sbv-engine
暂无简介
- License 许可
- AGPL-3.0
- 所属仓库
- QuanuX/Symphony
- 最近更新
- 2026-11-14
- Stars
- ★ 2
- Forks
- 0
02 资源文件 166
- 其他cmake/SymphonySbvInteropConfig.cmake.in496 B
- 其他cmake/SymphonySbvProviderConfig.cmake.in468 B
- 其他cmake/SymphonySbvSdkConfig.cmake.in590 B
- 其他cmake/uninstall.cmake.in199 B
- 参考文档CMakeLists.txt30.8 KB
- 参考文档FEATURES.md13.8 KB
- 参考文档FOLLOW-ON.md10.8 KB
- 其他include/symphony/sbv/feedback.h1.4 KB
- 其他include/symphony/sbv/interop.hpp6.2 KB
- 其他include/symphony/sbv/models.hpp2.2 KB
- 其他include/symphony/sbv/provider.h4.7 KB
- 其他include/symphony/sbv/sbv.hpp1.3 KB
- 其他include/symphony/sbv/sdk.h1.7 KB
- 其他include/symphony/sbv/sdk.hpp806 B
- 参考文档INSTALL.md7.1 KB
- 参考文档INTENT.md613 B
- 参考文档INTERFACE-GENERATOR.json729 B
- 参考文档M2-M3-READINESS.json21.7 KB
- 参考文档M3-QUALIFICATION.md12.0 KB
- 参考文档MANIFEST.md8.0 KB
- 参考文档OWNER-INTERFACE.json42.4 KB
- 参考文档schemas/v1/admin.schema.json2.6 MB
- 参考文档schemas/v1/admin.templates.json138.1 KB
- 参考文档schemas/v1/node-stream.schema.json5.3 KB
- 参考文档schemas/v1/partitioned-result.schema.json6.0 KB
- 参考文档schemas/v1/pointer-stream.schema.json2.0 KB
- 参考文档schemas/v1/result-page.schema.json6.8 KB
- 参考文档schemas/v1/result.schema.json2.4 KB
- 参考文档SPEC.md191.0 KB
- 其他src/account_simulate.cpp36.7 KB
- 其他src/allocation_economics.cpp19.8 KB
- 其他src/analyze.cpp23.9 KB
- 其他src/binary_studies.hpp16.5 KB
- 其他src/book.cpp21.1 KB
- 其他src/bundle_operations.cpp11.8 KB
- 其他src/census.cpp17.1 KB
- 其他src/census.hpp625 B
- 其他src/combinatorial.hpp4.2 KB
- 其他src/compare.cpp19.2 KB
- 其他src/compose_economics.cpp43.0 KB
- 其他src/compose_joint.cpp4.0 KB
- 其他src/compose.cpp4.4 KB
- 其他src/consumer_source.hpp9.7 KB
- 其他src/dataset.cpp12.2 KB
- 其他src/dataset.hpp1.4 KB
- 其他src/depth_sweep.hpp844 B
- 其他src/detail.hpp5.9 KB
- 其他src/economic_census_stream.cpp21.4 KB
- 其他src/economic_census_stream.hpp421 B
- 其他src/economics_kernel.hpp1007 B
- 其他src/economics_partitioned.cpp22.0 KB
- 其他src/economics.cpp27.4 KB
- 其他src/evaluate_partitioned.cpp24.3 KB
- 其他src/evaluate.cpp17.0 KB
- 其他src/exact_id_index.cpp3.5 KB
- 其他src/exact_id_index.hpp1.3 KB
- 其他src/experiment_paths.hpp4.8 KB
- 其他src/experiment_result.hpp5.4 KB
- 其他src/experiment.cpp42.5 KB
- 其他src/fit.cpp25.0 KB
- 其他src/generate_census.cpp11.4 KB
- 其他src/history.cpp21.3 KB
- 其他src/interface.generated.hpp16.1 KB
- 其他src/jobs.cpp19.8 KB
- 其他src/liquidity.cpp24.8 KB
- 其他src/logical_value.cpp13.6 KB
- 其他src/logical_value.hpp2.8 KB
- 其他src/main.cpp3.1 KB
- 其他src/measurement.hpp1.9 KB
- 其他src/model.cpp24.8 KB
- 其他src/operation_profiles.hpp12.4 KB
- 其他src/partitioned_output.cpp18.4 KB
- 其他src/partitioned_output.hpp2.2 KB
- 其他src/planning.cpp10.8 KB
- 其他src/provider.cpp30.8 KB
- 其他src/provider.hpp2.5 KB
- 其他src/queue_scenarios.cpp29.5 KB
- 其他src/rational.hpp1.3 KB
- 其他src/reader_extensions.inc16.6 KB
- 其他src/resample.cpp13.0 KB
- 其他src/resident.cpp33.5 KB
- 其他src/resource_admission.hpp1.9 KB
- 其他src/result_store.cpp61.4 KB
- 其他src/result_store.hpp7.3 KB
- 其他src/result.cpp19.9 KB
- 其他src/row_spool.cpp5.7 KB
- 其他src/row_spool.hpp1.7 KB
- 其他src/run_math.hpp9.0 KB
- 其他src/run_partitioned.cpp14.4 KB
- 其他src/run.cpp9.8 KB
- 其他src/sdk.cpp3.4 KB
- 其他src/select_view.cpp7.7 KB
- 其他src/select.cpp8.2 KB
- 其他src/selection_query.hpp5.5 KB
- 其他src/source_owners.cpp62.1 KB
- 其他src/source_owners.hpp1.6 KB
- 其他src/split.cpp30.3 KB
- 其他src/stream_census.cpp28.6 KB
- 其他src/stream_census.hpp1.1 KB
- 其他src/stream_model.cpp17.0 KB
- 其他src/stream_model.hpp1.9 KB
- 其他src/streaming_sha256.hpp5.2 KB
- 其他src/study_formula.hpp3.7 KB
- 其他src/wide_rational.hpp3.3 KB
- 其他tests/account_execution_test.cpp16.6 KB
- 其他tests/account_test.cpp11.0 KB
- 其他tests/adaptive_test.cpp18.3 KB
- 其他tests/administration_test.cpp9.3 KB
- 其他tests/allocation_economics_test.cpp14.1 KB
- 其他tests/analysis_test.cpp15.3 KB
- 其他tests/binary_studies_test.cpp13.0 KB
- 其他tests/book_test.cpp16.5 KB
- 其他tests/bundle_operations_test.cpp8.3 KB
- 其他tests/census_test.cpp25.9 KB
- 其他tests/combinatorial_test.cpp12.9 KB
- 其他tests/compose_economics_test.cpp32.5 KB
- 其他tests/dataset_test.cpp17.4 KB
- 其他tests/dependencies_test.cpp13.6 KB
- 其他tests/economics_partitioned_test.cpp26.0 KB
- 其他tests/economics_test.cpp23.3 KB
- 其他tests/evaluate_partitioned_test.cpp18.1 KB
- 其他tests/exact_id_index_test.cpp4.1 KB
- 其他tests/execution_partitioned_test.cpp16.4 KB
- 其他tests/experiment_partitioned_test.cpp12.8 KB
- 其他tests/experiment_plan_test.cpp9.3 KB
- 其他tests/feedback_fixture.cpp7.4 KB
- 其他tests/feedback_settlement_test.cpp13.5 KB
- 其他tests/fit_test.cpp12.7 KB
- 其他tests/fixture_requests.hpp3.9 KB
- 参考文档tests/fixtures/interface-history.v1.json19 B
- 其他tests/generate_partitioned_test.cpp10.1 KB
- 其他tests/history_test.cpp14.1 KB
- 其他tests/interop_consumer.cpp1.5 KB
- 其他tests/jobs_test.cpp11.6 KB
- 其他tests/liquidity_test.cpp17.0 KB
- 其他tests/logical_value_test.cpp8.8 KB
- 其他tests/model_partitioned_test.cpp14.2 KB
- 其他tests/models_test.cpp13.4 KB
- 其他tests/nested_split_test.cpp13.6 KB
- 其他tests/partitioned_output_test.cpp14.9 KB
- 其他tests/performance_test.cpp9.8 KB
- 其他tests/planning_test.cpp12.1 KB
- 其他tests/provider_abi_test.c1.1 KB
- 其他tests/provider_fixture.cpp15.6 KB
- 其他tests/provider_integration_test.cpp28.2 KB
- 其他tests/provider_runtime_test.cpp19.1 KB
- 其他tests/queue_test.cpp17.0 KB
- 其他tests/reader_extensions_test.cpp12.3 KB
- 其他tests/research_partitioned_test.cpp11.5 KB
- 其他tests/research_test.cpp11.6 KB
- 其他tests/resident_control_test.cpp12.4 KB
- 其他tests/resource_admission_test.cpp4.0 KB
- 其他tests/result_store_test.cpp12.8 KB
- 其他tests/row_spool_test.cpp9.9 KB
- 其他tests/run_partitioned_test.cpp11.7 KB
- 其他tests/sbv_test.cpp13.4 KB
- 其他tests/sdk_consumer.c651 B
- 其他tests/sdk_consumer.cpp483 B
- 其他tests/sdk_test.cpp4.7 KB
- 其他tests/select_view_test.cpp6.9 KB
- 其他tests/source_dataset_test.cpp20.6 KB
- 其他tests/source_owners_test.cpp28.6 KB
- 其他tests/split_test.cpp15.7 KB
- 其他tests/stream_census_test.cpp18.9 KB
- 其他tests/stream_model_test.cpp16.3 KB
- 其他tests/study_formula_test.cpp5.9 KB
03 SKILL.md 正文
SBV operational contract
Use the exact installed version and receipt-owned schema/template. Select all study, criteria, execution and replay choices explicitly. Run creates a new private result path; inspect a possibly committed destination before retrying an uncertain write. To read, supply the returned content SHA256. qxctl result export renders all retained fields; result query permits bounded traversal. Unavailable is not zero. Trade touch is not fill probability. No command contacts Databento or a live broker. Do not infer production readiness or full roadmap implementation from this experimental package.
For book, select an existing source-bound census, signal IDs, replay window, display depth, cadence, anomaly policy and optional checkpoint. Use its installed schema/template. Unknown initial state must remain unavailable until a reset/checkpoint; never relabel observed trade support or incomplete fragments as executable liquidity. All frame/checkpoint fields use the ordinary result query/export surface.
For liquidity, select a bound book result, explicit order/frame intents and all evidence policies. Activation is optional and remains an assumption. Independent orders use separate snapshots; shared-snapshot orders preserve input order under all-active conditioning. Do not turn unknown earlier shared outcomes into zero consumption, or conditional fill quantity into calibrated likelihood. Retained frames, replay, allocations and optional studies use ordinary result access.
For allocation_economics, select exact liquidity allocations and explicit mark/evidence, units, costs, return basis and conditional or activation-mixture mode. Use actual filled quantity, preserve active zero-fill costs and source shared conditioning, and keep hypothetical markout distinct from realized P&L. Query/export retains all source allocations and replay; no portfolio independence is inferred.
For source-series statistics, select analyze with explicit source role, unit, weighting, study list, variance/quantile/drawdown conventions and numerical profile. Do not infer returns from prices or equity from arbitrary row order. Inspect each study's status even when the operation completed. For supplied trial scoring, use compare; failed/pruned trial records and Pareto versus scalar weighting remain visible. Neither command executes a strategy search. Obtain exact templates with qxctl and preserve all source/result identities.
For reproducible sensitivity studies, resample requires an explicit uniform-row scheme, seed, block/sample lengths and worker count. Preserve index tapes where selected; empirical percentiles are not automatically confidence intervals. experiment executes only native local SBV producers from caller-supplied trial requests. Use a private existing directory and new summary destination on reconciliation. Preserve ambiguous claims and never claim that a missing completion means no child result was published. Inspect every trial and its exact artifact identity through qxctl. Remote dispatch and proposal algorithms remain separate integrations; local adaptive revisions are described below.
Independent programs may use the exact installed C ABI (Symphony::SbvSdk), optional C++ RAII wrapper or a separately supplied external language bridge. Native C++ owns the calculations; these bindings carry the same process envelope and typed result. Release response allocations with the originating SDK, retain exact release selection and inspect output artifacts after uncertain failures. A successful byte transfer does not establish a vendor tensor, zero-copy or CUDA integration.
Optional resident data: create a short caller-owned mode-0700 directory, obtain
sbv template --operation dataset_load, select source identity, fresh instance
id, RAM budget/residency, concurrency and idle policy, then call sbv dataset load. dataset_execute wraps run/evaluate/book/generate_census input without either output form;
the wrapper owns the destination. Use dataset inspect for counters and dataset
release after jobs finish. Select a job_id before execution to reconcile disconnects
with dataset jobs; dataset cancel requests cooperative cancellation.
dataset configure changes admission/idle settings using the inspected
configuration_generation, without reloading RAM. dataset stop explicitly
terminates the whole incarnation, including noncooperative providers; outputs
may be incomplete and are never automatically rolled back or retried.
Explicit file operations remain available. All
commands support text, JSON and NDJSON; future GUIs use identical payloads.
A transport failure may follow a completed write; inspect its destination.
For retained census reuse, pass an exact root or embedded run/evaluate result
reference to evaluate.census. Preserve the original native/external identity;
selecting another model does not regenerate membership. economics.source can
reference an evaluation result directly, including an experiment binding.
compose_economics selects economic rows by immutable source, signal and
conditioning. Explicitly choose independent or joint dependence, marginal policy,
state arithmetic/domain, optional resource constraints and studies. Inspect
unavailable distribution findings; never treat them as zero outcomes. Original
replay references remain observed evidence separate from economic trajectories.
For offline source retention, discover source_retain and source_export schemas/templates and capabilities.source_owner_profile. Supply actual declared capture/acquisition/access evidence and exact owner options; preserve them for exact reopen. Use one complete file source tuple or one retained_source selection. Acknowledge Dataset admission only when explicitly selected; this does not claim a completed backtest. Export original/capture/manifest bytes to new paths outside the selected store. On recovery_required, retain typed evidence and inspect/reconcile selected destinations before any caller-approved retry. SDK transport success is not publication success. This profile imposes no default deadline or general file/RAM dataset cap.
For scalable derived results, select the explicit bundle operations and both logical/manifest digests. qxctl sbv bundle import|inspect|query|verify|export uses the same native SDK contract. bundle export streams every retained field; use --receipt-only for its complete receipt. Native JSON is required for CLI JSON/text, native NDJSON for CLI NDJSON. Inspect/query do not imply full closure availability; use verify/export for that finding. Preserve incomplete-publication recovery evidence and inspect the selected destination before any explicit retry. Legacy producer/consumer limits are separate from the bundle format.
For native partitioned generation/run/evaluation/economics/composition, obtain sbv template --operation OP --variant partitioned. Select fresh disjoint bundle/workspace paths. The result receipt supplies exact storage.reference; use bundle query/export for all results and replay. A RAM-resident wrapper selects output itself. Scratch is intentionally retained, including on failure. Inspect parent/child recovery and cause before an explicit retry. max_signals:null selects an uncapped built-in census; deadlines and volume constraints remain caller choices. Bundle inputs verify complete outer closure before typed financial admission; captured ancestors are correspondence evidence, not authorship authentication.
For nested temporal validation, select split template variant nested or
nested_partitioned. Supply an inner plan for every outer fold and explicitly
choose outer_train isolation or source reuse plus its own interval policies.
Inspect nesting counts and original-source indices. Bind retained rows through
ordinary experiment/fit/predict requests; split performs no automatic winner
selection. Use adaptive experiment revisions below for proposal journaling. No untouched
holdout or overfitting claim follows from zero matching indices alone.
For adaptive local experiments, select the adaptive experiment template.
Append a contiguous revision with the exact predecessor plan hash, preserve the
cumulative trial/event prefix, and select fresh output and owned-plan paths.
Inspect proposal_states, original admission hashes and every trial receipt.
An ambiguous claim is never proof that work did not happen; an explicit new
trial identity can record a caller-directed retry without erasing the original.
Use compare's selection template to choose objectives and tie handling, then
bind /sections/selection/data/parameters/... with mode: value into a refit
request. Missing/ambiguous winners block dependents before claim. Reference mode
remains appropriate for dataset/model arrays. Users own graph structure,
proposals, constraints and continuation; no automatic quality gate is implied.
Use analyze's binary/binary_partitioned templates for supplied binary scores
and labels. Select studies, numeric label mapping, thresholds/equality, bins,
weights/normalization and tie credit explicitly. Inspect binary_input, study
statuses and full result exports. Excluded labels and undefined denominators
remain visible. A Brier score or empirical reliability table is not proof of
calibration, execution feasibility or untouched holdout history. Compare/experiment
can consume the resulting exact metrics without changing the user's objective.
For shared account experiments, use account_simulate with exact instruments,
cash units, initial state, nullable balance bounds and caller-ordered event streams.
Select fixed_intents or stateful_feedback and explicit guards. Supplied fills are
conditional scenarios, not matching evidence. Select snapshots when needed; start
a new branch from an exact snapshot reference and a caller-owned continuation tape.
Inspect dispositions and unavailable marks before interpreting terminal equity.
All controls and results use qxctl; independent experiments may execute concurrently.
For account-coupled depth sweeps, select liquidity_pools from exact retained
book-frame references and use liquidity_order at frame availability time.
Inspect execution_outcome.committed and account disposition before counting fills.
Select fixed quantities or stateful target_position rules and explicit fees/policies.
Pool-aware branches use v2 snapshots; unavailable skip_event results remain partial.
Use position_split only with explicit adjusted-unit evidence and no stale pool.
For native account feedback use the native_feedback template, replace its
library/digest and select explicit unavailable behavior. Keep all continuation
state in returned memory and inspect both trigger and resolved event/disposition.
Use v3 snapshots for feedback or linear_variation instruments. Supply explicit
basis and settlement events; this model supplies no exchange margin schedule.
The provider-inspect feedback variant exposes the matching selection template; all results remain queryable/exportable.
Use qxctl sbv job create to record an exact local calculation, then execute.
Retain the returned plan digest. inspect exposes current generation and immutable
state/result references; cancel and reconcile require that generation. An orphaned
claim is unknown and is not re-executed. A deliberate new attempt uses a new explicit
journal/output selection after caller review. All fields are exposed in owned job
schema/templates. Local journals do not establish SOV remote execution or SSIAG grants.