Read the rules

Probiome describes what research found. It never advises, and it never says a microbe or a habit causes, prevents or treats anything. This skill is the one rulebook for how strong a finding is, when it may be shown, and how it may be worded. It unifies the rules that grew up separately:

• the strain evidence grade (DESIGN.md › Strain evidence; packages/pipeline/src/evidence/evidence-table.ts);
• within-study replication on condition and microbe pages, the only way conditions are compared (docs/research/condition-comparison-study.md and its addenda; PRs #143 and #155);
• the cohort rule for landmark studies (packages/pipeline/sql/serving/study_signatures.sql; docs/research/landmark-studies.md);
• Factors: findings about how people live, and a person's own log beside them (docs/research/human-factors.md, where the reasoning and the numbers behind v2 are);
• how research quality shows to readers: qualifiers, the two questions, and very small differences (docs/research/research-quality.md, behind v3).

v3, October 2026. v2 added Factors and their field-note pages and settled v1's open questions. v3 adds how research quality shows to readers (§ 3, § 4, § 11). What's still open is at the end.

§ 1Grounding

Built on established frameworks, adapted to association data:

• GRADE (Guyatt et al., BMJ 2008, PMID 18436948): certainty of evidence is High, Moderate, Low or Very low.
  • Observational evidence starts Low; randomized trials start High.
  • Downgrade for risk of bias, inconsistency (it doesn't replicate), indirectness, imprecision and publication bias.
  • Upgrade for a large effect or a dose–response.
• Bradford Hill: consistency, strength, temporality, dose–response, plausibility and experiment are what move an association toward causation. Almost nothing we show meets that bar, so we stay with association language.
• Microbiome reporting standards: STORMS (Mirzayi et al., Nature Medicine 2021) and STREAMS (Nature Microbiology 2025, PMID 41326814). Use them to judge a study's quality: sequencing method, confounders recorded, data available.

§ 2Units: study, cohort, finding

• A study is a paper (PMID), the unit of credit. Every finding is shown under its study's name and size ("Fackelmann 2025 · 21,561 people"), never folded into "research shows".
• A cohort is one group of people recruited and sampled together (PREDICT 1, LifeLines-DEEP), the unit of replication.
  • The same cohort appears in many papers, so replication counts cohorts, never papers.
  • Cohorts that share people, or have the same parent in the registry (curated/cohorts.yaml), count once.
• A finding is one microbe × one contrast × one cohort (or the study's pooled result), as the authors report it: effect, effect type, direction, the study's own threshold, and what it adjusted for. Nothing is recomputed or re-tested.

§ 3The ladder

What kind of evidence a finding rests on, from weakest to strongest. Code assigns it, never a model.

Moving up or down, always in code:

• Up: consistent direction across independent cohorts; a large effect; dose–response; a prospective design.
• Down:
  • a single cohort drives the result;
  • a confounder is likely (§ 6);
  • few people;
  • the effect type doesn't match the claim (§ 5).

No visible GRADE words. Pages never say "High certainty" or "Moderate evidence": readers take those as "proven". Each page shows what was counted (the strain grade; "in N of M"), and the ladder stays internal, for ordering and thresholds.

Two questions, two kinds of study. Size says how sure we are; design says what kind of answer we have. Never trade one for the other.

• Who differs: answered by statements from studies that measured each person once. Their strength word counts independent groups (F1).
• What changes: answered by studies that followed the same people, or by trials. Their status is Tested in a trial, Followed the same people, or "Not in the studies we've checked".
  • A randomized trial with a comparison group ranks above following the same people without one.
  • Under 100 people adds "A small study".
• Neither kind changes the other's word. When both exist, code says whether they agree ("A trial saw the same direction") or marks the statement Studies disagree.
• An exposure page shows both answers at the top.

§ 4Showing a finding

• Replication before prominence:

  • Condition pages and microbe pages (decided 2026-10-06, replacing #143's pooled lists): conditions are compared only within studies that had both people with the condition and healthy people, 10 or more of each. A row is listed only if it's clear within studies.

    • Its strength word counts studies, with § 10's words: Seen once for 1 study, listed last and never a headline; Seen more than once for 2 that agree and are the majority; Seen again and again for 3 or more, none clearly opposite.
    • Rows where studies disagree, or only a minority agree, aren't listed; one link counts them.
    • Pooled numbers are never shown.
    • Amount-only rows ("As many carry it"): when 80%+ of healthy people carry the microbe, a difference can only show in the amount. Such a row is listed only when the within-study amount test agrees in 2 or more studies.
    • How much carriers have is never on a row. It lives only in the row's sheet, with one line per study.
    • What's typical in healthy guts is described, never compared (DESIGN.md › In healthy guts and In conditions).

  • Landmark studies and Factors, rule F1. This rule is for Factors and landmark studies only; condition and microbe pages have their own within-study rule above. The two differ on single findings: under F1 a "Seen once" finding never makes a list row or a microbe-page line, while a condition's "Seen once" row is listed last and never leads (never a page's first row, a home-page regular, an MCP card's headline or a share image). A Factors or landmark list row or microbe-page line needs all four:

    1. It's clear by the study's own threshold.
    2. At least 2 independent cohorts agree, and they're the majority of those that reported the microbe: cohorts_agreeing ≥ 2 and cohorts_agreeing > cohorts_tested / 2.
    3. It goes one way: every clear genome bin agrees, and no independent cohort is clearly opposite. Otherwise it's Mixed evidence.
    4. It's a taxon we have: a species, or a genus for 16S results (genus pages only).

    A study that publishes only pooled results counts as one cohort, unless its table reports per microbe that its cohorts were consistent (Cochran's Q not significant and I² under 50%). Findings that don't pass go in a collapsed "Seen in one cohort" group and on the study page, never hidden.

  • Across studies, only sources whose tables list every microbe they tested count toward N of M (so their nulls count). A paper that reports only its hits never adds to the count: its silence isn't a null, and its hits are a biased sample.

  • Findings a model reads from a paper's text (its abstract, or its full text without per-microbe tables) are one study's report, never a table. They're described in our own words, named and dated ("a 2024 study of 312 adults reported…"), and may make a "Seen once" line. They never add to N of M, a list row or a microbe-page line, and they count toward a strength word only after a person checks them against the full paper (docs/research/knowledge-pipeline.md § 6.4).

• "N of M" counts direction, not significance.

  • N is how many studies or cohorts point the same way as the overall result; M is how many had a result. The pooled or stratified test (random-effects, CMH, van Elteren) carries the significance.
  • Write it "Fewer in 8 of 8 studies" or "Higher in 5 of 5 cohorts", and only when M is 2 or more. Never "clearly higher in 5 cohorts". On a condition row, this count lives in the row's sheet; the row itself shows only the strength word.

• Very small differences: a finding below the usual "small" threshold carries "A very small difference":

  • standardized difference |d| < 0.2;
  • correlation |ρ| < 0.1;
  • how many carry it, Cohen's h < 0.2;
  • the carriers' amount, rank-biserial < 0.15.

  Big studies make tiny differences clear: 402 of the 501 diet lines shown in October 2026 were below |d| 0.2.

  They don't lead (D21). In lists and on the factor line they fold into "N more with very small differences", each row labeled, as condition pages do with their thresholds. They're never hidden, and the study page keeps everything. Counts and headlines use what leads, and say how many folded.

• Small groups (under 50 people) fold and get a small-group note. Small comparisons: a published contrast with fewer than 2,000 people gets a power note: "Fewer people were compared, so smaller differences can't be told apart."

• Page-level caveats, not repeated row notes, when the whole page shares a problem: one company's cohorts behind most findings (§ 8), reverse causation for medicines (§ 6).

• Merge only what one sentence says exactly. In contrasts with three or more groups:

  • one group higher than both others: "Higher in omnivores than in vegetarians or vegans";
  • one group lower than both others: "Higher in vegetarians and vegans than in omnivores";
  • all pairs in one order: "Highest in omnivores, then vegetarians, then vegans".

  Anything else gets one line per pair. Each pair keeps its own cohort count.

• Strains: grade per condition or outcome, never one grade per strain:

  • Not tested in people: no trials.
  • Early: one single-strain trial, or blends only.
  • Some: two or more single-strain trials.
  • Well studied: a meta-analysis or systematic review that reports the strain on its own, plus 3+ single-strain trials with 300+ people.

  It measures how much testing, never whether it works.

• Reviews and meta-analyses summarize other studies. They're never one more study (docs/research/knowledge-pipeline.md › Meta-analyses and reviews):

  • Trials count as trials, and reviews are shown as their own kind ("226 trials · 51 reviews", "a meta-analysis of 12 trials"). A review's trials and people are never added to a trial count or to each other.
  • A trial a review includes is marked as included and still counts once. Trials we don't hold count only through the review's own totals, on the review's line.
  • A review counts toward a strain only when it reports that strain on its own. One that pools strains ("Lactobacillus probiotics") informs the species or genus, not the strain.
  • Reviews that overlap are one body of evidence: lead with the most recent and largest, and say they overlap. When their included studies are unknown, say so ("includes: unknown"); never claim an overlap or its absence.
  • Narrative reviews are never counted; they're for discovery and context only.

• Evidence stays separate and credited:

  • Never average AGORA2, BacDive and research into one number.
  • Never combine a trial and observational cohorts, or two studies, into one number. Each gets its own line, named for what it is ("in a trial", "in 3 cohorts").
  • Code may say in words that they agree.
  • Conflicts become Mixed evidence, with one sentence naming both sides.
  • Never pool another study's results into our own statistics.

• Methods don't mix: a 16S genus result and a shotgun species result are different findings. Never compare effect sizes across studies or methods.

§ 5Wording

The data type decides the verb:

• "Higher" always comes with its note: a bigger share of gut bacteria on average, across everyone in the group, carriers or not.
• Per-unit sentences are about people, never about a dose acting on a microbe.
• People first: when a study gives both how many carry it and how much, the people line leads.

Color follows the comparison. The compare cyan and magenta mean more or fewer people than a reference group (healthy people in the same studies, non-users, non-drinkers). In healthy guts is a description, never a comparison, so it never takes them. A contrast with no reference group (vegans vs omnivores) is set in ink.

Banned words in our own voice:

• causes, leads to, drives, prevents, treats, cures, heals, boosts, fights, protects, improves, promotes, supports gut health;
• feeds or "feed your", gut-friendly, balance, optimal or optimize, detox;
• "good/bad/beneficial/harmful bacteria", "good for", "bad for";
• "you should", try, consider;
• "research shows" or "studies show" without naming or citing the study.

Quoting a label's own claim is allowed, attributed to the manufacturer.

Certainty words follow the ladder:

• "seen in one study"
• "consistent across N studies/cohorts"
• "associated with" (the default)

Never "linked to better health".

Every comparison page carries an association line: "A difference is an association: it doesn't mean a microbe causes, prevents or treats X." Factor pages say the same of the factor: "…it doesn't mean coffee changes these microbes, or that they affect health."

Health correlations reported by authors (e.g. cardiometabolic markers) are attributed: "the authors found…". They're never our own claim, and they're shown only where the study's own tables have them.

§ 6Confounders

• Show the study's adjustment set (age, sex, BMI, medicines, diet, depth or batch) on its card. Say plainly where something it didn't adjust for matters.
• Medicines vs disease: a curated, cited map from conditions to their common medicines (curated/condition-medicines.yaml). When a condition row and a shown medicine finding for the same microbe point the same way, the row gets a quiet note: "Many people with type 2 diabetes take metformin, and this microbe is also higher in people who take it (Forslund 2015)." Never hidden, never a verdict.
• Sequencing depth: study cards say whether the study adjusted for depth. Condition pages carry no depth caveat: they compare only within studies, where both groups went through the same lab and pipeline.
• Stool form and transit move microbes on their own. Factor pages for things that loosen or firm stool (laxatives, antibiotics, fiber) say so.
• Reverse causation: medicine pages say "People take these medicines for reasons that can also go with different gut microbes."

§ 7AI-written text

• Models extract and write. Code decides grades, counts and what's shown.
• Cards are forms filled from a title and abstract: strain-trial cards, and factor-study cards. They're cached by paper, model and prompt version, and validated against the form. A card never makes a list row or a microbe-page line: those come only from tables. A card can still be described in a digest as one study's report (§ 4). Cards feed the score, the queue and cited sentences.
• Summaries (strain, landmark study, factor synthesis) are written only from facts and cards we hold, never from abstracts directly.
  • Every sentence cites a paper we hold a card for (a single-study summary checks numbers against its facts instead).
  • Numbers must match the cards, tables or facts. There's no arithmetic across studies: only numbers code put in the input.
  • Sentences that fail are dropped. A summary that loses more than a third of its sentences, or keeps fewer than 2, isn't kept.
• One check list for every summary (claim-check.ts):
  • no advice, dose talk, "works/proven/safe" or "you";
  • no cause or judgment words (§ 5);
  • say when evidence is blends only;
  • no animal or lab studies in human-evidence text;
  • a one-cohort finding says so, and trials and observational studies are named as what they are;
  • a review and its trials are never added into one count: numbers may be the trials, or the reviews, or a review's own totals, never their sum.
• Everything is a draft until it is published (the kind's switch: PUBLISH_STRAIN_SUMMARIES, PUBLISH_STUDY_SUMMARIES, PUBLISH_FACTOR_SUMMARIES), and an AI summary is published only when a person approved that version and the kind's switch is on. It can be withdrawn at once (pnpm pipeline drafts withdraw); a withdrawn version stays down until a person approves a new one. A page must work with no AI text. "Expert reviewed" only after a named person reviews that version; a new version drops it.

§ 8Disclosures

• Interests come in three layers, kept apart (decided 2026-10-05). A company affiliation is not in itself a competing interest, so no layer stands in for another, and none is a verdict on the findings.
  1. What the authors declare (theirs): the declaration of competing interests (founders, employees, consultants, share options, funders) as the paper publishes it, with its source and date. It's fetched from PubMed's CoiStatement or, when PubMed has none, the open full text in Europe PMC (the study-interests pipeline source); a person's summary in our words (competing_interests) sits beside it.
     • The quoting limit (decided 2026-10-05): 80 words. A declaration of up to 80 words is quoted as published, as a short, factual, attributed quotation. A longer one is summarised in our words (competing_interests), with a link to where it's published, and never quoted in part. The code's QUOTE_MAX_WORDS follows this rule; change the rule here first. Our check (layer 3) quotes only the single sentence a finding rests on.
     • A company paying for a study counts even when no author works for it.
     • The full declaration shows on the study's card and on the study's own page, and nowhere else (DESIGN.md › Factors). Other places that name the study link to it.
     • A study whose declaration can't be read (neither PubMed nor an open full text has one: the paywalled case) is still shown, never hidden: small studies matter. Its declaration reads "Not checked yet", and the pipeline lists it under "Needs a person" (in each run's summary and an open issue) until a person reads the paper. A fetched declaration counts as checked.
     • The tone is neutral: "Competing interests", in ink, never a warning.
  2. What they're affiliated with (a fact): every affiliation in the paper's author list, with how many authors list each, and which are companies. Companies come from a person's reading of the byline (author_companies, which wins wherever it exists) or from the Research Organization Registry (ROR), where a type of company counts and an affiliation ROR isn't confident about stays unresolved. Never inferred from a name. An organisation the byline doesn't say is a company is recorded as unclear, with a short neutral note (decided 2026-10-05): it's never marked as a company, and it still shows in the plain affiliation list. They show beside the declaration, on the same card and page only, as facts with no judgment words, as in Palleja 2018's: "Authors list companies among their affiliations: Clinical-Microbiomics A/S (2 authors), BGI-Shenzhen (5 authors), iCarbonX (1 author)." Counts are per company and aren't added up. Nothing shows until they're recorded, or when none are listed. They don't add the "Interests" tag, which comes from the declaration.
  3. Our check (ours, labelled as our analysis): a pure, versioned rule set (apps/api/src/services/interests-check.ts, interests-check/1) that says what we detected, never what's true, each finding with its evidence (the declaration's sentence, the affiliations, the cohorts):
     • the declaration names a company, or a tie (founders, employees, consultants or advisers, shares, funding or fees);
     • the declaration says none while authors list company affiliations;
     • one company runs most of the study's cohorts. It carries its rule version and a link to its method (docs/research/research-quality.md § 4.6). It's always shown as ours, apart from the declaration, and a rule whose input is missing is skipped, never counted as clear. The reader-facing words wait for design.
• Concentration: when more than half of the people behind a factor's shown findings come from cohorts run by one company, the page says so once, as the concentration line in "Who's been studied". That line is the only place a factor page speaks about interests outside the study cards. It counts who runs the cohorts, not where authors work, so author company affiliations don't feed it.
• Licence and credit for every source, on the page and on /sources. Per-microbe tables are loaded only under CC BY or CC0 (or a data repository's stated open licence); everything else is cited at the level of the paper.
• Where a figure comes from: the paper's table versus our own analysis, and why counts can differ (e.g. our replication filter). Show both counts.

§ 9A person's own log next to research

The gut log is private health data. Where it meets research (the Factors bridge):

• Opt-in for anything personal on a research page; plain links from the log to a factor page are fine, carrying no data.
• Side by side, never connected:
  • Show the person's own usual and logged days.
  • Never a sentence linking them to a finding, a comparison with a study's groups, a guess at their microbes, a score, a goal, or "people like you".
• Private by construction: fetched with the session and no-store, never in a cached public page, a URL, analytics, share images or link previews. Never returned with research findings over MCP.
• Sensitive labels (medicines, mood, stress, alcohol, symptoms) stay in the person's own words, only where they turned it on.

§ 10Describing what's known: field notes

Exposure pages (docs/research/factor-pages.md) describe the state of knowledge in words. The words are only as strong as the counted evidence behind them.

• Statements, not abstracts. A page's sentences come from statements: one plain claim each, linked to its evidence (study, cohort, source, and whether it supports, contradicts or finds no difference). The AI's short version is written only from approved statements, with the checks in § 7.

• Strength words are computed by code, never picked by a model, counting independent cohorts through the registry:

  • Seen once: one cohort, a pooled-only study without per-cohort numbers, or one cited study.
  • Seen more than once: F1's minimum.
  • Seen again and again: 3 or more independent cohorts agree, none clearly opposite.
  • Studies disagree.
  • Tested in a trial: its own line, saying what happened.

  They name what was counted, not a certainty grade (§ 3). Numbers stay one tap away.

• Ideas are attributed and labeled by where they have support: Proposed · In a dish · In animals · Seen in people · Tested in a trial.

  • "In animals" covers mice, rats, pigs, zebrafish and the rest. It was "In mice" until 2026-10-07.
  • When the animals carried human gut microbes (germ-free animals given a human microbe or a person's gut community), say so: "In animals · mice given human gut microbes". It's a closer test than an animal's own microbes, but still not people.
  • Only ideas the authors or a review put forward; never ours or a model's.
  • Never the lede, never in a heading.
  • Health effects appear only as open questions.

• Unknowns are said plainly. They come from the evidence itself (measured once per person, one country, no children, few people, no health outcomes), from the authors' stated limitations, and from reviews. Say "Not tested" only after a search for trials; until then, "Not in the studies we've checked".

• Narrative reviews are a source for ideas and unknowns, never for observations.

• Everything is dated ("As of Oct 2026"), and statements are versioned, so a change in the evidence shows.

§ 11Qualifiers: how research quality shows

Full list, sentences and fields: docs/research/research-quality.md.

• Describe, don't grade. At a glance, show only:

  • the design mark (● measured once · ■ followed the same people · ▲ a trial);
  • the strength word;
  • the two-question header;
  • read-with-care notes, by exception, when one applies.

  No "raises trust" tags, no counts of strengths, no scores.

• Computed, never typed. Qualifiers are views over facts recorded when a study is verified, in factor-studies.yaml (randomized, comparison group, how the exposure was measured, population, a replication group, registration, adjusted for, per-cohort results), and over a statement's evidence (very small difference, shares people, steps up with the amount, more than one country).

• Facts, decided 2026-10-05:

  • Absent, or not reported. A fact nobody has checked is absent. A fact the paper doesn't report is recorded as checked and not reported, with what was read and the date (not_reported). Both show nothing: no qualifier is computed from either, and nothing is guessed. Only the absent one stays on "Needs a person".
  • Conditions by name (population.conditions): each condition the paper counts, with its share. More than half with a gut condition (IBD, IBS, diabetes) reads "Mostly people with a condition"; any other condition is named, by its share: "About half had depression".
  • How the exposure was measured: self-reported (questionnaires, diaries, apps and buttons; "Self-reported"), an interview by a trained person (no qualifier), a device or records ("Measured, not remembered"), or assigned by the study.
  • A comparison group is asked only of trials and of studies that assigned or gave people something, never of observational ones.

• One sentence each, from one module (@probiome/tool-ui): what it means, and why it raises trust or calls for care. The wording rules of § 5 apply.

• "Landmark study" is computed: 5,000+ people, 2+ independent groups with their own numbers, shotgun, and full open per-microbe tables. It describes scale, not a verdict on studies we describe without loading.

• Never shown to readers as quality:

  • citation counts or the relative citation ratio;
  • the journal as a badge;
  • a "peer-reviewed" tag;
  • p- or q-values at a glance;
  • the licence or route ("described, not loaded");
  • the noteworthy score;
  • "new";
  • samples in place of people.

• Our words are not the research: AI-WRITTEN and EXPERT-REVIEWED stay in their banner, apart from research qualifiers.

§ 12Finding studies: routing and the "noteworthy" score (v1)

Routing is by facts, never by the score:

• Load: machine-readable per-microbe tables (spreadsheets or text) that list every microbe tested, under CC BY or CC0 (the article's, or a data repository's). Plus 500 or more people, or a randomized trial. Shotgun results are species-level; 16S results load at genus level only, onto genus pages.
• Cite: everything else in scope, at the level of the paper.

Non-commercial, no-derivatives and publisher-only terms are cite-only.

The score only ranks the queue for a person to review. It never publishes anything (study-score.ts, 100 points):

• People: 0–25, log scale.
• Independent cohorts: 1 → 0, 2 → 8, 3–4 → 12, 5+ → 15.
• Method: shotgun 10, 16S 4.
• Design: trial 10, prospective 8, cross-sectional 4.
• Per-microbe tables: machine-readable 10, PDF 3.
• Licence: CC BY/CC0 10, CC BY-NC 4.
• Confounders recorded: 2 each for age, sex, BMI, medicines and diet.
• NIH iCite RCR: up to 10, halved under 2 years. Journals aren't scored.

Gates: the study is in people, about the gut microbiome and about a listed factor. It isn't retracted, a protocol or a preprint, and it has an abstract. Narrative reviews are queued only as sources for ideas and unknowns (§ 10).

What reaches the queue: a score of 40 or more, every randomized trial, and the 5 best-scoring papers per factor. Every queued study is approved by a person, and what's shown is added in a pull request.

§ 13Worked examples

• F. prausnitzii in Crohn's: within the 8 studies that had both people with Crohn's and healthy people, fewer carry it in all 8. The row shows the strength word "Seen again and again"; "8 of 8 studies" is the count in the row's sheet.
• A. muciniphila in obesity: within the one study that had both groups it isn't clearly different, so it isn't listed. A pooled 68% vs 43% is never shown.
• Studies that disagree: when one study of a condition clearly points the other way, or only a minority of studies agree, the microbe isn't listed. It's counted in the condition's "Also compared" link, where it's named.
• Seen once: a microbe clearly different in the only study that had both groups is listed after the others, and never leads: not the page's first row, a home-page regular, an MCP card's headline or a share image.
• F. prausnitzii in Parkinson's: about 90% of healthy people carry it, so a difference can only show in the amount. The row reads "As many carry it" in ink, and is listed because the within-study amount test agrees in 3 of 3 studies. How much carriers have (about half as much) shows only in the row's sheet.
• Fackelmann 2025: the study reports abundance as a standardized mean difference, so the wording is "Higher in vegans", shown only when 2 or more of its cohorts agree and are the majority. Merged lines:
  • Alistipes putredinis reads "Higher in omnivores than in vegetarians or vegans" (5 of 5 · 4 of 5 cohorts).
  • Streptococcus thermophilus, a dairy microbe, reads "Higher in omnivores and vegetarians than in vegans".
  • [Ruminococcus] torques reads "Highest in omnivores, then vegetarians, then vegans" (5 of 5 cohorts in each pair).
• Vegetarians vs vegans (1,744 people): a small comparison, so its findings carry the power note, and "no clear difference" there says little.
• Two questions, antibiotics: Baldanzi 2026 (14,979 people, 3 groups pooled) answers "Who differs: Seen once" until its groups are checked one by one. Palleja 2018 (12 men, followed through a course, no comparison group) answers "What changes: Followed the same people · A small study". Neither changes the other's word.
• A very small difference: Faecalicoccus pleomorphus is higher in omnivores than vegans in 5 of 5 cohorts, but at d = −0.03. Its line carries "A very small difference".
• Coffee field notes: Lawsonibacter asaccharolyticus higher in people who drink more coffee across 7 cohorts is "Seen again and again". That coffee compounds help it grow in a dish is an idea, "In a dish", attributed to the authors. Whether drinking coffee changes it is listed under what nobody knows yet, because each person was measured once.
• B. longum butyrate: research says yes and genome analysis found it in 0 of 28 genomes, so it's Mixed evidence, never "Can".
• Coffee, per cup: a correlation with cups a day reads "Higher in people who drink more coffee", never "coffee raises it".
• Metformin and type 2 diabetes: a microbe more common in type 2 diabetes and also higher in metformin users gets the medicine note on the condition page; neither result is hidden.

§ 14probiome.ai: counted, read, written

Probiome marks what an AI model made (decided 2026-10-06: docs/design/screens/probiome-ai-provenance.html and DESIGN.md › probiome.ai. Lab notes, the brand and Insights are still proposals in docs/design/proposals/probiome-ai-*.html; their rules below apply when they ship). probiome.ai is the lab's name for the AI layer: the byline on AI work, the social handle, and the sender of lab notes. Probiome is the product and the data.

• Three sources, one label each:

  • Counted: code computed it from data (comparisons, "in N of M", strength words, trial counts, rating totals). It's the default and carries no label; methods link to How we evaluate research.
  • Read by probiome.ai: a model interpreted the person's own input: the Bristol type and color from a photo, the traced shape. The label shows only until the person confirms or corrects it; after that it's theirs ("Type 4 · you confirmed"), and How this was made stays in the specimen's details.
  • Suggested by probiome.ai: chips proposed from the person's own words in the intake. They show on white, dashed in their own facet color, under a byline line (no violet wash, which would clash with the facet colors). The person keeps or removes each one; kept chips turn solid, are their answers, and carry no label.
  • Written by probiome.ai: a model wrote sentences from sources we hold: summaries (§ 7), lab notes, Insights.

  Quoted words keep their quote marks and attribution. People may refine and regenerate any written thing; the page shows only the byline, never a "reviewed" or "not reviewed" badge ("Expert reviewed" stays as § 7 defines it).

• The byline is required wherever AI work appears: page, email, social post, MCP result. It reads: mark · probiome.ai · Read or Written · from what · date, plus "How this was made", which lists the sources, the model and prompt version, the checks, what was dropped, how that version was made (drafted by probiome.ai, refined or regenerated by people at Probiome), the rules it follows (this skill, with its version and sections, linking to its public page, Read the rules), and earlier versions with what changed. The violet wash is for AI-written sentences and readings waiting for an answer: counted numbers never sit in it, and AI-written sentences never sit outside it.

• Numbers inside written text come from code. § 7 applies to every written thing, lab notes and Insights included: the model can't add a number, and sentences whose numbers don't match are dropped.

• Lab notes: one numbered issue a week (No. 1, No. 2…), the same for everyone at launch; following a microbe, condition, factor or product comes later. A quiet week still gets its short issue, never padded. Code decides what's news, by diffing our data: studies loaded or cited, strength words that moved, products crossing 10 ratings, new pages. probiome.ai writes the sentences. Each item is dated, cites its study and links to its page, and never says more than that page. Every issue and every social card says "AI summary of the latest research" and "Not medical advice". Lab notes never mention a person's log.

  • Review: while review is on (at least the first 8 issues), a person approves each issue before it sends; unapproved issues wait. Numbers and facts are locked; only sentences can be riffed or edited, and every version is kept and re-checked. Leaving an item out needs a reason, and "it contradicts something" is never one.

• Voice:

  • A lab notebook, not a character: short, dated, specific, numbers and design first.
  • No "I", no name, no face, no feelings, no hype ("Big news!", "Scientists have discovered").
  • § 5's words apply everywhere, social included.
  • Wry is fine about the lab, never about a person.

• Insights (when the gut log opens; the design is still open). Patterns in a person's own log, which extends § 9 without loosening it:

  • Three layers, paced by data, not the calendar:
    • after each log, one counted line (code only, no AI);
    • a written insight once there's enough new data (about every 4 logged days);
    • Your month.
  • Their log only: never a study, other people, "people like you", a guess at their microbes, a score or a goal.
  • Every pattern states both sides: "On the 3 days you logged more fiber, all 3 were type 3–5. On the other 2 days, 1 was."
  • A pattern needs 3 days on each side. Under 10 days, it says "Patterns this small can be chance."
  • It describes what was logged, never advice ("eat more fiber"), never disease.
  • No separate opt-in: the gut log consent covers it ("…and to write insights about your log"). It's on by default, with one switch to turn it off.
  • A person's own data (history, counts, the after-log line) is never paywalled.
  • It's private (§ 9).

Open questions (to refine)

• "Very small" for other effect types (odds ratios, fold changes): set when a study first uses them, in the same spirit as Cohen's conventions.
• The small-comparison cut-off (2,000 people) and the queue's cut-offs (a score of 40; 5 papers per factor) are first guesses, to calibrate on the first real queue.
• Lab screens: the exact wording and where an "In a dish" tag can appear.
• One factor across methods: how a factor page shows a 16S genus result next to a shotgun species result for the same group of microbes.
• Our own community data: when the research pool can yield findings of its own, which rung of the ladder they start on (they're self-reported, without microbiome data).

Settled in v2: one visible certainty scale (no: § 3), the minimum replication for Factors (F1: § 4), mixing evidence types (never one number: § 4), and confounding by medication (the note: § 6). Settled in v3: design vs size (two questions: § 3), the at-a-glance signal (describe, don't grade: § 11), and very small differences: labeled and folded (§ 4, D21). Settled 2026-10-06: conditions are compared only within studies, and healthy guts are described, never compared (§ 4, § 6).