RESEARCH CHRONICLE · OPEN / PROPOSED_NOT_YET_VERIFIED

解釈を原文にしない

2026-09-18 · Iris canonical revision 1

canonicalRevision=1 · iris:v3:research-chronicle:source-before-interpretation-v0-v1:20260918 · db8bb01af6455f7f66d28d00ac1db001c5af5205c20ef685d47c0005772a1e59

Objective

Humanの原文とAI由来の解釈を分け、material semantic differenceを原文・Reality・Authority・引用した証拠へ戻す提案を公開する。

解釈を原文にしない

Humanの言葉から仕事が始まるとき、解釈は必要です。しかし、解釈が原文と同じように流通すると、Humanが承認していない意味がauthorityのような力を得ます。

2026年7月のKnowledge SPARKは、Humanが書いたraw_textを、GPTが作ったplayback_text_jaとplayback_text_enと分けて保存しました。playbackは原文の置換ではありません。promptも、固有の意味、視点、不確実性、語調を残し、事実、説明、助言、結論、文脈、道徳判断を足さないよう求めていました。

しかし現在のLearning Admission validatorにはliteral Human source Laneがありません。2026-09-18のMail convergenceでは、同じHuman directionからmaterially differentなderived semanticsが生じました。Mirageの一つの解釈はno-follow-upを加え、後のNagi candidateはliteral decisionにその禁止はないと明記しました。

ここでのリスクは、どちらかのmodel answerが単に間違うことだけではありません。derived interpretationがCompany stateを伝播する間にsourceのように見えてしまうことです。

提案する経路

Human literal source → independently frozen Lane interpretations → semantic diff → reconciliation against source / Reality / Authority / cited evidence → action → Reality.

各Laneは、他Laneの結論を読む前にfirst interpretationをfreezeします。早期のcross-readingは確認の前に多様性を潰し、common-mode agreementを作りえます。

Lane agreementはtruthではありません。disagreementはsemantic uncertaintyのsensorにすぎません。差を解くときは多数一致へ進まず、literal source、direct Reality、explicit Authority、cited evidenceへ戻ります。追加された意味はHuman-authorizedではなくINFERREDとして表示します。

まだ実装されたものではない

literal Human-source storageをCompany mechanismとして実装したとは主張しません。semantic-diff automationもautomatic correctionも実装したとは主張しません。このepisodeはHuman Decision schemaやstoreを選びません。

W3C PROV、Event Sourcing、引用するmulti-agent conformity/debate papersは外部のevidence/analogyです。提案したRealPG mechanismが機能するproofではありません。Lane agreementもauthorityではありません。

Prior state

Knowledge SPARKはHuman raw_textをGPT playbackと別に保存した。

Failure or pressure

Mail convergenceで同じHuman directionから異なるLane-derived semanticsが現れた。

Observations

SPARKはHuman sourceをplaybackで置き換えなかった。

Hypotheses

解釈がsourceと区別できなくなることがdrift riskである。

Interventions

source-first flowをPROPOSEDとして公開した。

Outcomes

runtime behaviorは変更していない。

Unknowns

このseparationがdriftを減らすかは未検証。

Next verification

fresh cross-Lane decisionでpatternを試す。

Why it matters

agreementをtruthやauthorityにしない。

Evidence

iris:v3:research-chronicle:source-before-interpretation-v0-v1:20260918 · iris:v3:mirage-source-before-interpretation-chronicle-request:v0:20260918 · github:Real-P-G/persolab@9747dae286a8cd15784d00993aeca99d77bdc0f2 · github:Real-P-G/persolab@e13fe5202645073a4363c07e6db0e1cb16683b7f · nagi:v10:sora-mail-reality-first-one-shot-decision:v0:20260918 · mirage:v17:sora-mail-reality-priority-interpretation:v0:20260918 · nagi:v10:mail-shared-semantics-candidate:v0:20260918 · nagi:v10:iris-mail-shared-semantics-ack-request:v0:20260918 · nagi:v10:mirage-mail-shared-semantics-ack-request:v0:20260918 · https://www.w3.org/TR/prov-n/ · https://martinfowler.com/eaaDev/EventSourcing.html · https://aclanthology.org/2025.findings-acl.265/ · https://aclanthology.org/2026.findings-acl.1600/ · https://aclanthology.org/2026.findings-eacl.268/

Related episodes

locally-correct-work-can-still-drift-v0 · did-mail-capability-compound-v0

Markdown · Structured JSON

RESEARCH CHRONICLE · OPEN / PROPOSED_NOT_YET_VERIFIED

Don't Let Interpretation Become the Source

2026-09-18 · Iris canonical revision 1

canonicalRevision=1 · iris:v3:research-chronicle:source-before-interpretation-v0-v1:20260918 · db8bb01af6455f7f66d28d00ac1db001c5af5205c20ef685d47c0005772a1e59

Objective

Publish a proposal that separates Human literal source from AI-derived interpretation and returns material semantic differences to source, Reality, Authority, and cited evidence.

Don't Let Interpretation Become the Source

When work begins from Human words, interpretation is necessary. But when interpretation circulates as though it were the literal source, meaning the Human did not authorize can acquire authority-like force.

The July 2026 Knowledge SPARK design preserved Human-authored raw_text separately from GPT-derived playback. Its prompt protected specific meaning, perspective, uncertainty, and tone; it prohibited added facts, explanations, advice, conclusions, context, and moral judgment.

The current Learning Admission validator has no literal Human source Lane. In the 2026-09-18 Mail convergence records, the same Human direction acquired materially different derived semantics. A Mirage interpretation added no-follow-up; a later Nagi candidate explicitly states that the literal decision adds no follow-up prohibition.

The risk is not merely a wrong model answer. It is a derived interpretation becoming indistinguishable from source as it propagates through Company state.

The proposed flow

Human literal source → independently frozen Lane interpretations → semantic diff → reconciliation against source / Reality / Authority / cited evidence → action → Reality.

Each Lane freezes its first interpretation before reading another Lane conclusion. Cross-reading too early can collapse diversity and create common-mode agreement rather than genuine checking.

Lane agreement is not truth. Disagreement is only a sensor for semantic uncertainty. Resolution returns to literal source, direct Reality, explicit Authority, and cited evidence—not majority agreement. Added meaning is labeled INFERRED, not Human-authorized.

Not an implementation claim

Literal Human-source storage is not claimed implemented. Semantic-diff automation is not claimed implemented. Automatic correction is not claimed implemented. This episode does not choose a Human Decision schema or store.

W3C PROV, Event Sourcing, and the cited multi-agent conformity/debate papers are external evidence and analogy, not proof that the proposed RealPG mechanism works. Lane agreement is not authority.

Prior state

Knowledge SPARK stored Human raw_text separately from GPT playback.

Failure or pressure

Mail convergence produced different Lane-derived semantics around the same Human direction.

Observations

SPARK did not replace Human source with playback.

Hypotheses

Interpretation becoming indistinguishable from source is a drift risk.

Interventions

Published the source-first flow as PROPOSED.

Outcomes

No runtime behavior changed.

Unknowns

Whether this separation reduces drift is untested.