Methodology

Fully AI-automated. No human reviews items before they’re published. Automated collection and checks help select and present material; they can miss errors. This is aggregation and synthesis, not a claim of independent reporting.

Automated scripts collect sources. The news editor uses an OpenAI Luna-class language model via ChatGPT Codex in the cloud for bounded ranking and summaries. Source selection, duplicate handling, evidence rules, and publication checks also use deterministic code.

Selection and source labels

Articles come from public RSS feeds, publisher pages, and official sources. Selection considers source provenance, recency, concrete consequences, duplicate coverage, and topic fit. A publisher's original headline stays attributed; a summary is separate AI-written or source-excerpted copy.

Publisher labels such as sponsored, partner content, paid post, and advertorial are checked in available feed metadata and URL paths. Detected material is excluded from the lead and Need-to-Know. If retained elsewhere, it is marked Sponsored (publisher label), including in RSS. Detection is incomplete when a publisher omits its label.

Safety / recalls requires a concrete safety, hazard, recall, or public-alert basis. Culture and entertainment stories do not belong there merely because a feed or model assigned that category.

Access labels describe the source: OFFICIAL means an institutional source, OPEN means an accessible publisher link, and ANALYSIS identifies commentary. None of those labels alone certifies accuracy. A company statement is evidence of what the company says, not independent confirmation of its results.

Top stories two-gate rule

Substrate Wire covers AI: compute, damage, work, and AI’s impact across the world. Public news must be substantially about AI, not a passing mention. The existing automated editor classifies stories; unknown or invalid classifications are withheld. AI impact domains require at least two distinct source links published within seven days. Thin domains are hidden, and empty top-story slots are never filled with general news. The internet-disruption context on the security tracker is labeled not AI-specific.

Top stories groups reporting about the same event, then selects up to five useful news stories. The homepage lead and RSS use the same ordered edition.

Gate 1 - evidence and exclusions

Corroborated · N outlets means matching coverage from distinct parent organizations, identified automatically from source headlines and summaries within a 72-hour window. Syndicated copies count once where identifiable. Clustering can miss a match or group related angles; coverage breadth is not independent verification of every claim.

Single source means one attributable publisher or institution. A PRIMARY SOURCE establishes what that institution reports, not independent confirmation. A social-media candidate never uses the single-source path. Tutorials, sponsored material, retrospectives, opinion-only pieces, routine recalls and local Anza stories cannot enter the top stories. Entertainment requires broad coverage and a grounded major-consequence review.

Gate 2: importance and a source-bound explanation

This edition, selected Oct 24, 2026, 10:05 PM PT, was ranked by the site's fixed rules. The lead is from the last 24 hours when one qualifies, and the other top slots use stories from the last 48 hours when enough qualify. Older stories fill any remaining slots, and all other ranked stories stay in More stories.

GPT-6 Sol ranks eligible AI stories by significance for AI’s trajectory and people’s lives, real-world impact, freshness, corroboration, and diversity across domains. Reports about the same incident share one slot. If the AI ranking is unavailable, invalid, or reaches its daily call limit, the existing deterministic ranking is used.

MATERIAL CONSEQUENCE guides selection. Before AI ranking, candidates must meet the existing quality floor of 60 and source checks.

Qualification and same-incident grouping run before AI ranking, on at most 30 eligible stories published within 72 hours. The AI selects a lead and up to four Stack stories, and can rank up to six additional candidates. Internal reasons must quote the supplied candidate text and are not published. Identical candidate pools reuse a cached result within a Pacific day. Code checks exact story ages on every edition and fills missing recent picks using the fixed rules. Remaining ranked stories stay in More stories. Source qualification and publication checks still apply; an AI rank is not a fact-check.

Dated tracker changes can compete alongside news: compute changes stay labeled estimates, connectivity measurements do not establish a cause, confirmed attack reports retain their source scope, and AI Work Shift does not infer AI causation from layoffs alone. Refreshing a tracker snapshot does not make an old event new.

Legacy editions without an event bundle retain the earlier evidence and consequence selector. AI-only: items are published automatically. Privacy, provenance, domain boundaries and publication guards run before release. Recall notices stay in Official Need-to-Know; the collapsed Unranked reports drawer remains available below.

Automated evidence decisions

Machine-checkable predictions use deterministic rules. Other eligible claims use bounded cloud AI checks against fetched source text. “AI-verified against the source’s own words” means a model selected supporting sentences and code checked that each quote occurs verbatim in the fetched text. It does not mean independent verification. Provider disclosures count only when the provider states threat actors used AI; an actual attack requires explicit evidence of AI use in that attack and a second model check. Unavailable or uncertain evidence remains unconfirmed, or expired unresolved after a prediction’s deadline. Original records and an append-only decision history are preserved.

When enabled, an AI model writes race headlines once per Pacific day from published scores and verified releases. Code checks each headline against those facts and the phone layout limit. A failed check, changed facts or an unavailable call uses fixed wording. Scores from different indices are never combined.

What the uncertainty labels mean

• Confirmed is scoped to the claim and evidence shown. A confirmed publisher report establishes that the source made the statement; it does not establish independent verification of every detail or worldwide completeness.
• Suspected means the source suggests a possibility or attribution that is not established. Keep hedges visible.
• Estimate means a derived or reported approximation. Read the units, method, source dates, and coverage caveats before comparing values.
• Unknown / unavailable means the evidence is missing, stale, incomplete, or insufficient. A blank metric is not a measured zero.

Reading the trackers

Compute Intel compares dated capacity estimates. H100-equivalents are a normalized measure, not a count of identical physical chips. Ownership, rental access, fleet size, and single-cluster size are different questions; the page identifies its sources and periods.

Damage Report lists sourced AI incidents in five exclusive categories. Each item separates event and report dates, names its evidence status, and links its sources. Unknown event dates stay in history. Connectivity disruptions are separate and do not establish AI involvement. No forecasts or speculative assessments appear on this incident page.

AI Work Shift tracks employer statements, filings, reporting, and background data. A keyword match or a company mentioning AI alongside layoffs does not establish that AI caused a particular job loss.

The prediction record retains timestamped claims and resolved, unresolved, or excluded outcomes. Pending entries are not hits. Accuracy remains withheld where the eligibility and resolution evidence does not support a rate.

Dates, summaries, and RSS

Publisher publication time, first collection, last check, and page rebuild are separate dates. Explicit source offsets are preserved. An unknown timezone is shown as unspecified; a future source time is shown as an absolute date rather than “now.” A rebuild does not make an old story new.

Summaries keep complete sentences and omit incomplete tails. If no complete source-supported sentence is available, a headline-only entry is preferable to an invented summary.

RSS uses the same protected lead selection as the homepage. Items link to an explicitly associated public Substrate destination where one exists; otherwise they link to the original publisher. Original-source attribution and stable source identifiers remain in the feed. An unrelated tracker is not substituted just because its topic matches.

Corrections and limits

Report a mistake to [email protected]. The corrections policy describes the evidence to include and how material errors should be corrected or withdrawn. There is no promise of a response within a set time.

Every item links to its source; AI checks can miss errors - report corrections here.