Liti-agent-Development/Case_02_Comparison_Research/plans/s2-10-hybrid-implementation-v1.md

S2_10 Hybrid Prompt Implementation v1

Objective

Implement the S2_10 legal-resolution-map package defined by stage_2_optimal_update_strategy_v.5-1.md, S2_10_SOW_v.1.md, and S2_10_assets_v.1.md, while retaining the compatible legal/runtime boundaries from S2_10_SOW.md. Produce the versioned authoring YAML Stage_2_S2_10_v.1.yml, deploy its version-free projection and supporting assets under Default_Agent/Stage_2_Clean/, and leave a reproducible offline validation record.

Deliverables

• YAML_Prompts/2. Stage_2/Stage_2_S2_10_v.1.yml
• The fixed S2_10 physical asset set listed by S2_10_assets_v.1.md, deployed at its canonical paths
• Updated shared release/module manifests required to close the new hashes
• Byte-identical .py/.txt pairs for every Python asset
• Updated YAML_Prompts/2. Stage_2/MEMORY.md

Scope and Non-Scope

In scope: hybrid inline-static prompt ownership, two-field dynamic-item boundary, schema/binding/workflow/projection/receipt/release updates, fixtures, offline tests, and release resealing. Out of scope: live GPT invocation, native adapter execution, production admission, legal sign-off, and dependencies on Stage 2 v0-v3 YAML specifications.

Known Inputs

• stage_2_optimal_update_strategy_v.5-1.md
• S2_10_SOW.md and controlling hybrid delta S2_10_SOW_v.1.md
• S2_10_assets_v.1.md
• SKILL.md
• test_code_executor.ipynb
• Current Default_Agent/Stage_2_Clean/ S2_00/S2_10 assets

Material Assumptions

• The explicit requested output basename Stage_2_S2_10_v.1.yml is the authoring source; the deployment projection remains version-free.
• S2_10_SOW_v.1.md controls where it narrows or replaces the older S2_10_SOW.md external-prompt design.
• S2_10 contains exactly one direct LLM map task and no runtime Python or Code Executor task. Python files are offline build/validation/test assets only and must have byte-identical .txt mirrors.
• The 36-file fixed S2_10 inventory counts physical files; shared release manifests that require resealing are tracked separately.

Questions That Could Change the Outcome

• Whether the host platform's native LLM adapter supports the declared four-message order and structured-output contract remains a live-admission question, not an offline implementation blocker.
• Model benchmark and Korean-law review receipts remain pending until external evidence exists.

Workstreams and Dependencies

1. Freeze the 36-file inventory and retained/replaced/missing ledger.
2. In parallel, draft the hybrid authoring YAML and update independent assets.
3. Build the version-free projection, then bind hashes into receipts and the child release.
4. Reseal shared parent/module manifests without changing S2_00 business logic.
5. Run one asset-level review/fix cycle and exactly two YAML review/fix cycles.
6. Run offline validation, parity, negative-mutation, and forbidden-dependency checks.
7. Record results and residual live-admission gaps in project memory.

Source and Tool Plan

• Use local repository documents and assets only; no web research is required.
• Use apply_patch for authored changes and mechanical copy only for .py to .txt byte mirrors.
• Use repository Python validators/tests in offline mode. Do not invoke live models or adapters.
• Use independent subagents for inventory/design review and YAML validation; batch asset-level review assignments within the four-agent concurrency limit.

Validation Plan

• Parse all YAML/JSON and validate schemas and cross-file hashes.
• Assert one map task, zero deterministic/tool/reduce tasks, exact model/reasoning/verbosity, no tools, and preflight: false.
• Assert inline common/static prompts and exactly two permitted dynamic item placeholders.
• Reject legacy preassembled prompt fields and raw Stage 1 rereads.
• Verify local-reference output schema and adapter-owned permanent-ID rewrite boundary.
• Verify .py/.txt byte parity.
• Run positive fixtures and required negative mutations.
• Distinguish offline contract completion from live adapter/model/legal admission.

Approval Boundaries

No external writes, live model calls, production admission, or commits are authorized. Existing unrelated worktree changes are preserved.

Progress

• Read repository rules, local YAML skill, notebook guidance, strategy/SOW/assets, and project memory.
• Freeze the exact inventory and delta ledger (K=36).
• Draft and deploy all S2_10 assets.
• Complete one asset validation and fix cycle.
• Complete YAML validation and fix cycle 1.
• Complete YAML validation and fix cycle 2.
• Reseal shared manifests and run full offline validation.
• Update project memory and record residual uncertainty.

Decision Log

Date/Stage	Decision	Basis	Consequence
2026-08-30 / intake	Treat S2_10_SOW_v.1.md as the controlling hybrid delta over S2_10_SOW.md.	It was produced specifically to implement stage_2_optimal_update_strategy_v.5-1.md; the old SOW still assumes externally assembled full prompts.	Retain compatible legal/output/adapter contracts, replace prompt ownership and item boundaries.
2026-08-30 / intake	Use Stage_2_S2_10_v.1.yml as source and version-free deployment projection.	Explicit user filename plus existing deployment convention.	Builder/receipts must bind both source and projection identities without overwriting the old versioned source.
2026-08-30 / intake	No runtime Python in S2_10.	Strategy and SOW assign the single legal judgment to one native LLM map task.	Notebook Code Executor rules apply only as safety constraints to offline Python assets; no inline run_code belongs in the S2_10 YAML.
2026-08-30 / implementation	Exclude cluster_case_payload_sha256 from the payload's inner echo and bind it only in the top-level item/output echo.	A payload containing its own byte hash would require an impossible fixed point.	The inner payload echo has 19 fields; the top-level dispatch/model echo has all 20 fields.
2026-08-30 / asset review	Replace the regression manifest's generic oracle marker with fixture-specific exact oracle IDs and recompute manifest/module/release closure.	The one requested asset-review round found three affected physical assets plus stale module/test counts.	Tests now compare exact per-fixture oracle lists; module counts are builder-derived (76, 12).

Evidence Ledger

Claim/Issue	Source or Test	Status	Notes
Hybrid prompt ownership	strategy v5-1 and SOW v1	accepted	YAML owns stable common/task prompt; item owns case data only.
Physical asset count	assets v1 inventory and filesystem audit	verified	K=36; all 36 paths exist at the specified locations.
S2_10 offline tests	tests/s2_10	PASS	39/39, including exact fixture oracle seal and .py/.txt parity.
S2_00 compatibility	tests/s2_00, both builders in check mode	PASS	69/69; parent hash and opaque S2_10 Agent/binding handoff resealed.
Live adapter compatibility	external adapter evidence	pending	Must not be inferred from offline tests.
Model quality and legal correctness	benchmark/legal review receipts	pending	Receipts remain explicitly non-admitted.

Risks and Failure Modes

• Hash cycles between parent and child releases; prevent by keeping the parent one-way and not reverse-locking the child from the parent.
• YAML parser normalization changing prompt bytes; use explicit block scalars and hash the parsed canonical strings where specified.
• Dynamic item reintroducing full prompt text or PII; enforce schema and negative fixtures.
• Static tests being mistaken for live execution or legal approval; retain explicit status labels and pending receipts.
• Shared dirty worktree causing accidental scope expansion; edit only identified Stage 2 assets.

Results and Residual Uncertainty

The 36-file hybrid S2_10 package is implemented and offline verified. The authoring and deployment Agent are byte-identical and contain exactly one gpt-5.6-sol/xhigh/medium/responses map task, four ordered prompt blocks, and two dynamic JSON placeholders. Parent release raw hash is 8dcfe060af33e2d86a93f1c277c485b67255500a8cfb7f3024ffc427e802b948; child release raw hash is 5adb88482fc6e54ae7475570facf9a6ae87fa613fce65d867d16d5256604f14d. Both builders pass read-only check mode, S2_10 tests pass 39/39, S2_00 compatibility tests pass 69/69, and all seven Python assets have byte-identical .txt mirrors.

This is not live or legal admission. The platform adapter, model benchmark, official authority/profile release, and Korean-lawyer review receipts remain explicitly pending; the parent remains DRAFT_NOT_EXECUTABLE and production execution remains blocked. No live LLM/MCP adapter or end-to-end matter run was performed.