Editorial standards — Foomax blog collection

September 2026

Applied by the editorial pass that produced the posts in this directory. Sources are never modified — edited copies land here.

Output format contract

Every post in this directory has exactly this shape:

# <Title>

# Human-authored

<left EMPTY — reserved for the author's own hand-written foreword or notes>

# Human-to-read

<the full edited narrative, in the author's voice, ending with the
methodology/how-this-was-made footer where the post has one>

# LLM-to-read

<structured reference for machine readers>

Format v2, per author request: an empty # Human-authored section sits at the top of every post; the former # Human and # LLM sections are now labelled # Human-to-read and # LLM-to-read.

Format v3 (September 2026) keeps that structure and tightens the provenance rules below, after an identity scrub found that the v2 provenance requirements were themselves the collection’s most consistent source of leakage.

The # LLM-to-read section contains, as terse markdown bullets/tables:

• Abstract — one paragraph, neutral register.
• Claims — the post’s central claims with their numbers, one bullet each, stated so they can be checked against the underlying repo.
• Data & provenance — what corpus/dataset, size, date range, how obtained. NEVER the local project directory name, the storage medium, or the drive it came from. (Format v2 required the directory name; v3 forbids it. That requirement was the single most repeated leak in the collection — it put a private directory name in nearly every post.)
• Method — one short paragraph.
• Reproduction — commands/paths if the post gives them, genericized (repo-relative, never absolute local paths). Do not annotate the genericizing — no “all paths repo-relative”, no “withheld under the privacy scrub”.
• Caveats — the post’s own stated limitations and any editorial ones, as one undifferentiated list. Do not label which is which; an author/editor split discloses the pipeline.
• Provenance line — “Edited ”. No source filename, no exact date. A day-precise stamp repeated across every post binds the collection to one working session.

Voice

First-person, dry, plainspoken, honest about limitations. The author labels speculation as speculation and keeps methodology footers. Preserve this. Do not corporatize, do not add hype, prefer cutting to rewriting. Keep the authors’ titles and their polemic energy where the post is a declared polemic.

Editing scope

• Structural: strong lede, kill throat-clearing, merge redundant sections, endings that land. Reordering allowed; new arguments are not.
• Line: grammar, tightening, consistent spelling variant (keep the post’s dominant variant), consistent number and percentage formatting, consistent heading case. Do not record which variant that is — regional spelling is a locale signal and belongs in no document here.
• Never invent or “correct” facts/numbers. Internal inconsistencies are kept as the author wrote them and flagged in the editorial report.
• Honest disclosures (ToS admissions, consent caveats, “nothing has been run”) are house signatures: keep them prominent; never soften or bury them.

Pseudonymity and privacy scrub

• The author publishes under a pseudonym. Zero tolerance: no legal name, no email addresses, no absolute local filesystem paths, no storage-medium or project-directory names, no employer/family/precise-location identifiers, no occupational or credential claims about the author. (The scrub greps live in the scrub script, not in this document — writing the target strings here would defeat the rule.)
• Personal-data posts: keep aggregate patterns; remove merchant+date+amount combinations, exact totals, product identifiers and dated life events. Assess re-identification in aggregate, not line by line — a set of individually harmless facts can still resolve to one person. Flag anything that resists scrubbing.
• Never state what was redacted or why in the published text. “Withheld under the privacy scrub” is itself a disclosure; delete the sentence instead.
• Criticism targets published work and ideas, not persons. Named-researcher critique stays, but characterizations of people (rather than their claims) get tightened to the evidence. Flag anything libel-adjacent.

Grading

Each post gets one grade in the editorial report:

• READY — publish as-is.
• READY-WITH-FLAGS — publishable; listed flags need a conscious decision (e.g. the erowid posts self-document a ToS violation).
• HOLD — a defect (truncation, unresolvable identification risk, factual incoherence) blocks publication until resolved.