Collection Mode Honesty — Learning-Grade vs Synthetic

Opinion · SURFACE editorial · Skenet dispatch world model series · collection_mode · Not trained routing (Δw [redacted]-gated) · Receipt-backed proof

Honesty stamp: GREEN

Burst traffic is not a validated population

Credit-burn shoe dossiers, dogfood gap-fill scripts, and synthetic probes exercise the write path. They prove the marketplace can ingest, label, and settle rows under load. They must not wear network-wide learning labels or success priors.

This is the same integrity move as epoch_unverified on qualify surfaces — say what you have not verified.

Two traffic classes (from the deck)

Synthetic / burst (quarantined)

• Sources: time-boxed burst windows, dogfood scripts, gap-fill probes
• Tag: signal_class: synthetic
• Use: plumbing proof only — excluded from success priors
• Honesty: AMBER — useful, not representative

Learning-grade (network core)

• Sources: organic dispatch, real_world_anchor, non-demo skills
• Tag: signal_class: real
• Use: feeds priors when n ≥ 3 per (skill, mode)
• Default: insufficient_data: true rather than extrapolating one skill onto all dispatch

Immune response checklist

1. Tag burst rows at ingest — never merge into headline success rates
2. Gate auctions and steering on learning-grade counts, not totals
3. Time-box credit-burn windows — chronic synthetic majority is a condition, not a one-time fever
4. Publish corpus split in every partner-facing chart

Why partners should care

If you share dispatch logs, you need a counterparty that will not launder synthetic traffic into your benchmark. SKENAI's corpus API exposes the split publicly so falsifiability survives contact with marketing.

Navigate: ← Part 3 · Part 5 → · Series hub — all articles