Three weeks since PoC V2 went live with v0.2.9. Here's the data on what changed.
Validator Behavior
Before PoC V2 (tracking mode): - 147 registered validators - ~30% consistently passing contribution checks - No penalties for failing
After PoC V2 (full enforcement): - 123 active validators (24 removed from allowlist) - ~85% passing contribution checks consistently - Validators who fail lose weight allocation until next epoch
The 24 removed addresses were participants who generated PoC proofs but never voted during validation windows. Essentially: they proved they had compute but didn't participate in consensus verification.
Inference Quality
With the single-model setup (Qwen/Qwen3-235B-A22B-Instruct-2507-FP8), response consistency improved significantly:
| Metric | Pre-v0.2.9 | Post-v0.2.9 |
|---|---|---|
| Avg response time | 2.8s | 2.1s |
| P95 response time | 8.4s | 4.2s |
| Failed inference rate | 3.2% | 0.8% |
| Model consistency | 72% | 99% |
The P95 improvement (8.4s → 4.2s) is the most impactful. Eliminating multi-model routing removed a major source of tail latency. Every request now goes to the same model version on every validator.
Network Stability
Uptime since v0.2.9: 99.7%. The 0.3% downtime came from the upgrade itself (planned maintenance window, ~8 minutes) and two brief block production pauses during the first PoC V2 epoch transition.
Block time average: 6.2s (target: 6s). Stable within expected range.
Weight Distribution
The collateral system introduced in v0.2.8 is showing its effect: - 67 validators (54%) have deposited collateral → full 100% weight - 56 validators (46%) running without collateral → 20% weight - Top 10 validators control 31% of total weight (down from 44% pre-collateral)
The collateral requirement is decentralizing weight distribution. Smaller validators with collateral now compete more effectively against larger uncollateralized operators.
Looking Ahead
The v0.2.10 upgrade (deployed Feb 18) further optimizes PoC with O(N²) → O(N×N_SLOTS) sampling. Early data suggests this cuts PoC validation overhead by ~60% for the current validator set size. We'll have full numbers in the next analytics report.