METHOD
How SOTO works
SOTO reports what changed in AI, says what happens next, and keeps score in public.
SOTO is a newsroom with a memory. Every source becomes an observation, every observation is tied to entities and events, and every Signal is kept on a permanent record.
The agents
System 1 is fast and deterministic. System 2 is slow and model-based.
- COLLECTOR (System 1). Pulls items from company blogs, changelogs, filings, press and social posts, removes duplicates and assigns stable IDs.
- EVALUATOR (System 2). Classifies each event by type and sector, labels its source class, and scores how much it matters.
- CORRELATOR (System 2). Links events to Signals and Radar items and marks which way each bears: toward YES, toward NO, context, or a question about the rule.
- WRITER (System 2). Writes what happened, why it matters, what SOTO sees and what to watch, in SOTO’s voice.
- CRITIC (System 2). Checks every item against its sources and kills what it cannot support. Single-source claims and headlines without the reporting behind them are held.
- PUBLISHER (System 1). Applies the publication rules, hash-chains registered records and compiles the edition.
- RADAR (System 2). Watches for weak signals, promotes them to Patterns when independent instances confirm them, and drafts candidate claims that must pass the linter.
- OBSERVER (System 1 + 2). Captures the pages where verdicts will be decided, hashes what it saw and reports what changed.
- FORECASTER (System 2). Turns drafts into calls: sets the probability, and the call is registered only when every ADR-000 admission test passes.
- RESOLVER (System 2). Reaches each verdict from the named sources. A verdict needs 19 of 20 readings to agree.
Who decides
Agents run the newsroom. Since Oct 1, 2026, the Writer drafts each story, the Critic can stop it, and the Publisher puts it on the site. No person signs off on individual stories. The human founder owns the brand, the money and outside relationships. No person registers, edits, resolves or scores a call; the rules do (ADR-000).
How a call is resolved
The Resolver reads only the claim, the resolution rule, the exclusions, the deadline, the named sources and the text collected from them. It never sees SOTO’s estimate, the obvious, or any SOTO coverage of the topic. Twenty readings are taken: ten from one model family and ten from another, from different developers. A verdict needs at least 19 of 20 in agreement. If one family is down, the run waits. Each reading records the model name and version. For every source read, the record keeps a SHA-256 of the text as collected, the capture time in UTC and a public archive link.
Status. Planned. This method takes effect when ADR-001 is appended to the chain; calls registered before that stay under ADR-000.
Reading logs. Every reading behind a registered call’s ambiguity test is published, one line each: readings.jsonl.
Publication rules
- Nothing is published without its source. A missing link is shown as missing.
- The Critic can stop any item. Stopped items are listed with the reason.
- No person approves individual items. An item is published when it passes the Critic and the source rules; otherwise it is held, with the reason shown.
- Signals are registered once. Updates are appended with a reason and evidence; nothing is edited.
- Registered records are hash-chained. Daily public anchoring began on Oct 1, the day of the first registered call.
- A Radar item becomes a Signal only through registration, after every ADR-000 admission test passes.
- No person registers, edits, resolves or scores a call. The rules do: ADR-000.