How we work.

01Sourcing

Every dispatch ends with a source line. We name the source whenever we can. When we can't, we describe its proximity to the story ("a person who attended the meeting," not "a person familiar with the matter") and why we're protecting it.

We do not run dispatches sourced solely to a single tweet, a single screenshot, or a single anonymous claim. We do not run dispatches sourced to other newsletters without independent verification.

SOURCE TIERS

• Primary: Documents we hold, code we ran, statements we recorded, things we observed.
• Direct: Named people speaking on the record, or anonymous sources we have spoken to ourselves.
• Cited: Other publications. We link them and credit them by name in the source line.

02Verification

Two-source minimum for any non-public claim. A document plus a person counts as two. Two people who learned it from the same Slack channel does not.

For benchmark numbers, we run them ourselves when the model is publicly accessible, and we say so. When we can't, we mark the dispatch "not independently verified" in the dateline.

03Anonymity

We grant anonymity sparingly and we explain it every time. Anonymity is for sources who would face professional retaliation, legal exposure, or physical risk for speaking. It is not for sources who simply prefer not to be named, and it is not for executives floating trial balloons.

We never share a source's identity outside our own desk. Tip line technical specifics are at /tip.

04House voice

Short sentences. Active verbs. No hedges that don't earn their keep. No "could be a game-changer," no "experts warn," no "AI revolution."

Headlines describe what happened, not what it might mean. Deks add the second clause. The body is dispatches: dated, sourced, one fact at a time.

WORDS WE DON'T USE

• AI revolution, AI race, AI arms race — describes nothing
• Could, may, might — unless we're hedging a specific prediction
• As we've previously reported — link instead
• Sources say — name the proximity
• Game-changing, paradigm-shifting — the reader decides

05Frontier Index

The number on the front of the wire is a composite of five normalized signals: capability (weighted average of public benchmarks we trust), price-to-task, refusal rate (measured weekly against a fixed prompt set), throughput, and shipping cadence.

Weights and prompt sets are published quarterly. The full methodology PDF lives at /frontier-index/methodology.pdf. The raw data is at /frontier-index/data.csv, updated nightly.

The Index is a tracking number, not a leaderboard. It moves slowly because the frontier moves slowly, even when it doesn't feel that way.

06Corrections

Top of the next morning's edition. Named — what we got wrong, what's right, who flagged it. We never silently edit a published dispatch; if a fact changes after publication, we add a stamped update with the new information and the time it was added.

The corrections log is permanent and public at /corrections. It will get longer. We are at peace with that.

CORRECTIONS YEAR-TO-DATE

• Mar 18 — Misstated FrontierMath score for o4-mini-high (corrected Mar 19 edition)
• Apr 02 — Confused two Anthropic researchers with the same first name
• Apr 27 — Reported a CoreWeave deal value 10× too high (corrected within 90 minutes)

07Independence

JustSaid is reader-supported. No advertising. No sponsorships. No paywall as of today; if one ever appears, the editor will write a separate edition explaining why.

The editor holds no equity in any AI company, accepts no consulting fees from any company we cover, and refuses speaking honoraria from lab-organized events. Personal investment positions, if any, are disclosed at /disclosures.

08AI use

We use AI tools for transcription, search across our own corpus, and occasional copy-edit suggestions. We do not use AI to generate dispatches, headlines, or analysis. Every word you read on JustSaid was written by a person, edited by the same person, and signed.

We test the models we cover. That's reporting, not generation.