Prompt drift resistente: fluxo multi-lane de IA para evitar deriva em projetos longos

Goal: design and test a drift-resistant, multi-lane AI workflow to prevent memory-driven drift in long-running projects. The prompt asks the AI to define a lane-based architecture, specify memory management, guardrails, drift detection metrics, and provide concrete prompt templates for each lane, plus a demonstration scenario.

Instructions:
1) Define lanes and their roles:
- Rules & Constraints Lane: captures immutable rules, versioning scheme, and a mechanism to lock rules so they cannot be changed without a formal review. Include how to detect changes (drift) in rules and rollback procedures.
- Memory / Context Lane: governs what the model remembers across interactions, memory budget, tokenization limits, expiry rules, and a changelog of memory updates. Specify how to reference persistent project specs without polluting short-term context.
- Drift Detection Lane: monitors for deviations from defined rules, goals, or project state. Define drift signals, scoring (0-1), thresholds for alerting humans, and automatic correction options.
- Task Orchestration Lane: coordinates sub-prompts, tasks, and responses. Ensure deterministic sequencing, isolation between tasks, and clean handoffs between lanes.
- Validation & Output Lane: validates outputs against criteria, formats results for humans and machines, and logs decisions and drift evidence.

2) Memory management details:
- Memory tokens and references: how to tag, store, and retrieve persistent project information.
- Expiry and purging: rules for when and how to remove or refresh memory to minimize drift.
- Consistency checks: baseline checks to ensure memory aligns with the current Rules & Constraints.

3) Drift rules and evolution:
- Define what constitutes drift in each lane (e.g., rule changes, context leakage, misaligned goals).
- Provide a drift score computation method (0-1 scale) and escalation steps when drift exceeds thresholds.
- Specify rollback or versioning actions to revert to the last stable configuration.

4) Execution flow and prompts:
- Provide a step-by-step execution plan that ties the lanes together for a typical long-running project.
- Include example prompts for each lane to illustrate how a real session would proceed and how to maintain separation between lanes.
- Show how to extract a human-readable summary and a machine-readable JSON record after each run.

5) Output requirements:
- Human-readable recap of decisions, drift detected, and recommended actions.
- Machine-readable JSON including: project_id, rules_version, memory_state, drift_score, decisions, evidence, and lane_state.

6) Example demonstration:
- Use a hypothetical project named Project Aurora. Describe a drift scenario (e.g., a rule drift in the long-term spec and a memory mismatch) and illustrate how each lane would respond to contain and correct the drift.

7) Safety and boundaries:
- Include guidance on protecting sensitive data, avoiding leakage across lanes, and respecting privacy constraints.

8) Optional: architecture diagram ideas (Mermaid or ASCII) to visualize lanes and interactions.