RUN 2026-10-14115 scanned → 63 published94% claims clean63% ensemble agreementpeak 20.67 GB reasoning OFF generated locally with Qwen3.6 27B · Gemma 4 31B · Ministral 3 8B
Flaimify
Methodology

How the brief is made.

Flaimify is an experiment in doing editorial work with local models and letting you audit every step. This is the full rubric — what gets in, how it’s scored, and how each claim is checked. The numbers below are from the most recent run.

01 FETCH
Fetch
collect items from every source
02 RATE
Rate
ensemble of models scores each item for significance & depth
03 WRITE
Write
a local model drafts the brief, reasoning disabled
04 VERIFY
Verify
each claim is fact-checked against its own source
05 NARRATE
Narrate
the brief is read aloud by a local voice model
06 RENDER
Render
the email is built
07 PUBLISH SITE
Publish site
the web edition is generated
08 PUSH TO LIST
Push to list
the campaign is created as a draft
09 COLLECT
Collect
the issue is archived

01What counts as significant

Every candidate is scored on two independent axes. Heat is significance — would this change what a working ML engineer does this week? Depth is how technical the item is, so you can triage by the kind of reading you’re in the mood for. Neither is about novelty for its own sake.

Very hot
Shifts the field or your toolchain — a frontier release, a result that changes practice.
Hot
Clearly matters to practitioners in the area; worth reading this week.
Niche
Real but narrow — a specific tool, dataset, or incremental result.
V.TECH
Architecture, proofs, kernels — you’ll want the paper open.
TECH
Engineering-level; familiar with the fundamentals is enough.
POP
Industry and big-picture; no background needed.

02The rating ensemble

Scores aren’t one model’s opinion. 3 different local models each rate every item independently — Qwen3.6-27B-4bit, gemma-4-31b-it-4bit, Ministral-3-8B-Instruct-2512-4bit, and the median vote wins. When they disagree sharply the item is marked contested and the disagreement is recorded rather than hidden.

Last run: 115 items rated · 73 unanimous · 6 contested · 63% agreement.

03Fact-checking every claim

After the brief is drafted, each factual claim is checked back against its own linked source by a separate model. A claim is marked ok, unsupported (the source doesn’t say it), or overstated (the source says less). Anything not clean is rewritten or cut before the issue is sent — you never see the unverified draft.

Last run: 63 claims checked against source · 2 flagged · 2 rewritten. Checker: gemma-4-31b-it-4bit.

04Source credibility

Sources earn standing from their own track record on this feed, not from reputation. Each source accrues a signal from three things: the average heat of what it surfaces, how often its items are independently corroborated by another source, and how often its items clear the bar to publication. New sources start neutral and move as the evidence comes in.

Deliberately not included: any judgement about a source’s truthfulness. Calling a source unreliable is a claim that needs far more evidence than one feed can gather, so the score measures signal, not virtue.

05Why reasoning is switched off

The brief is written with the model’s hidden reasoning disabled. On this workload that isn’t a compromise — it’s faster and the output is tighter, because the model commits to prose instead of narrating a chain of thought it then has to summarise. The switch lives in the chat template (enable_thinking=False), which is why the brief is generated in-process rather than over a chat API that ignores it.

06Local and inspectable

Everything runs on one machine — no cloud API is called to read, rate, write, or check. The exact prompts that produced today’s issue are published verbatim on the home page, and every issue is available as raw markdown at /issues/<date>.md so you can feed it to your own tools.

Last run: 3906 tokens generated · 593s · peak 20.67 GB.