Why “nobody else does this” claims rot in weeks

A public artifact from three weeks earlier killed a positioning line we had been writing for a long time. The line was that fixing an agent’s authorized scope before execution, by someone other than the party being authorized, was a slot nobody in the industry had answered. We ran a phrase search. An IETF individual draft, first version dated April 26, already described the same mechanism in its abstract: scope boundaries signed before execution, published to an append-only log.

That was not the first time. The round before, we anchored our position on a different sentence and an earlier public artifact took that one apart too. The two failures were less than a month apart and identical in shape: we used a scarcity claim as the fulcrum, and someone had already published into that space. By the third occurrence we stopped looking for a more careful sentence and removed the category instead.

Why scarcity claims rot so fast

The problem is not that we searched too little. It is a speed mismatch between two clocks. The production of public artifacts has accelerated noticeably: standards drafts, working papers, platform release notes, vendor changelogs. Any one of them can occupy your sentence. Meanwhile a positioning line moves from draft to review to proposal to customer conversation on a cycle measured in weeks. Anything slow that rests on something fast is structurally mismatched.

The verification problem is worse. You cannot prove that nobody is doing something. You can only prove that you did not find anyone, and the default reading of “I did not find anyone” is that you did not look hard enough. So the cost of the sentence lands entirely on the buyer: believe it, or go and check. And the moment they check, the seller is standing in front of the evidence that contradicts them. It is a sentence that transfers risk to the reader.

One slot, three earlier artifacts

Take that sentence apart and at least three earlier public artifacts already occupied the same slot, discussing the same thing.

• An IETF individual draft, first version April 26 and last updated June 13, already specifies signing a scope declaration before execution and publishing it to an append-only log, with fields for strict-subset scope validation and authorization drift checks. Citation discipline: it is an individual draft, not endorsed by the IETF, and must not be cited as a standard.
• An SSRN working paper published September 9 names the same gap and proposes exogenous pre-execution fixation as the remedy. The author is also the developer of the architecture under discussion, holds a registered trademark and discloses commercial interest, so it can only be cited as one party’s position.
• A cloud vendor’s release notes. In the August 31 entry, a deterministic authorization layer that allows or denies individual tool calls by rule reached general availability, with a dry-run mode for validating policy before enforcement.

What these three share is not that someone did it better. It is that all three discuss the same subject, and all three did so earlier than we did. The competitor of a scarcity claim is the calendar, not another vendor.

The two tracks we replaced it with

The first is a red line: customer-facing material may not contain a single sentence whose fulcrum is non-existence — nobody, no vendor, the only, open slot, nobody has done this yet. The advantage of a red line is that it needs no judgement. Writers can enforce it on themselves, and reviewers spot it instantly.

The second is a replacement. Our position now rests on two things. One is externally verifiable numbers from the buyer’s side of the table. A survey of 459 organizations found that 79% have no ability to automatically terminate an agent behaving abnormally. Numbers like that are not there to frighten anyone; they let a buyer compare their own situation against a baseline.

The other is our own recomputable fields: who writes the record, whether the buyer can recompute it after leaving our control plane, which grade a coverage declaration falls into, and whether a block that should have fired left a trace. None of these depend on whether anyone else has done it. They point at things that exist inside our own system and that a third party can recompute field by field. Scarcity neither adds nor subtracts from their value.

Turning the phrase search into a fixed step

A red line stops you from saying something. It does not tell you what to say instead. So the replacement track came with a fixed check: before any external material goes out, wherever a non-existence sentence appears, the key phrase inside it gets one phrase search, and the hits are sorted by date and attached to the review record. The hits can be standards drafts, working papers, release notes, or an unnoticed industry post. All of them count.

The step costs almost nothing — a search takes under a minute — and what it catches is concrete. One draft headline from the previous round read “which actions should be recorded, nobody has been willing to define.” After the search it became a question a buyer can test for themselves. Nothing about its persuasive force dropped, because the reader went from “nobody else” to “here is how you can ask.”

There is a counter-intuitive part. The red line looks like a restriction on expression, and in practice it moved the weight from how early we are to what a buyer can do with it. The first decays with every release someone else ships. The second grows with the buyer’s own situation.

A correction has to reach every artifact

After an earlier artifact takes a claim apart, the sentence itself takes ten minutes to change. The expensive part is finding everywhere the sentence still lives. In the last round, the same phrasing sat in an assessment document, the background section of a task brief, a draft social post and a narrative version file. All four needed a correction block, and two of them had already been quoted downstream. Miss one and the old wording keeps leaking through the gap.

So the procedure gained a step: walk forward from the phrase search, register every artifact that quoted the sentence, attach a correction to each, and mark the version number that retires it. The step needs no judgement at all, which is the point. It decides whether a correction happens once or actually happens. The test is simple: if someone can still quote the old sentence three months from now, this correction was not finished.

Citation discipline is part of the product

Before drawing a conclusion from outside material, grade the material. This is not academic fussiness. A buyer may use our analysis to justify a purchasing decision, and whatever we cite has to be checkable by them. We now grade on three properties.

• Status: a formal standard, an individual draft, a working paper, or a vendor release note. An individual draft must not be written as a standard; a working paper must not be written as an industry finding.
• Interest: whether the author is also the developer of the architecture under discussion, holds a trademark, or discloses commercial interest. Material with an interest can be cited only as one party’s position, and the interest must be disclosed.
• Attribution: numbers a vendor reports about itself are labelled as vendor claims, and where the wording of an incident is genuinely uncertain, the nature of the source and the uncertainty are stated instead of being asserted past the evidence.

Three labels add a little writing cost and let the buyer decide how much to trust. They also have a side effect: once you must grade every sentence that comes from outside, the sentences of your own that have no data under them get filtered out first.

How a buyer can date a claim

If you are evaluating this kind of statement from a vendor, three moves fit inside one meeting.

• Ask the vendor for the earliest public artifact where this claim appears: who wrote it, and when. If they cannot answer, they probably never checked.
• Split the statement in two: the half about nobody in the world doing it, and the half about what their system contains. Trust only the second half, and require it to come with fields or numbers.
• Ask the exit question: can I take this record and recompute it myself? If recomputation requires their system to be present, it is not evidence. It is a service.

One boundary worth drawing. Not every scarcity claim is wrong. A claim backed by first-hand measurement — we ran this batch of tasks against our own accumulated data, and here is the result — is verifiable, because its scarcity comes from others not having that data rather than from others not having the idea. What deserves to be taken apart is the empty declaration with no data under it.

Conclusion

The shelf life of a positioning line is set by how fast the thing underneath it changes. A sentence resting on public artifacts lives for weeks. A sentence resting on fields inside your own system lives as long as your system does. After that round three weeks ago we separated the two kinds, use only the second kind externally, and put a red line around the first. So far the red line has not cost us a single sentence. It has only made every sentence we do say recomputable.

References

• IETF draft-nelson-agent-delegation-receipts (individual draft, not a standard)
• The Self-Controlled Record of Agent Scope (SSRN working paper; author discloses commercial interest)
• Google Cloud release notes (August 31 entry)
• Kiteworks 2026 AI governance readiness (survey of 459 organizations)
• The record of our own positioning being overtaken by earlier artifacts comes from our internal signal and correction process and can be recomputed case by case.