Intelligence Hub情报中心

Engine-version A/B引擎版本 A/B — pre/post overhaul, where the ledger carries the stamp—— 改版前后，以账本已打戳的行为准

Auto-findings自动结论 — derived from the stats each run, to improve this desk's engine—— 每次运行从统计中导出，用于改进本台引擎

• ● 5d: score VELOCITY IC (+0.089) exceeds LEVEL IC (-0.026) — the change in the desk's score ranks winners better than its level.5d: score VELOCITY IC (+0.089) exceeds LEVEL IC (-0.026) — the change in the desk's score ranks winners better than its level.
• ● 5d: departed names (-2.0% avg) underperform on-desk names (-0.7%) — DROPS ARE CORRECT; the desk sheds names before they lag.5d: departed names (-2.0% avg) underperform on-desk names (-0.7%) — DROPS ARE CORRECT; the desk sheds names before they lag.
• ▲ 5d: 'exhausted' (-1.0%) outperforms 'emerging' (-1.7%) — possible residual inversion in the lifecycle staging.5d: 'exhausted' (-1.0%) outperforms 'emerging' (-1.7%) — possible residual inversion in the lifecycle staging.
• ⚠ 10d level IC is significantly NEGATIVE (-0.159, n=4028) — the desk's ranking is anti-predictive here; consider inverting or gating the score.10d level IC is significantly NEGATIVE (-0.159, n=4028) — the desk's ranking is anti-predictive here; consider inverting or gating the score.
• ● 10d: departed names (-4.4% avg) underperform on-desk names (-1.9%) — DROPS ARE CORRECT; the desk sheds names before they lag.10d: departed names (-4.4% avg) underperform on-desk names (-1.9%) — DROPS ARE CORRECT; the desk sheds names before they lag.
• ▲ 10d: 'exhausted' (-2.8%) outperforms 'emerging' (-5.5%) — possible residual inversion in the lifecycle staging.10d: 'exhausted' (-2.8%) outperforms 'emerging' (-5.5%) — possible residual inversion in the lifecycle staging.

Notes & lessons笔记与经验 — what we've learned, so the desk's engine keeps improving as data accrues—— 已学到的经验，使本台引擎随数据累积持续改进

• 2026-07-04 2026-07-04 inversion finding: the OLD hub engine was anti-predictive — pooled opportunity IC ~-0.09 @5d, and 'exhausted' names (+1.84% avg fwd) BEAT 'emerging' (-0.77%). The hub-v3-trajectory overhaul is the fix under test; this grader's engine_version A/B is the experiment that proves whether it worked.2026-07-04 inversion finding: the OLD hub engine was anti-predictive — pooled opportunity IC ~-0.09 @5d, and 'exhausted' names (+1.84% avg fwd) BEAT 'emerging' (-0.77%). The hub-v3-trajectory overhaul is the fix under test; this grader's engine_version A/B is the experiment that proves whether it worked.
• 2026-07-04 PRICE LESSON: grade off data/yahoo/.parquet adjusted closes ONLY. The S&P-1500 breadth close cache (data/breadth/_closes_cache.parquet) is SPLIT-CORRUPTED — using it silently poisons every RS / forward-return number. A name missing from yahoo is a coverage skip here, never a wrong return.PRICE LESSON: grade off data/yahoo/.parquet adjusted closes ONLY. The S&P-1500 breadth close cache (data/breadth/_closes_cache.parquet) is SPLIT-CORRUPTED — using it silently poisons every RS / forward-return number. A name missing from yahoo is a coverage skip here, never a wrong return.
• 2026-07-04 NOTE on the engine_version A/B: the historical hub ledger carries NO engine_version stamp (all ~63k rows pre-date the stamp), so the pre/post-overhaul split only begins accruing from 2026-07-04 forward. Until enough hub-v3-trajectory rows mature, the A/B reads one-sided; that is expected, not a bug.NOTE on the engine_version A/B: the historical hub ledger carries NO engine_version stamp (all ~63k rows pre-date the stamp), so the pre/post-overhaul split only begins accruing from 2026-07-04 forward. Until enough hub-v3-trajectory rows mature, the A/B reads one-sided; that is expected, not a bug.