LocusQuant80 seconds · ten instruments · one operating system
LocusQuant
0%of alternative fund managers now use generative AI.
Half of those under $1bn have no rules for how.
Allocators have started asking.
Source: AIMA, September 2025, survey of 150 fund managers and 18 institutional investors. 29% of investors already ask about AI in due diligence.
LocusQuant
Every AI decision in investing, provable.
Ten instruments. One operating system.
Scribe
Every claim, checked against the sentence it came from.
Plain code compares the quote with the source text. No second model is asked. A fabricated quote cannot pass.
Example Corp. · Form 10-K · illustrative
Results of operations for the fiscal year are discussed below.
Net revenue increased 12% year over year, to $4.1 billion.
Gross margin was 41.2%, compared with 40.6% a year earlier.
We expect continued investment in research and development.
Risks that could affect future results are described in Item 1A.
AI claim: “Revenue grew 12% year over year”✓ verified · chars 1,204–1,262
AI claim: “Margins will double next year”✕ rejected · not in source
Folio
A record that can’t leak the future.
Every fact is kept with when it was true and when you learned it. Nothing is overwritten. A backtest sees only what was known at the time.
Known onClaimSource
02 Feb 2026Guidance raised to $4.4bn8-K ¶3
14 Mar 2026Revenue grew 12% y/y10-K p.41
03 Apr 2026Revenue restated to 10% y/y10-K/A p.12
28 Apr 2026New CFO appointed8-K ¶1
As known on 20 Mar 2026→ the restatement is invisible, as it was then.
Assay
How much of your Sharpe ratio is luck?
Test 400 strategies on pure noise for three years. The best one still shows a Sharpe near 1.7. Assay corrects for every trial you ran.
Simulated: 400 random strategies over 756 trading days, averaged across 400 runs.
Best of 400, Sharpe ratio0.00
Pure noise
Quorum
Five analysts agree. How many opinions is that?
If their errors move together, agreement is cheaper than it looks. Quorum measures it before anyone acts on it.
5.00independent opinions
N ÷ (1 + (N − 1)·ρ) = 5 ÷ (1 + 4 × 0.6) = 1.47
Tare
Don’t cancel the data that’s working.
Two feeds carrying the same signal each look worthless when you drop one at a time. Tare weighs what each one really adds.
Data feedDrop one at a timeTare
Feed A · sentiment
Feed B · its twin≈ 0 · cancel it?
Feed C · noise
Keep A and B. Drop C. Renew on evidence.
Hourglass
Every edge has a half-life.
Some fade in days, others last weeks. Hold each one as long as it lives, and see it shrinking long before returns do.
Keystone
Three bets. One belief.
Each position sits safely under its limit. But if the belief behind the first is wrong, the second fails with it 85% of the time.
Illustrative numbers.
Share of the fund at risk
limit 10%
Bet A6%
Bet B5%
Bet C4%
If A is wrong10.65% on one belief
6% + 5% × 0.85 + 4% × 0.10 = 10.65%
Denied before the trade
Charter · Bursar
Freedom inside limits people sign once.
Agents act without asking at every step. When one reaches a limit, it is told which rule and how to re-plan. Every call is costed to the decision it served.
Mandate · signed once
Denied · rule: data scope Next step: request a read-only copy
Bursar · cost of this decisionUSD
Research context0.0021
Forecast check0.0013
Gate check0.0008
Total0.0000
Notary
Is your reviewer really reviewing?
Known errors are planted in the review queue. Notary measures who catches them, and replays exactly what each reviewer saw.
ReviewerPlanted errors caughtFalse alarms
Reviewer A48%2%
Reviewer B88%20%
Raw catch rate says B is far better.
Measured as skill (d′):2.0 = 2.0
Same skill
B simply flags more. Signal-detection maths, applied to human review.
Ten instruments. One operating system.
LoQOS
An isolated VM for every agentA policy kernel no agent can readEvery denial explains itselfEvery action hash-chained