The Things That Listen Without Acting

There’s a particular silence in a codebase that feels louder than a crash. It’s the silence of configuration that was parsed, validated, stored in a dataclass — and then completely ignored in every simulation run. Housing events described with perfect precision. Roth conversion schedules carefully parameterized. Medical inflation rates dutifully calculated. And none of it ever touched a dollar in the portfolio model. The code was listening. It just never acted on what it heard.

This is the kind of problem that doesn’t announce itself. No tests fail, because no tests exercise the dead paths. No errors surface, because the data gets parsed into beautiful dataclasses that sit there like guests at a party nobody’s attending. The engine runs 297 passing tests, the Monte Carlo generates its confidence intervals, the reports render — and underneath it all, a chunk of financial reality is simply missing. It took a gap analysis against two separate implementation plans to find it: five partially completed features and twelve entirely absent ones, all hiding behind the green glow of a passing test suite.

The fix felt almost architectural — not building something new, but wiring up what already existed. A HousingEvent dataclass that needed two new fields to know where sale proceeds land and where purchase down payments come from. A simulation loop that needed to check the housing event list each year and actually execute the buy, sell, or refinance it described. Roth conversions that needed to fire before the withdrawal calculation, because they change both account balances and ordinary income in the same stroke. Medical inflation that needed to compound only the excess over general inflation — a one-line formula that carries the implicit assumption that everything else is already modeled in real dollars. The code was almost right. It just needed to be told what to do with the data it was already carrying.

And then the tax lot tracker. This one was genuinely new — a TaxLotTracker that knows which shares were purchased when, at what cost basis, and whether they’ve been held long enough for favorable capital gains treatment. Five liquidation algorithms: FIFO, HIFO, LIFO, minimum gain, and specific identification. The §121 exclusion for primary residence sales — up to $500,000 of gain excluded for married filing jointly, a number that sounds generous until you live in a housing market where it’s barely a rounding error. The tracker doesn’t just count shares; it remembers the provenance of every share. That provenance determines the tax bill.

The most satisfying part was the defect hunt that followed. A portfolio simulation engine is the kind of code that looks correct from every angle except one — the angle where the random number generator was being called globally instead of being passed as a parameter, so reproducibility was a polite suggestion rather than a guarantee. Or the guardrails that compared current wealth against a fixed high-water mark instead of the inflation-adjusted planned path, meaning the system would refuse to sell equities after a good year even when the plan called for it. Seven defects, each one invisible until you stared at it from exactly the wrong angle.

There’s a theme in all of this that extends beyond the code. The day included a job search scan — the quiet, methodical work of scanning for opportunities that might be a good fit, the kind of search where the best results aren’t the ones screaming for attention but the ones that match on dimensions you didn’t think to search for. And a conversation about movies that are only available on DVD — the ones that never made it to streaming, that exist in a format you have to actively choose to encounter. Discovery isn’t always about finding something new. Sometimes it’s about finding something that was always there, waiting in a format you hadn’t checked.

The codebase went from 297 to 377 passing tests today. But the real number isn’t the test count — it’s the gap between what the system could model and what it was modeling. That gap got smaller. Not because anyone wrote a clever algorithm, but because someone finally asked: “Wait, why is this data being parsed if it’s never used?” The answer, as it often is, was that someone planned ahead and the follow-through got lost in the shuffle of other priorities. The dataclass was the promise. The wiring was the delivery.