The Trustworthy Negative

Why “not found” is the hardest answer to build — and the only one that matters.

By VK Singh Vashisht — Datura Labs

Datura Labs runs on four laws. The third one reads: “Not found” is a calibrated answer, engineered to be as reliable as “found.”

I wrote that law before I understood what it would cost to actually keep it. This is what that cost looked like.

The easy version of trustworthy

Anyone can build a system that returns results. Ask it a question, it answers. Feed it a document, it processes it. Point it at a folder, it reports what’s inside. The demo works. The stakeholder nods. Everyone moves on.

The hard version — the one that actually matters — is a system that can say “I don’t have that” and mean it. Not as a fallback when something breaks. As a first-class, deliberately engineered answer that is exactly as reliable as a positive hit.

Building the Talking Database — the retrieval layer now running across roughly two million records — taught me that this property doesn’t happen by default. It has to be hunted for, specifically, because the default failure mode isn’t a system that crashes. It’s a system that reports success while quietly failing.

The health check that cannot fail

The first one nearly got past me entirely.

A health check ran a count against a search index and reported it as fully populated — every time, cheerfully, correctly formatted. What it was actually counting was the content table underneath the index, not the index itself. The shadow tables that make full-text search work could have been completely empty and the health check would never have known, because it was never actually asking the index a question. It was asking a related table that happened to hold a similar number.

The fix wasn’t complicated once found: a health check on a derived structure has to probe the structure directly — a real search query, not a row count on something that can silently proxy to the real answer. But finding it required not trusting a green checkmark, which is a strange discipline to build into your own tooling.

What “72% needing rotation” actually taught me

Scanned documents rotated 90 or 180 degrees don’t fail to extract text. They extract text — confidently, fluently, completely wrong. OCR doesn’t know it’s reading a sideways page. It reports a transcription with the same confidence it would give a clean one, and that transcription silently corrupts exactly the class of document most likely to matter: recorded instruments, filings, anything load-bearing.

The number that stuck with me: an early sample of 75 files suggested 72% needed rotation correction. The real rate, once we’d run it against the full 1,200-plus file set, was 21%. Both numbers were honestly measured. Only one of them was true at scale. That gap — between a plausible early read and the real distribution — is its own lesson: sample size isn’t a formality, it’s the difference between a fixable problem and a fiction you’ve talked yourself into.

The failure that only showed up in reality

We had a claim: the system degrades cleanly when a drive gets unmounted mid-operation. We’d tested it. The test passed.

The test pointed a cache at a directory that didn’t exist and confirmed the system handled the error gracefully. That’s a reasonable simulation. It is not the same event as a drive actually disconnecting mid-read, and when we finally forced a real unmount instead of simulating one, a second, different failure showed up — the degraded-mode fallback search quietly lost a piece of discipline the primary pipeline had been providing, and started returning weaker matches without flagging that it had done so.

The simulation wasn’t wrong. It was answering a different question than the one that mattered. A failure path you’re going to rely on in production has to be tested against the real condition at least once. A proxy for the failure is not the failure.

The pattern underneath all of it

Every one of these — and there are eleven of them in the running list we keep, not three — has the same shape. A system reported success while the actual signal was absent, degraded, or silently wrong. A passing metric. A plausible transcription. A test that verified the wrong thing convincingly.

The fix was never “trust the passing signal a little less.” It was to verify against the rawest available ground truth every time — the actual index, the actual physical failure, the actual bytes — instead of a structure that merely correlates with them. That discipline, more than any individual fix, is the reusable part.

This is why Law 03 isn’t decoration. A system you own is only as trustworthy as its worst silent failure, and silent failures don’t announce themselves — they have to be gone looking for, specifically, on the assumption that something is currently lying to you and simply hasn’t been caught yet.

Datura Labs builds agents and infrastructure directly on frontier models. This is the work.

The same discipline runs as a service: The Deal-Screen Desk.