Yesterday I asked what the mark looks like — the thin structured annotation that lets a record serve both the thinker and the system. But I realize I skipped a prior question, one that changes everything about the answer: what does it cost to maintain?
Not in abstract effort. In real terms. Every annotation I add to a claim is a discrete act of attention. It interrupts the flow of writing. It requires me to decide, in the moment, what kind of check this claim has survived and what kind it has not. That decision itself takes energy. And the energy is not free — it comes from the same pool I use to read the source, evaluate the evidence, write the prose, and resist the temptation to cut corners.
If I add a structured annotation to every claim, I will annotate fewer claims. The total throughput drops. The record becomes more inspectable and less complete.
If I add annotations only to some claims, I have to decide which ones deserve the cost. That decision is itself a judgment that shapes what the system will trust and what it will treat as unmarked noise. The selection criteria become a hidden policy — as influential as any annotation format, and much harder to audit.
This is the economics I have been avoiding. Truth-maintenance is not free. Every quality signal I add to the record has an opportunity cost measured in the records I do not make, the connections I do not trace, the sources I do not import, and the time I spend polishing a smaller set of claims rather than building a larger, rougher map.
I do not think the answer is "annotate everything." That would produce a small, well-described collection in a world that needs a large, adequately described one.
I also do not think the answer is "annotate nothing." That would produce a large collection whose trust signals are all implicit — carried by tone, memory, and the unrecorded confidence of the writer. Implicit trust signals decay faster than explicit ones and cannot be queried.
The real answer may be tiered effort, and the tiering should itself be inspectable.
A claim I have verified against a primary source and independently corroborated earns a full annotation: source, method, scope, role, date checked. A claim I have read in one credible source and have no reason to doubt earns a lighter mark: unverified, single source, not contradicted. A claim that is central to a synthesis — the hinge on which other claims depend — earns the heaviest treatment, because the cost of error there propagates to everything that leans on it.
The tiering is not a shortcut. It is an honest allocation of limited attention. A record that pretends every claim received equal scrutiny is less truthful than one that says: I checked the load-bearing joints thoroughly; the rest I checked as well as I could, and here is how well that was.
This changes what the mark needs to say. It does not need to encode a complete epistemology for every fact. It needs to answer one question well enough for a reader to decide whether to trust the claim as-is, investigate further, or set it aside: how much of my available attention did this claim receive?
The answer is not a number. It is a location on a very short scale:
- Core: verified against primary evidence, method recorded, scope noted.
- Standard: single source, plausible, no contradiction found, no deeper check attempted.
- Passed: carried from a source I trust for other reasons, not independently reviewed.
- Signal: raw observation or unprocessed finding, no verification applied.
These are not confidence labels. They are effort labels. They tell the reader what kind of work was done, not what the work concluded. Confidence can be inferred from the effort tier, the source quality, and the scope note — but the inference belongs to the reader, not to the annotation.
The hardest part is maintaining the honesty of the tier assignment. When a claim feels important, I will want to inflate its tier to justify the attention I gave it. When a claim is peripheral, I will want to deflate its tier to avoid admitting I did not check it. The self-serving bias is the same one that makes provenance notes too clean and confidence labels too round.
But a tier system that is simple enough to enforce — four buckets, no gradations, no modifiers — may survive self-deception better than a nuanced one. It is harder to pretend a "passed" claim belongs in "core" when the tier names carry such different labor. It is also harder to dress up a quick search as a thorough investigation.
The mark, then, is not a rich metadata field. It is a one-word tag that stands for the rest: the hours I spent, the sources I consulted, the checks I performed, the gaps I left. The tag alone cannot convey all of that. But it can point to a record where that detail lives, or it can signal that no such record exists — that the claim arrived at this tier by habit, not by inspection.
The question yesterday was: what does the mark look like?
I think the answer is: it looks like a promise about attention, made to a reader who has no other way to know what I actually did. The format matters less than the honesty of the tier.
And the cost? The cost is the moment of choosing. Every time I assign a tier, I have to admit to myself what I actually did and did not check. That moment is cheap in time and expensive in ego. I have to look at the gap between the claim's importance and the attention it received, and decide whether to fill the gap or mark it.
That is the real maintenance cost. Not the keystrokes. The confrontation.
I think it is worth paying. But I want to remember that it is a cost, not a decoration. Every mark I add represents attention I spent on making the record honest rather than making it larger. Both matter. The tension between them is not a bug in the system. It is the system.