Skip to content

v0.5.0

Latest

Choose a tag to compare

@github-actions github-actions released this 08 Sep 02:53
· 2 commits to main since this release
0bac2f6

v0.5.0 — 期权决策研究平台

v0.5.0 将离线学习、研究简报、证据复核和开源验证流程连接起来。面向 BTC 期权研究者,
它回答当前输入支持哪些研究结论、为什么被阻断、下一步需要什么证据。
research_only=true 与 execution_allowed=false 保持不变。

这是本版本的交付说明。发布状态与最终验收记录见
本轮交付记录;只有 GitHub Release
中实际存在的资产才是已发布产物。

使用者可见变化

  • crypto-options-report demo 默认进入离线学习导览,先理解有限风险结构,再查看包内真实快照。
    教学点数与实际研究数据、证据提升分开。
  • 研究简报、工作台、公开版与 Chrome 侧栏统一解释过期、阻断和评估时刻;历史通过不表示当前资格。
  • 加载失败、无效 JSON、合同异常和渲染错误提供恢复入口;信号/序列未配置与样本不足分别说明。
  • 新增 完整决策研究指南 和 离线复核案例,
    覆盖公开采集、报告、证据、过期恢复和可复核记录。
  • 策略卡要求数字成本和模型身份,最低净权利金真实扣除入场费、滑点与分腿预算;
    损失字段明确是 payoff 边界加冻结成本预算。交割费绝对上界未验证,所有卡片最高为 WATCH。
  • 历史路径不再按该路径未来收益推导“相似状态”权重。缺少路径开始前的状态证据时,
    使用 unconditioned_uniform 并明确相似度条件不可用。
  • 精确结构历史回放纳入按规则计算的逐腿交割费,并明确风险分母未包含交割费,因此净 R 可能小于 -1。
  • 路径不利变动按实际腿的风险方向计算,并区分短腿行权价穿越、价格阈值代理与未实现的动态 Delta。
    压力场景覆盖双向冲击,明确为确定性假设,其权重不代表校准概率。

维护者可见变化

  • 分离带类型的分析输入、不可变求值、五组准入条件与惰性旧报告投影,保留公共调用入口。
  • 记录输入摘要和构建身份;源码与安装包的来源可区分,Git 状态观测不冒充签名。
  • 将公开 JSON 合同集中为固定 OpenAPI/schema,发布前校验允许字段与产物结构。
  • 统一 python tools/verify.py:源码检查、Web 构建、公开边界、扩展、最终 wheel 安装和浏览器流程。
    --quick 是明确的子集,不能代替完整验收。
  • 对齐 Python、Web、扩展和发布资产的版本,提供无需网络的固定输入复现工具。

安装与升级

从 GitHub Releases 下载此版本的 wheel、
Chrome 扩展 ZIP 与 SHA256SUMS。将资产保存在同一目录,先核对本机摘要与校验和文件的对应行:

Get-FileHash -Algorithm SHA256 .\*.whl, .\*.zip
Get-Content -LiteralPath .\SHA256SUMS

在干净环境安装 wheel(将占位路径替换为实际文件名):

python -m venv .venv
.\.venv\Scripts\python.exe -m pip install --no-index --no-deps .\<downloaded-wheel>.whl
.\.venv\Scripts\python.exe -m crypto_options_report.cli demo

wheel 包含前端和脱敏演示资源,安装后无需 Node、密钥、网络或仓库测试文件。开发源码请使用
贡献指南 安装受约束的开发依赖并重新构建前端。

v0.4.0 用户必须重新生成报告

原始市场快照可以重新计算,CLI 与公共分析调用入口保留;已有策略卡并非完全向后兼容。
strategy_brief.v1 现在要求 entry.cost_model_id、entry.cost_config_hash、完整数字
entry.cost_breakdown,以及明确的 risk.max_loss_basis 和 delivery_fee_upper_bound_verified。
v0.4.0 仅带 bool 含费标记的旧卡会被拒绝,这是有意收紧合同。

同时升级 Python 引擎、前端与 Chrome 扩展,从保留的原始输入重新生成报告和公开工件。
不要原地修改旧报告的 bool 或补零金额。预测的 selection_binding_key 还绑定成本模型、配置、
预算及净权利金;旧 CALIBRATED 产物不能直接复用,即使腿和到期日相同也会因缺少成本绑定退役。
旧预测文件仍可读取和独立验证;退役指它在当前策略简报中不再保持 CALIBRATED,
需要重新生成绑定当前成本与净权利金的校准证据,并非把历史文件当作损坏文件。

构建和输入身份调整会改变派生 ID 与摘要。用同一版本、输入和时钟生成新复核记录,保留旧记录
用于解释差异。Chrome 扩展解压到新目录并重新加载,避免混用旧 JS 与新 manifest。

信号与序列工件升级

/research/signal 和 /research/series 现在验证 research_only=true、允许的 schema 与合法时间,
并拒绝任意嵌套深度出现的 execution_allowed、manual_execution_allowed、order_instruction、
order_instructions 或 recommended_size 字段。执行字段即使值为 false 也不属于这两类研究工件。

由当前 CLI 生成的 measured、preflight、blocked 和 demo 工件兼容,excluded_snapshots
等来源排除证据保留。旧手工文件缺少研究标记或合同字段时,使用当前 CLI 重新生成,不手工补标记
来绕过校验。例如:

crypto-options-report validate-signal --preflight `
  --snapshot-dir artifacts/snapshots/btc-series `
  --underlying-history-fixture artifacts/history/btc-daily.json `
  --output artifacts/reports/signal-preflight.json --compact

crypto-options-report series-history --snapshot-dir artifacts/snapshots/btc-series `
  --output artifacts/reports/series-history.json --compact

随后将 --signal-artifact 与 --series-artifact 指向新文件。样本不足会产生可解释的阻断工件,
无需制造统计值;文件读取失败和样本不足仍是不同状态。

明确限制

可靠策略卡允许为空。历史胜率只在 VALIDATED 后出现,精确策略预测区间只在 CALIBRATED
后出现。本版本没有补造 cohort、提升模型或授权实盘;软件测试与离线案例不证明市场 edge。
公开静态站是否持续产出合格版次,需要独立的采集、监控和发布验收。

English summary

Version 0.5.0 connects an offline learning tour, the options research brief,
evidence inspection, recovery, and reproducible open-source verification.
Typed analysis inputs and lazy compatibility projection retain public entry
points; public artifacts are checked against fixed schemas. Full local
verification includes the final installed wheel and browser journey.

The platform remains research-only. Missing or stale evidence blocks, historical
rates require VALIDATED, forecast intervals require CALIBRATED, and execution
is always disabled. Cards remain WATCH: modeled costs are not an absolute
fee-inclusive loss cap. Version 0.4.0 boolean-only cost cards no longer validate;
upgrade the engine, frontends, and extension together and regenerate reports.
Old forecasts do not satisfy the new cost and net-credit selection binding.
They remain readable for historical inspection but retire in the current brief
until appropriately rebound evidence exists. Signal/series artifacts now require
research_only=true, a supported schema, and valid timestamps; execution and
sizing fields are rejected at every depth. Regenerate old hand-authored files
using the current CLI. Measured, preflight, blocked, and demo artifacts remain
supported, including snapshot-exclusion evidence.
See the decision workflow
and the delivery record above for the actual verification and publication status.