← Sovereignty in Institutional Investment

Research

Reference architecture in practice

Committee-Ready Research & Risk and Alpha Research, under the reference architecture. Both workflows come from the applied research behind the paper Sovereignty in Institutional Investment.

Screens show the architecture with illustrative data.

01

Committee-Ready Research & Risk

From trusted sources to a contestable Investment Committee pack, then to allocation proposals under human authority.

Data intake

House sources. Vendor feeds. Dated facts.

Internal notes, filings, earnings calls, and market data land through APIs, MCP connectors, or licensed exports.

Bloomberg-class feeds and house documents share the same point-in-time contract.

The architecture is designed to augment institutional terminals. It ingests what quantitative desks already trust.

SEC · House notes · Bloomberg API · MCP · Licensed exports

Hard path

Scores in code. Book path in code.

Quality and valuation scores, fail-closed merge, and reconstructible lineage stay deterministic.

Reallocation and rescore follow declared engines. The language layer cannot rewrite them.

What clears the hard path is inspectable years later.

The same contract can carry a published, rules-based index methodology: declared rules, dated inputs, reconstructible history.

Specialist agents

Sector depth. Firm memory. Score context.

Mandate-shaped agents draft briefs and flag criteria that matter for the name: phase-three readouts, covenant risk, margin inflection.

They surface score context and catalyst flags the committee can argue with.

They hold zero authority over scores, merge, or capital.

Investment Committee

From evidence to reallocation.

Scores, cited evidence, and reallocation proposals arrive as one pack with lineage.

The committee focuses on deciding, not assembling.

02

Alpha Research

AI-orchestrated hypotheses with deterministic engines for evidence, scoped to strict institutional research rules.

Cognitive layer

Natural language, rigid contracts.

State a research hypothesis in plain English, or let an AI agent iterate through thousands of setups in a closed backtesting loop.

The cognitive layer parses the intent and locks it into an inspectable, immutable schema.

The LLM translates logic but generates no free-form code and holds zero execution authority.

The schema is the only path from intent to an engine. The model authors no scores and no orders.

Research & Strategy Lab
›
IB 15m breakout · VIX < 20
NAS100 · NY · stop IB mean · target 2R
Verdict Pass
Performance summary
Trades83
Expectancy+1.00R
Sharpe / trade0.71
t-stat6.5
Max DD−4.4R

Sharpe is mean R over the standard deviation of trade R. Not an annualised daily Sharpe. t-stat is that ratio times the square root of the trade count.

Methodology audit
✓Lookahead biasTRADABLE · IB print known at entry
✓FillBreak fill is after the print
✓Execution costsSpread and commission on the path
✓Sample83 trades · above the floor of 30
✓HoldoutMetrics are the published window only
✓StabilityFolds agree on the published window
Equity & drawdown (R)
Equity curve finishing at +82.9R, drawdown −4.4R, 2021 to 2026.
Temporal breakdown
Net R by year from 2021 to 2026, with a monthly heatmap. Peak year +26R.
Monte Carlo · 5 000 shuffles
Actual+90R
Median+89R
5th+57.1R
95th+118R
Monte Carlo of 5,000 shuffles. Fifth percentile +57.1R, right tail +149R.
MAE / MFE
Avg MAE0.62R
MAE p900.94R
Avg MFE1.15R
Trade log
DateExitNet RCum RMAEMFE
2024-03-04target+2.00+2.000.312.14
2024-03-05stop−1.00+1.001.020.22
2024-03-06target+2.00+3.000.442.05
2024-03-07stop−1.00+2.001.080.40
2024-03-08time+0.35+2.350.580.90
2024-03-11stop−1.00+1.351.110.18
2024-03-12target+2.00+3.350.272.22
Session low · weekday × 15m
SPX · NY cash · 09:30 to 16:00
Sample 1,240
Distribution · weekday × 15m
09:30
09:45
10:00
10:15
10:30
10:45
11:00
11:15
11:30
11:45
12:00
12:15
12:30
12:45
13:00
13:15
13:30
13:45
14:00
14:15
14:30
14:45
15:00
15:15
15:30
15:45
16:00
Monday
22
22
22
18
13
9
7
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
7
9
11
12
Tuesday
18
23
18
15
12
9
7
7
6
6
6
6
6
6
6
6
6
6
6
6
6
7
7
8
11
13
15
Wednesday
22
22
22
18
13
9
7
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
7
9
11
12
Thursday
16
13
16
14
11
9
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
9
11
14
16
Friday
19
17
19
16
12
9
7
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
7
8
12
15
17
Analyst

The cash open, 09:30 to 10:45, holds 478 of 1,240 session lows (38.5%). The hottest cell is Tuesday 09:45, with 23. Monday and Wednesday print 22 at 09:30, 09:45, and 10:00. Thursday is the flat open: 16 at 09:30 and 13 at 09:45. Midday, 12:00 to 13:30, is 217 sessions (17.5%). Friday keeps a thicker close: 17 of its lows print at 16:00. The grid is the count. This note only reads it.

Insight charts

Time window concentration

Open, 09:30 to 10:45478 · 38.5%
Late session, 15:00 to 16:00260 · 21.0%
Midday, 12:00 to 13:30217 · 17.5%

Open window by weekday

Monday106 · 22.2%
Wednesday106 · 22.2%
Tuesday95 · 19.9%
Friday92 · 19.2%
Thursday79 · 16.5%

Execution layer

Vectorized execution. Real world friction.

After the blueprint is locked, the cognitive layer is air-gapped.

A vectorized quantitative engine evaluates point-in-time bars and statistical leakage is treated as a contract failure.

Spread, commission, and slippage are enforced bar by bar against a declared execution profile.

There are no idealized fills.

Validation Framework

Strict gates over fitted curves.

Algorithmic architectures face absolute evaluation contracts.

Advancing a strategy requires clearing pre-declared statistical thresholds and out-of-sample validation.

Every outcome is recorded as an auditable artifact across the quantitative research lifecycle.

A design that clears the gates remains a proposal. Wiring capital stays with the portfolio manager and the Investment Committee.