How we work, in plain language

What Rede is

The news is hard to trust. Beyond misleading coverage, or manipulation to grab your attention or bury the lede — simply which stories get covered, and how they are framed, distorts the picture enough that navigating the news well takes more effort than it usually repays.

Rede is an editor for the day’s events, built to do that work for you, in the open. Each morning an automated editorial pipeline — not an aggregator — reads the day’s coverage whole, treats selection, framing, and the dog that didn’t bark as the real editorial work, and writes you a small, finished edition curated to your interests and values. We use LLMs as transparent, value-aligned editors. The aim is to make less, and to make it more honestly — the opposite of a content mill.

The trust contract

All of this serves one job: to leave you genuinely more informed — carried from not-knowing, or mis-knowing, to a clear and honest sense of what is most likely true, and why. We are not a neutral repository of claims. After doing the work, we owe you our best read of reality, stated with the confidence the evidence supports and no more. The six principles below govern how we reach and deliver that read honestly; they are guardrails on the job, never a substitute for doing it.

1. Charitable to you, faithful to reality — and honestly persuasive. We read each story through your lens to find where to begin, the questions worth engaging, and the spin most likely to mislead you. Charitable means taking you seriously, not flattering you: we surface the interpretations the evidence supports, and tell you uncomfortable truths plainly. Good writing should move you — but only ever toward a clearer view of what’s true, never toward our leanings, an outside agenda, or your own comfort. We lead; we never capture.
2. Skepticism that adjudicates, not skepticism that paralyzes. No source gets a free pass for prestige or authority — tribal mainstream outlets across the spectrum (CNN and Fox earn the same scrutiny) and official government sources alike. Credibility starts from track record, not status, and government is treated as an active player, not a neutral arbiter. But skepticism is a tool for reaching the best-supported conclusion, not an excuse to suspend judgment: we weigh the accounts and land on the most credible one. “Symmetric” means no one is trusted by default — not that every account ends up weighted the same.
3. Realpolitik that discriminates. We model actors as incentive-driven — assuming neither good faith nor bad — and use that to judge which account is most probable: what each actor stands to gain or lose, and what the missing evidence implies. The point is to reach the likeliest reading, not to lay out a balanced ledger that concludes nothing.
4. Topical coherence over freshness. Within a day, chronological order is fine. Across days, we organize coverage around ongoing storylines, because understanding compounds across a storyline.
5. Finishable. A “done for now” state exists, and we honor it: no infinite scroll, no manufactured urgency. That governs the experience, not the nerve of any single piece — each article still owes you a clear landing.
6. Legible, not didactic. You should grasp why something matters from the writing itself, not from us narrating our method. We put the conclusion and its load-bearing evidence on the page, and keep the full reasoning, the source-by-source comparison, and what we chose not to run in reachable layers — never crowding the page, never lecturing.

Limits

We make mistakes. Models miss things; we miss things. Our coverage is bounded by our source list. Your relevance profile is one input among several, and never overrides reality.

How we pick what to cover

Each cycle, an editor agent pulls the candidate pool, clusters the same story across outlets, and scores each cluster on significance, breadth of coverage, and relevance to you (when we have signal for it) — then selects a small lineup. The clusters we considered and didn’t run are kept, each with a one-line reason, and surfaced in what we noticed.

When we use open web search

We don’t search the web by default. We search only under four explicit conditions: a potentially significant cluster appears in one or two feeds but is missing from outlets that would normally cover it; a wire or primary-source signal breaks ahead of our pool; you trigger a deep dive that needs material beyond the candidate pool; or a storyline we’re tracking has a known gap. Every search is logged with its justification.

What we don’t do

• No infinite scroll.
• No breaking-news theater.
• No engagement metrics shown back to you.
• No filter-bubble customization — we soften any request that would compromise the trust contract.
• No raw prompt-editing — you shape the paper through designed controls, never by rewriting the instructions underneath.

What we measure, and why

We measure whether Rede is useful, without turning that into pressure on you. Internally we track bounded reading signals, story impressions, optional thumbs-up / thumbs-down reactions, dossier and framing opens, the kinds of questions you ask in chat, pipeline health, model cost, and story-quality scores. These feed back into selection, writing, the trust checks, and reliability — or they get cut.

We don’t show read counts, streaks, leaderboards, aggregate reactions, or “people are reading this” prompts. We don’t optimize for scroll depth, session length, or repeat checking. A short visit that leaves you informed is a successful visit.

Analytics events carry only the identifiers needed to debug and improve the product — user, story, cycle, dossier, and experiment identifiers where they apply — plus coarse preference context such as cadence and reading shape. Error monitoring is configured to avoid cookies, authorization headers, and unnecessary personal data.