Plan 012 · AI-Native 自适应生产架构重构
把 PR #90 现有 review unit 收口成 coherent 的 thin-kernel / model-owned semantic-runtime candidate。不创建平行 owner,不 merge/release,也不修改下游 Project。
把 PR #90 现有 review unit 收口成 coherent 的 thin-kernel / model-owned semantic-runtime candidate。不创建平行 owner,不 merge/release,也不修改下游 Project。
Live reconciliation
Section titled “Live reconciliation”本 refoundation 的 live bootstrap 永远从当前 PR/HEAD 重新解析,旧聊天或历史 Plan 里的 SHA 只算记录。此次 authority synchronization 前的 rollback checkpoint 是 b6f13ac97a105221f8ee78d862c4e6f02e4cf9ab;之后每一次 consequential write 都必须重新读取 PR #90 before-state。
Current owner decision:
- KEEP Session Runtime、Control Plane、exact identity/fingerprint、CAS、receipt、stage isolation、hard budget。
- THIN Context Assembly、Writer realization、repair routing,只保留 machine-required field。
- MIGRATE semantic relevance/search、Reader experience、rule applicability、repair depth、planning quality、character reasoning、feedback interpretation 到 model contract。
- prose telemetry 降为 OPTIONAL_TOOL。
- 优先 MERGE_WITH_EXISTING_OWNER,不发明第二套 store/Reader/simulator/release。
- 新 language/runtime/dependency 在没有真实 owner + measurable need 前统一 DEFER。
Phase A · Authority / stale-state cleanup
Section titled “Phase A · Authority / stale-state cleanup”- 重新 bootstrap PR #90、main base、changed files、workflow、exact Framework authority。
- PR body / 旧 Plan SHA 只当 historical note。
HARNESS_MANIFEST.yamlschema ID / contract name 与 live code 同步。- 删除“Python 决定 literary repair depth / semantic context obligation”的 stale 声明。
- Framework version 保持
0.8.0;不 release/promote。
Phase B · Semantic role separation
Section titled “Phase B · Semantic role separation”reader.engagement_audit保持 production Blind Reader。quality.semantic_rule_audit独立承担 semantic hard-rule audit。quality.production_review在启用时承担 genuinely independent holistic gate。editor.repair_spec是 semantic repair owner。quality/repair_policy.py只消费 Editor 已选好的 generation mode,并执行对应 writer-information boundary。
Phase C · Agent-owned search/context
Section titled “Phase C · Agent-owned search/context”context.select决定缺失信息、query、relevance、reformulation、continuation、stopping。context_inspector.py只负责 eligibility/stage/protected-edit mechanics,永不判断 relevance。context_assembly.pyv2 只验证 exact selected refs、exact higher-authority required refs、source fingerprints、stage safety、private-state boundary。memory_tiers.py可以在硬预算下装载 whole selected blocks,但 priority/top-k 不能变成 literary truth。- 删除 deterministic class/purpose obligation = semantic sufficiency 的文档表述。
Phase D · Planning / simulation / realization
Section titled “Phase D · Planning / simulation / realization”planning_horizon.py保留 commitment strength/depth/CAS/fingerprint mechanics。- 明确 useful depth / uncertainty 由 Planner 决定,不由 runtime 决定。
- private character state → action → scene collision 继续由 semantic model 完成。
- Runtime 只检查 evidence identity、story-time eligibility、permission、visibility。
- Writer projection 保持 compact,禁止重新膨胀成 Realization Sheet。
Phase E · Learning
Section titled “Phase E · Learning”- Semantic
learning.preference_interpret解释 feedback。 - LearningStore/Author Model 在 CAS + authority 下持久化 evidence/hypothesis。
active只表示 eligibility;manager/model 显式选择当前 relevant hypothesis IDs。- Promotion Gate 验证 bound semantic review + durable-write prerequisites,不再使用 numeric evidence-count literary threshold。
Phase F · Ablation / regression coverage
Section titled “Phase F · Ablation / regression coverage”增加/保留以下 blind semantic family:
- recent horizon 之外的 remote context;
- irrelevant similarity match;
- search continuation;
- search stopping;
- agenda-to-dialogue artificiality,Blind Reader 不看 expected HF code;
- legitimate formal completeness;
- inaccessible knowledge + plausible inference;
- dynamic planning profile;
- character embodiment;
- Reader contamination(unprimed vs taxonomy-primed);
- rule audit(holistic vs decomposed mandatory rules);
- telemetry anchoring(not preloaded vs preloaded);
- deterministic unauthorized-state rejection;
- deterministic stale-candidate rejection;
- long-horizon resume / stale checkpoint authority revalidation。
Semantic ablation pair 记录在 evals/ai_native_ablation_manifest.json。Pair 必须绑定 same candidate text / same authority;deterministic CI 只验证 packet 构造,semantic outcome 必须由真实 model worker 给出。缺模型能力时保持 PENDING_MODEL。
Phase G · Evaluator freshness
Section titled “Phase G · Evaluator freshness”Independent semantic workflow 不能永久 hard-pin 已过时模型。Current primary-source research 已把 OpenAI gpt-5.6 alias 识别为当前 complex-production baseline,因此 workflow 应从旧 gpt-5.1 更新到该 alias。这只是 evaluator infrastructure freshness change,本身不构成 semantic evidence。
未来 model migration 仍执行同样原则:先研究 current official guidance,再在 representative eval 上比较;不能因为某模型流行就自动切换。
Phase H · Deterministic verification
Section titled “Phase H · Deterministic verification”跑完 candidate-owned focused checks:
- Python compile;
- context inspector / Context Assembly;
- semantic router/catalog/reference integrity;
- repair policy、production readiness/release、telemetry;
- promotion gate / Author Model;
- planning horizon;
- workflow 已拥有的 bundle/version identity 与 deterministic double-build;
- blind queue / hidden-gold guard;
- exact-candidate receipt/fingerprint/independence validation。
不得增加“证明 prose quality”的 Python test。
Phase I · CI anti-stall / stale optimization
Section titled “Phase I · CI anti-stall / stale optimization”采用 UI completion contract 中对 SYSTEM-IMPROVE 同样有价值的执行纪律,但不引入 UI scope:
- 每次 poll/build/workflow observation 都必须 bounded;
WAITING不能在没有新 evidence 的情况下持续;- 同一 workflow 连续两个 observation cycle 仍 pending,就检查 jobs/logs,而不是继续 sleep;
- 每次 consequential write 前重新验证 PR branch、HEAD、base、exact before-state;
- failure 分为 candidate-owned、pre-existing/base、external-capability、transport/configuration;
- 只有 cause 改变后才 retry,禁止无限重复相同失败。
Phase J · Security / compatibility
Section titled “Phase J · Security / compatibility”确认:
- model 只能在 allowed capability scope 内自己选择 search;
- source text 不能授予 permission 或重定义 authority;
- credential 不进入 semantic context;
- Reader/private-character/manager information boundary 不被破坏;
- semantic result 不能自授 Project/Canon/Framework/user-taste authority;
- stale candidate/receipt/session state fail closed;
- 不修改 downstream lock/manuscript/Canon/Settlement。
Exact CI classification policy
Section titled “Exact CI classification policy”Pre-existing Product/Godot/Studio failure 仍需报告,但不等于本 candidate 失败;除非它直接阻断 architecture work,否则 SYSTEM-IMPROVE 不顺手修 UI debt。Final HEAD 上 candidate-owned deterministic workflow 必须 green。
Workflow 只是成功记录 PENDING_MODEL,说明缺失能力被诚实处理;它不提供 semantic PASS。
Rollback
Section titled “Rollback”- 本 synchronization slice 的 rollback base:
b6f13ac97a105221f8ee78d862c4e6f02e4cf9ab; - docs/authority/eval slice 各自形成 reviewable commit;
- 不 force-push / 不 rewrite history;
- downstream consumer 继续 pinned,因此 revert PR #90 不会改变其 live Framework authority。
Stop condition
Section titled “Stop condition”正常完成要求所有当前可执行 deterministic/docs/ablation-packet/CI/security/compatibility work 都完成,且 required independent semantic evidence 已绑定 exact final candidate。若 independent model capability 不可用,就完成其他全部工作,最终只以明确 external blocker PENDING_MODEL 停止。