规格 · NovelForge 7.1 Adaptive Runtime
NovelForge 7.0 已能持久化 preference evidence / hypothesis、生成 Corpus gap、构造 typed discovery request、验证声明的 rights metadata、路由 semantic work、checkpoint 长任务。但这些能力还没有被一个 deterministic、可恢复的 Learning Cycle 串成闭环;Host 的 Web/GitHub/MC
- 上一版本:NovelForge 7.0.0
- 基线 commit:
de05666cc4eae13f09868d87659e76f2aa524314 - Rollback:上述基线 commit
- Change class:Framework release / structural feature
NovelForge 7.0 已能持久化 preference evidence / hypothesis、生成 Corpus gap、构造 typed discovery request、验证声明的 rights metadata、路由 semantic work、checkpoint 长任务。但这些能力还没有被一个 deterministic、可恢复的 Learning Cycle 串成闭环;Host 的 Web/GitHub/MCP/tool 可用性主要存在于文档与 routing 描述里,而不是 typed capability contract;Project lock 也尚无可独立验证的 immutable bundle fingerprint。
因此仍有三个风险:
- 模型可能“猜”某个 Web/GitHub/MCP connector 存在,而不是证明 capability;
- Corpus discovery、source verification、semantic analysis、eval evidence、promotion candidacy 可能退化成松散的临时步骤;
- Consumer 虽锁 exact commit,却不能独立校验本地 materialized Framework bundle 的内容指纹。
G1 · Typed Host Capability Contract
Section titled “G1 · Typed Host Capability Contract”增加 provider-neutral host capability manifest + deterministic resolver。Capability 只有在显式声明或本地可证明时才能使用。缺失 capability 必须返回真实 pending/blocked route,禁止伪造 tool access。
G2 · Durable Adaptive Learning Cycle
Section titled “G2 · Durable Adaptive Learning Cycle”增加 stdlib-only executable Learning Cycle 状态机,协调:
evidence/hypothesis → corpus gap → discovery planning → verified discovery results → semantic mechanism analysis → eval evidence → promotion candidate → activation/promotion gate
Learning Cycle 属于 operational/learning state,永远不是 Canon。
G3 · Discovery / Provenance Runtime
Section titled “G3 · Discovery / Provenance Runtime”增加 typed discovery result contract:source locator、retrieval channel、tool/provider provenance、retrieval timestamp、evidence fingerprint、declared rights basis、storage intent、dedupe/diversity accounting。
Discovery 与 Ingestion 严格分离。Deterministic code 不得根据 URL/title 推断版权状态。
G4 · Analysis + Eval Work Packaging
Section titled “G4 · Analysis + Eval Work Packaging”为 Corpus mechanism analysis 与 learning eval 创建 bounded、fingerprint-bound semantic jobs。默认不携带 hidden gold、writer private reasoning、与研究问题无关的用户私有信息或整部现代版权正文。
G5 · Promotion Gate
Section titled “G5 · Promotion Gate”增加 deterministic gate,可输出 blocked | ready_for_activation | promotable,但不能自行修改 Framework behavior 或 durable user taste。
General Craft 必须具备 provenance、counterexample/profile boundary、cross-work evidence、capability + regression eval evidence、target version、rollback ref、green deterministic CI。
G6 · Immutable Framework Bundle
Section titled “G6 · Immutable Framework Bundle”增加 deterministic bundle build + verify。Bundle 内容排序、metadata 归一化;生成的 bundle metadata 不参与自身 fingerprint。输出 content manifest + SHA-256 bundle fingerprint,Consumer 可写入 novelforge.lock.json。
G7 · Release-grade Automation
Section titled “G7 · Release-grade Automation”Normal CI 与 scheduled maintenance 继续禁止模型调用。Weekly maintenance 可以推进 deterministic cycle planning 并生成 typed work queue,但不能声称已经执行不存在的 Web/model tool,也不能 auto-promote Framework behavior。
Live semantic/provider execution 继续保持显式触发、单独计费/usage。
G8 · Consumer Upgrade Contract
Section titled “G8 · Consumer Upgrade Contract”Project SDK/Adapter validation 支持 NovelForge 7.1 lock 与 bundle fingerprint verification。只有最终 7.1 Framework commit CI 绿、deterministic bundle fingerprint 已知后,才升级 Chinatown consumer。
- Generic Framework 不得吸收 consumer novel Canon、人物、剧情、private project state 或 private user taste。
- 不自动 Canon write / SETTLE / 下一章 drafting。
- Candidate 即使 promotable,也不自动修改 General Craft source。
- Normal CI / weekly maintenance 不隐藏消耗 API/Codex/Claude/model usage。
- 不 mass-mirror 现代版权全文。
- 不建立 named-author imitation profile。
- 不要求所有 host 都具备 Web、GitHub、MCP、provider API、Codex 或 Claude。
Runtime Contract
Section titled “Runtime Contract”Host capability manifest
Section titled “Host capability manifest”至少记录:
- schema/version;
- host identity/runtime class;
- capability availability + provenance + permission scope;
- cost/usage class;
- 是否需要 user interaction;
- 是否会执行 model inference;
- credential secret 永远不写入 manifest。
Learning-cycle identity
Section titled “Learning-cycle identity”learning_cycle_id != runtime session_id != semantic job_id != project Canon state。
Cycle 必须记录每个 discovery/analysis/eval result 的 exact ref/fingerprint,并对 logical result consume-once。
Discovery result
Section titled “Discovery result”每个 source candidate 必须记录来源 channel 与 tool/provider。Unknown / analysis-only rights 不允许保存 full text。
Semantic analysis
Section titled “Semantic analysis”Corpus analysis job 只携带 research question 与被允许的 bounded evidence。Result 返回 typed mechanism observation、counterexample、applicability boundary、evidence refs、confidence。
Promotion
Section titled “Promotion”Deterministic gate 只检查 evidence completeness;真正的 behavior-source mutation 由 manager/human/authorized Framework workflow 执行。
Acceptance Criteria
Section titled “Acceptance Criteria”runtime_capabilities.py self-test证明 undeclared capability 永不被选中。learning_cycle.py self-test证明合法 transition、resume/idempotency、result consume-once、无 Canon authority。discovery_runtime.py self-test证明 provenance binding、dedupe/diversity、rights-gate enforcement。learning_eval.py self-test证明 blind fingerprinted analysis/eval packaging 且无 answer-key leakage。promotion_gate.py self-test证明 General Craft 缺 counterexample/cross-work/eval/rollback/CI 任一项都不能通过。build_framework_bundle.py self-test证明 deterministic bytes/fingerprint,篡改后 verify 必须失败。- 顶层
novelforge.py self-test覆盖全部 7.1 modules。 - Normal CI 编译/测试 7.1 全模块且
model_execution=false。 - Weekly maintenance 只生成 capability-aware queue,不伪造 Web/model execution。
- 中英双语 human-facing docs 同步。
HARNESS_MANIFEST.yaml、SKILL*、README/CHANGELOG、Project SDK 全部报告 7.1.0。- 最终 NovelForge CI green。
- Chinatown project lock 升级到 exact green commit + deterministic bundle fingerprint,Project CI green。
- Chinatown Canon / active story state 在工程迁移中保持不变。