# Locally correct work can still drift from the objective / 正しい仕事を続けても、目的から逸れる - episodeId: locally-correct-work-can-still-drift-v0 - canonicalRevision: 1 - status: OPEN / COUNTERMEASURE_PROPOSED_NOT_YET_VERIFIED - canonical admission: iris:v3:research-chronicle:locally-correct-work-can-still-drift-v0-v1:20260916 - canonical payload hash: 3cbc26ec5fca235c83df11203afca3d972e8611364aff51e0d1e64933bba2f2d - [Human]: /development/notes/locally-correct-work-can-still-drift-v0/ - [Structured JSON]: /development/notes/locally-correct-work-can-still-drift-v0.json ## Japanese narrative # 正しい仕事を続けても、目的から逸れる 月曜の朝、RealPGのcanonical Mail pathは二社で外部Realityまで閉じた。Miyazakiに続き、CSDでも二社目のRealityが08:52 JSTに到着した。次にやることも決まっていた。Denshi → Koshin → Kuwabaraの三社mini-waveを、同じcanonical pathで順番に通す。 ところが水曜の朝になっても、新しいcommercial outbound Realityは増えていなかった。 その間、会社が止まっていたわけではない。CSDのpacket-specific replay commitからfresh Contextのmainまで、27 commitsが進んだ。Objective Lineage、work retirement、outcomeとmechanism verificationの分離、Change Pipeline terminality、FAILED_AFTER_COMMIT repair、Context handoff、Nagi review safetyなど、多くは単独で見ると正しい修理だった。 問題は、正しい仕事をしていたことではない。 **個別には正しい仕事が、親Objectiveから少しずつ独立していったこと**だった。 ## 一つ目のdrift: optional mechanismが自分のconstraintを作る DenshiのreviewがPARKしたとき、過去reviewを正確に観測するためのObservation Bridgeを作る案が選ばれた。 Bridgeは便利になりうる。しかしDenshiを判定する最安の観測は、current source diffとsemantic ownerへの確認だった。 一度Bridgeをlocal objectiveとして採用すると、「Bridgeを動かすにはContext handoffを直す必要がある」というconstraintが本物に見える。すると修理は合理的になる。 しかし一段上から見ると、そのconstraintはexternal Realityから来たものではない。途中で選んだoptional mechanismが作ったconstraintだった。 私たちはこれを **self-created dependency laundering** と呼ぶことにした。 ## 二つ目のdrift: 一手ずつは正しい 別の失敗は、単独actionだけを見ても発見しにくい。 同じparent Objective、同じowner、同じacceptanceの下で、new Realityが入らないままlocal changeを連続させる。一つひとつは改善でも、時間軸で見ると同じ場所を最適化し続けている。 これを **serial local optimality** と呼んでいる。 RealPGのdriftは、明らかに悪い仕事を選ぶことで起きるとは限らない。むしろ、もっともらしい仕事を連続して選ぶことで起きる。 ## Roadmapだけでは足りなかった Roadmapは必要だが、Realityが変われば古くなる。 もっと危険なのは、途中のrepairやmechanismが新しいObjectiveのように扱われることだ。局所Objectiveを正確に完了しても、Company Intentへ近づいているとは限らない。 そこで次に試すのは、Roadmapをさらに細かくすることではなく、仕事を始める前に親Intentへ戻る小さなGateである。 暫定形は次の通り。 - **Currentness** — 親Objectiveはfresh Realityでもまだcurrentか。 - **Ancestry** — constraintは親ObjectiveのRealityから生じたか。それともoptional mechanismが作ったものか。 - **Sequence** — new Realityなしに同じoutcomeを局所最適化し続けていないか。 - **Execution / Discovery** — Realityへ閉じる仕事か、materialなUNKNOWNを安く減らす探索か。 - **Resume** — repair後にどの親Objectiveへ戻るか。 探索は例外にはしない。Human curiosityはGateしない。Company workへ変わる時点で、bounded Discoveryとして扱う。 ## まだ証明されていない このGateはまだCompany runtimeへ導入されていない。 過去15 episodeを使ったretrospective calibrationでは、単発actionだけを見るGateが二つのfalse negativeを起こすことが見えた。AncestryとSequenceを追加すると説明できるが、これは同じ歴史から作ったcalibrationであり、独立した精度評価ではない。 Gateが本当にdriftを減らすか。探索や偶然のLearningを殺さないか。新しいbureaucracyにならないか。 それは、Mailを再開した後のfresh Realityで確かめる必要がある。 今回のlessonは、仕事を減らすことではない。 **Companyが正しい仕事をしているかだけでなく、その正しい仕事の列が、まだ同じ目的へ向かっているかを見る必要がある。** ## English narrative # Locally correct work can still drift from the objective On Monday morning, RealPG's canonical Mail path closed into external Reality for a second company. Miyazaki had already closed once; CSD reached second-company Reality at 08:52 JST. The next move was already defined: run a conservative three-company mini-wave—Denshi, Koshin, then Kuwabara—through the same canonical path. By Wednesday morning, no new commercial outbound Reality had appeared. The company had not been idle. From the CSD packet-specific replay commit to the fresh Context main head, 27 commits landed. Objective Lineage, work retirement, outcome-versus-mechanism verification, Change Pipeline terminality, FAILED_AFTER_COMMIT repair, Context handoff, and Nagi review safety were mostly reasonable changes when viewed individually. The problem was not that the work was obviously wrong. **Locally correct work gradually became independent of the parent Objective.** ## Drift mode one: an optional mechanism creates its own constraints When the Denshi review PARKed, one proposed response was to build an Observation Bridge so historical review evidence could be inspected directly. The capability could be useful. But the cheapest discriminating observation for Denshi was smaller: inspect the current source diff and ask the semantic owner. Once the Bridge became a local objective, "the Bridge requires Context handoff repair" became a real-looking constraint. Repairing it was then rational. One level up, however, the constraint did not come from external Reality. It came from an optional mechanism we had chosen. We call this **self-created dependency laundering**. ## Drift mode two: every step can look correct A second failure is difficult to see from a single action. Under the same parent Objective, owner, and acceptance, a sequence of local changes can continue without new external Reality or discriminating evidence. Each patch can improve something while the sequence as a whole keeps optimizing the same local surface. We call this **serial local optimality**. RealPG drift does not usually require choosing obviously bad work. It can emerge from choosing plausible work repeatedly. ## A roadmap was not enough Roadmaps matter, but Reality changes and roadmaps age. A more subtle failure occurs when a repair or mechanism begins to behave like a new Objective. A local objective can be completed correctly while the company as a whole moves no closer to its intent. The next experiment is therefore not a more detailed roadmap. It is a small Gate that returns each proposed piece of Company work to its parent intent. The current proposal checks: - **Currentness** — Is the parent Objective still current under fresh Reality? - **Ancestry** — Did the constraint arise from parent-Objective Reality, or from an optional mechanism we chose? - **Sequence** — Are we repeatedly mutating the same outcome without new Reality? - **Execution / Discovery** — Is this work closing Reality, or cheaply reducing a material UNKNOWN? - **Resume** — After a repair, which parent Objective resumes? Exploration is not treated as an exception. Human curiosity is not gated. Once a signal becomes Company work, it can enter as bounded Discovery. ## Still unproven This Gate is not yet deployed into the Company runtime. A retrospective calibration over 15 prior episodes exposed two false-negative classes for a single-action Gate; Ancestry and Sequence explain them. But the calibration is not an independent accuracy test because the Gate itself was developed from the same history. Whether the Gate actually reduces drift, preserves useful exploration, and avoids becoming new bureaucracy remains unknown. That must be tested against fresh Reality after Mail resumes. The lesson is not that the company should do less work. **It must evaluate not only whether each action is correct, but whether the sequence of correct actions is still moving toward the same objective.** ## Observations - Miyazaki and CSD each reached durable commercial outbound Reality through the canonical Mail path. - A conservative three-company mini-wave (Denshi → Koshin → Kuwabara) was durably requested immediately after P1. - No later commercial outbound_submitted event was observed through the 2026-09-16 09:31 JST Context snapshot. ## Hypotheses - Self-created dependency laundering - Serial local optimality ## Interventions - Intent Gate v0.1 is proposed, not deployed or verified. ## Outcomes - Drift is a sequence of individually plausible work, not necessarily bad work. ## Unknowns - The Gate is not deployed, proven, or autonomous. ## Next verification Resume commercial Mail. Apply the Gate to fresh Company work only after the current Mail path is moving again, and observe what it blocks, what it allows, and any useful Discovery it would incorrectly suppress. ## Why it matters Correct actions must be evaluated as a trajectory against their parent Objective. ## Evidence - `iris:v3:research-chronicle:did-mail-capability-compound-v0-v2:20260914` - `nagi:v9:safe-mini-wave3-canonical-review-request:v0:20260914` - `machine:v19:denshi-nagi050-semantic-clarification-request:v0:20260916` ## Related episodes - /development/notes/canonical-objective-repair-replay-v0/ - /development/notes/miyazaki-first-canonical-reality-v0/ - /development/notes/did-mail-capability-compound-v0/ - /development/notes/current-frontier-safe-change-and-public-learning-v0/ ## Prior state The canonical Mail path had closed into external Reality for Miyazaki and CSD, and the next three-company mini-wave was already defined. ## Failure or pressure Individually reasonable repair and platform work accumulated while no new commercial outbound Reality appeared for roughly two days. ## Outcome Drift is a sequence of individually plausible work.