The agentic equity research desk
From headline
to verdict.
On the record.
Discover → validate → execute → review — one evidence chain, end to end. A six-lens council cross-examines every thesis, a judge holds veto power, and failed methods go to a public graveyard.
01 / 04
Discovery · /newsscan
One headline. The whole supply chain.
Paste in a news item and the discovery engine runs five methods in parallel — explicit mentions, supply-chain back-tracing, causal chains, thematic recall, vector expansion — ranking beneficiaries by multi-method resonance. Strong names go straight to validation.
Real batch 2026-06-25 · one scan → 16 beneficiary briefs
02 / 04
Validate · council + judge
Four disciplines. Six lenses. One veto.
Technicals, fundamentals, valuation and options flow build four independent evidence lines; six investor lenses debate them; a judge rules last — including survival vetoes that have nothing to do with being right. The output isn't a score. It's a falsifiable battle card: levels, triggers, invalidation.
03 / 04
Market radar
The market isn't backdrop. It's a gate.
GEX structure, dark-pool DIX, sector rotation, industry KPIs — market-level data is computed once and shared by everyone. The regime gate rules Easy or Difficult, and that decides what you're allowed to do today.
Sector rotation · RS vs SPY · 2026-07-31
Options structure · MARKET card · 2026-07-31
SPY GEX
1.87bn
QQQ GEX
-2.94bn
DIX
45.3%
Industry KPIs · latest per series
04 / 04
Research ledger
Every failure, on the record.
Every method is pre-registered before it's tested: pass the line and it ships, fail and it goes to the graveyard — permanently public. This isn't marketing copy. The table below is the ledger itself.
| Date | Method | Trials | Outcome | Note |
|---|---|---|---|---|
| 2026-07-05 | vrp_timing | 1 | ⚰ graveyard | 五分位差+0.3pp(07-04) |
| 2026-07-05 | dsr_family_recheck | 0 | ✓ pending | N=12复检: connors DSR 0.13/voltarget 0.34 均<0.95 → 两规则降级'条件保留,前瞻监控裁决'(rule_monitor) |
| 2026-07-05 | bollinger_upper_fade | 1 | ⚰ graveyard | 事件研究:P(5日跌)=49%(07-02) |
| 2026-07-05 | connors_pctb_entry | 3 | ✓ deployed | 3变体,59%胜率,DSR0.71@N=3(07-02) |
| 2026-07-05 | brooks_failed_breakout | 1 | ⚰ graveyard | 量级仅日波1/4(07-03) |
| 2026-07-05 | brooks_80_rule | 1 | ⚰ graveyard | range成功率48%非20%(07-03) |
| 2026-07-05 | brooks_wedge_fade | 1 | ⚰ graveyard | 三推后f10高于基线(07-03) |
| 2026-07-05 | kalman_trend_turn | 1 | ⚰ graveyard | 精确率26%召回5%(07-04) |
| 2026-07-05 | vol_targeting | 2 | ✓ deployed | Calmar0.34→0.40(07-04)待复检 |
| 2026-07-05 | gap_quantile_bands | 1 | ⚰ graveyard | 样本外违约9.4%(07-04) |
Verdicts can be wrong.
The process has to stand.
This engine runs on a real-money personal book first. Every number on this page is a byproduct of it running.
Enter the CRCL war room →