This is the maintainer’s map, a personal project. The deep argument map covers one lens, the market and economics of AI; the Signal Board files developments under six audience lenses. The confidence levels are one person’s judgment, not a consensus or a measurement.
Collection is automatic; judgment is not. Weekday crons run the collection engines (the news scan, the Signal Board’s discovery pipeline, company intel, arXiv research, and a weekly tooling scan), and the Daily Edition is written from what they stored, on schedule, with no one present. The argument map is different: no confidence moves without the maintainer writing a reason, so the map reflects what the maintainer has judged and can lag the feed by days or weeks.
The recommend-only AI works from the text it is handed: a grounded call is auditable and its inputs are on the record. The live web reaches the system in two ways. The collection engines’ search legs run mostly on Tavily, a model-free search API, beside RSS feeds and primary sources (SEC EDGAR, FDIC, CFPB, arXiv, GitHub, Hacker News). Anthropic’s web_search, which lets the model search for itself, runs in a few bounded features: Scout’s company discovery, competitor scan, and intel sweeps, the Tooling Monitor’s deep dives and its big-pull category enumeration, and the Ask workspace’s web toggle, which lets an answer fill gaps the records leave, with the sources listed and no faithfulness guarantee on the web half; the collection engines fall back to it only when their model-free providers are not configured. With that toggle off, deep research searches the Atlas corpus, with one exception: a fresh question the records do not cover turns web search on for that turn, and the answer says so. A dossier or a draft is only as good as the text behind it and can be wrong or out of date.
Every item the engines collect keeps the text behind it: articles, filings, papers, and the text extracted from uploaded documents, retained with no retention limit today. That text grounds the AI’s answers and reports. Public visitors get the summary, the finding, and a link to the original. Readers with an access key can open the retained text beside an answer to check a quote and download it in the key-gated datasets; it is not republished on the public pages. Some outlets block fetching entirely, so their items carry summaries and links rather than text.
Some dynamics here feed on themselves, like money flowing into something because people expect it to pay off. Those claims get flagged, but a static map cannot fully capture a feedback loop.
The guardrails reduce the obvious problems, they do not remove them. Confidence levels can carry the maintainer’s bias. Stances may be argued unevenly. Evidence can stay one-sided if the looking stops. Claims can go stale between updates. The scout’s company facts come from the web and can be thin or wrong until enriched and reviewed. The Daily Edition, the scheduled reports, and the engines’ summaries and tags are model-written with no human review step, so they can misread a source. The structure pushes against all of this; it guarantees none of it.
The Daily Edition, the research roundup, Savant, the weekly tooling entrants report and the company intel deck publish on a schedule without a person reading them first. Every one passes a citation gate (a link that does not resolve to a record is removed) and Savant also passes an editor model, but a gate strips a bad link; it does not repair a bad argument.
Ask finds records by their words and by their meaning. On a set of forty test questions, the right record lands in the top ten about 86 percent of the time, so roughly one question in seven misses something it should have found. An answer that says the Atlas does not track a topic can be wrong.
When a report passage is shown to someone without an access key, sentences naming a tracked company are removed by matching the registry’s names. An alias, an abbreviation or a misspelling gets through.
The usual time and cost shown on a run come from past runs of the same kind. A first run, a slow provider, or an unusual week can run long or cost more.