# Prompt／Instruction Configuration Management

CURRENT_STATUS: UNKNOWN
OBSERVATION: SPEC_ONLY
FRESHNESS: UNKNOWN
PAGE_TYPE: reference
CANONICAL_URL: /tools/ea-058-prompt-instruction-configuration-management
CLAIM_IDS: none
EVIDENCE_IDS: EVD-EA-058-URL-118388C854B6, EVD-EA-058-URL-07789C42B49A, EVD-EA-058-URL-B5D112A68AF1, EVD-EA-058-URL-843C089305C7, EVD-EA-058-URL-86C7EB613982, EVD-EA-058-URL-D0A28C3433C2, EA-058-R11-RESEARCH, EA-058-R11-WEB-READBACK
UNKNOWN_IDS: EA-058-UNKNOWN-01, EA-058-UNKNOWN-CURRENTNESS
OPS_IDS: none

## Summary
「Promptやsystem instructionは単なる文章ではなく、本番挙動を変える構成資産としてversion管理する。再現・監査に必要なのはprompt本文だけではなく、実行時に解決されたprompt revision、model、generation設定、tool/retrieval/policy等の組合せである。`production`のようなaliasは運用に便利だが、監査ログにはalias名だけでなく実際に参照したimmutable versionを残す。」

## Current judgment
1. Prompt versioning is an existing implementation practice, not merely a proposal. MLflow provides immutable prompt versions, commit messages, tags, environment aliases, response format and model configuration; aliases can move between versions for promotion/rollback.
2. Google Vertex AI also exposes saved prompt resources and versions and documents version listing/restoration. Current Python samples use `vertexai.preview.prompts`; therefore API maturity/availability must be rechecked at publication-update time.
3. Prompt text alone is insufficient for reproducibility. MLflow’s Prompt object itself recognizes model_config, and Vertex examples bind prompt data with model_name/system_instruction and optional generation/safety settings. From these implementations plus EA-055 replay requirements, Core should treat the deployment unit as an execution-configuration bundle rather than only text. This is INFERENCE/PROPOSAL, not a universal standard.
4. Recommended bundle identity: immutable prompt/instruction revision + model ID/revision/alias resolution + generation parameters + response schema + tool definitions/versions + retrieval/index/policy references + safety/guardrail config + environment + evaluation dataset/version + owner/approval/change reason. Runtime traces should record the resolved immutable revisions actually used.
5. Environment aliases (`production`, `staging`, etc.) improve operability but are mutable pointers. Audit/replay records must resolve and persist the target version at execution time; retaining only the alias is insufficient forensic provenance.
6. Rollback should create or select a known configuration revision and preserve history; “edit the current prompt back by hand” destroys provenance. Google’s restore sample creating a restored version and MLflow’s immutable-version/alias model both support this operational pattern.

## Scope
- Enterprise AI Core: versioned configuration / immutable revision / promotion / rollback / provenance
- PromptOps / implementation: prompt registry、environment alias、CI/eval gate
- Provider mapping: Google Vertex AI Prompt Management等
- OSS practice: MLflow Prompt Registry
- Governance / audit: owner、review、approval、change reason、effective deployment version

## Confirmed
- Prompt versioning is an existing implementation practice, not merely a proposal. MLflow provides immutable prompt versions, commit messages, tags, environment aliases, response format and model configuration; aliases can move between versions for promotion/rollback.
- Google Vertex AI also exposes saved prompt resources and versions and documents version listing/restoration. Current Python samples use `vertexai.preview.prompts`; therefore API maturity/availability must be rechecked at publication-update time.
- Prompt text alone is insufficient for reproducibility. MLflow’s Prompt object itself recognizes model_config, and Vertex examples bind prompt data with model_name/system_instruction and optional generation/safety settings. From these implementations plus EA-055 replay requirements, Core should treat the deployment unit as an execution-configuration bundle rather than only text. This is INFERENCE/PROPOSAL, not a universal standard.
- Recommended bundle identity: immutable prompt/instruction revision + model ID/revision/alias resolution + generation parameters + response schema + tool definitions/versions + retrieval/index/policy references + safety/guardrail config + environment + evaluation dataset/version + owner/approval/change reason. Runtime traces should record the resolved immutable revisions actually used.
- Environment aliases (`production`, `staging`, etc.) improve operability but are mutable pointers. Audit/replay records must resolve and persist the target version at execution time; retaining only the alias is insufficient forensic provenance.
- Rollback should create or select a known configuration revision and preserve history; “edit the current prompt back by hand” destroys provenance. Google’s restore sample creating a restored version and MLflow’s immutable-version/alias model both support this operational pattern.

## UNKNOWN
- EA-058-UNKNOWN-01: OPEN: provider横断configuration bundleの最小共通schema。OPEN: first-party prompt registriesのprovider/plan差とretention/export portability。OPEN: promptだけでなくtool schema/retrieval policy/guardrailを同一atomic releaseとして扱う標準的manifestの有無。STOP CONDITION: versioning・promotion・rollbackの実装precedent、prompt-only再現性の限界、Core候補bundle、publication-time recheck点まで確立したためEA-058初回packetはbounded complete。

---
- EA-058-UNKNOWN-CURRENTNESS: 外部根拠の再確認日がcanonical body/control-planeで確定していないため、currentnessはUNKNOWNとして保持する。日付を推測補完しない。

## Evidence
- EVD-EA-058-URL-118388C854B6 (RESEARCH): 1) [author-maintained OSS docs] Prompt Registry, MLflow AI Platform. Canonical: — https://mlflow.org/docs/latest/genai/prompt-registry/
- EVD-EA-058-URL-07789C42B49A (RESEARCH): 2) [author-maintained OSS docs] Manage Prompt Lifecycles, MLflow AI Platform. Canonical: — https://mlflow.org/docs/latest/genai/prompt-registry/manage-prompt-lifecycles-with-aliases/
- EVD-EA-058-URL-B5D112A68AF1 (RESEARCH): 3) [author-maintained OSS docs] Use Prompts in Apps, MLflow AI Platform. Canonical: — https://mlflow.org/docs/latest/genai/prompt-registry/use-prompts-in-apps/
- EVD-EA-058-URL-843C089305C7 (RESEARCH): 4) [official product docs] List available prompt versions, Generative AI on Vertex AI, Google Cloud. Canonical: — https://docs.cloud.google.com/vertex-ai/generative-ai/docs/samples/generativeaionvertexai-prompt-list-prompt-version
- EVD-EA-058-URL-86C7EB613982 (RESEARCH): 5) [official product docs] Restore a version of a prompt, Generative AI on Vertex AI, Google Cloud. Canonical: — https://docs.cloud.google.com/vertex-ai/generative-ai/docs/samples/generativeaionvertexai-prompt-restore-version
- EVD-EA-058-URL-D0A28C3433C2 (RESEARCH): 6) [official product docs] Create and use a Vertex AI Prompt to generate text, Google Cloud. Canonical: — https://docs.cloud.google.com/vertex-ai/generative-ai/docs/samples/generativeaionvertexai-prompt-template-create-generate-save
- EA-058-R11-RESEARCH (RESEARCH): EA-058 監査済みresearch / canonical revision binding
- EA-058-R11-WEB-READBACK (IMPL): EA-058 Web body identity / canonical readback

## Failure / limits
Versioning a prompt cannot guarantee identical model output; model/provider revisions, nondeterminism, tools, retrieval state and external data may change.
## Failure / limits
A prompt registry does not by itself enforce review, evaluation quality, separation of duties or safe promotion. Those are organization workflow controls.
## Failure / limits
Mutable aliases simplify deployment but introduce time-of-use ambiguity if execution telemetry records only the alias.
## Failure / limits
Provider-native prompt resources differ in schema and lifecycle; do not claim portability of prompt IDs/versions across providers.
## Failure / limits
Google sample APIs shown under `vertexai.preview.prompts` are PRODUCT_VOLATILE and must be revalidated before future publication.
## Failure / limits
Registry deletion/retention settings can conflict with forensic needs; production retention policy must be designed explicitly.
## Failure / limits
OPEN: provider横断configuration bundleの最小共通schema。OPEN: first-party prompt registriesのprovider/plan差とretention/export portability。OPEN: promptだけでなくtool schema/retrieval policy/guardrailを同一atomic releaseとして扱う標準的manifestの有無。STOP CONDITION: versioning・promotion・rollbackの実装precedent、prompt-only再現性の限界、Core候補bundle、publication-time recheck点まで確立したためEA-058初回packetはbounded complete。

---
## Practice
MLflow Prompt Registry behavior depends on deployed MLflow version/backend and organization auth configuration; docs establish OSS capability, not any managed-service parity.
## Practice
Vertex AI Prompt APIs require Google Cloud project/auth/location and the current SDK/API surface; preview namespace usage means publication-time status must be rechecked.
## Practice
No claim is made that Anthropic/OpenAI or every provider offers an equivalent first-party registry in every plan/region; provider mapping remains separate.
## Practice
IMPLEMENTABLE NOW. Core pattern: immutable artifact revision -> evaluation -> approval -> environment alias/pointer -> deployment -> runtime logs capture resolved revision -> rollback by alias/promotion to an approved immutable version. Store prompt/instruction assets in EA-056 registry and link runtime revision IDs into EA-051 telemetry and EA-055 forensic records.
## Practice
Configuration management quality should be tested by change gates rather than a generic model benchmark: regression eval set, policy/safety checks, tool-call contract tests, output-schema tests, cost/latency comparison and canary/rollback exercise. Evaluation dataset/scorer versions are themselves provenance. Passing evaluation does not guarantee production safety.
## Practice
Security: prompts can contain policy logic, hidden instructions, tool constraints and sometimes secrets; secrets should be referenced through secret management rather than embedded in prompt artifacts.
## Practice
Privacy: test examples and prompt variables can contain sensitive data; registry access and retention need data classification.
## Practice
Audit: record author, reviewer, immutable version/hash or provider version ID, change reason, evaluation artifact, approval, promotion time, alias movement and resolved runtime version.
## Practice
Cost: regression evaluation and dual-running/canarying incur inference cost; risk-tier changes rather than every wording edit can determine gate depth.
## Practice
Operability: registry/alias caching can delay pointer changes. MLflow documents different cache behavior for immutable versions vs mutable aliases, so rollback RTO must account for cache/propagation semantics.
## Practice
EA-056 AI Asset Registryへ prompt/instruction asset type、immutable version、owner、environment alias、evaluation/approval linkageを追加候補。
## Practice
EA-051 telemetryへ `prompt.asset_id`, `prompt.version`, `configuration_bundle_id` 相当をprovider-neutral fieldとして検討。
## Practice
EA-054 rollbackではprompt alias切替だけでなくmodel/tool/retrieval/policyとの整合を確認するcompound rollback checklistへ接続。
## Practice
EA-055 replayではmutable aliasを再解決せず、当時記録したimmutable revisionを参照する。
## Practice
MLflow prompt templates/versions are immutable but model_config on a prompt version is mutable after creation. Vertex AI prompt versioning/restoration remains PRODUCT_VOLATILE. A prompt-version ID alone is not an immutable full execution configuration.
## WHY_THIS_MATTERS_TO_UPDATE
Prompt/system instructionはAI挙動を変える実行構成であり、コードを変えなくても安全性、tool選択、出力形式、business rule遵守が変わる。prompt本文だけをGit断片や管理画面に散在させると、incident/replay時に“何が本番で使われたか”を復元できない。EA-054 rollback、EA-055 replay/forensics、EA-056 asset inventoryと接続する構成管理が必要。

## SCOPE_ROUTING
- Enterprise AI Core: versioned configuration / immutable revision / promotion / rollback / provenance
- PromptOps / implementation: prompt registry、environment alias、CI/eval gate
- Provider mapping: Google Vertex AI Prompt Management等
- OSS practice: MLflow Prompt Registry
- Governance / audit: owner、review、approval、change reason、effective deployment version

## CLAIM_OR_RESEARCH_QUESTION
Promptをversioned artifactとして管理するだけで再現性は得られるか。model、generation settings、tools、retrieval/policy等をどこまで同じconfiguration bundleとして固定する必要があるか。

## EVIDENCE TABLE
1) [author-maintained OSS docs] Prompt Registry, MLflow AI Platform. Canonical: https://mlflow.org/docs/latest/genai/prompt-registry/ . Accessed 2026-09-14. Locator: Prompt object attributes; sequential Version, Commit Message, Tags, mutable Alias, response_format, model_config; existing prompt versions are immutable. Label: OSS_PRACTICE. Relation: support.
2) [author-maintained OSS docs] Manage Prompt Lifecycles, MLflow AI Platform. Canonical: https://mlflow.org/docs/latest/genai/prompt-registry/manage-prompt-lifecycles-with-aliases/ . Accessed 2026-09-14. Locator: staging/production aliases, alias switching, rollback/A-B use. Label: OSS_PRACTICE. Relation: support.
3) [author-maintained OSS docs] Use Prompts in Apps, MLflow AI Platform. Canonical: https://mlflow.org/docs/latest/genai/prompt-registry/use-prompts-in-apps/ . Accessed 2026-09-14. Locator: `prompts:/<prompt_name>/<version_or_alias>` and model lineage linkage. Label: OSS_PRACTICE. Relation: support/context.
4) [official product docs] List available prompt versions, Generative AI on Vertex AI, Google Cloud. Canonical: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/samples/generativeaionvertexai-prompt-list-prompt-version . Accessed 2026-09-14. Locator: `prompts.create_version`, `prompts.list_versions`, `prompts.get`. Example uses `vertexai.preview.prompts`. Label: VERIFIED_SPEC / PRODUCT_VOLATILE. Relation: support/availability.
5) [official product docs] Restore a version of a prompt, Generative AI on Vertex AI, Google Cloud. Canonical: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/samples/generativeaionvertexai-prompt-restore-version . Accessed 2026-09-14. Locator: `prompts.restore_version`; restored older content becomes a new version. Label: VERIFIED_SPEC / PRODUCT_VOLATILE. Relation: support.
6) [official product docs] Create and use a Vertex AI Prompt to generate text, Google Cloud. Canonical: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/samples/generativeaionvertexai-prompt-template-create-generate-save . Accessed 2026-09-14. Locator: Prompt includes prompt_data, variables, model_name, system_instruction, optional generation/safety configuration and saves a version. Label: VERIFIED_SPEC / PRODUCT_VOLATILE. Relation: support/context.

## FINDINGS
1. Prompt versioning is an existing implementation practice, not merely a proposal. MLflow provides immutable prompt versions, commit messages, tags, environment aliases, response format and model configuration; aliases can move between versions for promotion/rollback.
2. Google Vertex AI also exposes saved prompt resources and versions and documents version listing/restoration. Current Python samples use `vertexai.preview.prompts`; therefore API maturity/availability must be rechecked at publication-update time.
3. Prompt text alone is insufficient for reproducibility. MLflow’s Prompt object itself recognizes model_config, and Vertex examples bind prompt data with model_name/system_instruction and optional generation/safety settings. From these implementations plus EA-055 replay requirements, Core should treat the deployment unit as an execution-configuration bundle rather than only text. This is INFERENCE/PROPOSAL, not a universal standard.
4. Recommended bundle identity: immutable prompt/instruction revision + model ID/revision/alias resolution + generation parameters + response schema + tool definitions/versions + retrieval/index/policy references + safety/guardrail config + environment + evaluation dataset/version + owner/approval/change reason. Runtime traces should record the resolved immutable revisions actually used.
5. Environment aliases (`production`, `staging`, etc.) improve operability but are mutable pointers. Audit/replay records must resolve and persist the target version at execution time; retaining only the alias is insufficient forensic provenance.
6. Rollback should create or select a known configuration revision and preserve history; “edit the current prompt back by hand” destroys provenance. Google’s restore sample creating a restored version and MLflow’s immutable-version/alias model both support this operational pattern.

## LIMITS / COUNTEREVIDENCE / FAILURE ENVELOPE
- Versioning a prompt cannot guarantee identical model output; model/provider revisions, nondeterminism, tools, retrieval state and external data may change.
- A prompt registry does not by itself enforce review, evaluation quality, separation of duties or safe promotion. Those are organization workflow controls.
- Mutable aliases simplify deployment but introduce time-of-use ambiguity if execution telemetry records only the alias.
- Provider-native prompt resources differ in schema and lifecycle; do not claim portability of prompt IDs/versions across providers.
- Google sample APIs shown under `vertexai.preview.prompts` are PRODUCT_VOLATILE and must be revalidated before future publication.
- Registry deletion/retention settings can conflict with forensic needs; production retention policy must be designed explicitly.

## AVAILABILITY_GATES
- MLflow Prompt Registry behavior depends on deployed MLflow version/backend and organization auth configuration; docs establish OSS capability, not any managed-service parity.
- Vertex AI Prompt APIs require Google Cloud project/auth/location and the current SDK/API surface; preview namespace usage means publication-time status must be rechecked.
- No claim is made that Anthropic/OpenAI or every provider offers an equivalent first-party registry in every plan/region; provider mapping remains separate.

## IMPLEMENTATION_STATE
IMPLEMENTABLE NOW. Core pattern: immutable artifact revision -> evaluation -> approval -> environment alias/pointer -> deployment -> runtime logs capture resolved revision -> rollback by alias/promotion to an approved immutable version. Store prompt/instruction assets in EA-056 registry and link runtime revision IDs into EA-051 telemetry and EA-055 forensic records.

## BENCHMARK_OR_EVAL
Configuration management quality should be tested by change gates rather than a generic model benchmark: regression eval set, policy/safety checks, tool-call contract tests, output-schema tests, cost/latency comparison and canary/rollback exercise. Evaluation dataset/scorer versions are themselves provenance. Passing evaluation does not guarantee production safety.

## SECURITY / PRIVACY / AUDIT / COST / OPERABILITY
- Security: prompts can contain policy logic, hidden instructions, tool constraints and sometimes secrets; secrets should be referenced through secret management rather than embedded in prompt artifacts.
- Privacy: test examples and prompt variables can contain sensitive data; registry access and retention need data classification.
- Audit: record author, reviewer, immutable version/hash or provider version ID, change reason, evaluation artifact, approval, promotion time, alias movement and resolved runtime version.
- Cost: regression evaluation and dual-running/canarying incur inference cost; risk-tier changes rather than every wording edit can determine gate depth.
- Operability: registry/alias caching can delay pointer changes. MLflow documents different cache behavior for immutable versions vs mutable aliases, so rollback RTO must account for cache/propagation semantics.

## SAFE_WORDING_FOR_FUTURE_BOOK
「Promptやsystem instructionは単なる文章ではなく、本番挙動を変える構成資産としてversion管理する。再現・監査に必要なのはprompt本文だけではなく、実行時に解決されたprompt revision、model、generation設定、tool/retrieval/policy等の組合せである。`production`のようなaliasは運用に便利だが、監査ログにはalias名だけでなく実際に参照したimmutable versionを残す。」

## UPDATE_WORK_HANDOFF
- EA-056 AI Asset Registryへ prompt/instruction asset type、immutable version、owner、environment alias、evaluation/approval linkageを追加候補。
- EA-051 telemetryへ `prompt.asset_id`, `prompt.version`, `configuration_bundle_id` 相当をprovider-neutral fieldとして検討。
- EA-054 rollbackではprompt alias切替だけでなくmodel/tool/retrieval/policyとの整合を確認するcompound rollback checklistへ接続。
- EA-055 replayではmutable aliasを再解決せず、当時記録したimmutable revisionを参照する。

## OPEN_UNKNOWN / STOP_CONDITION
OPEN: provider横断configuration bundleの最小共通schema。OPEN: first-party prompt registriesのprovider/plan差とretention/export portability。OPEN: promptだけでなくtool schema/retrieval policy/guardrailを同一atomic releaseとして扱う標準的manifestの有無。STOP CONDITION: versioning・promotion・rollbackの実装precedent、prompt-only再現性の限界、Core候補bundle、publication-time recheck点まで確立したためEA-058初回packetはbounded complete。

---

## CURRENTNESS_AUDIT
MLflow prompt templates/versions are immutable but model_config on a prompt version is mutable after creation. Vertex AI prompt versioning/restoration remains PRODUCT_VOLATILE. A prompt-version ID alone is not an immutable full execution configuration.

NEXT: /agent/start.json
STOP: This is a knowledge projection, not an execution order.