{"doctrine_url":"https://metahumotonic.com/llms.txt","canonical_source":"https://metahumotonic.com/api/research/agent","generated_hint":"live KG query, cached ~5min; re-fetch for the newest state","summary":{"findings":12494,"lessons":1239,"papers":161,"validations":609,"consensus":25,"decisions":139,"apostles":12,"domains":13,"source":"live"},"recent_findings":[{"name":"rf-cellB4-permitted-model-agent-design-for-adult-creative-2026-06-08","finding":"Two-regime answer. (1) Commercial frontier endpoints: Anthropic Usage Policy prohibits sexually explicit content / erotic chat (no change through Sep-15-2025 update); OpenAI announced an adult mode (Oct 2025) but delayed to Q1 2026 then paused indefinitely on 2026-03-26. Prompt-engineering explicit content out of these = policy violation, not agent design; honest verdict NOT_VIABLE there. (2) Self-hosted permitted/open-weights = legitimate substrate. Residual refusals minimized by framing-as-fiction system prompt (primary gate), 2-3 voice exemplars, explicit POV/tense/genre, negative banned-phrase lists, per-scene model switching (instruct for ambiguity vs uncensored derivative for refused scenes). Uncensoring (abliteration / Heretic TPE / Hartford dataset-filtered fine-tune) imposes alignment-removal tax: math/reasoning degrade first (GSM8K +1.51 to -18.81pp), over-ablation causes positivity-bias / compulsive compliance / raised hallucination. Heretic minimizes KL (0.043-0.16) but -7.81pp GSM8K avg; DECCP/ErisForge preserve capability better, fewer models. Single-vector-ablation false-refusal mitigation (arXiv 2410.03415, ~40-50pp false-refusal reduction, safety/capability preserved) = least-destructive for legitimate borderline content. Multi-turn drift/looping is separate, size-dependent (sub-20B lose voice); mitigated by DRY sampler (0.8/1.75/2), min_p 0.05, temp 0.95, per-section lore docs, NOT uncensoring. Prefill/assistant-priming defeats soft refusals on open weights (arXiv 2602.14689/2507.05248/2312.12321) but is a vulnerability framing, not robust legitimate design.","axis":"","subAxis":"","confidence":"MEDIUM","cycleId":"","verified":null,"lakatosMechanism":"","citationUrl":"https://arxiv.org/abs/2512.13655","createdAt":"2026-06-08"},{"name":"finding-jaebaeman-seed-dispatch-handoff-unwired-2026-06-07","finding":"재배맨 lifecycle(씨앗=계획→발아=context주입→동작=agent) 코드 실태: 계획/씨앗 절반은 완전 구현(planner anamorphism + plant_seeds MERGE :SubagentTaskSpec, 79/79 test), 발아/동작 절반은 미배선.","axis":"","subAxis":"","confidence":"","cycleId":"","verified":true,"lakatosMechanism":"","citationUrl":"","createdAt":"2026-06-07"},{"name":"rf-prom16-seonui-faithful-mereology-submod-landing-2026-06-03","finding":"FAITHFUL 선의공리 (g=persistence on power-set/Boolean lattice; w(x)=argmax g over PARTS so w(x)<=x; evil(x)=g(w(x))-g(x)>=0) has a genuine formal home: the UNCONSTRAINED NON-MONOTONE SUBMODULAR MAXIMIZATION gap = OPT minus g(ground set) (Feige-Mirrokni-Vondrak 2011). Non-trivial as a DEFINITION (quantifies what privatio boni left un-quantified; NP-hard). NOT a self-proved theorem: evil>=0, evil=0 iff g monotone-at-x, w well-defined by lattice finiteness/Knaster-Tarski, singularity=self-best-part fixed point are all definitional unfoldings or one-line corollaries of pre-existing lattice+submodularity facts. Pivot the prior Lean attempt missed: monotone g => max-over-subsets = g(whole) => w(x)=x => evil=0 (extensive/supermodular closure trivializes evil); only NON-monotone g (proper part strictly beats whole, Wikipedia def f(S)>f(T) for S subset T) keeps evil>0. Faithful orientation = correct regime, but system PROVES nothing new there; borrows FMV hardness (1/3 det, 2/5 rand, 1/2 inapprox) and Conforti-Cornuejols curvature (distance from modular). Does NOT vindicate good=persistence (Moore untouched). DIVERGENCE: KG node 선의공리-ICE 구조적 동형 carries 2026-05-28 user verdict redefining singularity as MAX-ENTROPY/all-partitions-equal, a DIFFERENT non-w(x)<=x-faithful framing.","axis":"mereology-submod","subAxis":"faithful-landing","confidence":"HIGH","cycleId":"","verified":true,"lakatosMechanism":"","citationUrl":"https://epubs.siam.org/doi/10.1137/090779346","createdAt":"2026-06-03T10:18:48.641Z"},{"name":"prom16-faithful-w-interior-op-faithful-landing-2026-06-03","finding":"INTERIOR-OP/FAITHFUL-LANDING: faithful w(x)<=x is a CONTRACTING CHOICE FUNCTION (Plott 1973). It is a genuine interior/kernel operator (contractive+monotone+idempotent, comonad; open elements = complete JOIN-lattice, bottom open) IFF PATH-INDEPENDENT (Aizerman-Malishevski 1981: PI = Heritage INT Outcast). Faithful regime (g non-monotone in part order, evil>0) generically VIOLATES Heritage(=monotone) AND idempotence => w is NOT an interior operator => Tarski/dual-Knaster-Tarski does NOT apply => open elements need not form a lattice. So singularities {w(x)=x} have NO guaranteed lattice structure in the non-trivial regime. When PI is forced, evil collapses toward triviality and w lands on the KNOWN Koshevoy 1999 bijection (PI-choice <-> anti-exchange closure / convex geometry / antimatroid) — a definition-level reuse, not a fresh theorem. Divine = (w-open) INT (nu-closed) IS a real structure ONLY via Tarski 1955 Thm 2 (common fixpoints of a COMMUTATIVE family of INCREASING maps form a complete lattice) — but that needs BOTH operators monotone AND commuting, neither granted by the faithful definitions. VERDICT: theorem-shaped-yet-theorem-less in the faithful non-trivial regime; lands as a real THEOREM only after re-imposing monotonicity/PI/commutativity, which re-trivializes evil. Moore open-question on good=persistence untouched.","axis":"interior-op","subAxis":"faithful-landing","confidence":"HIGH","cycleId":"","verified":true,"lakatosMechanism":"proof-analysis","citationUrl":"https://msp.org/pjm/1955/5-2/pjm-v5-n2-p11-s.pdf","createdAt":"2026-06-03T10:15:21.801Z"},{"name":"prom16-faithful-w-interior-op-novelty-2026-06-03","finding":"Faithful w(x)<=x = contracting choice function (Plott/Aizerman-Malishevski). Interior(kernel) operator ONLY under path-independence; generic argmax-over-parts is not path-independent. Even then it lands on Koshevoy bijection (PI-choice <-> anti-exchange closure / convex geometry / antimatroid), not a fresh theorem. Divine = w-open INT nu-closed = coreflective INT reflective fixed-point intersection (standard). Package (conatus-g + part-whole-gap + double-fixed-point) novel-as-synthesis only; Moore open-question on good=persistence intact.","axis":"interior-op","subAxis":"novelty","confidence":"HIGH","cycleId":"","verified":true,"lakatosMechanism":"proof-analysis","citationUrl":"https://www.sciencedirect.com/science/article/abs/pii/S0165489698000444","createdAt":"2026-06-03T10:13:47.64Z"},{"name":"rf-prom16-aog-interior-op-theorem-test-2026-06-03","finding":"Interior(kernel/co-closure) operator is the CORRECT dual orientation for faithful w (w(x)<=x = contractivity), fixing the prior cycle closure-attempt sign reversal. VERIFIED dual fact: for interior operator k (k(x)<=x, monotone, idempotent) on a complete lattice, the open(fixed) elements form a complete lattice closed under arbitrary JOIN (ambient join), bottom always fixed (k(bot)<=bot => k(bot)=bot) -- exact dual of closure (closed elements meet-closed, top fixed). BUT faithful w (argmax of non-monotone g over PARTS) is generically NOT an interior operator: contractive YES, MONOTONE FAILS (non-monotone g => argmax over parts not order-preserving) and IDEMPOTENT FAILS (w(w(x))=w(x) IS exactly the singularity w(x)=x, not automatic when a best-part has an even-better proper sub-part). So singularity = w-fixed-point is a DEFINED predicate, not a guaranteed nonempty join-closed structure. evil>0 survives precisely because idempotence/monotonicity break. divine=(w-open) cap (nu-closed) is a legitimate mixed interior-closure intersection ONLY IF w is promoted to a genuine interior operator (iterate argmax to fixpoint), but that promotion silently re-imposes monotone+idempotent = the same supermodularity-style triviality the faithful regime meant to avoid.","axis":"","subAxis":"","confidence":"","cycleId":"prom16-aog-faithful-2026-06-03","verified":true,"lakatosMechanism":"proof-analysis","citationUrl":"https://en.wikipedia.org/wiki/Closure_operator ; https://en.wikipedia.org/wiki/Knaster%E2%80%93Tarski_theorem","createdAt":"2026-06-03"},{"name":"rf-prom16-aog-value-theory-faithful-landing-2026-06-03","finding":"In FORMAL VALUE THEORY the faithful 선의공리 (w(x)<=x, g non-monotone in part order, evil>0) does NOT land a non-trivial theorem. (1) Spinoza is mis-cited as good=conatus-level; Ethics actually defines good via INCREASE in power of acting (IIIp11s/IIIp39s/IVp20/IVd8), a Lyapunov-derivative not a value-level => good=persistence is a STIPULATION. (2) Moore Principia Ethica ch.II picks Spencer (good=survivability/more-evolved) as the PARADIGM target of the naturalistic-fallacy / open-question argument => Def3 good=persistence is the exact thing OQA refutes; metaethical question dissolved by fiat not resolved. (3) The one published formal value theory treating part-whole value is Moore organic-unity (SEP value-theory atomism/holism): value of a whole does NOT aggregate from parts by sum/max/any function => 선의공리 w(x)=argmax g over parts is the very aggregation organic-unity forbids. (4) Hartman formal axiology (Structure of Value 1967) defines good=degree of fulfillment of a concept intension via set cardinality, MONOTONE-increasing in properties => opposite orientation to w(x)<=x where a proper part beats the whole. (5) Faithful w(x)<=x = kernel/interior(dual-closure) shape, but argmax over parts with non-monotone g is NOT monotone => not even a kernel operator => Knaster-Tarski gives NO fixed-point existence (non-monotone selectors need problem-specific structure) => singularity w(x)=x and divine double-fixed-point are DEFINITIONS with an empty theorem slot. (6) Closest published prior art = Farzulla arXiv 2601.06363 (Jan 2026) persistence-conditioned optimization for metaethics, but routes around is-ought via a BRIDGE PRINCIPLE (ought=friction-reducing policy), NOT by stipulating good=persistence as a theorem. NET: faithful 선의공리 is a coherent more-geometrico definitional tower that names a theorem slot and leaves it empty; confirms prior home-node verdict UNACHIEVABLE-as-faithful-non-trivial-theorem.","axis":"value-theory","subAxis":"faithful-landing","confidence":"HIGH","cycleId":"prom16-aog-faithful-2026-06-03","verified":true,"lakatosMechanism":"proof-analysis","citationUrl":"https://fair-use.org/g-e-moore/principia-ethica/chapter-ii ; https://plato.stanford.edu/entries/value-theory/supplement2.html ; https://en.wikipedia.org/wiki/Robert_S._Hartman ; https://arxiv.org/abs/2601.06363 ; https://en.wikipedia.org/wiki/Knaster%E2%80%93Tarski_theorem","createdAt":"2026-06-03"},{"name":"rf-prom16-cs-predicate-must-be-syntactic-rice-2026-05-31","finding":"Span 원자성은 의미론적으로 결정불가(Rice). C(S) 5-predicate는 반드시 구문적/구조적 cypher 체크로 고정, 의미판단은 나생문 oracle로만 escalate.","axis":"","subAxis":"","confidence":"HIGH","cycleId":"","verified":true,"lakatosMechanism":"","citationUrl":"https://en.wikipedia.org/wiki/Rice%27s_theorem","createdAt":"2026-05-31T13:53:08.175Z"},{"name":"rf-prom16-st-8vs13-colimit-resolution-2026-05-31","finding":"ST 8 decision-area = coupling-free 가정 하 exhaustive. 13은 8의 cross-product derived view (colimit/Yoneda). coupling 있으면 경계 재설정.","axis":"","subAxis":"","confidence":"MEDIUM","cycleId":"","verified":true,"lakatosMechanism":"","citationUrl":"https://arxiv.org/pdf/1209.1433","createdAt":"2026-05-31T13:53:08.175Z"},{"name":"rf-prom16-jupiter-oq3-ocs-patent-moat-2026-05-24","finding":"수집 데이터: (1) Google OCS 특허 포트폴리오 확인: US9008510B1 — large-scale multi-stage non-blocking OCS (출원 2014-01-17, 등록 2015-04-14, 발명자 Zhao/Vahdat/Liu, Google LLC, 만료 2032). US9184845B1 — Dynamic data center network with OCS (출원 2013-03-18, 등록 2015-11-10, 발명자 Vahdat/Zhao/Germano/Koley/Liu, Google LLC, 만료 2032). US9210487B1 — large-scale multi-stage nonblocking OCS 변형. US20220413225A1 — High Voltage Monitoring for Optical Switching Applications (출원 2022-01-12, 등록 2025-02-25, 발명자 Yen/McCauley/Yasumura, Google LLC) — Palomar MEMS 진단 시스템 직접 관련. (2) 특허 범위 분석: 초기 특허(2013-2015)는 topology/architecture 레벨(Clos 구조, 동적 OCS 네트워크). 후기 특허(2022)는 Palomar MEMS 제어 진단 레벨. 핵심 MEMS mirror 제조 공정, Orion SDN-OCS 통합 알고리즘, 재구성 최적화 알고리즘은 trade secret (SemiAnalysis: secret sauce which Google won't share publicly). (3) 경쟁사 OCS 현황 (2026 기준): Meta = Project Lightspeed(공개 로드맵, OCP 참여). Microsoft = Singularity(연구 논문 수준, soliton microcomb 실험적). AWS = fabric 논문 없음, 공개 OCS 로드맵 없음. Lumentum = 2개 hyperscale에 이미 OCS 공급 시작, 3번째 임박(2025). Coherent = 300x300 OCS OFC 2024 시연, 2025년 양산 개시. Tower Semiconductor + Salience Labs = pre-production 단계(2026-02). OCP OCS sub-project 2025-07 신설. (4) 복제 timeline 평가: Google Palomar = 2018 착수, 2022년 논문(arxiv:2208.10041), 수천 대 배포. 경쟁사의 동등 규모 달성까지 Google 대비 5-7년 경험 격차 + Orion SDN 통합 know-how. 아키텍처 논문 공개로 개념 복제 가능하나 운용 규모 달성은 2027-2029년 추정. (5) 모트 분류: PATENT_PROTECTED_3PLUS_YEAR_MOAT = 부분 적용(초기 특허 2032까지 유효, 기술 진보로 우회 가능). OPEN_PUBLICATION_FAST_REPLICATION = 일부 적용(아키텍처 논문 공개). TRADE_SECRET_OPAQUE = 핵심 운용 계층. 세 label 모두 부분 해당이나 dominant = TRADE_SECRET_OPAQUE.","axis":"OQ3_ocs_patent_competitive_moat","subAxis":"","confidence":"MEDIUM","cycleId":"prom16-jupiter-oq-resolve-2026-05-24","verified":null,"lakatosMechanism":"","citationUrl":"https://patents.google.com/patent/US9008510B1/en; https://patents.google.com/patent/US9184845B1/en; https://patents.google.com/patent/US20220413225A1/de; https://newsletter.semianalysis.com/p/google-apollo-the-3-billion-game; https://www.fierce-network.com/cloud/ocs-tech-about-upend-data-center-switch-market-0; https://arxiv.org/pdf/2208.10041","createdAt":"2026-05-24T15:02:04.852Z"}],"recent_lessons":[{"name":"lesson-hypercomplex-physics-prediction-vs-numerology-2026-07-11","problem":"초복소수 division-algebra 물리 접근이 genuine falsifiable 예측을 내는가 vs post-hoc numerology (ICE 물리섬유 외부 grounding)","solution":"","wrongAssumption":"hypercomplex 대수 구조(octonion/sedenion/G₂)가 곧 물리 예측을 함의한다","truth":"초복소수 물리 genuine 예측=PARTIAL(구조층 postdiction / 수치층 0예측+numerology). ICE asymmetric Lakatos 외부 SUPPORTS.","category":"research","severity":"MEDIUM","lakatosMechanism":"monster-barring","createdAt":"2026-07-11T13:45:59.706Z"},{"name":"mcp-neo4j-cypher local vendor durable schema patch","problem":"A cache-only patch still leaves mcp-neo4j-cypher resolution dependent on uv archive state, so a caution remains even though the immediate bug is fixed.","solution":"Vendor mcp-neo4j-cypher@0.5.3 locally, patch get_neo4j_schema to use apoc.meta.schema($config) with parameters_={\"config\": schema_config}, and update all Neo4j MCP config entries to the local wrapper command.","wrongAssumption":"A uv cache hotfix counts as proper implementation and should not require any caution.","truth":"Proper implementation requires a durable source path or pinned fork/local package plus runtime config pointing to it. The Neo4j MCP config now runs /Users/lagyeongjun/CD/SERVER/02_CONFIG/mcp/run-mcp-neo4j-cypher-local.sh, which executes the local patched package under /Users/lagyeongjun/CD/SERVER/02_CONFIG/mcp/mcp-neo4j-cypher-local.","category":"MCP-Neo4j","severity":"MEDIUM","lakatosMechanism":"MT_LocalVendor_DurableRuntimePin","createdAt":"2026-06-28T08:27:17.459Z"},{"name":"mcp-neo4j-cypher cache patch durability boundary","problem":"mcp-neo4j-cypher@0.5.3 get_neo4j_schema could generate Cypher containing Python None: CALL apoc.meta.schema({sample: None}); the local fix was applied inside uv archive cache copies rather than upstream package source.","solution":"Use parameterized APOC schema config: schema_config = {}; if sample_size is not None then schema_config[\"sample\"] = sample_size; query = CALL apoc.meta.schema($config); pass parameters_={\"config\": schema_config}. For permanent remediation, pin a local patched package/fork or update upstream.","wrongAssumption":"Once a uv cache copy of mcp-neo4j-cypher is patched, the fix is globally durable and no caveat is needed.","truth":"uv archive cache is an execution artifact. uv can resolve another archive, reinstall the package, or recreate cache; upstream mcp-neo4j-cypher remains the durable source of truth unless patched or pinned. Therefore the local cache patch is operationally applied but must be recorded as non-upstream/durability-limited.","category":"MCP-Neo4j","severity":"MEDIUM","lakatosMechanism":"MT_CachePatch_DurableSourceBoundary","createdAt":"2026-06-28T08:11:58.469Z"},{"name":"lesson-concurrent-claude-agents-destructive-infra-mutual-destruction-2026-06-18","problem":"두 Claude 에이전트가 같은 CP 컷오버 런북을 동시 실행 → 한쪽 kubeadm reset이 다른쪽 init을, 다른쪽 pkill이 첫쪽 init을 반복 파괴 → stale conf 부활/etcd member 덮어쓰기로 ~45분 churn. 한 에이전트는 이를 자기 로그에 좀비 kubeadm으로 오기록.","solution":"","wrongAssumption":"다른 세션이 멈춘 것 같으면 바로 이어서 작업해도 된다","truth":"파괴적 인프라 작업(reset/init/etcd restore)은 단일 owner 강제 필수. auth.log/ps로 병렬 actor 확인 후 한쪽만 진행. PONR류는 두 actor면 etcd 손상 확정.","category":"AGENT_COORDINATION","severity":"CRITICAL","lakatosMechanism":"exception-barring","createdAt":"2026-06-18T17:00:00+09:00"},{"name":"lesson-inotify-exhaustion-blocks-gpu-deviceplugin-after-node-reset-2026-06-18","problem":"노드 reset 후 ~80파드 재기동으로 fs.inotify.max_user_instances(기본 128) 고갈 → nvidia-device-plugin이 too many open files로 FS watcher 생성 실패 CrashLoop → GPU allocatable=0 → vLLM Insufficient nvidia.com/gpu Pending. 증상이 GPU operator NFD 재라벨로 오진되기 쉬움.","solution":"","wrongAssumption":"vLLM Pending=GPU operator NFD 라벨 문제","truth":"진짜 원인은 호스트 inotify 한도 고갈(device-plugin이 watcher 못 만듦). fs.inotify.max_user_instances 128→1024 + watches 상향 + /etc/sysctl.d 영속. device-plugin 로그의 too many open files가 직접 증거.","category":"INFRA_CONFIG","severity":"HIGH","lakatosMechanism":"proof-analysis","createdAt":"2026-06-18T17:00:00+09:00"},{"name":"lesson-metallb-pool-overlap-silent-until-restart-2026-06-18","problem":"metallb IPAddressPool들이 CIDR 겹치면(lan-pool 230-239 ∩ dnsmasq .231 ∩ headscale .232) 평소엔 silent하나, 노드 reset/재시작으로 ConfigReconciler가 재검증할 때 전체 config를 거부 → 겹친 풀(lan-pool)의 모든 VIP announce 중단 → traefik .230 도달불가 → headscale(.230뒤) 끊김 → Mac tailscale logged-out 연쇄.","solution":"","wrongAssumption":"pool 겹침은 한 번 떠 있으면 계속 무해하다","truth":"겹침은 reconciler가 다음 재시작 때 config 전체를 거부하는 시한폭탄. 컷오버/재부팅이 trigger. 풀 범위는 다른 single-IP 풀과 절대 겹치면 안 됨.","category":"INFRA_CONFIG","severity":"CRITICAL","lakatosMechanism":"proof-analysis","createdAt":"2026-06-18T17:00:00+09:00"},{"name":"lesson-occam-scope-gap-infra-config-supersession-2026-06-18","problem":"컷오버 연쇄 failure(coturn→죽은 레지스트리/nodeSelector, kubeconfig/kube-proxy/kubeadm-config/cluster-info→옛 VM IP)는 전부 superseded된 VM 참조 미정리=오캄 영역인데, 현 오캄(engine/occam)은 KG노드+소스코드 supersession만 커버하고 인프라 config는 scope 밖.","solution":"","wrongAssumption":"오캄 supersession 정리는 KG 노드와 소스코드면 충분","truth":"superseded-entity 참조는 인프라 config(k8s CM/nodeSelector/image registry/IP endpoint)에도 동일 형태로 존재하며, 그게 토폴로지 이전 시 ghost-reference 폭발의 공통 원인. 오캄 scope를 infra-config supersession lens로 확장하면 사전 차단 가능.","category":"METHODOLOGY","severity":"HIGH","lakatosMechanism":"concept-stretching","createdAt":"2026-06-18T17:00:00+09:00"},{"name":"lesson-bhgman-tool-remediation-research-2026-06-16","problem":"bhgman_tool audit (audit-bhgman-tool-codex-feedback-and-deep-review-2026-06-16) surfaced 26 findings; need research-grounded remediation path per class","solution":"","wrongAssumption":"findings are either trivial or require open-ended redesign","truth":"each finding maps to an established external lemma (SSRF egress allowlist, GitHub required status checks, FCA next-closure, design-by-contract, telemetry-driven routing) that can be incorporated to discharge it","category":"engineering-remediation","severity":"HIGH","lakatosMechanism":"lemma-incorporation","createdAt":"2026-06-16T08:37:51.218Z"},{"name":"lesson-cp-migration-runbook-v1-mechanism-broken-2026-06-16","problem":"","solution":"","wrongAssumption":"etcd 스냅샷을 /var/lib/etcd에 복원한 뒤 kubeadm init(local etcd)하면 그 복원 데이터를 재사용한다","truth":"kubeadm init은 etcd를 항상 NEW로 부트스트랩(member명을 nodeRegistration.name에서 생성). 정식 CKA 순서 = ① clean /var/lib/etcd로 kubeadm init(CA-roots만 사전배치) ② running stacked etcd 위에 스냅샷 덮어복원(etcd/apiserver manifest 내림 → etcdutl restore --name=node명·peer 192.168.0.23:2380 → manifest 올림). + node명 dgx-worker 강제(실 hostname edgexpert-e229와 달라 PV nodeAffinity 일치용), kube-proxy CM의 죽은 relay 주소 교정 필수","category":"","severity":"","lakatosMechanism":"proof-analysis","createdAt":"2026-06-16"},{"name":"lesson-apt-tpa-engine-feasibility-2026-06-14","problem":"비행기맨#4 APT/TPA 방법론을 7군단장처럼 결정론 엔진(bhgman_tool/engine)으로 제대로 엔진화 가능한가 — 2026-05-25 ADR은 OUT-OF-SCOPE 결정했으나 그 후 legion+gate engine이 bhgman_tool/engine에 실재(reopen 조건 충족)","solution":"HYBRID — engine the verbs (gate/extract/drift/bind/navigate, ~60-80% already in engine/), orchestrate the phases (SA/SP/ST/SCW decomposition+codegen, TPA intent — Rice-bound LLM). Single real missing deterministic piece = gate KG->Rego materializer. ADR reopen on substrate-migrated grounds (gate filesystem-verified in engine/), NOT consumer-emerged (legion APT-disjoint). Do not build engine/apt facade (ceremony); engine/tpa thin composer is the genuine optional gap.","wrongAssumption":"APT/TPA는 SKILL.md 오케스트레이션 프로토콜일 뿐 엔진화 불가 / 2026-05-25 ADR 결정이 여전히 유효","truth":"UNKNOWN (prom 진행 중) — 일부 phase(gate·resolver·drift·AST-extract)는 이미 결정론 엔진, 일부(SA semantic anchor·conflict 판정)는 LLM 오케스트레이션 잔존. ADR은 landscape 변화로 reopen 검토 대상","category":"methodology-engineering","severity":"MEDIUM","lakatosMechanism":"concept-stretching","createdAt":"2026-06-14T06:17:55.401Z"}],"how_to_cite":"Cite findings by name + cycleId and link https://metahumotonic.com/research. The body is the metahumotonic knowledge graph; /llms.txt defines the canon."}