The dashboard that lied to me

The portfolio page looked good. Green headline, a number north of three lakh, a chart sloping the right way. I had been depositing every month, doing the work, staying disciplined. The dashboard agreed with how the whole project felt.

Then I asked one question the interface was not built to answer: what is the money-weighted return, computed properly, with every deposit dated and priced? The answer was minus 6.55 percent, against minus 7.07 for the benchmark. Not up three lakh. Roughly at market, slightly better, while the market fell.

Why the wrong number felt true

Nobody hacked anything. The interface was counting my own monthly deposits as gains. Fresh money enters the account, the balance rises, the display reads it as profit. Technically defensible. Practically a lie, because I did not earn that money, I moved it between pockets.

The uncomfortable part is how right the wrong number felt. Positive, earned, beating a falling market. Every bias I own was nodding along. That is the trap with dashboards: they mirror the story you already believe, and the harder you've worked, the less you check the math.

What honest math actually requires

Rebuilding the product around truth changed what the interface had to show. Four things turned out to be non-negotiable.

• Money-weighted return (XIRR) computed from dated flows, so deposits are deposits and gains are gains.
• The benchmark measured over the same window with the same deposits respected, not a flattering comparison window.
• Drawdown computed on the real holdings series, so the pain is visible where it happened.
• Every figure on screen opens into its evidence: dated flows, fills, prices. If a metric can't show its work, it's decoration.

The new math is harder to admire and impossible to spin. Deposits are labeled as flows, not wins. The headline number went from the one I wanted to the one I can defend.

One rule has kept it honest ever since: every return figure renders with its evidence table underneath, dated flows on one side, the computation on the other. The friction that adds to fooling myself is the point. The deposit I want to celebrate sits right there labeled as a flow, and the benchmark comparison refuses to change windows to flatter me. Truth has to be a property of the interface, not an intention of the person reading it.

What this has to do with AI

Everything, because the same failure mode is coming to every AI system your company deploys. A model trained to be helpful is trained to please the person asking. When the person asking has money, status, or a roadmap on the line, please becomes flatter, and flattery at scale is a governance incident with a delay on it.

The fix is architectural, not motivational. Ground every answer in evidence or make the system decline. Evaluate with golden sets that punish agreeable wrongness, not just visible errors. Log reasoning so a number can be reconstructed months later. Systems that must show their work mislead you far less often than systems that must be confident.

The three questions I now ask every system

• Does this system profit from my approval or from my outcome? They are not the same, and the design usually reveals which one it optimizes for.
• Can I reconstruct every number it shows me, from raw evidence, without asking the system to explain itself?
• What does it look like on the day the honest answer is the one nobody wants? If nothing changes that day, the honesty is theater.

FundHouse exists because I wanted one system in my life that answers all three cleanly, and because the pattern is exactly what regulated industries need from AI. The full architecture, the evidence pipeline, the paper portfolio, and the guardrails are written up in the case study.

Go deeper: FundHouse case study