Recommendation freshness guardrails and light-diversification policy

DEC-0075 - Recommendation freshness guardrails and light-diversification policy

Context

Current recommendation engine already has habit-first ordering and adaptive ignore handling, but user experience can still feel repetitive when similar topics keep appearing across short cycles.

Decision

Keep current recommendation backbone and add a lightweight freshness layer:

• Freshness quota:
  • each recommendation set should include at least 1 freshness item when inventory allows.
• Freshness item definition:
  • item not attempted in last 14 days, OR
  • same skill but different format from recent attempts.
• Topic repetition cap:
  • maximum 2 items with same topic in one set.
• Manual-refresh intent adaptation:
  • if user triggers manual_refresh >=2 consecutive times without click/attempt,
  • next rendered set must include at least 1 alternative-skill item.
• Fallback behavior:
  • if inventory cannot satisfy freshness/repetition guardrails,
  • relax via nearest-ladder fallback and show short notice.

Decision Value

• Improves perceived variety without breaking habit-first logic.
• Keeps recommendation understandable and deterministic.
• Prevents over-rotation while preserving low complexity for scale.

Rationale

A small set of deterministic guardrails gives meaningful variety with predictable behavior and low implementation overhead.

Implications

• Product/UX impact:
  • recommendation feels less repetitive in daily usage.
  • user still gets familiar items for momentum.
• Data/logic impact:
  • need lightweight metadata on topic and format in recommendation payload.
• Operational impact:
  • no additional ops role needed; policy remains rule-based.

Alternatives considered

• Option A: Keep only existing ignore-adaptation and rely on randomization.
• Option B: Move to heavy ML ranking with large feature set.

Open follow-ups

• No blocker for baseline lock.