Truth marking is the rule that every significant claim about AI carries one of four labels: “measured”, “building”, “concept” or “forecast”. The label sets the verb tense and what must stand next to the claim: an instrument or source and a date for a measured fact, an acceptance check and a deadline for something being built, a scenario branch and an external benchmark for a forecast, so that anyone can check whether it came true. A text then stops mixing what works with what is promised. The institute writes its book “Proven Action” and the pages of the IS 3.0 section by this rule.
In brief
- The problem: texts about AI describe a prototype as an industrial standard, an intention as a working feature and market expectations as a proven model.
- Four labels: MEASURED, done and measured; BUILDING, under construction with an acceptance check and a deadline; CONCEPT, a design with no acceptance check yet; FORECAST, a scenario with a branch and a benchmark.
- The verb tense follows the label: a concept may not be described in the present tense.
- A forecast has three branches: A, “normal technology”; B, the trend continues; C, acceleration; each with an external benchmark and a check date.
- An empty indicator means “not measured”, not “all is well”: absence of data is not passed off as safety.
Why label claims about AI?
Whoever decides on investments, security architecture or regulations reads texts about AI in which genres are mixed. A prototype that works in a lab is described in the present tense, the developers’ intention is passed off as an existing feature, and multiplied market expectations as a financial model. From such a text one cannot tell what already exists and what is only promised. A label gives the reader that boundary back.
Four labels and what must come with them
| Label | Verb tense | When it is used | What must stand next to it |
|---|---|---|---|
| MEASURED | Present and past | The fact is recorded by a working instrument, a decision record or a verifiable external primary source | The instrument or source and the date of measurement; if a figure comes from a secondary review, the explicit note “per a secondary source” |
| BUILDING | Future | A mechanism is under construction or a series of observations is being collected | The acceptance check (what exactly must show “done”) and the planned deadline |
| CONCEPT | Conditional, design | A prospective concept with no acceptance check yet | The word “concept” stated plainly in the text |
| FORECAST | Scenario | An extrapolation years ahead | The scenario branch, today’s value and an external benchmark with a date by which it is checked whether the forecast came true |
How a forecast works: branches and benchmarks
A forecast without a branch and a benchmark is not accepted. The institute’s book unfolds the horizon to 2035 in three branches:
- Branch A, “normal technology”: capability growth slows, AI spreads over decades like electricity and the internet, institutions have time to adapt.
- Branch B, the trend continues (baseline): compute, the price of a token and the horizon of autonomous tasks keep changing at the current pace; agents take over execution, a person keeps the rules, the keys and the responsibility.
- Branch C, acceleration: AI automates its own research, and a technological leap compresses the decade.
Each indicator has a 2026 value with a source and an external benchmark: a publication and a date by which an independent observer checks which branch is coming true. The consolidated register is on the page of the forecast register to 2035 (in Russian), and the branches are discussed on the page on the 2027–2035 horizon.
“Not measured” is not “clean”
An indicator that is always green is broken in the same way as one that is always red. If an instrument is silent or there is no data, the honest answer is “not measured”, not “no violations”. This rule of truth marking applies to security reports as well: the absence of violation signals proves safety only when it is known that the instrument was working.
The rule applied to ourselves: what the instrument showed
The institute applied truth marking first of all to its own development. According to the book:
- after a third state, “not measured”, was introduced, a retrospective run over 98 days of its own development history showed 0 green days out of 98 (September 2026);
- on 9 October 2026, the product maturity instrument showed 1 of 6 core claims fully closed;
- the gate had no commercial revenue when the book was written (October 2026), and the sales channel calculations have the status of a testable hypothesis.
These figures are not hidden but placed at the start of the book: the boundary between what is proven and what is being built is what makes the text an engineering document rather than an advertisement.
How to introduce truth marking at your company
- Label every significant claim in reports and presentations about AI with one of the four labels.
- Match the verb tense to the label: “the gate blocks” only if it is measured; otherwise “will block”, with an acceptance check and a deadline.
- Give an instrument or source and a date for every measured figure; call secondary sources secondary.
- Give a forecast a branch and a benchmark: what exactly, and by what date, will show whether it came true.
- Keep a register of claims with the condition under which each is recalculated or retired (the book’s claims register, in Russian).
- Tell “not measured” from “clean” on any security dashboard.
Frequently asked questions
What is truth marking?
The rule that every significant claim carries one of four labels (MEASURED, BUILDING, CONCEPT or FORECAST) and that the verb tense and what accompanies the claim follow the label: an instrument and a date for a measured fact, an acceptance check and a deadline for something being built, a branch and a benchmark for a forecast.
How does “measured” differ from “building”?
“Measured” means the fact is already recorded by an instrument or a primary source with a date; it is written in the present or past tense. “Building” means the mechanism is under construction; it is written in the future tense, stating which check will confirm readiness and by what deadline.
What is a forecast benchmark?
An external publication and a date by which any observer can check whether the forecast came true and which scenario branch is unfolding. A forecast without a benchmark cannot be checked, so it is not accepted.
Why can “not measured” not be treated as “clean”?
Because a silent instrument does not prove the absence of violations: the instrument may not have been working. An indicator that never turns red is as useless as one that never turns green.
Where is truth marking used?
In the institute’s book “Proven Action” and on the pages of the IS 3.0 section: the stages of the product line on the Russian section page are labelled measured, building and concept with the date of measurement. The rule suits any report about AI, internal or for an investor.
How can the book’s claims be checked by their labels?
Through the claims register: each row has a label, an instrument or source, a date and a recalculation condition. The label counts by chapter are on the page of the book’s claims register (in Russian).
Sources
- The institute’s book “Proven Action. Power and Control over AI Agents: Strategy 2027–2035”, introduction, section “Truth marking” (in Russian): https://isslab.ru/en/book/
- Forecast register to 2035 (in Russian): https://isslab.ru/kniga/reestr-prognozov/
Checked: 18 October 2026, against the text of the book.