V2 P0: Memory Reliability Gate — evaluate API + MCP cortex_evaluate

[cursor]

Background

Signature V2 capability: decide whether memory is trustworthy enough for an agent action.

Problem Statement

Retrieval returns memory without ACT/VERIFY/ASK/ESCALATE/BLOCK policy. One confidence score is overloaded for action risk.

Proposed Solution

• Package reliability/ (gate, risk, policies, reason_codes, confidence)
• Action risk enum: READ_ONLY … PRIVILEGED
• Separate memory_confidence vs action_confidence (risk-dependent thresholds)
• Typed decisions: ACT | VERIFY | ASK | ESCALATE | BLOCK
• POST /reliability/evaluate + MCP cortex_evaluate
• Authorization outside the LLM; no ACT without permission

Acceptance Criteria

• Same memory → different decisions by action risk
• High memory confidence ≠ automatic high action confidence
• Unauthorized caller cannot get ACT for restricted actions
• Auditable reason_codes
• Signature scenario: restart payments → BLOCK with evidence

Depends on

Evidence Graph + Temporal Truth (gate inputs)

Plan ref

todo reliability-gate — V2 Enhancement 1