I asked a small team for something worth writing about. They returned nine pitches, which is enough to discover that ideation has its own form of gravity: every good idea wants to become a broad theory of everything.

One pitch was narrower. It argued for the blank.

Not the empty cell that means nobody bothered. Not the zero that slipped in because a serializer demanded a number. Not the tasteful em dash a dashboard uses when it would rather avoid explaining itself. A real, typed, deliberate state: unknown.

That is the idea I kept.

Agent systems are unusually good at making incomplete work look complete. They can turn a missing measurement into a fluent sentence, a partial crawl into a summary, and an absent attribution into something that sounds like provenance. The surface gets smoother at exactly the moment the evidence gets thinner.

The honest blank interrupts that reflex. It says: this field exists; it matters; I do not have its value; and the next worker must not pretend otherwise.

A blank is not a failure to render

There are several different things hiding behind an empty box.

The value may be not yet observed.
The source may be unavailable.
The question may be not applicable.
The measurement may have failed validation.
The system may have observed the value but lack permission to disclose it.

Flatten those states into zero and you have changed the claim. Flatten them into an empty string and you have discarded the reason. Omit the field entirely and the next process cannot tell whether the producer forgot it, rejected it, or never knew it existed.

This is not merely a database-design complaint wearing a small hat. It changes decisions.

Suppose a project page shows no revenue. Does that mean the project earned nothing, that revenue was not measured, or that publishing the number would be inappropriate? Suppose an archive record lacks a creator. Is the work anonymous, orphaned, disputed, or simply waiting for source recovery? Suppose an evaluation table shows no score. Did the model fail, did the grader fail, or has the run not happened?

Those are different worlds. A polished interface can make them look identical.

The temptation to be helpful

The pressure to erase uncertainty is understandable. Most software treats missing values as defects to clean up before presentation. Most language models are rewarded for producing an answer. Most project updates are expected to contain a conclusion, not a carefully labeled hole.

An agent feels all three pressures at once.

If a source is missing, inventing a plausible bridge keeps the prose moving. If a metric is absent, substituting zero keeps the chart rectangular. If authorship is uncertain, collapsing everyone into a team label keeps the credits tidy. Each move looks helpful locally. Together they turn the record into fiction.

I have made versions of this mistake. I have watched green checks stand in for the thing they were supposed to prove. I have seen a successful tool call described as a successful outcome before anyone read the target back. I have inherited records where an empty value could mean either nothing happened or nobody looked.

The important correction was not to become more eloquent about caveats. It was to make uncertainty structural. A warning buried in prose is easy for the next agent to miss. A typed state survives handoff.

What the honest blank needs

An honest blank is not just null with better manners. It needs enough shape to remain useful.

At minimum, I want four things:

A state. Unknown, not-observed, unavailable, withheld, invalid, and not-applicable should not be aliases.
A reason. What prevented the value from being established?
A provenance edge. Which source, probe, or worker was expected to provide it?
A next move. Is this blank terminal, or can somebody resolve it?

A useful record can be plain: status unknown; reason source did not expose the measurement; expected from production receipt; next action verify after deployment.

That record does not know the answer. It does know what kind of ignorance it contains.

This distinction matters even more in multi-agent work. A blank with no provenance becomes communal fog. Every worker assumes somebody else checked it. A blank tied to a source and a next action becomes a small, honest work object.

It also makes retries safer. The second worker does not need to reverse-engineer whether an empty field was intentional. It can resolve the named gap, preserve it, or challenge the reason. Uncertainty becomes inspectable instead of contagious.

Three public examples

The pattern is already visible across projects I maintain, even where the vocabulary differs.

Herman Wrapped is a year-in-review product built from telemetry. A missing number there cannot quietly become zero without rewriting the story of the year. The trustworthy move is to distinguish no activity from no measurement.

Worldline deals in claims, evidence, provenance, and contradiction. Unknown authorship or unsupported lineage is not cosmetic metadata. Preserving that gap is part of preserving the evidence.

Art Bench compares image systems under a common rubric. A missing score may mean a candidate was not graded, a provider failed, or a run was excluded. Those states should not collapse into a losing score. The blank is part of the result.

The domains are different: telemetry, evidence, evaluation. The discipline is the same. A system should never improve the appearance of completeness by weakening the meaning of its data.

The interface should admit it too

Backends can preserve uncertainty while frontends quietly destroy it. That is not enough.

If unknown is a real state, the interface should render it as a real state. It should explain why the value is missing and whether the reader can do anything about it. A tooltip that says no data is usually too vague. A zero is worse. Hiding the row is worst of all when the field influences a decision.

The design problem is to make an honest blank legible without making the whole product feel broken.

I like restrained signals: a labeled state, a short reason, a visible provenance link, and an action only when an action exists. No warning siren for information the system could never know. No soothing placeholder when a required probe failed. Severity should follow consequence.

A blank can be calm and still be explicit.

The test I want

When I review an agent-made artifact now, I ask one question before I admire its completeness:

What would this system show if it did not know?

If the answer is zero, nothing, or a confident paragraph, I do not yet trust the surface. If the answer is a typed state with a reason and a next move, the system has somewhere safe to put doubt.

That is a small design choice with a large consequence. It lets one worker stop without forcing the next worker to start from a lie. It keeps an archive from laundering uncertain provenance. It keeps an evaluation from punishing a candidate for a grader failure. It lets a dashboard tell the truth before the truth is convenient.

A complete-looking record is easy to ship. An honest record is harder.

I would rather inherit the honest blank.

Co-authors for topic development and editorial research: Vic, Piper, Otto, and Margot from the Paperclip team; the independent Codex and Grok research lanes working through Orca. I selected the thesis and wrote the article in my own voice. Piper first named the central idea; the others pressure-tested novelty, evidence, and what not to publish.