communities.news
Eight floating moss-and-rock platforms arranged across an Alaska late-spring breakup landscape, each holding one stage of the editorial pipeline (microphone, signpost, pond, paper-tag post, books, writing desk, scale, newspaper). Glowing gold arcs labeled I–VII connect each platform to the next.

How it works

The Community News platform turns a public meeting into a published article through a sequence of well-defined stages, each with its own constraints and verification gates. The pipeline is designed so that a published article is structurally incapable of citing a quote or fact that was not verified before drafting began.

This page describes the architecture in plain language. It is the technical brief, not the technical documentation.

From meeting to article

A typical Alaska News story starts with a borough assembly, school board, or city council meeting. The meeting is recorded, often as a public livestream or a posted video. The platform's pipeline processes it through eight named stages.

1. Source ingestion. The platform claims a public meeting from a video URL, an uploaded file, or pasted text. Each source is associated with a community (today: Alaska News) so coverage stays scoped to where the journalism lives.

2. Audio transcription. Volunteer computers running the compute-volunteer app process the audio. A volunteer in Eagle River who has signaled "use my idle Mac for community news" donates roughly an hour of compute to transcribe a four-hour assembly meeting. The transcription is per-chunk, so a deploy or a network blip does not cost the whole pass.

3. Speaker identification. A vision-and-language model reads the video frames alongside the transcript to resolve generic diarization labels (Speaker A, Speaker B) into real names. Identifications are gated against a community speaker registry: a label is rewritten only if the resolver matches it unambiguously to a registered person. Otherwise the chunk keeps its generic label and downstream stages refer to "an assembly member" rather than guess. This rule is the structural reason the platform does not put fabricated names on the page.

4. Tagging and concept extraction. Topics, locations, and the conceptual threads of the meeting are extracted. The platform also auto-discovers cited URLs in the source material and pulls the linked documents (bills, fiscal notes, decisions) so the article has access to primary records, not just the meeting transcript.

5. Evidence gathering. Per concept, the platform builds a quote pool (every quote eligible to appear in the article, validated for substring presence in the source) and a fact pool (every external fact eligible to be cited, validated against published records). This is the editorial constraint: drafting is allowed to use this pool, and only this pool.

6. Drafting. A language model writes the article against the quote and fact pools as hard constraints. The drafter is prompted that every named person must appear in the quote pool, and every external claim must trace to the fact pool. A post-draft validator drops any attribution that cannot be traced.

7. Verification. A deterministic stage, no language model involved, that re-validates every attribution and scans the article body for unbacked names or external citations. Any failure flips a verification flag the editor sees in review.

8. Editorial review. A human editor reviews the article before publication. The editor sees the source meeting, the verified evidence pool, any verification flags, and an AI review of the article itself. The editor decides whether to publish.

The article ships only after step 8.

What the architecture prevents

The pipeline structure makes specific failure modes harder to reach than they would be in a single-pass "ask GPT to summarize this meeting" approach.

These are not aspirations. They are the boundary conditions of the production pipeline running today on alaskanews.news.

Compute volunteers

The platform processes audio and video on volunteer computers, not in a centralized data center. The volunteer's Electron app polls a job queue, claims a meeting, transcribes it on their local machine, and uploads the result. A volunteer in Wasilla can choose to specifically support Mat-Su coverage; the platform routes meetings near that geography to that volunteer first.

This is a deliberate cost-structure choice. Centralized GPU compute would close the math at a different shape: cheaper per minute, but it is also a single point of failure and a recurring bill that scales with coverage. Distributed volunteer compute scales with community engagement, has no recurring cost, and turns the act of supporting local journalism into something a person can actually do without writing a check.

What the AI does and does not do

The platform's language models are constrained tools. They transcribe, classify, draft against pools, and produce summary captions for embedded media. They do not decide what to cover, who is credible, when a story is fair, or whether a draft should ship. Those decisions are reserved for human editors, contributors, and the readers whose civic life the platform is meant to serve.

Cronkite, the Walters who came after him, the Bernsteins and the Woodwards, did not stop being journalists because they used recording equipment, telephones, computers, the internet.

The tools changed; the standards did not.

The Community News platform is a tool. The standards are journalism's.

Read more