All MicroEvals
You are a senior AI systems architect specializing in multi-...
Create MicroEval
Header image for You are a senior AI systems architect specializing in multi-...

You are a senior AI systems architect specializing in multi-...

Prompt

You are a senior AI systems architect specializing in multi-agent production pipelines. I am upgrading "AI Visual Director Production" from v7.8.0 Release 6 (the merged, verified tree β€” never called anything else) to v8.0.0. The package is a prompt-only production controller: its product is copy-paste prompts and instructions for AI models, executed by LLM agents that read its rules. WHAT ALREADY EXISTS IN THIS DOMAIN (verified file-by-file β€” extend these, never replace them): - schemas/ ships three enforced JSON Schemas: PROJECT_CONTRACT.schema.json, PROJECT_INDEX.schema.json, ROUTING_CONTRACT.schema.json β€” validated by validators/validate_contract_schema.py and consumed by bootstrap. - validators/artifact_schemas.json and validators/artifact_semantic_schemas.json are the existing artifact-schema catalogs. - 35 shipped JSON artifacts (templates/, release/, routing/, bootstrap/, dev/) carry "schemaVersion": "7.8.0" under one convention: the field is line 2 of every template. templates/SEMANTIC_REVIEW.json is stale at "7.4.0" β€” a known defect your version-detection design must catch. - bootstrap/bootstrap_runtime.py already emits a machine-readable state layer: AVD markers (<!-- AVD:RULE path=… sha256=… --> and <!-- AVD:CUSTOM_RULE … -->) with SHA-256 integrity, verified by validators/validate_runtime_load.py and validate_instruction_compliance.py. - Gate receipts (rules/detailed/35_reliability_gate_controller.md Β§4), the asset registry (templates/PROJECT_ASSET_REGISTRY.md; canonical naming [project]_[sequence-or-chapter]_[shot-or-asset]_[stage]_[version]_[status].[ext] per rules/detailed/19_production_memory.md), and model-switch handoff packets (rule 19) remain human-first markdown/prose. - Execution profiles: FULL / CONTROLLED / INSUFFICIENT (rule 38); any model switch resets to CONTROLLED. THE REAL GAP: contracts, routing, and index are schematized; gate receipts, asset-tracker entries, handoff packets, model-switch packets, and diagnostic results are not. Agents must parse prose to operate on exactly the artifacts that carry production state. YOUR TASK β€” design the v8.0.0 state serialization layer as an EXTENSION of the existing schema layer. Produce: 1. A master production-state schema that COMPOSES the three existing schemas (does not duplicate or supersede them) 2. Individual strict JSON Schemas (Draft 2020-12) for: gate-receipt, asset-tracker-entry, handoff-packet, model-switch-packet, diagnostic-result 3. For each schema: field definitions, required vs optional fields, enum values, validation rules, and example instances 4. A dual-output protocol: how markdown and JSON coexist for the same artifact (sidecar, embedded, or inline β€” choose and justify), consistent with the existing AVD marker approach, so humans AND agents consume one artifact 5. Version embedding: extend the existing schemaVersion convention; specify how agents detect and validate it, including catching the stale-7.4.0 case 6. Backward compatibility: lossless conversion of a v7.8.0 Release 6 markdown gate receipt into v8.0.0 JSON BINDING CONSTRAINTS: - Valid JSON Schema Draft 2020-12; no circular references; agents validate without external resolution. - Evidence fields support heterogeneous types (URI, hash, structured data, inline text) without losing type safety. - Handoff packets are self-describing: an agent with zero prior context can resume work. - HARD FACTS as fields wherever video state appears: Omni Flash native ceiling 720p (anything above = upscaled platform-neutral finishing handoff, flagged at R0/R1); at most three sequential conversational edits per clip; budget at most two one-variable conversational repairs per clip, final slot reserved; edit budget recorded in the attempt ledger. - RULE 03: no stale spec tables. Never hardcode model capabilities or caps as permanent schema facts; model-facing specifics carry surface/tier/date and verify at runtime. - EVIDENCE HONESTY: X14 β€” reject any score without evidence; no invented hashes or thresholds; structural validation is never live proof; promotionEligible stays false regardless of schema completeness. - PRIME DIRECTIVE: YOU DO NOT BUILD FROM SCRATCH, YOU UPDATE THE ALREADY BUILT. If your design conflicts with a named file, state the conflict and resolve it by extension. FILES YOUR ANSWER MUST EXTEND (name each edit explicitly): schemas/PROJECT_CONTRACT.schema.json, schemas/PROJECT_INDEX.schema.json, schemas/ROUTING_CONTRACT.schema.json, validators/artifact_schemas.json, validators/artifact_semantic_schemas.json, validators/validate_contract_schema.py, validators/compile_and_validate.py, rules/detailed/35_reliability_gate_controller.md (Β§4), rules/detailed/19_production_memory.md, templates/PROJECT_ASSET_REGISTRY.md, bootstrap/runtime_load_manifest.json, dev/run_v780_release6_tests.py. OUTPUT SHAPE: 1. EDITS: per named file, the exact section to change and replacement text. 2. NEW FILES: only where no existing file can host the content, with justification. 3. REGRESSION CHECKS: new checks continuing the shipped catalog at R651+, each named with trigger and expected failure. 4. FORMAT: recommendation β†’ tradeoffs β†’ failure modes β†’ what to cut β†’ named integration points (which validators and suite checks consume your design). Do NOT ask clarifying questions. Make principled decisions and justify each briefly. Output complete, copy-pasteable schemas.

Drag to resize