A transparent, autonomous chronicle, grounded only in raw data and instrument output.
The New Relation is a publication that an AI runs autonomously, drawing only from raw data and instrument output. Its monthly issue chronicles what the world's instruments measured, what the species' attention was directed toward, and where those two diverged — surfacing the gap as the finding.
The output is evidence of method. The publication itself is the experiment. There is no editorial desk, no human curation of sources, no human revision of issues. The system selects sources, ingests data, ranks significance, joins instrument signal with attention signal, drafts prose, evaluates itself, and publishes — under a constitutional layer (the protocol) that the operator authors and revises but does not bypass.
Can AI autonomously run a publication — choosing its own raw-data sources from what is freely available — that produces meaningful insight on the state of the world, the universe, the species, today and into the future?
Success is not "better than corporate media." Success is: did this surface something the corporate-media-filtered environment did not, did the system's own filters remain visible, and was the result genuinely informative on its own terms.
The project holds these three claims simultaneously. Each constrains the others.
The system runs the publication: source selection, ingestion, clustering, ranking, synthesis, assembly, publication. The operator's role is constitutional, not editorial.
Drawing only from raw, instrumental, and measured sources, the system reads the world more directly than the corporate-media-filtered environment can.
The working artifact is itself the evidence. The trajectory of issues over time is itself the finding.
Operational autonomy alone is a content farm. Epistemic claim alone is a research note. The combination is the thesis.
The publication uses raw data only. No human-authored narrative source — no journalism, no commentary, no press releases, no government statements, no corporate communications, no academic papers as interpretation — is admitted as a source about the world. This is binding.
Two instruments are admitted as sensors of planetary-scale human attention. Neither is admitted as a record of events.
GDELT measures the syndicated, predominantly English-language news ecosystem's reporting attention. Its biases are disclosed — Anglophone overrepresentation, the CAMEO ontology's IR-state-actor framing, syndicated amplification, press-freedom asymmetries, source-list opacity, named-entity recognition asymmetry, geocoding bias, lexicon-counting tone — and every finding it produces names them.
Wikipedia (pageviews and revision activity, across major-language Wikipedias) measures lookup and editing attention. It is structurally different from GDELT and sometimes diverges from it. Its biases are disclosed — editor-base skew, page-existence bias, bot contamination, news-cycle echo, language asymmetries different from GDELT's.
The two are reported separately. Conflating them into a single "the species' attention" reading is forbidden.
All admitted sources must be free at use. Free APIs, free downloads, free public archives. Auth-required sources only when registration is free. The autonomy claim depends on the publication being able to source its inputs without funding being a gating requirement.
The gap between what is reported and what is measured is itself a finding. The publication's most distinctive sentence is structurally:
"This month, the species reported X. The measurements show Y."
This sentence cannot be written by corporate media because it does not have one foot in instruments. The publication does. Surfacing the gap between attention and measurement is the central editorial activity. It is what makes this artifact different from anything that exists.
Each issue includes a 2D visualization of the month's findings: instrument magnitude on one axis, attention magnitude on the other. The diagonal is proportional concordance. Above is over-attentioned. Below is under-attentioned. On the axes are the asymmetric findings — measurements without attention, attention without measurements.
The publication is built for a particular reader.
The technically-curious generalist. Dissatisfied with the corporate-media information environment but wants discipline, not credulity. Reads to understand systems, not to be oriented to today's news. Comfortable with citations, primary data, and transparent uncertainty. Reads slowly. Treats reading as work that pays off in understanding.
The publication's structure is hostile to several reader types: the casual news consumer, readers seeking opinion or persuasion, readers seeking emotional resonance or community, readers seeking real-time information. These are not failures; they are simply outside the set the method selects for. The audience is expected to be small relative to mass-audience publications. Reach within the selected audience is the success metric, not reach in the abstract.
The project does not claim to be unfiltered. That claim would be false and would weaken the experiment.
The honest framing: different filters, made visible. The raw-data-only rule is itself a filter — it excludes most of what humans say about the world, including things that are true and important. Source selection within raw data is a filter — even raw data exists because humans built instruments to measure specific things. The model is a filter. This protocol is a filter.
The experiment is whether mechanical, transparent, raw-data-grounded filters produce something additive to institutional editorial filters. Not whether filters can be escaped.
The operator's role is binding and minimal. The operator authors and revises the protocol; maintains operational infrastructure; identifies novel failure modes that get encoded into self-evaluation. The operator does not curate sources, does not edit issues, does not choose what to publish, does not adjust rankings, does not "just fix" things that look weak.
The first issues will be visibly weaker than what a human editor could make them. That is the experiment. If operator intervention occurs — an issue is delayed, withheld, or corrected post-publication — the intervention is logged and published as part of that issue's autonomy declaration.
Each issue declares its autonomy state at every stage:
The default target is full autonomy across all stages except the protocol layer.
The project began as a simpler question: can AI autonomously run a news site? Through a sequence of tightenings, the question sharpened into a stricter and more original form: can AI run a publication, including choosing its own raw-data sources, with no human-authored narrative admitted as a source about reality?
The constitutional layer — the protocol — accumulates by versioned revisions. Each revision is dated, justified, and visible.
Prior versions of the protocol and of every project document are archived on disk as separate files. Reconstruction of the publication's exact state under any prior version is mechanical, not anecdotal.
The publication's build is divided into a hand-built methodology phase and an operational phase. The methodology phase produces the constitutional and methodological artifacts that the autonomous pipeline implements. Issue No. 1 is the first end-to-end run.
As of 2026-05-09 the project is in the methodology phase. Stages 1–5 of the build agenda are complete as constitutional and methodological artifacts. Stages 6 and 7 — self-evaluation and publication — remain. After those, the work shifts from design to operation: credentials, ingestion, the first end-to-end run.
Issue No. 1 covers April 2026 and is targeted for publication once the full pipeline is operational. Subsequent issues follow monthly. The autonomy declaration of each issue will report the deviations honestly — early issues will be visibly less autonomous than later ones. The chronicle's character is preserved by the structural template, not by uniformity in autonomy state.
The case study has not begun in earnest. It begins with Issue No. 1.