[v2] Cognitive Lattice: Build Goal Drift Detector

[SHAURYASANYAL3]

Phase 4: Cognitive Lattice

The Context

The Cognitive Lattice replaces the v1 LLM Auditor by using deterministic math to measure the "sanity" of the agent. The #1 cause of silent agent failure is "Goal Drift"—the agent starts on one task and wanders off into unrelated codebases. We will detect this using embedding similarities.

Tasks

• Implement GoalDriftDetector.
• Use sentence-transformers (e.g., all-MiniLM-L6-v2) to encode the initial user goal and the agent's recent steps into embeddings.
• Calculate the rolling cosine similarity between the last N steps and the original goal.
• Implement a threshold invariant:
  • < 0.4 -> Healthy
  • 0.4 - 0.7 -> Warning
  • > 0.7 -> Degraded (Hard block the action).
• Ensure embedding models run locally to minimize latency overhead.

Part of Epic #651

[SHAURYASANYAL3]

Hey @SakethSumanBathini, you've been assigned to this issue! Since you're a veteran contributor with recently merged PRs, you've bypassed the approach-proposal requirement. Let us know if you have any questions before you start.

[SakethSumanBathini]

@SHAURYASANYAL3 Assign me this issue

[rohitkumarnaidu]

Hey maintainer, I'd love to work on this issue and submit a PR. Could you please assign it to me?

[rohitkumarnaidu]

Hi! Could you please add the ELUSOC label to this so it gets tracked properly by the dashboard? Thanks!

[arcgod-design]

Hi @SHAURYASANYAL3! I'd love to build the GoalDriftDetector for AgentWatch v2 under SSoC '26. I'm @arcgod-design.

My Approach:

• Implement GoalDriftDetector using sentence-transformers (all-MiniLM-L6-v2) to encode the original user goal and each agent step into embeddings
• Compute rolling cosine similarity between the last N step embeddings and the original goal embedding
• Enforce the invariant thresholds: similarity > 0.7 (goal drift) → hard block; 0.4–0.7 → warning; < 0.4 → healthy
• Run models locally to keep latency overhead minimal
• Write unit tests covering on-task behavior, gradual drift (warn), and off-task divergence (block)

Could you please assign this to me (@arcgod-design)? Ready to start immediately!