About & method.

Lantern is what I'd rather have read instead. One page, already filtered. An agent that read the week of research for me, dropped the announcements and the disguised pitches, scored the rest on what matters — relevance, evidence quality, actionability — and handed me a short selection with honest summaries. No marketing, no rephrased teaser.

I build it solo, evenings and weekends, alongside the workshop where I put these systems into service for teams. It's an author project, not a startup. That means three things:

• No ads, no trackers. Lantern doesn't follow you, doesn't sell your data, and loads zero third-party script.
• No growth hacking. No push notif, no gamification, no "streak". You read when you want.
• Explicit curation. Sources are public, scoring is documented below, and the AI never reorders your reading: the feed is chronological by default, score sort is opt-in.

Every article that lands in your feed traversed five steps. Most take a few seconds; LLM scoring is the bottleneck. The pipeline runs once a day — a chosen rhythm, not a technical limit: the ~10 active sources don't publish around the clock, one daily pass is plenty.

+-- 01 -----+    +-- 02 ------+    +-- 03 ------+    +-- 04 -----+    +-- 05 -----+
|  FETCH    | -> |  EXTRACT   | -> |  DEDUPE    | -> |  SCORE    | -> |  PUBLISH  |
+-----------+    +------------+    +------------+    +-----------+    +-----------+
rss-parser       readability       simhash 64b       gpt-5-nano     ranking + feed

What follows details each step — what it gets, what it produces, what can fail.

Scoring is the only place where AI decides something on your behalf. To keep it honest, here's the exact system prompt and the grid it applies. And you can always ignore the score: the feed is chronological by default — score sort and the min-score filter are opt-in.

The grid

• 5 ★ — Core target. The author shows what they built, what broke, or how to reproduce it — directly transferable to a system in service.
• 4 ★ — Very good. System actually deployed, operating numbers (cost per run, latency, failure rate) or an explicit eval method.
• 3 ★ — Solid but adjacent. A release, protocol, or library described with enough technical substance to be judged.
• 2 ★ — Marginal or promotional. Model announcement or benchmark, prompt listicle, vendor feature dressed as a guide — even excellent ones.
• 1 ★ — Noise. No AI involved at all, claim without evidence, AI-generated prose.

Publish bar: the final score is the weighted composite of the three axes (0.4·relevance + 0.3·rigor + 0.3·actionability, 1.0–5.0 scale). An article passes when that composite is ≥ 3.5 / 5. Below that, it stays in the database (audit, calibration) but doesn't reach the public feed.

The system prompt (excerpt)

The prompt runs in French (the scoring model reads both, the rubric is authored in French). Translated here for reference:

# role: AI systems research evaluator
You are a research evaluator for a technical practitioner who puts
AI systems into production and has to operate them afterwards
(agents, workflow automation, memory and context, permissions and
guardrails, evals, integrations). You output a structured score,
not a marketing recap.

# 3 axes, integers 1-5
- RELEVANCE       Does it help BUILD or OPERATE a running AI system?
                1 = off-topic · 5 = core target AND directly
                transferable to a real system.
                GATE 1 — AI PREREQUISITE: purely technical content
                with NO AI (database, infra, network, front-end,
                server perf) is off-topic → 1, however good it is.
                GATE 2 — BUILD OR OPERATE: even with AI, a model
                announcement/review/benchmark, a prompt listicle,
                a vendor feature, or generic consumer AI use
                → 1-2, EVEN excellent ones, EVEN from a reference
                source. When in doubt, if the reader walks away
                with nothing to implement → 2 max.
- RIGOR           Thesis grounded in real implementation (code,
                measurements, operating reports, edge cases hit) —
                or asserted? A purely promotional teaser OR a post
                where a vendor pitches its own product with
                unverifiable internal numbers caps at 2, even from
                a reputable source.
- ACTIONABILITY   Applicable to a real system this week?
                1 = theoretical · 5 = ready-to-use code, config,
                or procedure.

# adjustments (the LLM applies them itself)
- anti-patterns (-1): empty listicle, disguised vendor content, toy
demo passed off as a production system, gain claimed without
measurement, AI boilerplate prose.
- bonus (+1): real operating report (what broke, over what period),
operating numbers, code/architecture shown, explicit treatment of
permissions or failure behaviour.

Strict JSON output: { depth, novelty, applicability, type,
tags, summary, reasons }

The three axes are weighted in code (0.4 relevance · 0.3 rigor · 0.3 actionability) into the final score — they're stored under their legacy JSON keys depth / novelty / applicability, which is why the prompt output above uses those names. The prompt is versioned and frozen to stay comparable week to week; when we bump it (rare), it's noted in the changelog.

What I don't do

• No per-user personalization. Everyone sees the same editorial score. Lantern doesn't learn from your clicks.
• No A/B testing on titles. The displayed title is the source's, period.
• No commercial boost. No source pays to appear. No source is excluded for criticizing us.