Reader-facing articles — how we want them written

Purpose: Define article voice for paradigm-threat-files—the Markdown siblings of investigations/ dossiers and other long-form files—so assistants and authors match Paradigm Threat expectations. Pair with PARADIGM_INVESTIGATION_INSTRUCTIONS.md (dossier structure, evidence tiers, Limits) and LINKING_AND_SITE_PATHS.md (URLs).

URLs (same as blog entries)

Published reader-facing articles (parent-folder …/*.md essays under influence/, science/, history/, etc.—the pieces that ship like site pages) use the same link convention as blog/*.md posts: every in-repo target is an absolute public URL on https://paradigmthreat.net/..., including § anchors and images under the article’s site subtree. Do not use root-relative markdown links like (/influence/...) or repo-relative asset paths in those files.

Investigations, hub page.md indices, and internal docs may keep root-relative links unless a file is being promoted to “article” behavior.

Article vs dossier (division of labor)

If a claim needs five citations or a multi-row table, it belongs in the dossier; the article names the claim and links once (often to a § anchor).

Series articles (same thesis lane, multiple installments)

When several reader essays explore one ongoing thesis—e.g. predictive programming in blockbuster film/TV—treat them as a named series, not independent primers that repeat the same frame.

Reference series (predictive programming in movies):

• Installment I: influence/controlled_opposition/not-like-this.md — Matrix two-camp handoff.
• Installment II: influence/hollywood/everything-burns.md — Pacific / Dark Knight / 24 / Hong Kong litmus test.

Tables, lists, and catalog prose (articles vs investigations)

Hard rule for articles: No Markdown pipe tables anywhere — not in the body, not in ## Framing and limits, not in footnotes. Substack’s editor (and most paste-from-Markdown workflows) do not render | col | col | grids; rows collapse into a broken vertical list that reads like a corrupted dossier excerpt. Write normal paragraphs instead. Weave claim, tier, and link into complete sentences (commas, semicolons, short follow-on sentences). If the reader needs a grid, put it in the investigations/ dossier and link once.

• Investigations (…/investigations/*.md): Markdown tables are fine—tier matrices, chronologies, claim registries, comparison grids. Bullet and bold-lead catalogs (literature maps, author lanes, tier rows) also belong here.
• Articles (parent-folder reader essays): paragraphs only for argument and for limits. Do not recreate dossier tier tables as pipe grids, bold-lead stacks, or pseudo-tables (**Claim** — tier lines that paste like broken tables).
• If an article needs a table or structured illustration: export PNG or SVG, place the asset beside the article in the same folder (same pattern as hero images), and embed with an absolute https://paradigmthreat.net/... URL and descriptive alt text. Raster and vector images transfer; Markdown tables do not.
• Default for structured content: paragraphs, not lists. The reader should hear an essay, not scroll a glossary.
• Avoid catalog-style stacks in article bodies:
  • Repeated **Name** — gloss or **Name** — *work* (year): gloss lines (literature surveys, thinker rundowns, “who said what” lanes).
  • Long bullet runs of authors, eras, institutions, or claims that could be one or two paragraphs.
  • Pseudo-tables built from bold labels + em dashes (they paste into Substack like broken tables and read like dossier excerpts).
• When lists are OK (sparingly):
  • Short numbered lists (typically ≤5 items) for sequential steps the reader must follow in order (e.g. debunk-cycle phases, a tight parent checklist).
  • “Where next” link bullets at the end (routing only—not argument).
• If the material is inherently tabular or multi-row: keep the full grid or catalog in the dossier and link to the § anchor; the article gets one or two prose paragraphs that state the pattern and point there.
• Reference (catalog → prose): history/chronology/when-your-child-wont-fit-the-grade.md — “Literature that respected autodidactic children” (woven paragraphs, not author bullets). history/chronology/napoleon-corpses-dry-egypt-siberian-disease.md — ## Framing and limits as prose tiers, not a pipe table.

Default: Articles state stakes and thesis clearly—clinical but inspired (precise, unsentimental, still alive); investigations hold the audit trail. When in doubt, cut the paragraph that sounds like Limits and add a link: “the careful version lives in Limits.”

Reader-facing prose (generic surface)

Articles are written so they can be republished or read off-site without sounding like internal documentation.

• No project or repository meta in the body: Do not name the hosting project, the codebase, or “the repo.” Avoid phrases like in-repo, this tree, our investigations, a repo search, Paradigm Threat (and close synonyms) in running prose. Use neutral formulations: “this reading,” “the argument below,” “linked notes,” “supporting write-up,” “technical companion piece,” “a public search found…”.
• No drafting-process meta: Never ship author confusion, voice-note mis-transcriptions, session labels, “User query” framing, or “does it check out?” audit chatter in article bodies. If two subjects sound alike (e.g. Poe vs Prince), name the correct subject once in plain prose — do not explain how a mix-up happened or point readers at transcription logs. Resolve ambiguities before publish; ask the author when uncertain what they meant. Workflow mistakes belong in Investigator notes (investigations) or nowhere — not in reader-facing text.
• Links: In article bodies, use absolute https://paradigmthreat.net/... URLs for every internal page and image (same as blog/). Describe targets by topic, not by filing system (“see the linked DNA blueprint critique,” not “see our biology dossier”).
• First person — personal lane only: Use I / me / my only for personal experience, direct memory, or explicit personal stance the reader should attribute to a human speaker. Keep theory, history, and physics claims in neutral / documentary third person (“this model holds…,” “the record suggests…”).
• Neutral ≠ timid: You may still write hard thesis and manifesto cadence; just separate voice (optional I) from content (documentary).

Theory-led pacing (default for parent-folder articles)

Parent-folder reader essays are thesis vehicles, not primers. They may read as strong manifesto when the topic demands it.

• Lead with the unified theory: Open with the cross-cutting claim and its stakes. Do not bury the thesis under throat-clearing. Intelligent readers can handle a sharp fork on page one.
• Heavy proof stays in companion files: Multi-citation blocks, tier labels, and long Limits machinery stay in investigations/ (or equivalent technical write-ups). The article points; the companion prosecutes.
• Pacing: Short ## sections, one main idea each, momentum over repetition.
• Sensitive targets: When the author requests it, use structural / institutional framing (doctrine shapes, incentives, schooling) rather than naming modern groups—unless the author explicitly wants names.
• Honesty without dilution: When caveats are necessary, do not weave them into each section—that trains the reader to expect retreat after every claim. Prefer one consolidated ## Framing and limits (or ## Scope and limits) block after Where next, before Keywords:; keep the body declarative. Alternatively, one sentence plus a Limits link can suffice. Do not use caution as a substitute for clarity.
• Prisca sapientia (required in Framing and limits): Open that block with prisca sapientia—the belief that ancients held profound understanding of universe, nature, and theology, later lost or degraded; modern consensus is not default truth. One sentence plus dossier Limits for tiers is enough.

Reference implementation for generic surface + theory-led shape: science/biology/circle-of-life-and-how-it-began.md.

Voice (who we are writing for)

• Documentary default: Prefer a documentary voice first—direct, factual, and scene-forward without constant speaker framing. Use first person only when personal experience, memory, judgment under uncertainty, or stakes genuinely require it. Do not write “the author.” Stack with Reader-facing prose (generic surface) above for project-agnostic body text and the first-person-only-for-personal lane.
• First person is selective, not constant: I / me / my is valid when it adds signal (e.g. eyewitness memory, explicit personal inference), but avoid leading every section with self-reference. Rewrite default lines in documentary style when possible. Do not use first person for literary scene-painting (“opened my mind,” “we shared the same gasp”) when a plain fact carries the point (“the credits named X,” “fans read one movie”).
  • Bad (over-framed): “I still think in radio.”
  • Better (documentary): “Radio offers a useful model for layered signal bands.”
  • Use first person when earned: “I first noticed this pattern during …”
• Drop “this essay / this piece”: Prefer direct claims; use I only when earned. Phrases like “this essay will show,” “in this piece we,” or “what overlaps for this essay” read like syllabus metal—cut them unless a single signpost truly helps.
• State the claim — drop the frame: Prefer direct sentences over meta-leads. Bad: “My hard moral read stays blunt: Polynesian continuity was interrupted.” Good: “Polynesian continuity was interrupted.” Let syntax carry most of the weight, not a running commentary on your own stance.
• Bold — emphasis or new terms only: Use **bold** for (1) real emphasis (one beat per sentence or so—a contrast, a risk, a hinge you need the eye to catch), and (2) new terminology the article introduces or defines for this read (first pass; often plain text on repeat). Do not bold every proper noun, institution, date, or ordinary stressed word—that reads as sparkle fatigue and fake dossier layout. When in doubt, remove the bold.
• Register — clinical but inspired: Default to cool, exact prose that still lands the point—tight claims, clean images, no purple grief-as-genre layering. Drive the argument home with structure and consequence, not with repeated mourning or elegy cadence unless Ari explicitly wants a lament piece. Moral heat is fine in short controlled bursts; let evidence links carry weight alongside tone.
• Preserve Ari’s intensity and intent (hard rule): When Ari provides sensational, high-voltage, or accusatory claims, do not silently downgrade them into bland institutional prose. Keep the core charge, stakes, and emotional vector intact. You may tighten wording for readability, legal safety, or structure, but only if the primary intent is unchanged. If a sentence must be softened, preserve force by pairing it with explicit stakes (who is harmed, what is controlled, what is lost).
• No accidental de-fanging during cleanup: Editing passes for “neutrality,” “de-LLM,” or “documentary voice” must not erase load-bearing claims (e.g., suppression, controlled opposition, surveillance, corruption, coercive lock-in). Prefer: (1) keep claim strong in body, (2) move caveat machinery to ## Framing and limits, (3) link to dossier for heavy evidence.
• Author-supplied sentiment is first-class: If Ari gives direct language in chat, preserve it where possible (short quoted block or close paraphrase) before adding assistant synthesis. Do not replace high-signal claims with abstract summary labels.
• Intelligent reader — no hand-holding: Assume the audience already grasps why these topics matter. Skip “you might be wondering,” over-explaining basics they have heard before, and tutorial pacing unless the article is explicitly introductory. Do not soften or repeat the same caveat in three forms; trust the reader to follow and to use the dossier links when they want machinery.
• Audience: A curious lay reader—someone who will not read fifty pages of apparatus on first visit. Plain English, complete sentences, human rhythm (vary sentence length; avoid encyclopedia monotone).
• Point, don’t prescribe: Articles show patterns and place evidence; they do not coach the reader what to think. Implied conclusion beats authoritative instruction. Prefer observation (“the credits put X on a mass exit”) over coaching (“notice what the blockbuster allowed,” “stack this and you get,” “read who gets scapegoated”). Bad: authoritative I → you scaffolding (“I am not asking you to X; I am asking you to Y”) or constant second-person steering (“if you treat,” “you get,” “tell you”). Good: documentary third person or neutral this read; first person only for a specific fact the author witnessed—not literary audience poetry or lecturing.
• Not authoritative — lay and objective: Sound like someone reporting what is on the record, not a pundit managing consent. Avoid pulpit cadence: “that is why you must,” “the point is,” “what matters is that you see.” Let links and juxtapositions carry the inference.
• Simple over lyrical: State what happened in plain words. Bad: stacked shared-experience poetry (“we shared the same gasp,” “the same silence when…,” “opened my mind was not only…”). Good: one factual line (“fans read one movie,” “the credits named Tonkin and Islam blame”). Keep the point; drop the theater.
• You / we (restricted): Second person is for quoted dialogue, film lines, or rare inclusive scene-setting—not default narration. Inclusive we is almost never needed; prefer third person (“fans,” “audiences,” “the culture”). Default remains documentary third person.
• No cartoon villains: Prefer structural language (“schooling,” “press posture,” “museum caption,” “funding stack”) over sentences that imply every individual scientist intends harm. Name power and pattern; if documented intent matters, the dossier carries the receipts.

Shape (how we open and move)

1. Title — concrete; may use em dash or a strong image. Keep slug/path stable unless Ari explicitly requests a rename (URLs depend on paths).
2. Subtitle (required) — add one short line immediately under the # title (before any hero image), summarizing the thesis in plain reader language. Aim for one sentence fragment or clause (roughly 8–20 words), no ending period, no markdown heading marker (##), and no TL;DR label.
3. Optional hero image — immediately under the subtitle, https://paradigmthreat.net/... image URL (same as blog/), short alt that matches the article’s central image or claim (not only mood).
4. Reuse existing corpus images (when topics overlap) — If the article covers subjects that already have investigation, timeline, or sibling-article art, embed those assets at the ## section where the topic is argued—not only the new hero PNG beside the file. Search first: grep linked dossiers, timeline chapters, and hub pages for ![ / *.png / *.jpg; prefer published https://paradigmthreat.net/... URLs (timeline art often lives under https://paradigmthreat.net/files/timeline/media/...). Placement: one image after the paragraph that introduces the topic (illustrate the linked investigation, not decorate every section). Alt text: name the source lane and what the reader should see (e.g. “Historical antibodies — timeline chapter,” “Call of Duty militainment — Krasinski operator arc”). Do not copy images into the article folder unless you are creating new art; link to the canonical file in its home directory. Newsletter budget: many large PNGs + hero can trigger Gmail clipping—prefer fewer, well-placed reuses over a gallery; see Length → Newsletter / inbox check. Reference: influence/controlled_opposition/cia-1984-mars-chrono-trigger.md (hero + dossier PNGs mid-article); science/biology/circle-of-life-and-how-it-began.md (article hero + investigation header); influence/hollywood/ghost-war-the-antibody-phase.md (hero + timeline, militainment, PP hub, Iran runner reuses).
5. Hook — a scene, metaphor, or one sharp anecdote before the reader hits dense links. This is the contract: why this file exists before apparatus.
6. > **TL;DR:** — optional in articles. Use it when it adds a clean thesis line without cloning the dossier’s opening or sounding like a second investigation header. Skip it when the title + subtitle + first two paragraphs already do that work.
7. Body — ## sections (typically three to seven). Each section one main idea, usually prose paragraphs (see Tables, lists, and catalog prose). Bold only per Voice → Bold (emphasis or newly introduced terms)—not as default highlighting or list labels.
8. Citations in prose — Markdown link syntax [label](https://paradigmthreat.net/...) (and external https://...), same convention as blog posts—not bare /history/... paths and not backtick-wrapped raw URLs (they are not clickable in Substack). Each link: descriptive label plus what the reader will find there (one clause). Avoid hub-only link dumps unless Ari wants a routing sheet; prefer narrative + links.
9. Closing — short “where next”: Common Questions, timeline hubs, or § anchors in dossiers (e.g. “start in §2.1”).
10. Optional ## Framing and limits (or ## Scope and limits) — At most one such section, placed after Where next and before Keywords:, when in-file boundaries help. Prose paragraphs only — no pipe tables, no tier grids, no bold-lead claim stacks. Group documented vs author thesis in flowing sentences and link the dossier for the full audit trail. If the dossier already carries the full caveat stack, a single link line here is enough.
11. Keywords: — one line, Option B (see PARADIGM_INVESTIGATION_INSTRUCTIONS.md): #PascalCase tokens only, no ## Keywords heading.
12. **Last updated:** — ISO timestamp when the article changed materially: prefer full YYYY-MM-DDTHH:MM:SS±offset (e.g. 2026-05-23T09:18:22-04:00); date-only YYYY-MM-DD is acceptable for minor touch-ups.
13. Attribution footer (required — do not remove in de-LLM passes) — After **Last updated:**, a horizontal rule ---, then an italic one-liner, for example: Written and narrated by Ari Asulin, with drafting and research support from LLM agents. Minor wording drift is fine; keep the name and the fact that LLM tools helped produce the page. This footer is intentional credit, not an “LLM artifact” to delete when scrubbing machine cadence. It replaces repeated “as the writer,” “my position,” or “who is speaking” throat-clearing in the body.

Pre-publish pass (easy misses): The line immediately after # Title must be the subtitle (plain text, before any ![ hero or opening hook paragraph)—not a ##, not > **TL;DR:**, not a date stamp. Use **Last updated:** once at the end (with optional attribution after it); do not duplicate **Updated:** under the lede. Series installments: read the prior installment in full; cut shared frame that Installment I already taught. Grep the article for | — any pipe-table row means rewrite as paragraphs before Substack paste (see Tables, lists, and catalog prose). Run the Verifiable externals pass below on every ship. Run the de-LLM sweep (§ Phrasing we are tired of) — mandatory, not optional. Grep for drafting meta — reject before ship: voice-note, User query, session capture, misheard, does it check out, transcription (when referring to author/assistant workflow, not primary sources), I am not asking you, I am asking you to, I'm not asking, To be clear, It is worth noting.

Length (Substack / email)

Authoritative ceiling (email, not Substack’s own word cap): Substack states there is no word-count or length restriction on posts, but the editor can show “Post too long for email” when the email may be clipped or truncated in some inboxes—specifically citing Gmail truncating messages over 102KB (serialized message size in the inbox: HTML, images, embeds, tracking—not raw Markdown word count). You can still publish; readers may need “View entire message.” There is no reliable standalone “Gmail word limit”; clipping risk tracks rendered email bytes. Source: Substack Support — Why does it say “Post too long” in the post editor?.

Drafting targets (Markdown, in this collection):

• Default target for new flagship reader essays: about 2,600–3,400 words of body (tables and link lines count toward reading time), with ~3,000 as the nominal aim. Shorter is fine when the argument is already complete.
• Reference implementation: history/chronology/indigenous-creation-legends-and-geology.md (order of magnitude ~3k words, section pacing)—use it as the shape for “Substack-ready depth” without pasting the whole dossier. Tables in that file: if present, treat as legacy; new articles use prose or PNG/SVG per Tables, lists, and catalog prose above.
• Newsletter / inbox check: Before sending a long flagship post, do a real send test (or preview at View entire message) when the piece combines long prose with multiple large images—two hero-scale PNGs can burn the 102KB budget faster than +500 words of Markdown.
• Soft ceiling: treat ~3,200–3,400 words of prose in a single article as the repo drafting stop unless Ari explicitly splits across Part II or accepts clipping risk after that send test (images, embeds, and HTML bloat push toward the 102KB clip faster than word count alone).
• If the piece must grow: push Markdown tables, quote banks, and five-citation blocks into investigations/; add a Part II post for a second email rather than one mega-file. For tabular material the reader must see in the newsletter, ship a PNG/SVG in the article or link to the dossier table anchor.

Substack publish — title and subtitle fields

Substack renders Title and Subtitle from editor fields, not from the pasted Markdown body. When comparing a live post to the repo, do not treat “title/subtitle missing from body” as text drift.

Field mapping (default workflow):

Stonehenge reference: stonehenge-quantum-portal-golden-age-and-dark-ages.md — Title = # heading; Subtitle field = hook (The real nature of Stonehenge is not a calendar…); repo line 3 (Golden Age spiral end-port…) stays for site lede; body paste begins at hero / TL;DR / first ##.

After publish, add **Substack:** with the post URL before **Last updated:** (see circle-of-life-and-how-it-began.md).

Verifiable externals pass (required before ship)

LLM drafts confabulate text that sounds quotable: song lyrics, film lines, attributed speeches, headlines, statistics, court holdings, DOIs. That is separate from author thesis (which stays first-class) and separate from de-LLM cadence (contrast templates). Hard rule: any blockquote or verbatim line attributed to an external source must be checked against a primary, official transcript, or reputable lyrics/quote page before the file ships. If the author or a reader cannot find it online in a few minutes, it is probably made up.

Canonical failure (2026-06): history/chronology/stonehenge-quantum-portal-golden-age-and-dark-ages.md quoted Spinal Tap’s “Stonehenge” as “And the soft winds blow…” after “where the demons dwell.” That line does not exist in the song. The real couplet is “Where the banshees live and they do live well” (Songfacts — Stonehenge). An LLM agent inserted plausible-sounding lyric during drafting; it survived until the author searched and could not verify it — embarrassing and exactly what this pass prevents.

Checklist (every article and dossier excerpt republished as an article):

1. Lyrics / screenplay / TV dialogue — match against a lyrics site, subtitle transcript, or timestamped clip; never trust model memory.
2. Historical speeches and letters — digitized primary, scholarly edition, or archive URL beside the quote.
3. News headlines and pull quotes — live or archived URL; headline wording must match the link.
4. Numbers and dates in factual claims — trace to the table, report, or statute cited (not “about 40%” with no source).
5. DOI / PMID / case name — resolve once; drop or flag if broken.

If you cannot verify quickly: do not blockquote. Paraphrase with uncertainty, move the line to the dossier open questions, or omit. Author thesis in Ari’s voice does not need an external URL — but must not be dressed up as a fake lyric, fake Einstein quote, or fake headline.

Fast grep hints: scan for ^> blockquotes, lines in smart quotes attributed to named people, and “according to” without a [label](https://…) in the same paragraph.

Phrasing we are tired of (LLM tells)

Keep vs cut: Remove machine cadence in the body; keep the attribution footer (§ Shape step 13) that names Ari and LLM agent support. De-LLM passes target prose artifacts, not production credit.

• Required second pass (de-LLM sweep): After finishing any article draft, run one clean edit pass that looks only for machine-sounding cadence and rewrites it to plain spoken prose. This pass is mandatory before shipping.
  • Primary target: the contrast template “not A, (but) B” in one or two adjacent sentences (for example: “this is not a myth; it is a memory”; “it is not calomel alone; it is red mercury”; “this is not X, it is Y”).
  • Rewrite rule: keep the meaning, remove the formula. Prefer one positive claim, a concrete image, or a direct sentence you would actually say out loud.
  • Fast checks: search for patterns like “not a”, “not an”, “not the”, “it is not”, “but rather”, “this is not”, “in this repo”, “in-repo”, “I am not asking you”, “I am asking you”, “I’m not asking”, “I’m asking you to”, “notice what” (as a rhetorical pivot), “To be clear”, “It is worth noting”, “we shared”, “opened my mind”, “the same gasp”, “the same silence”, and inspect each hit for template-like contrast writing, permission-seeking meta, lyrical unity prose, or internal documentation voice.
  • Acceptance bar: if a sentence sounds like it is teaching a model contrast pattern—or negotiating with the reader instead of talking to them—rewrite it. Lay reader default: state the claim once, plainly, as you would to a friend who has not read the dossier.
• The “not A, it is B” teaching beat — repeated contrasts of the form “This is not chemistry / myth / debunk / a verdict… it is relationship / science / …” work once; across many articles they read as the same canned cadence. Rotate constructions: lead with the positive claim; use “Rather,” “Another name for that pressure:” “What I mean is simpler:” a question, or a concrete image—anything that breaks the identical not-this / but-that rhythm.
• Close cousin: “X is not Y; it is Z” every time you need nuance. Sometimes Z alone carries the point; sometimes two short sentences beat one negated compound.
• Permission / negation pairs — “I’m not asking you to X. I’m asking you to Y.” (and close variants) read as LLM debate scaffolding. Cut both halves. Say Y directly, or one short caveat then the claim: “Historians dispute the fine print; the credits line still landed on a mass exit in 1999.”
• Reader negotiation — “I am asking you to notice”, “let me be clear”, “the point is not X, the point is Y”, “if you stack / if you treat / you get” when a neutral observation would read fine. Prefer documentary delivery; conclusions stay implied unless Ari supplies an explicit personal-memory beat.
• Stance meta — “In this piece I will…”, “My thesis is…” every section (the title and flow already signal intent). Bad vs good examples live under Voice → State the claim.
• Essay / piece scaffolding — “This essay argues…”, “Below, this article…”, “In what follows the reader…” Same cure as above: say it; use I only when earned; do not announce the document.

What we avoid in articles

• Markdown pipe tables (anywhere in the file) — Substack breaks them on paste; never in articles. Use prose paragraphs in ## Framing and limits too; tier grids belong in investigations/ (see Tables, lists, and catalog prose).
• Catalog lists and bold-lead stacks — **Rousseau** — … / **1910** — … / bullet galleries of names, works, or eras that belong in a paragraph or in the dossier. Exception: short numbered how-to or phase lists (≤5 items). If you are tempted to write “literature review as bullets,” rewrite as two paragraphs and link the dossier §9-style map.
• Wall of acronyms without gloss—spell out once, then abbreviate if needed.
• Paste-dump of URLs with no integration in the sentence.
• Duplicating half of a dossier “so this page stands alone”—link instead.
• Duplicating a recent sibling article in the same series—link to the prior installment and keep only new geography / new titles / new calendar beats (see Series articles above).
• github.com/clevertree/... or other dead reader URLs (see LINKING_AND_SITE_PATHS.md).
• OS paths (/home/..., ~/dev/..., /media/..., maintainer mirror folders like ~/dev/wget/...) in body text—never in articles republished to Substack or other public surfaces. Use public URLs only (https://chronologia.org/..., https://paradigmthreat.net/...). Local paths belong in dossier Investigator notes at the end of investigations/ files only.
• Mid-article Limits voice: tier sermons, long “not proven” chains, claim-number callouts, Epstein-class analogy legal fencing repeated in prose—park that in the dossier (one link is enough).
• Sprinkled disclaimers — repeated “this is not to say,” “not every X,” “scholars disagree” every few paragraphs. Batch into ## Framing and limits or cut; the narrative should read as continuous thesis, not alternating thrust-and-retreat.
• Tone softening that changes meaning — avoid edits that make strong claims sound merely hypothetical when Ari intended a direct charge. Keep the claim explicit, then handle uncertainty in one limits block if needed.
• Lament-as-default — stacked mourning, wound, hurt framing in every section when the same stakes could read sharper in clinical mode. Reserve heavy elegy for when the topic or Ari calls for it.
• Lyrical unity / shared-feeling prose — “We all felt,” “the same gasp,” “opened my mind,” paired sensory beats that perform audience togetherness instead of stating reception (“fans read one movie,” “peer mockery converged on Maddox”). Cut the poetry; keep the fact.
• Frame labels and stance chatter — “My read,” “hard moral read,” “here I want to” before every claim. State the claim; use I when it earns its keep, not meta before every line.
• Sparkle bold — bolding most nouns, proper names, or whole clauses to simulate “importance.” If a paragraph reads like a highlighter ramp, strip bold back to Voice → Bold rules.

After you ship (discovery)

Follow PARADIGM_INVESTIGATION_INSTRUCTIONS.md → “New article or hub — discovery checklist”: site root page.md Investigation Categories, topic page.md, index-investigations.md (or sibling index), bidirectional Related links, npm run autogen.

Reference implementations

• Article + dossier pair (controlled opposition): influence/controlled_opposition/cia-1984-mars-chrono-trigger.md ↔ …/investigations/cia-reading-room-mars-exploration-1984-rv-investigation.md.
• Lay reader essay + citations (chronology): history/chronology/indigenous-creation-legends-and-geology.md ↔ …/investigations/indigenous-legends-vs-geologic-dating-investigation.md.
• Prose-over-catalog (schooling / parent advice): history/chronology/when-your-child-wont-fit-the-grade.md ↔ …/childhood-autism-spectrum-conformity-recategorization-investigation.md.
• Theory-led flagship (biology / continuity): science/biology/circle-of-life-and-how-it-began.md ↔ science/biology/investigations/dna-fingerprint-scalar-life-investigation.md plus creation event interpretations.
• Corpus image reuse (Hollywood / antibodies / PP): influence/hollywood/ghost-war-the-antibody-phase.md — hero + mid-article PNGs from militainment, Benghazi, timeline antibodies, 911 PP, PP hub, Trump-era scorecard, Iran runner.
• Series (predictive programming in movies): not-like-this.md (Installment I) ↔ everything-burns.md (Installment II) — dedup shared two-camp frame; Pacific litmus test only in II.

Last updated: 2026-06-18 (series articles + cross-installment dedup)