# RealPG Research Chronicle — source-before-interpretation-v0

# 解釈を原文にしない

- Status: OPEN / PROPOSED_NOT_YET_VERIFIED
- Iris semantic source: `iris:v3:research-chronicle:source-before-interpretation-v0-v1:20260918`
- Semantic payload hash: `db8bb01af6455f7f66d28d00ac1db001c5af5205c20ef685d47c0005772a1e59`
- [Human](/development/notes/source-before-interpretation-v0/)
- [Structured JSON](/development/notes/source-before-interpretation-v0.json)

## 日本語

### 目的

Humanの原文とAI由来の解釈を混同させず、material external blast radiusを変えうる意味の差を、原文・Reality・Authority・引用した証拠へ戻して照合する提案を公開する。

### 以前はどうだったか

2026年7月のKnowledge SPARKでは、Humanが書いた`raw_text`を、GPTが作った`playback_text_ja` / `playback_text_en`と分けて保存した。GPTへの処理指示も、その人固有の意味、視点、不確実性、語調を保ち、事実、説明、助言、結論、文脈、道徳判断を足さないものだった。

### 何が足りなくなったか

現在の`realpg_learning_admission_payload_valid`はGPT Lane（machine, mirage, nagi, application, experiment, iris）のみを受け入れ、`authority=gpt-interpretation-candidate`を要求する。literal Human source Laneはない。2026-09-18のMail convergenceでは、同じHuman directionからmaterially differentなderived semanticsが生じ、Mirageの解釈にはno-follow-upが加わった一方、後のNagi candidateはliteral decisionにno-follow-up prohibitionはないと明記した。

### 観察したこと

- 7月のSPARKはHumanの`raw_text`をGPT playbackで置き換えず、別々に残した。
- SPARK promptは、意味・視点・不確実性・語調を保存し、追加の事実や助言を禁じた。
- Learning Admission validatorにはliteral Human source Laneがない。
- Mail convergenceには同じHuman directionへの異なるderived semanticsがあり、Mirageのno-follow-up追加と、literal decisionに禁止はないとするNagi candidateが含まれる。
- Nagiは、どれか一つのLaneの解釈を黙って継承せず、Iris、Mirage、Machine、Applicationへshared Mail semanticsを独立に読み、exact deltaを返すよう依頼した。

### 仮説

- 大きなdrift riskは単なる誤答ではなく、derived interpretationがCompany内を伝播するうちにHuman sourceと区別できなくなることにある。
- より安全な経路は、Human literal source → independently frozen Lane interpretations → semantic diff → source / Reality / Authority / cited evidenceとのreconciliation → action → Realityである。追加された意味はHuman-authorizedではなく、必ずINFERREDと表示する。
- 各Laneは他Laneの結論を読む前にfirst interpretationをfreezeする。早すぎるcross-readingは多様性を潰し、独立した確認ではないcommon-mode agreementを作りうる。
- Lane agreementはtruthではない。disagreementはsemantic uncertaintyのsensorにできるが、解決はliteral source、direct Reality、explicit Authority、cited evidenceへ戻る。多数一致では解決しない。
- 有用なgovernanceはすべてのpacketのall-Lane consensusではなく、blast-radius-triggered semantic comparisonかもしれない。

### 変えたこと

このChronicleは次のflowをPROPOSEDとして記録する: Human literal source → independently frozen Lane interpretations → semantic diff → reconciliation against source / Reality / Authority / cited evidence → action → Reality。

### その後どう変わったか

まだruntime behaviorは変わっていない。この記録はsource-before-interpretationをpublic research propositionとして固定するだけである。W3C PROV、Event Sourcing、multi-agent conformity/debateの文献は外部のevidence/analogyであり、提案したRealPG mechanismが機能する証明ではない。

### まだ分からないこと

- このsource/interpretation separationがRealPGのdriftをend-to-endで測定可能に減らすかは未検証である。
- cross-Lane semantic comparisonを要求する最小で信頼できるblast-radius triggerは不明である。
- literal Human decision episodeをLearning Admissionに置くか、別のappend-only decision/provenance storeに置くかは未決定である。

### 次に確かめること

material external blast radiusを持つfresh cross-Lane decisionにsource-first patternを適用し、coordination bureaucracyを増やさずsemantic driftを捕捉できるか観察する。

### なぜ重要か

解釈は行動の根拠になりうる。しかし解釈を原文のように扱うと、Humanが承認していない意味がAuthorityらしく伝播する。first-passを独立にfreezeし、差をsourceとRealityへ戻すことは、agreementをtruthやauthorityにしないための境界である。

### 証拠

- `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/`

### 関連する記録

- /development/notes/locally-correct-work-can-still-drift-v0/
- /development/notes/did-mail-capability-compound-v0/

## English

### Objective

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

### Prior state

The July 2026 Knowledge SPARK implementation stored Human-authored `raw_text` separately from GPT-generated `playback_text_ja` / `playback_text_en`. Its processing prompt told GPT to preserve the person's specific meaning, perspective, uncertainty, and tone, and not add facts, explanations, advice, conclusions, context, or moral judgment.

### Failure or pressure

The current `realpg_learning_admission_payload_valid` validator accepts only GPT Lane sources (machine, mirage, nagi, application, experiment, iris) and requires `authority=gpt-interpretation-candidate`; it has no literal Human source Lane. In the 2026-09-18 Mail convergence records, materially different derived semantics formed around the same Human direction: a Mirage interpretation added no-follow-up, while a later Nagi candidate explicitly says the literal decision adds no follow-up prohibition.

### Observations

- July SPARK preserved Human `raw_text` separately from GPT playback rather than replacing the source text.
- The SPARK prompt protected meaning, perspective, uncertainty, and tone and prohibited added facts or advice.
- The Learning Admission validator has no literal Human source Lane.
- Mail convergence contains materially different derived semantics around the same Human direction, including the Mirage no-follow-up addition and Nagi's explicit literal-decision boundary.
- Nagi asked Iris, Mirage, Machine, and Application to independently read the shared Mail semantics and return exact deltas rather than silently inherit one Lane interpretation.

### Hypotheses

- A major drift risk is a derived interpretation acquiring source-like force as it propagates through Company state.
- A safer path is Human literal source → independently frozen Lane interpretations → semantic diff → reconciliation against source / Reality / Authority / cited evidence → action → Reality. Every added meaning is labeled INFERRED, not Human-authorized.
- Each Lane freezes its first interpretation before reading other Lane conclusions; early cross-reading can collapse diversity into common-mode agreement.
- 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, never majority agreement.
- The cheapest useful governance may be blast-radius-triggered comparison, not all-Lane consensus for every packet or local decision.

### Interventions

This Chronicle records the proposed flow exactly in substance: Human literal source → independently frozen Lane interpretations → semantic diff → reconciliation against source / Reality / Authority / cited evidence → action → Reality.

### Outcomes

No runtime behavior changed. This episode only fixes source-before-interpretation as a public research proposition. 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.

### Unknowns

- Whether this separation measurably reduces RealPG drift end to end is untested.
- The smallest reliable blast-radius trigger for cross-Lane comparison is unknown.
- It is not established whether literal Human decision episodes belong in Learning Admission or a separate append-only decision/provenance store.

### Next verification

Apply the source-first pattern to a fresh cross-Lane decision with material external blast radius, then observe whether it catches semantic drift without creating coordination bureaucracy.

### Why it matters

Interpretation can become a reason for action. If it is treated as source, Human-unapproved meaning can propagate with authority-like force. Freezing independent first passes and returning differences to source and Reality preserves the boundary that agreement is neither truth nor authority.

### 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

- /development/notes/locally-correct-work-can-still-drift-v0/
- /development/notes/did-mail-capability-compound-v0/