How ArkTelos Lab separates a conclusion from its evidence

Why a separate method is necessary

An engineering article may look convincing and still leave its reader with no way to determine which part is an observation, which part is an interpretation, and which result depends on an environment that no longer exists. ArkTelos Lab treats that loss of context as a publication defect rather than a stylistic detail.

The method therefore starts before an article enters the site. The article, its English and Russian editions, the claim ledger, the evidence ledger, declared assets, and experiment descriptors form one versioned package. A material change creates a new Git commit and a new immutable Lab revision.

The Evidence Signature

Every public record carries an Evidence Signature derived from its immutable revision. It identifies the ATL record, declared type and maturity, verification date, languages, claims, public sources, and reproducible records. The signature is not a score and does not claim that every conclusion is universally true. It makes the actual boundary of the published result visible.

A claim may be supported, partially supported, unsupported, or inconclusive. Partial and inconclusive claims require an explicit limitation. Evidence that has not passed verification cannot silently become proof because it appears next to a confident paragraph.

Reproducibility has a concrete meaning

When a conclusion depends on code, a benchmark, or an agent run, the public record includes the exact repository revision, environment, preparation and execution commands, input and expected-output references, observed result digest, repetition count, and known limitations.

An AI or agent experiment adds the provider, model and version, relevant parameters, tool set, prompt and input digests, evaluation method, observed variance, and repetitions. These fields do not eliminate stochastic behaviour. They define what another engineer needs in order to examine the same run instead of reproducing only its marketing description.

Two languages are two edited editions

English and Russian are reviewed as separate engineering texts. Literal translation is not accepted as evidence of parity. The review compares factual claims, limitations, maturity, commitments, and the meaning of conclusions. Both linguistic gates and the material parity gate must pass before the pair can be published.

Maturity describes the result, not the ambition

• Exploratory — the question and initial observation are useful, but the method or evidence is still incomplete.
• Experimental — the method is declared and the result can be examined, while its applicability remains narrow or unstable.
• Repeatable — the declared procedure has produced a consistent result across the recorded repetitions and environment.
• Stable — the result has a maintained scope, sufficient evidence, and a correction policy appropriate for continued reference.
• Superseded — a newer record or method has replaced the result without erasing its history.

Corrections preserve history

Published Markdown is never silently edited inside the CMS. A material correction is imported from a new private content commit, passes the complete review again, and atomically replaces both language editions. The private repository identity is not exposed to readers; the previous publication revision remains in internal history. A removed record returns an explicit archival statement and, when possible, a maintained replacement instead of becoming an unexplained 404.

What the method does not promise

The presence of an Evidence Signature does not make an experiment universal, a package production-ready, or an agent deterministic. It promises something narrower and more useful: the public conclusion remains attached to the exact material, environment, verification state, and limitations that were accepted for publication.