CAAC;IACD,oFAAoF;IACpF,yFAAyF;IACzF,MAAM,OAAO,GAAG,YAAY,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC,GAAG,QAAQ,CAAC,KAAK,EAAE,GAAG,CAAC,OAAO,QAAQ,CAAC,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;SACtJ,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,UAAU,CAAC,UAAU,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,SAAS,EAAE,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IAC3F,IAAI,OAAO,CAAC,MAAM;QAAE,OAAO,EAAE,GAAG,MAAM,CAAC,SAAS,EAAE,2BAA2B,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,QAAQ,EAAE,CAAC;IACxI,MAAM,QAAQ,GAAuB,EAAE,CAAC;IACxC,MAAM,YAAY,GAAgC,EAAE,CAAC;IACrD,MAAM,QAAQ,GAAG,OAAO,CAAC,QAAQ,CAAC,SAAS,CAAE,KAAK,CAAC,KAAgB,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC;IAC7F,KAAK,MAAM,IAAI,IAAI,OAAO,CAAC,KAAK,EAAE,CAAC;QACjC,IAAI,CAAC,OAAO,CAAC,aAAa;YAAE,MAAM,IAAI,KAAK,CAAC,4CAA4C,CAAC,CAAC;QAC1F,MAAM,SAAS,GAAG,MAAM,OAAO,CAAC,aAAa,CAAC,IAAI,CAAC,KAAK,EAAE,QAAQ,CAAC,CAAC;QACpE,QAAQ,CAAC,IAAI,CAAC,EAAE,GAAG,SAAS,EAAE,EAAE,EAAE,IAAI,CAAC,EAAE,EAAE,CAAC,CAAC;QAC7C,QAAQ,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC,QAAQ,CAAC,CAAC;QACrC,YAAY,CAAC,IAAI,CAAC,EAAE,CAAC,GAAG,SAAS,CAAC,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,MAAM,CAAC;QACtF,IAAI,SAAS,CAAC,MAAM,KAAK,MAAM;YAAE,MAAM;IACzC,CAAC;IACD,IAAI,MAAM,CAAC;IACX,IAAI,CAAC;QAAC,MAAM,GAAG,MAAM,WAAW,CAAC,EAAE,GAAG,MAAM,EAAE,MAAM,EAAE,QAAQ,CAAC,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,YAAY,EAAE,MAAM,EAAE,OAAO,CAAC,MAAM,EAAE,CAAC,CAAC;IAAC,CAAC;IAChI,OAAO,KAAK,EAAE,CAAC;QACb,IAAI,CAAC,CAAC,KAAK,YAAY,uBAAuB,CAAC;YAAE,MAAM,KAAK,CAAC;QAC7D,MAAM,UAAU,GAAG,iBAAiB,CAAC,QAAQ,CAAC,CAAC;QAC/C,OAAO,EAAE,GAAG,MAAM,CAAC,OAAO,EAAE,KAAK,CAAC,OAAO,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,QAAQ,EAAE,QAAQ,EAAE,UAAU;YACjG,kBAAkB,EAAE,KAAK,CAAC,kBAAkB;YAC5C,KAAK,EAAE,UAAU,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC;IACxE,CAAC;IACD,MAAM,UAAU,GAAG,iBAAiB,CAAC,CAAC,GAAG,QAAQ,EAAE,EAAE,GAAG,MAAM,CAAC,MAAM,CAAC,MAAM,EAAE,QAAQ,CAAC,EAAE,UAAU,EAAE,MAAM,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC;IAC3H,OAAO,EAAE,GAAG,MAAM,CAAC,MAAM,CAAC,MAAM,EAAE,UAAU,MAAM,CAAC,MAAM,CAAC,EAAE,cAAc,MAAM,CAAC,QAAQ,EAAE,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC;QACnH,KAAK,EAAE,MAAM,CAAC,KAAK,EAAE,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,QAAQ,EAAE,UAAU;QACpE,KAAK,EAAE,UAAU,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC;AACxE,CAAC"}
|
|
@@ -0,0 +1,10 @@
|
|
|
1
|
+
export { createEvaluationRegistry } from './registry.js';
|
|
2
|
+
export { evaluateNode, evaluatePlan, mayStop } from './engine.js';
|
|
3
|
+
export { selectEvidence, evidenceText, observedAbsence, ownPath } from './selectors.js';
|
|
4
|
+
export { result, combine, negate, weightedScore, checkThreshold } from './logic/index.js';
|
|
5
|
+
export { predicate, bindingValue } from './mechanical/index.js';
|
|
6
|
+
export { createJudgeCache, evidenceUnits } from './semantic/index.js';
|
|
7
|
+
export { renderJudgeTemplate, validateJudgeTemplate } from './semantic/index.js';
|
|
8
|
+
export type { EvidenceUnit, EvidencePacket } from './semantic/index.js';
|
|
9
|
+
export type { PlanMeasurementRepair } from './repair.js';
|
|
10
|
+
export { validateMeasurementRepair } from './repair.js';
|
|
@@ -0,0 +1,9 @@
|
|
|
1
|
+
export { createEvaluationRegistry } from './registry.js';
|
|
2
|
+
export { evaluateNode, evaluatePlan, mayStop } from './engine.js';
|
|
3
|
+
export { selectEvidence, evidenceText, observedAbsence, ownPath } from './selectors.js';
|
|
4
|
+
export { result, combine, negate, weightedScore, checkThreshold } from './logic/index.js';
|
|
5
|
+
export { predicate, bindingValue } from './mechanical/index.js';
|
|
6
|
+
export { createJudgeCache, evidenceUnits } from './semantic/index.js';
|
|
7
|
+
export { renderJudgeTemplate, validateJudgeTemplate } from './semantic/index.js';
|
|
8
|
+
export { validateMeasurementRepair } from './repair.js';
|
|
9
|
+
//# sourceMappingURL=index.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../src/training/evaluation/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,wBAAwB,EAAE,MAAM,eAAe,CAAC;AACzD,OAAO,EAAE,YAAY,EAAE,YAAY,EAAE,OAAO,EAAE,MAAM,aAAa,CAAC;AAClE,OAAO,EAAE,cAAc,EAAE,YAAY,EAAE,eAAe,EAAE,OAAO,EAAE,MAAM,gBAAgB,CAAC;AACxF,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,aAAa,EAAE,cAAc,EAAE,MAAM,kBAAkB,CAAC;AAC1F,OAAO,EAAE,SAAS,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAC;AAChE,OAAO,EAAE,gBAAgB,EAAE,aAAa,EAAE,MAAM,qBAAqB,CAAC;AACtE,OAAO,EAAE,mBAAmB,EAAE,qBAAqB,EAAE,MAAM,qBAAqB,CAAC;AAGjF,OAAO,EAAE,yBAAyB,EAAE,MAAM,aAAa,CAAC"}
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
export { result, negate, combine, weightedScore, checkThreshold, collectJudgeCalls } from './results.js';
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../../src/training/evaluation/logic/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,cAAc,EAAE,iBAAiB,EAAE,MAAM,cAAc,CAAC"}
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
import type { EvaluationResult, JudgeCall, Verdict, WeightedScore } from '../../contracts/evaluation.js';
|
|
2
|
+
export declare function result(status: Verdict, reason: string, final?: boolean): EvaluationResult;
|
|
3
|
+
export declare function negate(value: EvaluationResult): EvaluationResult;
|
|
4
|
+
export declare function collectJudgeCalls(results: readonly EvaluationResult[]): JudgeCall[];
|
|
5
|
+
export declare function combine(operation: 'all' | 'any', children: EvaluationResult[]): EvaluationResult;
|
|
6
|
+
export declare function weightedScore(weights: Readonly<Record<string, number>>, checks: Readonly<Record<string, EvaluationResult>>): WeightedScore;
|
|
7
|
+
export declare function checkThreshold(score: WeightedScore | undefined, minimum: number): EvaluationResult;
|
|
@@ -0,0 +1,70 @@
|
|
|
1
|
+
export function result(status, reason, final = true) {
|
|
2
|
+
return { status, finality: final ? 'final' : 'provisional', reason, evidence: [] };
|
|
3
|
+
}
|
|
4
|
+
export function negate(value) {
|
|
5
|
+
return { ...value, status: value.status === 'pass' ? 'fail' : value.status === 'fail' ? 'pass' : value.status,
|
|
6
|
+
reason: `not: ${value.reason}`, children: [value],
|
|
7
|
+
...(value.score === undefined ? {} : { score: 1 - value.score }) };
|
|
8
|
+
}
|
|
9
|
+
export function collectJudgeCalls(results) {
|
|
10
|
+
const calls = new Map();
|
|
11
|
+
for (const value of results) {
|
|
12
|
+
for (const call of value.judgeCalls ?? [])
|
|
13
|
+
if (!calls.has(call.id))
|
|
14
|
+
calls.set(call.id, call);
|
|
15
|
+
}
|
|
16
|
+
return [...calls.values()];
|
|
17
|
+
}
|
|
18
|
+
export function combine(operation, children) {
|
|
19
|
+
if (children.length === 0)
|
|
20
|
+
throw new Error('A logical composition cannot be empty');
|
|
21
|
+
const decisive = operation === 'all' ? 'fail' : 'pass';
|
|
22
|
+
const unanimous = operation === 'all' ? 'pass' : 'fail';
|
|
23
|
+
const witnesses = children.filter(child => child.status === decisive);
|
|
24
|
+
const status = witnesses.length ? decisive
|
|
25
|
+
: children.every(child => child.status === unanimous) ? unanimous
|
|
26
|
+
: children.some(child => child.status === 'error') ? 'error'
|
|
27
|
+
: children.some(child => child.status === 'unknown') ? 'unknown' : 'skipped';
|
|
28
|
+
const final = witnesses.length ? witnesses.some(child => child.finality === 'final')
|
|
29
|
+
: children.every(child => child.finality === 'final');
|
|
30
|
+
const judgeCalls = collectJudgeCalls(children);
|
|
31
|
+
// A named judge failure nested under all:/any: is the same finding as a top-level
|
|
32
|
+
// one; dropping it here would make composition the new place errors go anonymous.
|
|
33
|
+
const failure = status === 'error' ? children.find(child => child.judgeFailure)?.judgeFailure : undefined;
|
|
34
|
+
return { ...result(status, `${operation}: ${children.map(child => child.status).join(', ')}`, final),
|
|
35
|
+
children, evidence: children.flatMap(child => child.evidence), ...(judgeCalls.length ? { judgeCalls } : {}),
|
|
36
|
+
...(failure ? { judgeFailure: structuredClone(failure) } : {}) };
|
|
37
|
+
}
|
|
38
|
+
export function weightedScore(weights, checks) {
|
|
39
|
+
const entries = Object.entries(weights);
|
|
40
|
+
const total = entries.reduce((sum, [, weight]) => sum + weight, 0);
|
|
41
|
+
if (entries.length === 0 || entries.some(([, weight]) => !Number.isFinite(weight) || weight < 0) ||
|
|
42
|
+
!Number.isFinite(total) || total <= 0) {
|
|
43
|
+
throw new Error('Score weights must be finite, nonnegative with a positive total');
|
|
44
|
+
}
|
|
45
|
+
let lower = 0;
|
|
46
|
+
let assessedWeight = 0;
|
|
47
|
+
for (const [id, weight] of entries) {
|
|
48
|
+
const check = checks[id];
|
|
49
|
+
if (!check || (check.status !== 'pass' && check.status !== 'fail'))
|
|
50
|
+
continue;
|
|
51
|
+
const score = check.score ?? (check.status === 'pass' ? 1 : 0);
|
|
52
|
+
if (!Number.isFinite(score) || score < 0 || score > 1)
|
|
53
|
+
throw new Error('Invalid evaluator score');
|
|
54
|
+
lower += (weight / total) * score;
|
|
55
|
+
assessedWeight += weight / total;
|
|
56
|
+
}
|
|
57
|
+
const missingWeight = Math.max(0, 1 - assessedWeight);
|
|
58
|
+
const finality = entries.every(([id, weight]) => weight === 0 || checks[id]?.finality === 'final')
|
|
59
|
+
? 'final' : 'provisional';
|
|
60
|
+
return { lower, upper: lower + missingWeight, assessedWeight, missingWeight, finality };
|
|
61
|
+
}
|
|
62
|
+
export function checkThreshold(score, minimum) {
|
|
63
|
+
if (!Number.isFinite(minimum) || minimum < 0 || minimum > 1)
|
|
64
|
+
throw new Error('Invalid score threshold');
|
|
65
|
+
if (!score || score.missingWeight > 1e-9)
|
|
66
|
+
return result('unknown', 'Scoring coverage is incomplete', score?.finality === 'final');
|
|
67
|
+
return { ...result(score.lower >= minimum ? 'pass' : 'fail', `Score ${score.lower}; minimum ${minimum}`, score.finality === 'final'),
|
|
68
|
+
score: score.lower };
|
|
69
|
+
}
|
|
70
|
+
//# sourceMappingURL=results.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"results.js","sourceRoot":"","sources":["../../../../src/training/evaluation/logic/results.ts"],"names":[],"mappings":"AAEA,MAAM,UAAU,MAAM,CAAC,MAAe,EAAE,MAAc,EAAE,KAAK,GAAG,IAAI;IAClE,OAAO,EAAE,MAAM,EAAE,QAAQ,EAAE,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,EAAE,EAAE,CAAC;AACrF,CAAC;AAED,MAAM,UAAU,MAAM,CAAC,KAAuB;IAC5C,OAAO,EAAE,GAAG,KAAK,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,KAAK,CAAC,MAAM;QAC3G,MAAM,EAAE,QAAQ,KAAK,CAAC,MAAM,EAAE,EAAE,QAAQ,EAAE,CAAC,KAAK,CAAC;QACjD,GAAG,CAAC,KAAK,CAAC,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,GAAG,KAAK,CAAC,KAAK,EAAE,CAAC,EAAE,CAAC;AACvE,CAAC;AAED,MAAM,UAAU,iBAAiB,CAAC,OAAoC;IACpE,MAAM,KAAK,GAAG,IAAI,GAAG,EAAqB,CAAC;IAC3C,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC5B,KAAK,MAAM,IAAI,IAAI,KAAK,CAAC,UAAU,IAAI,EAAE;YAAE,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;gBAAE,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,EAAE,IAAI,CAAC,CAAC;IAC/F,CAAC;IACD,OAAO,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,CAAC;AAC7B,CAAC;AAED,MAAM,UAAU,OAAO,CAAC,SAAwB,EAAE,QAA4B;IAC5E,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,uCAAuC,CAAC,CAAC;IACpF,MAAM,QAAQ,GAAY,SAAS,KAAK,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC;IAChE,MAAM,SAAS,GAAY,SAAS,KAAK,KAAK,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC;IACjE,MAAM,SAAS,GAAG,QAAQ,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC;IACtE,MAAM,MAAM,GAAG,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC,QAAQ;QACxC,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,CAAC,SAAS;YACjE,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,MAAM,KAAK,OAAO,CAAC,CAAC,CAAC,CAAC,OAAO;gBAC5D,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,MAAM,KAAK,SAAS,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC;IAC/E,MAAM,KAAK,GAAG,SAAS,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,QAAQ,KAAK,OAAO,CAAC;QAClF,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,QAAQ,KAAK,OAAO,CAAC,CAAC;IACxD,MAAM,UAAU,GAAG,iBAAiB,CAAC,QAAQ,CAAC,CAAC;IAC/C,kFAAkF;IAClF,kFAAkF;IAClF,MAAM,OAAO,GAAG,MAAM,KAAK,OAAO,CAAC,CAAC,CAAC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,YAAY,CAAC,EAAE,YAAY,CAAC,CAAC,CAAC,SAAS,CAAC;IAC1G,OAAO,EAAE,GAAG,MAAM,CAAC,MAAM,EAAE,GAAG,SAAS,KAAK,QAAQ,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,EAAE,KAAK,CAAC;QAClG,QAAQ,EAAE,QAAQ,EAAE,QAAQ,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,QAAQ,CAAC,EAAE,GAAG,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,UAAU,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;QAC3G,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,YAAY,EAAE,eAAe,CAAC,OAAO,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC;AACrE,CAAC;AAED,MAAM,UAAU,aAAa,CAAC,OAAyC,EACrE,MAAkD;IAClD,MAAM,OAAO,GAAG,MAAM,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC;IACxC,MAAM,KAAK,GAAG,OAAO,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,MAAM,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,MAAM,EAAE,CAAC,CAAC,CAAC;IACnE,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC,IAAI,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,MAAM,CAAC,EAAE,EAAE,CAAC,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,MAAM,GAAG,CAAC,CAAC;QAC5F,CAAC,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,EAAE,CAAC;QAC1C,MAAM,IAAI,KAAK,CAAC,iEAAiE,CAAC,CAAC;IACrF,CAAC;IACD,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,IAAI,cAAc,GAAG,CAAC,CAAC;IACvB,KAAK,MAAM,CAAC,EAAE,EAAE,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACnC,MAAM,KAAK,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC;QACzB,IAAI,CAAC,KAAK,IAAI,CAAC,KAAK,CAAC,MAAM,KAAK,MAAM,IAAI,KAAK,CAAC,MAAM,KAAK,MAAM,CAAC;YAAE,SAAS;QAC7E,MAAM,KAAK,GAAG,KAAK,CAAC,KAAK,IAAI,CAAC,KAAK,CAAC,MAAM,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAC/D,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,IAAI,KAAK,GAAG,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,yBAAyB,CAAC,CAAC;QAClG,KAAK,IAAI,CAAC,MAAM,GAAG,KAAK,CAAC,GAAG,KAAK,CAAC;QAClC,cAAc,IAAI,MAAM,GAAG,KAAK,CAAC;IACnC,CAAC;IACD,MAAM,aAAa,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,CAAC,GAAG,cAAc,CAAC,CAAC;IACtD,MAAM,QAAQ,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,EAAE,MAAM,CAAC,EAAE,EAAE,CAAC,MAAM,KAAK,CAAC,IAAI,MAAM,CAAC,EAAE,CAAC,EAAE,QAAQ,KAAK,OAAO,CAAC;QAChG,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,aAAa,CAAC;IAC5B,OAAO,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,GAAG,aAAa,EAAE,cAAc,EAAE,aAAa,EAAE,QAAQ,EAAE,CAAC;AAC1F,CAAC;AAED,MAAM,UAAU,cAAc,CAAC,KAAgC,EAAE,OAAe;IAC9E,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,OAAO,GAAG,CAAC,IAAI,OAAO,GAAG,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,yBAAyB,CAAC,CAAC;IACxG,IAAI,CAAC,KAAK,IAAI,KAAK,CAAC,aAAa,GAAG,IAAI;QAAE,OAAO,MAAM,CAAC,SAAS,EAAE,gCAAgC,EAAE,KAAK,EAAE,QAAQ,KAAK,OAAO,CAAC,CAAC;IAClI,OAAO,EAAE,GAAG,MAAM,CAAC,KAAK,CAAC,KAAK,IAAI,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,KAAK,aAAa,OAAO,EAAE,EAAE,KAAK,CAAC,QAAQ,KAAK,OAAO,CAAC;QAClI,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC;AACzB,CAAC"}
|
|
@@ -0,0 +1,6 @@
|
|
|
1
|
+
export { evaluateText } from './text.js';
|
|
2
|
+
export { evaluateTool } from './tools.js';
|
|
3
|
+
export { evaluateSequence } from './sequence.js';
|
|
4
|
+
export { evaluateTokenEfficiency } from './token-efficiency.js';
|
|
5
|
+
export { predicate, bindingValue } from './predicates.js';
|
|
6
|
+
export { evaluateTextCheck } from './text-checks.js';
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
export { evaluateText } from './text.js';
|
|
2
|
+
export { evaluateTool } from './tools.js';
|
|
3
|
+
export { evaluateSequence } from './sequence.js';
|
|
4
|
+
export { evaluateTokenEfficiency } from './token-efficiency.js';
|
|
5
|
+
export { predicate, bindingValue } from './predicates.js';
|
|
6
|
+
export { evaluateTextCheck } from './text-checks.js';
|
|
7
|
+
//# sourceMappingURL=index.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"index.js","sourceRoot":"","sources":["../../../../src/training/evaluation/mechanical/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,YAAY,EAAE,MAAM,WAAW,CAAC;AACzC,OAAO,EAAE,YAAY,EAAE,MAAM,YAAY,CAAC;AAC1C,OAAO,EAAE,gBAAgB,EAAE,MAAM,eAAe,CAAC;AACjD,OAAO,EAAE,uBAAuB,EAAE,MAAM,uBAAuB,CAAC;AAChE,OAAO,EAAE,SAAS,EAAE,YAAY,EAAE,MAAM,iBAAiB,CAAC;AAC1D,OAAO,EAAE,iBAAiB,EAAE,MAAM,kBAAkB,CAAC"}
|
|
@@ -0,0 +1,27 @@
|
|
|
1
|
+
import { Environment } from '@marcbachmann/cel-js';
|
|
2
|
+
import { ownPath } from '../selectors.js';
|
|
3
|
+
const environment = new Environment({ limits: { maxAstNodes: 1_000, maxDepth: 40,
|
|
4
|
+
maxListElements: 1_000, maxMapEntries: 1_000, maxCallArguments: 32 } })
|
|
5
|
+
.registerVariable('call', 'dyn').registerVariable('context', 'dyn')
|
|
6
|
+
.registerVariable('bindings', 'dyn').registerVariable('event', 'dyn');
|
|
7
|
+
export function predicate(expression, values) {
|
|
8
|
+
if (expression.length > 16_384)
|
|
9
|
+
throw new Error('CEL expression exceeds limit');
|
|
10
|
+
const parsed = environment.parse(expression);
|
|
11
|
+
try {
|
|
12
|
+
const value = parsed(values);
|
|
13
|
+
if (typeof value !== 'boolean')
|
|
14
|
+
throw new Error('CEL predicate must return boolean');
|
|
15
|
+
return value;
|
|
16
|
+
}
|
|
17
|
+
catch (error) {
|
|
18
|
+
// Missing observation data is uncertainty, not a successful match or invented field.
|
|
19
|
+
if (error instanceof Error && /no such key|not found|undefined|undeclared|missing/i.test(error.message))
|
|
20
|
+
return undefined;
|
|
21
|
+
throw error;
|
|
22
|
+
}
|
|
23
|
+
}
|
|
24
|
+
export function bindingValue(selector, values) {
|
|
25
|
+
return ownPath(values, selector.split('.'));
|
|
26
|
+
}
|
|
27
|
+
//# sourceMappingURL=predicates.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"predicates.js","sourceRoot":"","sources":["../../../../src/training/evaluation/mechanical/predicates.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAC;AACnD,OAAO,EAAE,OAAO,EAAE,MAAM,iBAAiB,CAAC;AAE1C,MAAM,WAAW,GAAG,IAAI,WAAW,CAAC,EAAE,MAAM,EAAE,EAAE,WAAW,EAAE,KAAK,EAAE,QAAQ,EAAE,EAAE;QAC9E,eAAe,EAAE,KAAK,EAAE,aAAa,EAAE,KAAK,EAAE,gBAAgB,EAAE,EAAE,EAAE,EAAE,CAAC;KACtE,gBAAgB,CAAC,MAAM,EAAE,KAAK,CAAC,CAAC,gBAAgB,CAAC,SAAS,EAAE,KAAK,CAAC;KAClE,gBAAgB,CAAC,UAAU,EAAE,KAAK,CAAC,CAAC,gBAAgB,CAAC,OAAO,EAAE,KAAK,CAAC,CAAC;AAExE,MAAM,UAAU,SAAS,CAAC,UAAkB,EAAE,MAA+B;IAC3E,IAAI,UAAU,CAAC,MAAM,GAAG,MAAM;QAAE,MAAM,IAAI,KAAK,CAAC,8BAA8B,CAAC,CAAC;IAChF,MAAM,MAAM,GAAG,WAAW,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC;IAC7C,IAAI,CAAC;QACH,MAAM,KAAK,GAAY,MAAM,CAAC,MAAM,CAAC,CAAC;QACtC,IAAI,OAAO,KAAK,KAAK,SAAS;YAAE,MAAM,IAAI,KAAK,CAAC,mCAAmC,CAAC,CAAC;QACrF,OAAO,KAAK,CAAC;IACf,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QACf,qFAAqF;QACrF,IAAI,KAAK,YAAY,KAAK,IAAI,qDAAqD,CAAC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC;YAAE,OAAO,SAAS,CAAC;QAC1H,MAAM,KAAK,CAAC;IACd,CAAC;AACH,CAAC;AAED,MAAM,UAAU,YAAY,CAAC,QAAgB,EAAE,MAA+B;IAC5E,OAAO,OAAO,CAAC,MAAM,EAAE,QAAQ,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC;AAC9C,CAAC"}
|
|
@@ -0,0 +1,75 @@
|
|
|
1
|
+
import { capturedCalls, completeToolScope, happensBefore } from '../../capture/index.js';
|
|
2
|
+
import { result } from '../logic/index.js';
|
|
3
|
+
import { bindingValue, predicate } from './predicates.js';
|
|
4
|
+
export async function evaluateSequence(check, context) {
|
|
5
|
+
const { snapshot } = context;
|
|
6
|
+
const steps = check.steps;
|
|
7
|
+
const ending = check.ending_at;
|
|
8
|
+
const artifact = ending ? snapshot.artifacts[ending.artifact] : undefined;
|
|
9
|
+
const accepted = artifact?.acceptedCallId;
|
|
10
|
+
const evidence = artifact ? [{ kind: 'artifact', id: artifact.checkpointId, digest: artifact.digest }] : [];
|
|
11
|
+
if (ending && !accepted)
|
|
12
|
+
return { ...result('unknown', 'Accepted artifact has no authentic call lineage', snapshot.closed), evidence };
|
|
13
|
+
const calls = capturedCalls(snapshot.events).filter(call => call.success === true && call.dispatch && call.completion)
|
|
14
|
+
.sort((a, b) => (a.dispatch.source < b.dispatch.source ? -1 : a.dispatch.source > b.dispatch.source ? 1 : 0) || a.dispatch.seq - b.dispatch.seq);
|
|
15
|
+
const matches = [];
|
|
16
|
+
let uncertain = false;
|
|
17
|
+
let visited = 0;
|
|
18
|
+
function search(index, prior, bindings) {
|
|
19
|
+
if (++visited > 20_000)
|
|
20
|
+
throw new Error('Sequence matching budget exceeded');
|
|
21
|
+
const step = steps[index];
|
|
22
|
+
for (const call of calls) {
|
|
23
|
+
if (call.tool !== step.tool || prior.some(item => item.id === call.id))
|
|
24
|
+
continue;
|
|
25
|
+
if (index === steps.length - 1 && accepted && call.id !== accepted)
|
|
26
|
+
continue;
|
|
27
|
+
const previous = prior.at(-1);
|
|
28
|
+
if (previous) {
|
|
29
|
+
const order = happensBefore(previous.completion, call.dispatch, snapshot.events, snapshot.coverage);
|
|
30
|
+
if (order === undefined)
|
|
31
|
+
uncertain = true;
|
|
32
|
+
if (order !== true)
|
|
33
|
+
continue;
|
|
34
|
+
}
|
|
35
|
+
const values = { call, context: snapshot.context, bindings };
|
|
36
|
+
const pass = step.where ? predicate(step.where.cel, values) : true;
|
|
37
|
+
if (pass === undefined)
|
|
38
|
+
uncertain = true;
|
|
39
|
+
if (pass !== true)
|
|
40
|
+
continue;
|
|
41
|
+
const nextBindings = { ...bindings };
|
|
42
|
+
let missing = false;
|
|
43
|
+
for (const [name, selector] of Object.entries(step.bind ?? {})) {
|
|
44
|
+
const value = bindingValue(selector, values);
|
|
45
|
+
if (value === undefined) {
|
|
46
|
+
uncertain = true;
|
|
47
|
+
missing = true;
|
|
48
|
+
break;
|
|
49
|
+
}
|
|
50
|
+
nextBindings[name] = value;
|
|
51
|
+
}
|
|
52
|
+
if (missing)
|
|
53
|
+
continue;
|
|
54
|
+
const tuple = [...prior, call];
|
|
55
|
+
if (index + 1 === steps.length)
|
|
56
|
+
matches.push(tuple);
|
|
57
|
+
else
|
|
58
|
+
search(index + 1, tuple, nextBindings);
|
|
59
|
+
}
|
|
60
|
+
}
|
|
61
|
+
if (!steps.length)
|
|
62
|
+
throw new Error('Sequence must have steps');
|
|
63
|
+
search(0, [], {});
|
|
64
|
+
if (matches.length) {
|
|
65
|
+
// Every matching tuple is retained; the first witness is deterministic by native source/sequence.
|
|
66
|
+
return { ...result('pass', `${matches.length} valid ordered tuple(s)`, !ending || snapshot.closed), score: 1,
|
|
67
|
+
evidence: [...evidence, ...matches.flatMap(tuple => tuple.flatMap(call => [
|
|
68
|
+
{ kind: 'event', id: call.dispatch.id }, { kind: 'event', id: call.completion.id },
|
|
69
|
+
]))] };
|
|
70
|
+
}
|
|
71
|
+
const complete = completeToolScope(snapshot.closed, snapshot.coverage);
|
|
72
|
+
return { ...result(complete && !uncertain ? 'fail' : 'unknown', uncertain
|
|
73
|
+
? 'Call ordering or predicate evidence is incomplete' : 'No successful bound sequence', snapshot.closed), evidence };
|
|
74
|
+
}
|
|
75
|
+
//# sourceMappingURL=sequence.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"sequence.js","sourceRoot":"","sources":["../../../../src/training/evaluation/mechanical/sequence.ts"],"names":[],"mappings":"AAEA,OAAO,EAAE,aAAa,EAAE,iBAAiB,EAAE,aAAa,EAAqB,MAAM,wBAAwB,CAAC;AAC5G,OAAO,EAAE,MAAM,EAAE,MAAM,mBAAmB,CAAC;AAC3C,OAAO,EAAE,YAAY,EAAE,SAAS,EAAE,MAAM,iBAAiB,CAAC;AAI1D,MAAM,CAAC,KAAK,UAAU,gBAAgB,CAAC,KAAgB,EAAE,OAA0B;IACjF,MAAM,EAAE,QAAQ,EAAE,GAAG,OAAO,CAAC;IAC7B,MAAM,KAAK,GAAG,KAAK,CAAC,KAAe,CAAC;IACpC,MAAM,MAAM,GAAG,KAAK,CAAC,SAA4D,CAAC;IAClF,MAAM,QAAQ,GAAG,MAAM,CAAC,CAAC,CAAC,QAAQ,CAAC,SAAS,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;IAC1E,MAAM,QAAQ,GAAG,QAAQ,EAAE,cAAc,CAAC;IAC1C,MAAM,QAAQ,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,UAAmB,EAAE,EAAE,EAAE,QAAQ,CAAC,YAAY,EAAE,MAAM,EAAE,QAAQ,CAAC,MAAM,EAAE,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;IACrH,IAAI,MAAM,IAAI,CAAC,QAAQ;QAAE,OAAO,EAAE,GAAG,MAAM,CAAC,SAAS,EAAE,iDAAiD,EAAE,QAAQ,CAAC,MAAM,CAAC,EAAE,QAAQ,EAAE,CAAC;IACvI,MAAM,KAAK,GAAG,aAAa,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,OAAO,KAAK,IAAI,IAAI,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAC,UAAU,CAAC;SACnH,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,QAAS,CAAC,MAAM,GAAG,CAAC,CAAC,QAAS,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,QAAS,CAAC,MAAM,GAAG,CAAC,CAAC,QAAS,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,QAAS,CAAC,GAAG,GAAG,CAAC,CAAC,QAAS,CAAC,GAAG,CAAC,CAAC;IACzJ,MAAM,OAAO,GAAqB,EAAE,CAAC;IACrC,IAAI,SAAS,GAAG,KAAK,CAAC;IACtB,IAAI,OAAO,GAAG,CAAC,CAAC;IAChB,SAAS,MAAM,CAAC,KAAa,EAAE,KAAqB,EAAE,QAAiC;QACrF,IAAI,EAAE,OAAO,GAAG,MAAM;YAAE,MAAM,IAAI,KAAK,CAAC,mCAAmC,CAAC,CAAC;QAC7E,MAAM,IAAI,GAAG,KAAK,CAAC,KAAK,CAAE,CAAC;QAC3B,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;YACzB,IAAI,IAAI,CAAC,IAAI,KAAK,IAAI,CAAC,IAAI,IAAI,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,KAAK,IAAI,CAAC,EAAE,CAAC;gBAAE,SAAS;YACjF,IAAI,KAAK,KAAK,KAAK,CAAC,MAAM,GAAG,CAAC,IAAI,QAAQ,IAAI,IAAI,CAAC,EAAE,KAAK,QAAQ;gBAAE,SAAS;YAC7E,MAAM,QAAQ,GAAG,KAAK,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;YAC9B,IAAI,QAAQ,EAAE,CAAC;gBACb,MAAM,KAAK,GAAG,aAAa,CAAC,QAAQ,CAAC,UAAW,EAAE,IAAI,CAAC,QAAS,EAAE,QAAQ,CAAC,MAAM,EAAE,QAAQ,CAAC,QAAQ,CAAC,CAAC;gBACtG,IAAI,KAAK,KAAK,SAAS;oBAAE,SAAS,GAAG,IAAI,CAAC;gBAC1C,IAAI,KAAK,KAAK,IAAI;oBAAE,SAAS;YAC/B,CAAC;YACD,MAAM,MAAM,GAAG,EAAE,IAAI,EAAE,OAAO,EAAE,QAAQ,CAAC,OAAO,EAAE,QAAQ,EAAE,CAAC;YAC7D,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC;YACnE,IAAI,IAAI,KAAK,SAAS;gBAAE,SAAS,GAAG,IAAI,CAAC;YACzC,IAAI,IAAI,KAAK,IAAI;gBAAE,SAAS;YAC5B,MAAM,YAAY,GAAG,EAAE,GAAG,QAAQ,EAAE,CAAC;YACrC,IAAI,OAAO,GAAG,KAAK,CAAC;YACpB,KAAK,MAAM,CAAC,IAAI,EAAE,QAAQ,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,IAAI,CAAC,IAAI,IAAI,EAAE,CAAC,EAAE,CAAC;gBAC/D,MAAM,KAAK,GAAG,YAAY,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC;gBAC7C,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;oBAAC,SAAS,GAAG,IAAI,CAAC;oBAAC,OAAO,GAAG,IAAI,CAAC;oBAAC,MAAM;gBAAC,CAAC;gBACrE,YAAY,CAAC,IAAI,CAAC,GAAG,KAAK,CAAC;YAC7B,CAAC;YACD,IAAI,OAAO;gBAAE,SAAS;YACtB,MAAM,KAAK,GAAG,CAAC,GAAG,KAAK,EAAE,IAAI,CAAC,CAAC;YAC/B,IAAI,KAAK,GAAG,CAAC,KAAK,KAAK,CAAC,MAAM;gBAAE,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;;gBAC/C,MAAM,CAAC,KAAK,GAAG,CAAC,EAAE,KAAK,EAAE,YAAY,CAAC,CAAC;QAC9C,CAAC;IACH,CAAC;IACD,IAAI,CAAC,KAAK,CAAC,MAAM;QAAE,MAAM,IAAI,KAAK,CAAC,0BAA0B,CAAC,CAAC;IAC/D,MAAM,CAAC,CAAC,EAAE,EAAE,EAAE,EAAE,CAAC,CAAC;IAClB,IAAI,OAAO,CAAC,MAAM,EAAE,CAAC;QACnB,kGAAkG;QAClG,OAAO,EAAE,GAAG,MAAM,CAAC,MAAM,EAAE,GAAG,OAAO,CAAC,MAAM,yBAAyB,EAAE,CAAC,MAAM,IAAI,QAAQ,CAAC,MAAM,CAAC,EAAE,KAAK,EAAE,CAAC;YAC1G,QAAQ,EAAE,CAAC,GAAG,QAAQ,EAAE,GAAG,OAAO,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,OAAO,CAAC,IAAI,CAAC,EAAE,CAAC;oBACxE,EAAE,IAAI,EAAE,OAAgB,EAAE,EAAE,EAAE,IAAI,CAAC,QAAS,CAAC,EAAE,EAAE,EAAE,EAAE,IAAI,EAAE,OAAgB,EAAE,EAAE,EAAE,IAAI,CAAC,UAAW,CAAC,EAAE,EAAE;iBACvG,CAAC,CAAC,CAAC,EAAE,CAAC;IACX,CAAC;IACD,MAAM,QAAQ,GAAG,iBAAiB,CAAC,QAAQ,CAAC,MAAM,EAAE,QAAQ,CAAC,QAAQ,CAAC,CAAC;IACvE,OAAO,EAAE,GAAG,MAAM,CAAC,QAAQ,IAAI,CAAC,SAAS,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,EAAE,SAAS;YACvE,CAAC,CAAC,mDAAmD,CAAC,CAAC,CAAC,8BAA8B,EAAE,QAAQ,CAAC,MAAM,CAAC,EAAE,QAAQ,EAAE,CAAC;AACzH,CAAC"}
|
|
@@ -0,0 +1,65 @@
|
|
|
1
|
+
import { isTextSelector, textSelectors } from '../../cases/index.js';
|
|
2
|
+
import { result } from '../logic/index.js';
|
|
3
|
+
import { observedAbsence, selectEvidence, textReference } from '../selectors.js';
|
|
4
|
+
import { extractedNumbers, normalizeQuote, quotedSpans, stripIgnored } from './text-values.js';
|
|
5
|
+
function textLeaves(value) {
|
|
6
|
+
if (typeof value === 'string')
|
|
7
|
+
return [value];
|
|
8
|
+
if (typeof value === 'number' && Number.isFinite(value))
|
|
9
|
+
return [String(value)];
|
|
10
|
+
if (Array.isArray(value))
|
|
11
|
+
return value.flatMap(textLeaves);
|
|
12
|
+
if (value && typeof value === 'object')
|
|
13
|
+
return Object.values(value).flatMap(textLeaves);
|
|
14
|
+
return [];
|
|
15
|
+
}
|
|
16
|
+
async function selectTexts(selectors, context, sources = false) {
|
|
17
|
+
const texts = [], evidence = [];
|
|
18
|
+
for (const selector of selectors) {
|
|
19
|
+
if (sources && !isTextSelector(selector)) {
|
|
20
|
+
const text = await context.readText(selector);
|
|
21
|
+
texts.push(text);
|
|
22
|
+
evidence.push(textReference('sidecar', selector, text));
|
|
23
|
+
continue;
|
|
24
|
+
}
|
|
25
|
+
const selected = selectEvidence(selector, context.snapshot);
|
|
26
|
+
if (!selected.found) {
|
|
27
|
+
const missingInput = !sources && observedAbsence(selector, context.snapshot);
|
|
28
|
+
return { failure: { ...result(missingInput ? 'fail' : 'unknown', `Missing ${selector}`, context.snapshot.closed),
|
|
29
|
+
...(missingInput ? { score: 0 } : {}), evidence: [...evidence, ...selected.evidence] } };
|
|
30
|
+
}
|
|
31
|
+
const strings = textLeaves(selected.value);
|
|
32
|
+
if (!strings.length)
|
|
33
|
+
return { failure: result('error', `Text selector requires string or number leaves: ${selector}`, context.snapshot.closed) };
|
|
34
|
+
texts.push(...strings);
|
|
35
|
+
evidence.push(...selected.evidence);
|
|
36
|
+
}
|
|
37
|
+
return { text: texts.join('\n'), evidence };
|
|
38
|
+
}
|
|
39
|
+
export async function evaluateTextCheck(check, context) {
|
|
40
|
+
const selected = await selectTexts(textSelectors(check.input), context);
|
|
41
|
+
if ('failure' in selected)
|
|
42
|
+
return selected.failure;
|
|
43
|
+
const ignore = (check.ignore ?? []);
|
|
44
|
+
const input = stripIgnored(selected.text, ignore);
|
|
45
|
+
if (check.eval === 'word-count') {
|
|
46
|
+
const count = input.match(/[\p{L}\p{N}]+(?:['’-][\p{L}\p{N}]+)*/gu)?.length ?? 0;
|
|
47
|
+
const passed = count >= (check.min ?? 0) && count <= (check.max ?? Infinity);
|
|
48
|
+
return { ...result(passed ? 'pass' : 'fail', `word-count: ${count}`, context.snapshot.closed), score: Number(passed),
|
|
49
|
+
evidence: selected.evidence, mechanicalEvidence: { count } };
|
|
50
|
+
}
|
|
51
|
+
const sources = await selectTexts(textSelectors(check.source), context, true);
|
|
52
|
+
if ('failure' in sources)
|
|
53
|
+
return sources.failure;
|
|
54
|
+
const source = stripIgnored(sources.text, ignore);
|
|
55
|
+
const numbers = extractedNumbers(input), quotes = quotedSpans(input);
|
|
56
|
+
const available = new Set(extractedNumbers(source).map(value => value.normalized));
|
|
57
|
+
const missing = numbers.filter(value => !available.has(value.normalized)).map(value => value.raw);
|
|
58
|
+
const normalizedSource = normalizeQuote(source);
|
|
59
|
+
missing.push(...quotes.filter(value => !normalizeQuote(value) || !normalizedSource.includes(normalizeQuote(value))).map(value => `"${value}"`));
|
|
60
|
+
const total = numbers.length + quotes.length, found = total - missing.length;
|
|
61
|
+
return { ...result(missing.length <= (check.max_missing ?? 0) ? 'pass' : 'fail', `values-in-source: ${found}/${total}; missing: ${JSON.stringify(missing)}`, context.snapshot.closed),
|
|
62
|
+
score: total ? found / total : 1, evidence: [...selected.evidence, ...sources.evidence],
|
|
63
|
+
mechanicalEvidence: { found, total, missing } };
|
|
64
|
+
}
|
|
65
|
+
//# sourceMappingURL=text-checks.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"text-checks.js","sourceRoot":"","sources":["../../../../src/training/evaluation/mechanical/text-checks.ts"],"names":[],"mappings":"AAEA,OAAO,EAAE,cAAc,EAAE,aAAa,EAAE,MAAM,sBAAsB,CAAC;AACrE,OAAO,EAAE,MAAM,EAAE,MAAM,mBAAmB,CAAC;AAC3C,OAAO,EAAE,eAAe,EAAE,cAAc,EAAE,aAAa,EAAE,MAAM,iBAAiB,CAAC;AACjF,OAAO,EAAE,gBAAgB,EAAE,cAAc,EAAE,WAAW,EAAE,YAAY,EAAE,MAAM,kBAAkB,CAAC;AAG/F,SAAS,UAAU,CAAC,KAAc;IAChC,IAAI,OAAO,KAAK,KAAK,QAAQ;QAAE,OAAO,CAAC,KAAK,CAAC,CAAC;IAC9C,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC;QAAE,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;IAChF,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;IAC3D,IAAI,KAAK,IAAI,OAAO,KAAK,KAAK,QAAQ;QAAE,OAAO,MAAM,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC;IACxF,OAAO,EAAE,CAAC;AACZ,CAAC;AACD,KAAK,UAAU,WAAW,CAAC,SAAmB,EAAE,OAA0B,EAAE,OAAO,GAAG,KAAK;IACzF,MAAM,KAAK,GAAa,EAAE,EAAE,QAAQ,GAAwB,EAAE,CAAC;IAC/D,KAAK,MAAM,QAAQ,IAAI,SAAS,EAAE,CAAC;QACjC,IAAI,OAAO,IAAI,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,CAAC;YACzC,MAAM,IAAI,GAAG,MAAM,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAC,CAAC;YAC9C,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;YAAC,QAAQ,CAAC,IAAI,CAAC,aAAa,CAAC,SAAS,EAAE,QAAQ,EAAE,IAAI,CAAC,CAAC,CAAC;YAC1E,SAAS;QACX,CAAC;QACD,MAAM,QAAQ,GAAG,cAAc,CAAC,QAAQ,EAAE,OAAO,CAAC,QAAQ,CAAC,CAAC;QAC5D,IAAI,CAAC,QAAQ,CAAC,KAAK,EAAE,CAAC;YACpB,MAAM,YAAY,GAAG,CAAC,OAAO,IAAI,eAAe,CAAC,QAAQ,EAAE,OAAO,CAAC,QAAQ,CAAC,CAAC;YAC7E,OAAO,EAAE,OAAO,EAAE,EAAE,GAAG,MAAM,CAAC,YAAY,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,EAAE,WAAW,QAAQ,EAAE,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC;oBAC9G,GAAG,CAAC,YAAY,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,QAAQ,EAAE,CAAC,GAAG,QAAQ,EAAE,GAAG,QAAQ,CAAC,QAAQ,CAAC,EAAE,EAAE,CAAC;QAC7F,CAAC;QACD,MAAM,OAAO,GAAG,UAAU,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;QAC3C,IAAI,CAAC,OAAO,CAAC,MAAM;YAAE,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,OAAO,EAAE,mDAAmD,QAAQ,EAAE,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;QACjJ,KAAK,CAAC,IAAI,CAAC,GAAG,OAAO,CAAC,CAAC;QAAC,QAAQ,CAAC,IAAI,CAAC,GAAG,QAAQ,CAAC,QAAQ,CAAC,CAAC;IAC9D,CAAC;IACD,OAAO,EAAE,IAAI,EAAE,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,QAAQ,EAAE,CAAC;AAC9C,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,iBAAiB,CAAC,KAAgB,EAAE,OAA0B;IAClF,MAAM,QAAQ,GAAG,MAAM,WAAW,CAAC,aAAa,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,OAAO,CAAC,CAAC;IACxE,IAAI,SAAS,IAAI,QAAQ;QAAE,OAAO,QAAQ,CAAC,OAAO,CAAC;IACnD,MAAM,MAAM,GAAG,CAAC,KAAK,CAAC,MAAM,IAAI,EAAE,CAAa,CAAC;IAChD,MAAM,KAAK,GAAG,YAAY,CAAC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;IAClD,IAAI,KAAK,CAAC,IAAI,KAAK,YAAY,EAAE,CAAC;QAChC,MAAM,KAAK,GAAG,KAAK,CAAC,KAAK,CAAC,wCAAwC,CAAC,EAAE,MAAM,IAAI,CAAC,CAAC;QACjF,MAAM,MAAM,GAAG,KAAK,IAAI,CAAE,KAAK,CAAC,GAA0B,IAAI,CAAC,CAAC,IAAI,KAAK,IAAI,CAAE,KAAK,CAAC,GAA0B,IAAI,QAAQ,CAAC,CAAC;QAC7H,OAAO,EAAE,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,eAAe,KAAK,EAAE,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,KAAK,EAAE,MAAM,CAAC,MAAM,CAAC;YAClH,QAAQ,EAAE,QAAQ,CAAC,QAAQ,EAAE,kBAAkB,EAAE,EAAE,KAAK,EAAE,EAAE,CAAC;IACjE,CAAC;IACD,MAAM,OAAO,GAAG,MAAM,WAAW,CAAC,aAAa,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,OAAO,EAAE,IAAI,CAAC,CAAC;IAC9E,IAAI,SAAS,IAAI,OAAO;QAAE,OAAO,OAAO,CAAC,OAAO,CAAC;IACjD,MAAM,MAAM,GAAG,YAAY,CAAC,OAAO,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;IAClD,MAAM,OAAO,GAAG,gBAAgB,CAAC,KAAK,CAAC,EAAE,MAAM,GAAG,WAAW,CAAC,KAAK,CAAC,CAAC;IACrE,MAAM,SAAS,GAAG,IAAI,GAAG,CAAC,gBAAgB,CAAC,MAAM,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC,CAAC;IACnF,MAAM,OAAO,GAAG,OAAO,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,SAAS,CAAC,GAAG,CAAC,KAAK,CAAC,UAAU,CAAC,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;IAClG,MAAM,gBAAgB,GAAG,cAAc,CAAC,MAAM,CAAC,CAAC;IAChD,OAAO,CAAC,IAAI,CAAC,GAAG,MAAM,CAAC,MAAM,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,cAAc,CAAC,KAAK,CAAC,IAAI,CAAC,gBAAgB,CAAC,QAAQ,CAAC,cAAc,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,IAAI,KAAK,GAAG,CAAC,CAAC,CAAC;IAChJ,MAAM,KAAK,GAAG,OAAO,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM,EAAE,KAAK,GAAG,KAAK,GAAG,OAAO,CAAC,MAAM,CAAC;IAC7E,OAAO,EAAE,GAAG,MAAM,CAAC,OAAO,CAAC,MAAM,IAAI,CAAE,KAAK,CAAC,WAAkC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EACrG,qBAAqB,KAAK,IAAI,KAAK,cAAc,IAAI,CAAC,SAAS,CAAC,OAAO,CAAC,EAAE,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC;QACpG,KAAK,EAAE,KAAK,CAAC,CAAC,CAAC,KAAK,GAAG,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,QAAQ,EAAE,CAAC,GAAG,QAAQ,CAAC,QAAQ,EAAE,GAAG,OAAO,CAAC,QAAQ,CAAC;QACvF,kBAAkB,EAAE,EAAE,KAAK,EAAE,KAAK,EAAE,OAAO,EAAE,EAAE,CAAC;AACpD,CAAC"}
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
export declare function stripIgnored(text: string, patterns: readonly string[]): string;
|
|
2
|
+
export declare function extractedNumbers(text: string): {
|
|
3
|
+
raw: string;
|
|
4
|
+
normalized: string;
|
|
5
|
+
}[];
|
|
6
|
+
export declare function quotedSpans(text: string): string[];
|
|
7
|
+
export declare function normalizeQuote(text: string): string;
|
|
@@ -0,0 +1,28 @@
|
|
|
1
|
+
/** Deliberately lexical: English grouped numbers, dot decimals, units and quoted spans. */
|
|
2
|
+
const currency = { '$': 'USD', '€': 'EUR', '£': 'GBP', '¥': 'JPY' };
|
|
3
|
+
const unit = '(?:USD|EUR|GBP|JPY|[$€£¥])';
|
|
4
|
+
const digits = "(?:\\d{1,3}(?:[, '\\u00a0\\u202f]\\d{3})+|\\d+)(?:\\.\\d+)?";
|
|
5
|
+
const numeric = new RegExp(`(?<![\\p{L}\\p{N}_])(?:${unit}\\s*)?[+−-]?${digits}(?:\\s*(?:%|${unit}))?(?![\\p{L}\\p{N}_])`, 'gu');
|
|
6
|
+
export function stripIgnored(text, patterns) {
|
|
7
|
+
return patterns.reduce((value, pattern) => value.replace(new RegExp(pattern, 'gu'), ' '), text);
|
|
8
|
+
}
|
|
9
|
+
function normalizeNumber(raw) {
|
|
10
|
+
const units = raw.match(/USD|EUR|GBP|JPY|[$€£¥%]/g) ?? [];
|
|
11
|
+
const value = raw.replace(/USD|EUR|GBP|JPY|[$€£¥%]|[, '\u00a0\u202f\s]/g, '').replace('−', '-');
|
|
12
|
+
// String arithmetic preserves integers above Number.MAX_SAFE_INTEGER.
|
|
13
|
+
const negative = value.startsWith('-');
|
|
14
|
+
const [whole, fraction = ''] = value.replace(/^[+-]/, '').split('.');
|
|
15
|
+
const normalized = `${whole.replace(/^0+(?=\d)/, '')}${fraction.replace(/0+$/, '') ? `.${fraction.replace(/0+$/, '')}` : ''}`;
|
|
16
|
+
return `${negative && normalized !== '0' ? '-' : ''}${normalized}:${units.map(item => currency[item] ?? item).join(':')}`;
|
|
17
|
+
}
|
|
18
|
+
export function extractedNumbers(text) {
|
|
19
|
+
return [...text.normalize('NFKC').matchAll(numeric)].map(match => ({ raw: match[0], normalized: normalizeNumber(match[0]) }));
|
|
20
|
+
}
|
|
21
|
+
export function quotedSpans(text) {
|
|
22
|
+
return [...text.matchAll(/"([^"\r\n]+)"|“([^”\r\n]+)”/gu)].map(match => (match[1] ?? match[2]));
|
|
23
|
+
}
|
|
24
|
+
export function normalizeQuote(text) {
|
|
25
|
+
return text.normalize('NFKC').replace(/[‘’]/gu, "'").replace(/[“”]/gu, '"').replace(/[‐‑‒–—−]/gu, '-')
|
|
26
|
+
.replace(/\s+/gu, ' ').replace(/^[\p{P}\s]+|[\p{P}\s]+$/gu, '');
|
|
27
|
+
}
|
|
28
|
+
//# sourceMappingURL=text-values.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"text-values.js","sourceRoot":"","sources":["../../../../src/training/evaluation/mechanical/text-values.ts"],"names":[],"mappings":"AAAA,2FAA2F;AAC3F,MAAM,QAAQ,GAAqC,EAAE,GAAG,EAAE,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,GAAG,EAAE,KAAK,EAAE,CAAC;AACtG,MAAM,IAAI,GAAG,4BAA4B,CAAC;AAC1C,MAAM,MAAM,GAAG,6DAA6D,CAAC;AAC7E,MAAM,OAAO,GAAG,IAAI,MAAM,CAAC,0BAA0B,IAAI,eAAe,MAAM,eAAe,IAAI,wBAAwB,EAAE,IAAI,CAAC,CAAC;AAEjI,MAAM,UAAU,YAAY,CAAC,IAAY,EAAE,QAA2B;IACpE,OAAO,QAAQ,CAAC,MAAM,CAAC,CAAC,KAAK,EAAE,OAAO,EAAE,EAAE,CAAC,KAAK,CAAC,OAAO,CAAC,IAAI,MAAM,CAAC,OAAO,EAAE,IAAI,CAAC,EAAE,GAAG,CAAC,EAAE,IAAI,CAAC,CAAC;AAClG,CAAC;AACD,SAAS,eAAe,CAAC,GAAW;IAClC,MAAM,KAAK,GAAG,GAAG,CAAC,KAAK,CAAC,0BAA0B,CAAC,IAAI,EAAE,CAAC;IAC1D,MAAM,KAAK,GAAG,GAAG,CAAC,OAAO,CAAC,8CAA8C,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC;IAChG,sEAAsE;IACtE,MAAM,QAAQ,GAAG,KAAK,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC;IACvC,MAAM,CAAC,KAAK,EAAE,QAAQ,GAAG,EAAE,CAAC,GAAG,KAAK,CAAC,OAAO,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC;IACrE,MAAM,UAAU,GAAG,GAAG,KAAM,CAAC,OAAO,CAAC,WAAW,EAAE,EAAE,CAAC,GAAG,QAAQ,CAAC,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,QAAQ,CAAC,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC;IAC/H,OAAO,GAAG,QAAQ,IAAI,UAAU,KAAK,GAAG,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,GAAG,UAAU,IAAI,KAAK,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC,QAAQ,CAAC,IAAI,CAAC,IAAI,IAAI,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC;AAC5H,CAAC;AACD,MAAM,UAAU,gBAAgB,CAAC,IAAY;IAC3C,OAAO,CAAC,GAAG,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC,QAAQ,CAAC,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,EAAE,GAAG,EAAE,KAAK,CAAC,CAAC,CAAC,EAAE,UAAU,EAAE,eAAe,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;AAChI,CAAC;AACD,MAAM,UAAU,WAAW,CAAC,IAAY;IACtC,OAAO,CAAC,GAAG,IAAI,CAAC,QAAQ,CAAC,+BAA+B,CAAC,CAAC,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,IAAI,KAAK,CAAC,CAAC,CAAC,CAAE,CAAC,CAAC;AACnG,CAAC;AACD,MAAM,UAAU,cAAc,CAAC,IAAY;IACzC,OAAO,IAAI,CAAC,SAAS,CAAC,MAAM,CAAC,CAAC,OAAO,CAAC,QAAQ,EAAE,GAAG,CAAC,CAAC,OAAO,CAAC,QAAQ,EAAE,GAAG,CAAC,CAAC,OAAO,CAAC,YAAY,EAAE,GAAG,CAAC;SACnG,OAAO,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC,OAAO,CAAC,2BAA2B,EAAE,EAAE,CAAC,CAAC;AACpE,CAAC"}
|
|
@@ -0,0 +1,40 @@
|
|
|
1
|
+
import { isDeepStrictEqual } from 'node:util';
|
|
2
|
+
import { compileJsonSchema } from '../../cases/index.js';
|
|
3
|
+
import { result } from '../logic/index.js';
|
|
4
|
+
import { observedAbsence, selectEvidence, textReference } from '../selectors.js';
|
|
5
|
+
export async function evaluateText(check, context) {
|
|
6
|
+
const selected = selectEvidence(check.input, context.snapshot);
|
|
7
|
+
if (!selected.found) {
|
|
8
|
+
// An absence the run observed is a behavioural failure, not missing evidence: the
|
|
9
|
+
// subject closed its own execution, every channel it declared is accounted for, and
|
|
10
|
+
// the artifact it had to produce is simply not there. It fails as a measured zero and
|
|
11
|
+
// keeps its weight, so nothing is scored better by omission. An absence the run could
|
|
12
|
+
// not observe stays unknown and keeps the trial unscorable, exactly as before.
|
|
13
|
+
if (observedAbsence(check.input, context.snapshot)) {
|
|
14
|
+
return { ...result('fail', `Missing ${String(check.input)} in a completed, fully accounted subject execution`), score: 0 };
|
|
15
|
+
}
|
|
16
|
+
return result('unknown', `Missing ${String(check.input)}`, context.snapshot.closed);
|
|
17
|
+
}
|
|
18
|
+
let passed;
|
|
19
|
+
if (check.eval === 'equals')
|
|
20
|
+
passed = isDeepStrictEqual(selected.value, check.value);
|
|
21
|
+
else if (check.eval === 'contains-any') {
|
|
22
|
+
const strings = typeof selected.value === 'string' ? [selected.value] : selected.value;
|
|
23
|
+
if (!Array.isArray(strings) || strings.some(item => typeof item !== 'string')) {
|
|
24
|
+
return result('error', 'contains-any requires a string or array of strings', context.snapshot.closed);
|
|
25
|
+
}
|
|
26
|
+
passed = strings.some(item => check.values.some(value => item.includes(value)));
|
|
27
|
+
}
|
|
28
|
+
else if (check.eval === 'json-schema') {
|
|
29
|
+
const schemaBytes = typeof check.schema === 'string' ? await context.readText(check.schema) : undefined;
|
|
30
|
+
if (schemaBytes !== undefined)
|
|
31
|
+
selected.evidence.push(textReference('sidecar', check.schema, schemaBytes));
|
|
32
|
+
const schema = schemaBytes === undefined ? check.schema : JSON.parse(schemaBytes);
|
|
33
|
+
const validator = compileJsonSchema(schema);
|
|
34
|
+
passed = validator(selected.value);
|
|
35
|
+
}
|
|
36
|
+
else
|
|
37
|
+
throw new Error(`Unknown text evaluator ${check.eval}`);
|
|
38
|
+
return { ...result(passed ? 'pass' : 'fail', `${check.eval}: ${passed ? 'matched' : 'not matched'}`, context.snapshot.closed), evidence: selected.evidence, score: passed ? 1 : 0 };
|
|
39
|
+
}
|
|
40
|
+
//# sourceMappingURL=text.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"text.js","sourceRoot":"","sources":["../../../../src/training/evaluation/mechanical/text.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,iBAAiB,EAAE,MAAM,WAAW,CAAC;AAC9C,OAAO,EAAE,iBAAiB,EAAE,MAAM,sBAAsB,CAAC;AAGzD,OAAO,EAAE,MAAM,EAAE,MAAM,mBAAmB,CAAC;AAC3C,OAAO,EAAE,eAAe,EAAE,cAAc,EAAE,aAAa,EAAE,MAAM,iBAAiB,CAAC;AAEjF,MAAM,CAAC,KAAK,UAAU,YAAY,CAAC,KAAgB,EAAE,OAA0B;IAC7E,MAAM,QAAQ,GAAG,cAAc,CAAC,KAAK,CAAC,KAAe,EAAE,OAAO,CAAC,QAAQ,CAAC,CAAC;IACzE,IAAI,CAAC,QAAQ,CAAC,KAAK,EAAE,CAAC;QACpB,kFAAkF;QAClF,oFAAoF;QACpF,sFAAsF;QACtF,sFAAsF;QACtF,+EAA+E;QAC/E,IAAI,eAAe,CAAC,KAAK,CAAC,KAAe,EAAE,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC;YAC7D,OAAO,EAAE,GAAG,MAAM,CAAC,MAAM,EAAE,WAAW,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC,oDAAoD,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC;QAC7H,CAAC;QACD,OAAO,MAAM,CAAC,SAAS,EAAE,WAAW,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC;IACtF,CAAC;IACD,IAAI,MAAe,CAAC;IACpB,IAAI,KAAK,CAAC,IAAI,KAAK,QAAQ;QAAE,MAAM,GAAG,iBAAiB,CAAC,QAAQ,CAAC,KAAK,EAAE,KAAK,CAAC,KAAK,CAAC,CAAC;SAChF,IAAI,KAAK,CAAC,IAAI,KAAK,cAAc,EAAE,CAAC;QACvC,MAAM,OAAO,GAAG,OAAO,QAAQ,CAAC,KAAK,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC;QACvF,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,OAAO,CAAC,IAAI,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,OAAO,IAAI,KAAK,QAAQ,CAAC,EAAE,CAAC;YAC9E,OAAO,MAAM,CAAC,OAAO,EAAE,oDAAoD,EAAE,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC;QACxG,CAAC;QACD,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAE,KAAK,CAAC,MAAmB,CAAC,IAAI,CAAC,KAAK,CAAC,EAAE,CAAC,IAAI,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC;IAChG,CAAC;SAAM,IAAI,KAAK,CAAC,IAAI,KAAK,aAAa,EAAE,CAAC;QACxC,MAAM,WAAW,GAAG,OAAO,KAAK,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QACxG,IAAI,WAAW,KAAK,SAAS;YAAE,QAAQ,CAAC,QAAQ,CAAC,IAAI,CAAC,aAAa,CAAC,SAAS,EAAE,KAAK,CAAC,MAAgB,EAAE,WAAW,CAAC,CAAC,CAAC;QACrH,MAAM,MAAM,GAAG,WAAW,KAAK,SAAS,CAAC,CAAC,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC;QAClF,MAAM,SAAS,GAAG,iBAAiB,CAAC,MAAM,CAAC,CAAC;QAC5C,MAAM,GAAG,SAAS,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;IACrC,CAAC;;QAAM,MAAM,IAAI,KAAK,CAAC,0BAA0B,KAAK,CAAC,IAAI,EAAE,CAAC,CAAC;IAC/D,OAAO,EAAE,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,MAAM,EAAE,GAAG,KAAK,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,aAAa,EAAE,EACjG,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,QAAQ,EAAE,QAAQ,CAAC,QAAQ,EAAE,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC;AACnF,CAAC"}
|
|
@@ -0,0 +1,4 @@
|
|
|
1
|
+
import type { CheckNode } from '../../contracts/dataset.js';
|
|
2
|
+
import type { EvaluationContext, EvaluationResult } from '../../contracts/evaluation.js';
|
|
3
|
+
/** A normalized reward over complete native totals, never an execution budget. */
|
|
4
|
+
export declare function evaluateTokenEfficiency(check: CheckNode, context: EvaluationContext): Promise<EvaluationResult>;
|
|
@@ -0,0 +1,28 @@
|
|
|
1
|
+
import { result } from '../logic/index.js';
|
|
2
|
+
const count = (value) => typeof value === 'number' && Number.isSafeInteger(value) && value >= 0;
|
|
3
|
+
/** A normalized reward over complete native totals, never an execution budget. */
|
|
4
|
+
export async function evaluateTokenEfficiency(check, context) {
|
|
5
|
+
const { snapshot } = context;
|
|
6
|
+
const target = check.target_tokens, metric = check.metric;
|
|
7
|
+
if (!count(target) || target === 0 || typeof metric !== 'string' || !['input', 'output', 'total'].includes(metric)) {
|
|
8
|
+
return result('error', 'Invalid token-efficiency parameters', snapshot.closed);
|
|
9
|
+
}
|
|
10
|
+
if (!snapshot.closed)
|
|
11
|
+
return result('unknown', 'Subject token usage is not final', false);
|
|
12
|
+
const subject = snapshot.subjectUsage;
|
|
13
|
+
if (subject?.status !== 'complete' || !subject.usage) {
|
|
14
|
+
return result('unknown', 'Complete subject token usage is unavailable');
|
|
15
|
+
}
|
|
16
|
+
const usage = subject.usage;
|
|
17
|
+
const values = metric === 'input' ? [usage.inputTokens] : metric === 'output' ? [usage.outputTokens]
|
|
18
|
+
: [usage.inputTokens, usage.outputTokens];
|
|
19
|
+
if (!values.every(count))
|
|
20
|
+
return result('unknown', 'Relevant subject token totals are unknown or invalid');
|
|
21
|
+
const measured = values.reduce((sum, value) => sum + value, 0);
|
|
22
|
+
if (!count(measured))
|
|
23
|
+
return result('unknown', 'Subject token total exceeds the safe integer range');
|
|
24
|
+
// Ratio form avoids overflowing the denominator; cache/reasoning are subsets.
|
|
25
|
+
return { ...result('pass', `token-efficiency: ${metric} subject tokens=${measured}; reference=${target}`),
|
|
26
|
+
score: 1 / (1 + measured / target) };
|
|
27
|
+
}
|
|
28
|
+
//# sourceMappingURL=token-efficiency.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"file":"token-efficiency.js","sourceRoot":"","sources":["../../../../src/training/evaluation/mechanical/token-efficiency.ts"],"names":[],"mappings":"AAEA,OAAO,EAAE,MAAM,EAAE,MAAM,mBAAmB,CAAC;AAE3C,MAAM,KAAK,GAAG,CAAC,KAAc,EAAmB,EAAE,CAChD,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,aAAa,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC,CAAC;AAEzE,kFAAkF;AAClF,MAAM,CAAC,KAAK,UAAU,uBAAuB,CAAC,KAAgB,EAAE,OAA0B;IACxF,MAAM,EAAE,QAAQ,EAAE,GAAG,OAAO,CAAC;IAC7B,MAAM,MAAM,GAAG,KAAK,CAAC,aAAa,EAAE,MAAM,GAAG,KAAK,CAAC,MAAM,CAAC;IAC1D,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,IAAI,MAAM,KAAK,CAAC,IAAI,OAAO,MAAM,KAAK,QAAQ,IAAI,CAAC,CAAC,OAAO,EAAE,QAAQ,EAAE,OAAO,CAAC,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;QACnH,OAAO,MAAM,CAAC,OAAO,EAAE,qCAAqC,EAAE,QAAQ,CAAC,MAAM,CAAC,CAAC;IACjF,CAAC;IACD,IAAI,CAAC,QAAQ,CAAC,MAAM;QAAE,OAAO,MAAM,CAAC,SAAS,EAAE,kCAAkC,EAAE,KAAK,CAAC,CAAC;IAC1F,MAAM,OAAO,GAAG,QAAQ,CAAC,YAAY,CAAC;IACtC,IAAI,OAAO,EAAE,MAAM,KAAK,UAAU,IAAI,CAAC,OAAO,CAAC,KAAK,EAAE,CAAC;QACrD,OAAO,MAAM,CAAC,SAAS,EAAE,6CAA6C,CAAC,CAAC;IAC1E,CAAC;IACD,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;IAC5B,MAAM,MAAM,GAAG,MAAM,KAAK,OAAO,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC,CAAC,CAAC,MAAM,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,YAAY,CAAC;QAClG,CAAC,CAAC,CAAC,KAAK,CAAC,WAAW,EAAE,KAAK,CAAC,YAAY,CAAC,CAAC;IAC5C,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,KAAK,CAAC;QAAE,OAAO,MAAM,CAAC,SAAS,EAAE,sDAAsD,CAAC,CAAC;IAC3G,MAAM,QAAQ,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,KAAK,EAAE,EAAE,CAAC,GAAG,GAAG,KAAK,EAAE,CAAC,CAAC,CAAC;IAC/D,IAAI,CAAC,KAAK,CAAC,QAAQ,CAAC;QAAE,OAAO,MAAM,CAAC,SAAS,EAAE,oDAAoD,CAAC,CAAC;IACrG,8EAA8E;IAC9E,OAAO,EAAE,GAAG,MAAM,CAAC,MAAM,EAAE,qBAAqB,MAAM,mBAAmB,QAAQ,eAAe,MAAM,EAAE,CAAC;QACvG,KAAK,EAAE,CAAC,GAAG,CAAC,CAAC,GAAG,QAAQ,GAAG,MAAM,CAAC,EAAE,CAAC;AACzC,CAAC"}
|