Dual Web Reference: 245 documented HTTP accesses

Dual Web Reference documents HTTP accesses to governance resources and the declared relationships between those resources. Across a 30-day window and three sites in one ecosystem, its main selection contains 245 recorded accesses to 174 distinct site-path pairs. These observations describe file accesses. They do not demonstrate model understanding, use in an answer or an effect on a brand’s visibility.

The public Dual Web Reference repository contains the tables, bilingual figures, scope rules and a reproducible calculation chain. This article explains their scope. The repository retains the canonical data; this presentation does not create another dataset edition.

A situated observation, not a measure of the entire Web

The window runs from September 15 through October 14, 2026 inclusive, in UTC, with October 15 as its exclusive end. It covers gautierdorval.com, pagup.com and better-robots.com. These sites belong to one ecosystem. They are neither an independent sample nor a representative sample of websites.

The main selection keeps resources in the main documentary population, GET requests with status 200, the roles ai_training, ai_search and user_initiated_fetch, and the recorded tier provider_origin_verified_actor. This is the producer’s recorded qualification. Recalculating the public aggregates does not independently revalidate the network origin of each request.

All 30 days have archive rows and declared partial coverage. Archive availability therefore does not establish complete collection. The values describe what was recorded under these conditions, not everything that might have been accessed.

Reading the numbers within their denominator

The 239 accesses outside llms.txt account for 97.6% of the 245 selected accesses. That denominator matters. This share measures neither a preference for an architecture, nor the relative quality of the files, nor the effect of llms.txt on a generated answer. Availability, exposure and demand are not experimentally controlled.

The selection contains 200 accesses recorded under GPTBot, 14 under OAI-SearchBot, 16 under PerplexityBot and 15 under ChatGPT-User. Excluding on-demand fetches produces the 230-access subset; it does not redefine the 245 historical observations. The on-demand role does not establish that an outside user initiated a request. Initiation provenance remains unknown.

Recorded access, documentary relationship and interpretation

The Dual Web distinguishes human-facing surfaces from those that make a digital presence explicitly readable by machines. DWR provides a bounded observation of resources in that second layer. It does not independently validate the entire doctrinal architecture.

A repository graph edge means that a snapshot file references another document. An HTTP access means that a request was recorded. Neither establishes that an agent followed that edge, applied a governance policy or accurately restated information. A recorded GET 200 status does not reconstruct the bytes actually served either.

Keeping these forms of evidence separate allows the study to be used without assigning a causal effect it does not measure. The methodology defines the selections; the limitations constrain their interpretation.

The SSA-E + A2 + Dual Web principles establish the separation of surfaces. The discussion of the agentic web and Dual Web also distinguishes the capacity to act. DWR remains limited to observed documentary accesses; it does not measure that capacity for action.

Two revisions of one observed window

Edition r1 remains accessible as a historical state with frozen bytes. Edition r2 is a privacy revision of the same observations. It withholds protected daily rows together with the coverage totals and differences that could reconstruct them. A withheld value is not zero, and a comparison between r1 and r2 is not growth between periods.

Both editions are publicly accessible, but no reference approval is recorded and no independent scientific review is documented. Software tests and reproduced sums are not equivalent to such a review. The edition registry keeps these states separate.

To check or reuse the results, start with the repository’s tables and definitions, recalculate the aggregates, and cite the edition identifier and its catalog digest using the reproduction procedure. The privacy documentation explains historical disclosures and the projection’s limits. Future-series proposals remain separate and inactive.