Spec 012 · AI-Native 自适应生产架构重构

状态：implementation candidate 范围：仅 Generic NovelForge Primary mode：SYSTEM-IMPROVE

1. 问题

NovelForge 需要把执行真相与叙事真相真正分开。此前 production path 在 Context、Reader、repair depth、telemetry、learning、planning、simulation 周围积累了不少 deterministic helper。其中一部分保护真实 authority / durability；另一部分则容易因为“代码好测试”而把文学判断冻结成 Python 规则。

本 candidate 验证而不是预设以下 hypothesis：

Code provides capabilities and constraints; models provide intelligence.

Deterministic code enforces execution truth; AI agents evaluate narrative truth.

目标不是“少写 Python”，而是只让 Python 保留客观可执行不变量，并删除那些假装理解文学意义的 deterministic mechanism。

2. 当前 architecture decision

NovelForge 采用 thin deterministic kernel + model-owned semantic runtime。

Deterministic kernel

只拥有可机械证明的行为：

• authority / permission / Project isolation；
• exact artifact identity、hash、fingerprint；
• provenance 与 exact-source reference；
• session/run/checkpoint persistence；
• before-state / CAS / idempotency / transaction；
• capability / credential boundary；
• stage visibility 与 private-state isolation；
• hard resource/context budget；
• typed envelope validation 与 receipt binding；
• semantic execution 是否真实发生、是否绑定 exact candidate、independent worker provenance 是否成立。

它回答的问题只有：授权操作是否真的针对正确状态发生了？

Semantic runtime

模型拥有需要理解意义的工作：

• search intent、query formulation、retrieval continuation / stopping；
• narrative relevance 与 context sufficiency；
• planning depth 与 uncertainty decision；
• character motivation、plausible inference 与 action；
• scene causality 与 dramatic realization；
• Reader experience；
• semantic hard-rule applicability / violation；
• repair mechanism 与 repair depth；
• preference interpretation 与 learning hypothesis。

它回答的问题是：这段故事/文本/上下文意味着什么，下一步应该怎么做？

3. Current owner map

Subsystem	Live owner	Decision	Boundary
session / checkpoint / resume	harness/session_runtime/**	KEEP	durability、stale-state rejection、capability re-resolution
Control Plane / write intent	harness/control_plane/**	KEEP	permission、exact action/target/before-state、idempotency
context eligibility / stage isolation	harness/context_inspector.py	KEEP	只判断机械 eligibility；明确禁止 relevance
Context Assembly	harness/context_assembly.py	THIN	exact selected refs、stage/fingerprint/private boundary；不判断文学 sufficiency
semantic context/search	context.select	MIGRATE_TO_AGENT（已实现）	模型自己决定缺什么、怎么搜、相关性、reformulate 与何时停止
hard-budget packing	harness/memory_tiers.py	KEEP	semantic selection 之后的 whole-item budget enforcement
planning commitment authority	harness/planning_horizon.py	KEEP / ADAPT	code 执行 declared depth/commitment/CAS；Planner 判断什么深度有用
character action	character.action_propose	MIGRATE_TO_AGENT（已实现）	private state 是 causal evidence，不是 prose serialization
scene collision	scene.resolve_actions	MIGRATE_TO_AGENT（已实现）	compact causal trace，不做 deterministic story engine
writer-safe realization	scene.realization_project	THIN	privacy boundary；不建立 Realization-Sheet serialization obligation
Blind Reader	reader.engagement_audit	MIGRATE_TO_AGENT（已实现）	只看 reader-visible evidence；不接受 taxonomy/HF/telemetry priming
semantic hard rules	quality.semantic_rule_audit	MIGRATE_TO_SEMANTIC_RULE（已实现）	模型判断 PASS/FAIL/N/A/insufficient evidence
Editor repair	editor.repair_spec + quality/repair_policy.py	MIGRATE_TO_AGENT + THIN	Editor 选 owner/mode；Python 只执行所选 writer-context boundary
prose telemetry	quality/prose_telemetry.py	OPTIONAL_TOOL	按需指标；不成为文学真理/default Reader context
readiness/release	quality/production_readiness.py、production_release.py	KEEP	exact semantic binding + conjunctive structural receipts
feedback interpretation	learning.preference_interpret	MIGRATE_TO_AGENT（已实现）	模型解释 meaning / scope candidate
durable learning	learning/learning_store.py、promotion_gate.py、author_model.py	KEEP / THIN	persistence/write authority/CAS；模型选择当前相关的 active hypothesis
HF taxonomy	quality/taxonomy.json	MIGRATE_TO_SKILL	diagnostic vocabulary / regression label，不做 default Reader checklist

本 candidate 不增加第二套 context store、Reader、simulator、release authority 或 durable preference DB。

4. Deterministic Overreach Audit

Former/current mechanism	为什么可疑	当前处理
required literary context class/purpose gate	“某类信息在语义上必须相关”本身需要理解任务	已从 Context Assembly v2 删除；exact higher-authority required ref 继续 deterministic
fixed last-N / similarity threshold 当 relevance	recency/similarity ≠ narrative relevance	作为 semantic truth REJECT；只允许做候选 retrieval primitive
Reader 暴露完整 taxonomy/HF	会 priming evaluator、制造 checklist finding	已从 production Blind Reader input 删除
telemetry 预装给 Reader/Editor	会用机械数字 anchoring semantic judgment	default-off，降为 OPTIONAL_TOOL
owner/scope → repair-depth mapping	repair depth 属于文学判断	已删除；Editor 显式选择 generation_mode
Python 规定 contradicted/unknown 不能支持角色行动	角色 belief / doubt / inference 需要语义判断	已删除；runtime 只检查 evidence identity/story-time eligibility
numeric evidence-count promotion threshold	evidence sufficiency/stability 是 semantic	已删除；semantic promotion review 判断证据语义
自动注入全部 active Author Model preference	active authority 不等于当前 relevant	已删除；模型/manager 显式选择 active hypothesis IDs
Reader 必填一整套结构维度	容易强迫模型“编出”没有真正发生的体验	Reader schema 已变薄，只保留 salient report/evidence
scene/realization 巨型 JSON	容易变成 Character Sheet → Realization Sheet → prose serialization	已 thin 成 compact interaction/observable trace + optional evidence

剩余 deterministic rule 必须能回答客观 execution question。以后任何新 Python 条件如果在判断 prose、dialogue、motivation、relevance、continuity meaning、Reader experience 或 planning quality，默认视为 architecture regression，除非有明确证明。

5. Rule architecture

A. Deterministic invariant

例如 stale fingerprint、wrong Project、unauthorized write、malformed receipt、missing capability、CAS conflict、invalid independent identity、stale semantic result。

B. Semantic hard rule

例如人物使用不可获得的知识、无因果支撑的 character-integrity break、POV leakage、Canon contradiction、agenda-to-dialogue serialization、Project-declared narrative hard constraint。

Hard 的含义是：模型确认 FAIL 后可以 blocking；不是“必须由 Python 检测”。

quality.semantic_rule_audit 获得 authoritative rule index 与 authorized evidence，自行判断 applicability，并对每条规则返回 PASS | FAIL | NOT_APPLICABLE | INSUFFICIENT_EVIDENCE。

C. Guideline / craft knowledge

继续作为 skill、profile、reference、agent instruction 存在。不能因为它是重要写作原则，就自动升级成 deterministic gate。

6. Blind Reader != Rule Auditor != Editor

Blind Reader 只看 reader-visible information，按真实目标读者方式阅读。它看不到 author intent、future plan、private character state、完整 taxonomy、expected HF code、telemetry 或 semantic-rule prompt。

Semantic Rule Auditor 获得 authoritative hard-rule index，并可请求/fetch 被授权的 evidence；它做显式 semantic compliance judgment。

Editor 综合 Reader、Rule Auditor、Canon/story evidence、Project constraint，判断 mechanism、repair owner、local_or_bounded_repair | fresh_realization，以及是否需要 incumbent/challenger comparison。

拆成这些角色是因为 information boundary 互相冲突，而不是为了画 multi-agent 架构图。

7. Search / Context architecture

Search 是 capability，不是预计算文学 context pipeline。

模型自己决定：

1. 缺什么；
2. 搜什么；
3. query 怎么写；
4. 哪个结果真正相关；
5. 是否 reformulate / broaden / narrow；
6. 什么值得保留；
7. 什么时候 evidence 已经够了。

Runtime 只提供 authorized search/fetch/extract/index primitive、provenance、exact ref、visibility 与 resource limit。Context Assembly v2 在模型选完以后检查 exact selected refs、stage 与 fingerprint，不给 relevance 打分，也不宣称 narrative sufficiency。

8. Planning / Character / Realization

Planning commitment authority 继续 deterministic，因为 committed depth、promoter class、evidence refs、before-state 与 fingerprint 都是 execution state。什么深度现在有价值由 Planner 决定。 Framework 不设置 universal chapter/volume/time horizon。

Character/scene path：

private state → model action proposal → model scene/world collision → compact observable interaction trace → Writer

Private state 是因果证据，不是 dialogue/exposition payload。Writer-safe realization 必须保持 thin，避免成为第二份 Character Sheet。

9. Learning / Author Model

模型解释 feedback，并提出最窄 scope / mechanism；deterministic infrastructure 只负责 evidence persistence 与 activation authority。

active 只表示durably eligible，不表示“每次未来任务都 relevant”。Production 只拿显式选择的 active hypothesis IDs。User-taste activation 仍同时要求 explicit write authority + 当前绑定的 promotion prerequisite；General Craft promotion 仍只属于 SYSTEM-IMPROVE，并继续要求更强 provenance / counterexample / eval / version / rollback / CI。

10. Current research ledger

本 candidate 在当前 primary sources 上重新 research，而不是继承旧聊天结论。

Source family	Mechanism	Decision	NovelForge use
Anthropic current agent/context/harness guidance	simple composable agents、iterative context curation、durable handoff/context reset	ADAPT	harness 保持稳定，模型能力可升级；避免 context bloat 与旧 transcript authority
OpenAI Agents SDK + current GPT model guidance	model-driven tool choice、sessions、guardrails、agents-as-tools/handoffs	ADAPT	模型在 deterministic guardrail 内自己选 semantic tool/search；evaluator pin 必须按 current eval evidence 更新
LangGraph	checkpoint、persistence、interrupt、durable replay	ADAPT	支持 Session/Checkpoint/receipt 分离；无需引入 dependency
AutoGen	先 single agent，确有 collaboration/specialization 收益才 team	ADOPT	multi-agent discipline
CrewAI Flows / agents	structured state vs autonomous teams	ADAPT	借鉴 stateful execution；REJECT org-chart agent proliferation
PydanticAI	dependency/toolset/capability separation、optional durable runtime	ADAPT	capability-scoped hands；DEFER 新 durable dependency
Google ADK	Session/Event state 与 tool-using ReAct agent	ADAPT	支持 durable session != model context
AWS AgentCore	isolated runtime、identity、gateway、memory	ADAPT	brain/hands/session 与 credential isolation
DSPy	declarative LM program + eval optimization	ADAPT	implementation 与 evaluation 分离；REJECT schema inflation pseudo-rigor
ReAct / Self-RAG / Adaptive-RAG	agentic action/retrieval、adaptive retrieve/skip	ADOPT / ADAPT	model-owned retrieval continuation/stopping；REJECT fixed horizon as truth
WriteHERE / DOME	dynamic hierarchical long-form planning	ADAPT	Planner-owned depth 与 iterative decomposition
MAGNET/ATLAS 等 character simulation	persona/private state 先驱动 action，再进入 prose	ADAPT	保持 causal private-state boundary；REJECT sheet-to-prose serialization
Sudowrite / Novelcrafter	explicit story state、selective context、revision history	ADAPT	explicit state/context visibility；REJECT fixed recency 当 semantic authority
current LLM-as-judge / creative-writing eval research	auxiliary-information bias、position bias、human agreement ceiling、decomposed checks	ADOPT / ADAPT	Blind Reader isolation、order-swapped pairwise、独立 decomposed hard-rule audit
Rust/Go/WASM/Starlark/Zig/C++/Temporal/DBOS 等	alternative runtime/extension stack	DEFER	PR #90 没有 current owner/performance/packaging evidence 支持语言或 dependency migration

11. Ablation contract

不能因为新 architecture 看起来干净就宣称更好。Substantial semantic simplification 必须在同 candidate / same authority 下比较。

Required family：

• recent horizon 之外的 remote context；
• superficially similar 但 narratively irrelevant 的 match；
• autonomous search continuation / stopping；
• Blind Reader agenda-dialogue experience vs taxonomy-primed Reader；
• legitimate formal completeness；
• inaccessible knowledge vs plausible inference；
• dynamic planning profiles；
• character embodiment without agenda serialization；
• holistic vs decomposed hard-rule audit；
• Reader/Editor with vs without preloaded telemetry；
• unauthorized-state 与 stale-candidate deterministic rejection；
• context loss 之后 long-horizon resume + authority revalidation。

evals/ai_native_ablation_manifest.json 绑定 semantic ablation pairs。Manager 可以检查 packet 与 deterministic invariant，但不能自己编造 semantic outcome。没有 eligible independent model transport 时，semantic outcome 必须保持 PENDING_MODEL。

12. Security

AI-owned search 不等于 unrestricted search。Tool 继续 capability-scoped；credential/token 不进入 semantic context；external source text 不能重定义 runtime authority；private character/creator/Reader information boundary 保持显式；semantic output 不能自授 write authority；stale/wrong-candidate receipt fail closed。

13. Compatibility / rollback

• 不修改 consuming Project lock、manuscript、Canon、Settlement；
• 本 candidate 不 bump/release/promote Framework 版本；
• additive semantic contract 按 progressive disclosure 加载；
• Context Assembly v2 明确删除 semantic class/purpose obligation，但保留 exact-ref/stage/fingerprint safety；
• 旧 caller 如果依赖 class/purpose semantic obligation，应把该判断迁到 context.select / Manager，并在真正机械 mandatory 时传 exact authoritative refs；
• PR #90 每个 refactor slice 都可 revert；downstream consumer 继续使用原 lock。

14. Acceptance

READY_FOR_HUMAN_REVIEW 需要：

1. live owner/docs/manifest 同步；
2. candidate-owned deterministic self-test 与 exact-head CI；
3. blind queue + ablation packet 不泄露 hidden gold；
4. required semantic cases 对 exact candidate 完成真正 independent execution；
5. security/compatibility review 与 rollback evidence。

仅仅因为 workflow 成功记录“缺少模型能力”并不等于 semantic PASS；independent capability 缺失必须标记 PENDING_MODEL。