Production Hub · WikiDesignCo

The classifier sits below its baseline for the first sustained stretch since launch. The drop is concentrated in cold-start prompts: queries that arrive without conversational context. The system has auto-generated an investigation; the next step is yours.

Investigation · auto-generated

Failing categories

· cold-start prompts (no prior context), 38% failure rate
· prompts referencing post-2025 memes, 22% failure rate
· prompts in informal register, 14% failure rate

Hypothesized cause

Training data ends Q1 2026; post-Q1 meme corpus has shifted register significantly. Brand-fit constraint inherited the older register's assumptions. Two prompt revisions and an evaluation-set expansion would likely close the gap.

§ IV Optimization backlog profile-guided, sorted by impact

Expand evaluation set with post-Q1 2026 meme register

memetic-resonance-scorer · would lift cold-start accuracy by an estimated +5 to +8%

+5–8% acc · 2h effort

Add cold-start constraint to brand-fit prompt

memetic-resonance-scorer · explicit handling when no prior context exists

+3% acc · 30m effort

Cache RAPTOR summaries by KB to cut retrieval latency

/retrieve/v2 · would reduce p50 by ~18% on repeat queries

−18% p50 · 4h effort

Promote claim-temporal-arc to platinum tier

12 weeks of stable performance · 3 improvement cycles complete

infra priority · 5m effort

Reduce token output cap on contradiction-explainer

analyzed runs · 142 outputs were ≥80% redundant with the structured fields

−12% cost · 15m effort