--- # trust.md — a portable, human- and machine-readable EPISTEMIC TRUST manifest # Proposed convention (v0.1). Reference implementation: https://neuronautix.com/trust.md # # Distinct from JournalList's trust.txt, which declares an organisation's TRUSTED # RELATIONSHIPS (memberships, ownership, vendors). trust.md instead declares the # EPISTEMIC STATUS and CONFIDENCE of the knowledge a repository publishes: how it # was produced (human / AI), how each claim is graded, how it is reviewed, and how # trustworthy the corpus is — the repository-level companion to per-claim, # inline epistemic markup (here: the "Trust Lens"; see /notes/EPISTEMIC-MARKUP.md). # # Aligns with established assertion/provenance vocabularies: W3C PROV, PAV, # nanopublications (assertion + provenance + pubinfo), SEPIO (Monarch Initiative), # the Evidence & Conclusion Ontology (ECO), and schema.org ClaimReview. trust_md_version: "0.1" title: "Neuronautix — Trust & Epistemic Provenance Declaration" description: > How knowledge and inferences in this repository are produced, graded, and reviewed — and how confident the reader should be in them. canonical: "https://neuronautix.com/trust.md" license: "Apache-2.0" # matching the DESIGN.md format this repo follows companions: fair: "/fair.md" inline_markup_spec: "/notes/EPISTEMIC-MARKUP.md" # ── Who and what produces the knowledge (provenance of authorship) ── produced_by: humans: - name: "Damien Huzard, PhD" orcid: "0000-0003-4820-7951" role: "author, domain reviewer, accountable signatory" org: "Neuronautix" agents: - name: "Claude (Anthropic), via Claude Code" role: "retrieval, drafting, epistemic markup, summarisation" oversight: "human-reviewed" # no agent output is published without human review # ── Governance ── governance: source_of_truth: > The human-approved record — not raw model output — is the source of truth. AI-assisted text is retrieval-grounded and verified before publication. no_fabricated_citations: true review_policy: > Every published claim is human-reviewed. Claims are tagged with an epistemic category and a 0–100 confidence score (see epistemic_model) that the reviewer can audit and adjust. correction_policy: > Errors are corrected in place with a note in the page's changelog / git history. Substantive corrections are dated. conflict_of_interest: > Neuronautix is a scientific consultancy. Promotional or normative statements are explicitly labelled as the 'view' category, not presented as cited fact. # ── The epistemic model (mirrors the inline Trust Lens) ── epistemic_model: categories: - {id: "cited", label: "Cited fact", definition: "Directly supported by a cited source"} - {id: "consensus", label: "Established consensus", definition: "Widely accepted domain knowledge / standard definitions"} - {id: "inference", label: "Inference", definition: "Reasoned from one or more sources; not stated verbatim"} - {id: "hypothesis", label: "Hypothesis", definition: "Forward-looking or speculative claim"} - {id: "view", label: "Neuronautix view", definition: "Explicit interpretation, position, or normative conclusion"} confidence_scale: type: "integer" range: [0, 100] independent_of_category: true # a 'view' may be sincere but low evidentiary trust bands: - {range: "90-100", label: "Very high", meaning: "directly stated in a primary/peer-reviewed/regulatory source"} - {range: "70-89", label: "High", meaning: "stated in a cited source, secondary or lightly interpreted"} - {range: "50-69", label: "Moderate", meaning: "reasonable inference, or consensus without a pinpoint citation"} - {range: "30-49", label: "Tentative", meaning: "plausible forward-looking claim with partial support"} - {range: "0-29", label: "Speculative", meaning: "normative/opinion/vision with little direct evidence"} encoding: inline: 'HTML spans: ' reader_tooling: "Trust Lens (assets/js/trust-lens.js) — toggleable highlight, color-blind-safe palette, per-note summary" alignment: - "W3C PROV / PAV — provenance & authoring of assertions" - "Nanopublications — assertion + provenance + publication info" - "SEPIO (Scientific Evidence & Provenance Information Ontology)" - "Evidence & Conclusion Ontology (ECO)" - "schema.org ClaimReview" # ── Corpus-level trust profile (auto-derivable from inline markup) ── corpus: notes_with_markup: 19 total_claims: 440 average_trust: 74 category_distribution: {cited: 233, view: 97, inference: 85, consensus: 13, hypothesis: 12} # ── Per-artifact trust profiles (machine-readable; derived from inline markup) ── # avg = mean data-trust; dist = claim counts per category. artifacts: - {path: "/notes/2026-05-namo-new-approach-methodology-ontology/", claims: 40, avg: 88, dist: {cited: 36, inference: 3, view: 1}} - {path: "/notes/2026-05-nam-data-management-ind-submissions/", claims: 29, avg: 83, dist: {cited: 22, inference: 4, view: 3}} - {path: "/notes/2026-05-llm-kg-metadata-architecture/", claims: 43, avg: 81, dist: {cited: 38, consensus: 1, inference: 2, view: 2}} - {path: "/notes/2026-06-data-welfare-animal-welfare/", claims: 25, avg: 80, dist: {cited: 15, consensus: 2, inference: 6, hypothesis: 1, view: 1}} - {path: "/notes/2026-05-after-the-knowledge-graph/", claims: 46, avg: 78, dist: {cited: 21, consensus: 5, inference: 5, view: 15}} - {path: "/notes/2026-05-neuronautix-as-research-infrastructure/", claims: 23, avg: 77, dist: {cited: 15, inference: 3, view: 5}} - {path: "/notes/2026-05-ai-ready-nam-data/", claims: 20, avg: 76, dist: {cited: 10, inference: 5, view: 5}} - {path: "/notes/2026-05-meet-the-founder-damien-huzard/", claims: 15, avg: 76, dist: {cited: 12, view: 3}} - {path: "/notes/2026-05-fairscape-biocompute-nam-provenance/", claims: 15, avg: 74, dist: {cited: 4, inference: 6, view: 5}} - {path: "/notes/2026-05-metadata-as-a-nam/", claims: 35, avg: 74, dist: {cited: 14, consensus: 2, inference: 9, view: 10}} - {path: "/notes/2026-05-data-first-fair-by-design/", claims: 12, avg: 72, dist: {cited: 5, inference: 2, view: 5}} - {path: "/notes/2026-05-metadata-as-nam/", claims: 11, avg: 72, dist: {cited: 2, consensus: 3, inference: 3, view: 3}} - {path: "/notes/2026-05-schema-first-metadata-agent-nam/", claims: 11, avg: 72, dist: {cited: 4, inference: 4, view: 3}} - {path: "/notes/2026-05-post-hoc-curation-fails-nam-platforms/", claims: 15, avg: 70, dist: {cited: 6, inference: 5, view: 4}} - {path: "/notes/2026-05-linkml-preclinical-metadata/", claims: 8, avg: 68, dist: {cited: 4, inference: 2, view: 2}} - {path: "/notes/2026-05-nam-metadata-mandatory-qualification/", claims: 21, avg: 68, dist: {cited: 7, inference: 10, view: 4}} - {path: "/notes/2026-05-human-loop-ai-ready-preclinical-data/", claims: 11, avg: 59, dist: {cited: 4, inference: 3, view: 4}} - {path: "/notes/2026-05-nam-evidence-commons-tool-stack/", claims: 34, avg: 55, dist: {cited: 7, inference: 10, hypothesis: 1, view: 16}} - {path: "/notes/2026-06-hcm-2050-vision/", claims: 26, avg: 54, dist: {cited: 7, inference: 3, hypothesis: 10, view: 6}} limitations: - "Confidence scores are author/reviewer judgements, not statistical measures." - "Self-declared; not independently audited." - "A single 0–100 score currently conflates evidence strength and stated confidence (a planned second axis would separate them)." last_reviewed: "2026-06-06" --- # trust.md — Neuronautix Trust & Epistemic Provenance This file declares **how much you should trust the knowledge in this repository, and why**. It is the repository-level companion to the inline *Trust Lens* — the per-claim epistemic markup readers can toggle on any note. The YAML block above is machine-readable; this prose is for people. It is a **proposed convention (v0.1)**, with this repository as its reference implementation. ## Why trust.md? Scientific writing — and now AI-assisted scientific writing — mixes *cited fact*, *reasoned inference*, *forward-looking hypothesis*, and *the author's own position*, usually with no visible distinction between them. As large language models help draft more of the literature, the question "**what kind of statement is this, and how confident is the author?**" becomes urgent and, today, largely unanswerable from the page itself. `trust.md` answers it at the level of the whole repository: - **Provenance of authorship** — which parts are human-written, which are AI-assisted, and what oversight applies. - **A grading model** — five epistemic categories and a 0–100 confidence scale. - **A corpus profile** — honest aggregate statistics, derivable directly from the inline markup. It is deliberately *distinct from* JournalList's **`trust.txt`**, which declares an organisation's trusted *relationships* (memberships, ownership). trust.md is about the *epistemic status of the content itself*. ## How knowledge here is produced This repository is authored by a human domain expert (Damien Huzard, PhD) with AI assistance (Claude, via Claude Code) for retrieval, drafting, and epistemic markup. The governing rule is simple: **the human-approved record, not raw model output, is the source of truth.** AI-assisted text is retrieval-grounded and verified before publication; fabricated citations are not permitted. Promotional or normative statements — unavoidable for a consultancy — are explicitly tagged as the *view* category rather than dressed up as fact. ## The grading model Every marked claim carries two independent axes: 1. **Category** (*what kind of statement*): `cited` · `consensus` · `inference` · `hypothesis` · `view`. 2. **Confidence** (*how well supported*): an integer **0–100**, bucketed into five bands from *Speculative* to *Very high*. These are independent: a *view* can be sincerely held but low in evidentiary trust; a *hypothesis* can be well-motivated yet tentative. The same model is encoded inline (`data-epi` + `data-trust` on each claim) and rendered by the Trust Lens; see **[`/notes/EPISTEMIC-MARKUP.md`](/notes/EPISTEMIC-MARKUP.md)**. ## Corpus trust profile As of the last review, the marked corpus is: - **19 notes · 440 graded claims · average confidence 74 / 100.** - **Category mix:** 53% cited · 22% Neuronautix view · 19% inference · 3% consensus · 3% hypothesis. The most evidence-dense notes (e.g. the NAMO and IND-submission notes) average **88** and **83**; the explicitly forward-looking ones — the *HCM 2050 vision* (avg 54, ten hypotheses) and the *NAM Evidence Commons* build-plan (avg 55, sixteen *views*) — score lowest **by design**. A low average here is a feature: it tells the reader the piece is a position or a prediction, not a literature report. The per-artifact table in the YAML above makes this auditable. ## Relationship to existing standards trust.md is a lightweight *front door*, not a competitor, to the formal assertion/provenance stack: - **W3C PROV / PAV** — provenance and authoring of assertions. - **Nanopublications** — the assertion + provenance + publication-info pattern; trust.md's per-claim model is a pragmatic, web-native cousin. - **SEPIO** (Monarch Initiative — the same project behind NAMO, which this repo documents) and the **Evidence & Conclusion Ontology (ECO)** — formal evidence and assertion modelling. - **schema.org `ClaimReview`** — a path to emit each graded claim as harvestable JSON-LD (planned). ## How to adopt trust.md 1. Copy this file to your repo / site root as `trust.md`; serve it at `https://yourdomain/trust.md`. 2. Fill in `produced_by` and `governance` honestly — especially the human/AI split and the review policy. 3. Adopt the `epistemic_model` (or your own), and mark claims inline so the `corpus` and `artifacts` profiles can be generated automatically rather than asserted. 4. Pair it with **[`/fair.md`](/fair.md)** so a reader gets both halves: *can I find and reuse this?* (FAIR) and *how much should I trust it?* (trust). ## Limitations - Confidence scores are expert judgements, not statistical measures. - This is a self-declaration, not an independent audit. - The single 0–100 score currently blends *evidence strength* and *stated confidence*; a planned second axis would separate them. ## Changelog - **v0.1 (2026-06-06)** — first draft of the trust.md convention and this reference implementation, derived from the repository's inline Trust Lens markup.