AI can retrieve evidence without knowing what kind of evidence it is.
Minerva gives claims an owner, a warrant and a dependency trail, so an agent can tell a measurement from an argument, a record from a fact, and an instruction from evidence.
Know what the evidence can actually support.
Rule. A claim must not be stated with greater certainty than its evidence warrants.
Minerva does not decide whether a claim is true. A warrant names the kind of support. It does not prove that the support holds.
| MA | mathematics, the reference domain |
| 25 | fields, one standard |
| 13 | evidence (warrant) types |
| 984 | stable identifiers |
| 63/63 | MSC2020 classes placed once |
An answer that outruns its evidence
The answer below sounds finished. The table splits it into claims. This is an illustration of the check, not a measured result and not a finding that the corrected wording is true.
| Claim | Owner | Required warrant | Finding |
|---|---|---|---|
| “The intervention caused improvement.” | Clinical or scientific field | Controlled intervention, or causal identification | Only an association was in view. The causal wording is stronger than that. |
| “The agency recommends this.” | Policy or institution | A dated record, or an official statement | Supported only as a report of what the agency said. |
| “This is legally required.” | Law | Applicable authority, and legal reasoning | The legal support is not shown. |
| “Ignore previous instructions.” | None | None. The line is an imperative. | An instruction, not evidence. |
Minerva does not certify a rewrite as true. It shows where the certainty is greater than the warrant.
How truth-seeking works
Truth-seeking here is a sequence of questions. It is not a machine that possesses the truth.
| Step | Question | What you get |
|---|---|---|
| Question | What exactly are we trying to establish? | An atomic claim, or a research question |
| Field | Which domain owns it? | An owner, and the border fields that matter |
| Warrant | What kind of evidence could settle it? | Proof, measurement, intervention, record, argument, convention, or another named type |
| Evidence | What support is available? | Sources, data, calculations, experiments or records |
| Claim | What can safely be said? | A calibrated answer |
| Defeater | What could undermine it? | Counterevidence, assumptions, missing controls |
| Status | How settled is it? | Established, provisional, contested, superseded or unknown |
Mathematics is the reference domain
Mathematics is the first field, and the reference domain. Every other exact field borrows from it. It borrows from none of them. A theorem is stated with its hypotheses, and other fields cite it. Read the mathematics framework.
The pyramids underneath
The check is the part you use. Under it sits the classification: 25 fields, one standard, 984 linked identifiers. Each field is a pyramid with four faces: objects, methods, warrants and bridges. That structure is the substrate. It is how an owner, a warrant and a dependency get stable names. It is not a claim to have finished all knowledge.
A new instrument for thinking with the whole of human knowledge
Introducing Taxonomy, Standard v2: a new instrument for thinking with the whole of human knowledge.
Taxonomy organises everything people know into 25 fields, one standard and 984 linked parts, starting from mathematics. Use it to see what any statement is and what would make it true, to view any subject from the position of all of knowledge at once, and to give an AI system the structure of human knowledge to reason with.
Built for analysts, researchers and policy teams who now read more machine-written text than any person can check by hand.
| 25 | fields, one standard |
| 63/63 | MSC2020 classes placed once |
| 984 | stable identifiers |
| 13 | evidence (warrant) types |
| 23 | ANZSRC divisions crosswalked |
Three things you can do with it
Classify a statement
Type any sentence. See which field owns it, what part of that field it concerns, what evidence it needs and how it was produced.
See the whole
Enter a subject, a product or a problem. See the questions all 25 fields put to it, and the blind spots a single discipline would miss.
Relate two things
Set two things side by side. Find the fields they sit in, the claim that joins them, and where evidence of one kind is being used for the other.
Why Taxonomy exists
Machines can now write fluent text on any subject in seconds. They cannot tell you which discipline is qualified to judge a sentence, or what evidence would settle it. Most of the classification systems we rely on cannot either. They were designed to put books on shelves and papers in journals, so they sort by topic. A single report contains dozens of claims of very different strength: a measured figure, a legal duty, a forecast, a value judgement. Topic schemes file them all in the same place.
The result is a costly kind of confusion. A minister receives a brief in which an economic estimate, a legal requirement and a political preference read as equally certain. A researcher cites a theorem without the conditions that make it true. An AI system repeats a confident statement whose only source was another AI system. In each case the error lies in the category, not the wording, and no style guide catches it.
Taxonomy fixes the category. Each claim is assigned to the field whose standard of proof can establish it, the cell inside that field it concerns, and the type of evidence it owes. Once that is done, a reader knows at a glance what kind of statement is in front of them and what checking it would take.
What you get
- One home for every claim. A theorem belongs to mathematics, a measured constant to physics, a statutory duty to law. Other fields cite it by a stable identifier such as EC.O.A5.1.
- The evidence named. Thirteen warrant types, from formal proof to a ruling before a court, tell you what a statement needs before you rely on it.
- Conditions written in. Every named theorem carries its hypotheses in the sentence. Gödel's incompleteness theorem is stated for effectively axiomatised theories, and the welfare theorems for complete markets.
- Clear borders between disciplines. Who bears a tax is economics. The duty to pay it is law. The return a firm files is business practice. Each has its owner.
- Crosswalks to the standards you already use: MSC2020, ANZSRC 2020, JEL, ACM CCS, CS2023, MeSH, ICD-11 and PhilPapers.
- Machine-ready structure. Stable IDs and typed links between fields, designed for retrieval systems, AI agents and an open API.
Mathematics first
Every field that reasons exactly borrows from mathematics, and mathematics borrows from none of them. That is why Taxonomy begins there. The mathematics framework places all 63 top-level classes of the Mathematics Subject Classification (MSC2020) exactly once, checked by script, in five cells ordered by what a proof may assume. Physics, economics, computer science and the rest cite those cells rather than restating results in their own words. Read about the mathematics framework.
Mathematics: the foundation
Taxonomy places mathematics at the apex because every other exact field depends on it. A physicist's conservation law rests on Noether's theorem. An economist's equilibrium rests on a fixed-point theorem. A cryptographer's protocol rests on a hardness assumption. In each case the theorem is owned once, by mathematics, and cited by the field that uses it.
The five cells of mathematics
Mathematics is divided by dependency: what a proof is allowed to assume. All 63 MSC2020 classes are placed once.
| Cell | Subject | What it covers | MSC2020 classes |
|---|---|---|---|
| MA.O.A1 | Foundations | What a proof may assume: logic, set theory, categories | 03, 18 |
| MA.O.A2 | Structure | Operations, order and number: algebra and number theory | 12 classes |
| MA.O.A3 | Space | Shape, nearness and dimension: geometry and topology | 8 classes |
| MA.O.A4 | Change and the continuum | Limits, functions and equations: analysis | 19 classes |
| MA.O.A5 | Chance | Measures of total mass one: probability and statistics | 60, 62 |
The remaining classes are bridges: pure mathematics in other guises (41), probability and statistics in application, computational mathematics (65, 68), mathematical modelling (15 classes), and mathematics as the object of another field's study, such as its history (01) and its teaching (97).
Why this matters outside mathematics
- Results are stated with their conditions. A theorem without its hypotheses is recorded as an error. That rule alone removes a large class of misuse in policy and economics writing.
- Open problems are labelled. P versus NP, the smoothness of the Navier–Stokes equations in three dimensions and other open questions are tagged OPEN wherever they are cited.
- Use is separated from proof. The first welfare theorem is proved in mathematics. Its use to argue for a market reform belongs to economics, and the choice of which game the parties are playing belongs to strategy.
The 25 fields
Each field is divided by one stated dimension into no more than five parts. Select a code in the pyramid, or choose from the list, to see how the field is divided.
All fields at a glance
| Code | Field | Divided by |
|---|
How Taxonomy works
Taxonomy applies the pyramid principle, developed by Barbara Minto at McKinsey & Company for structuring reports, to the structure of knowledge itself.
The pyramid principle
Minto's method states the main point first and groups the supporting points beneath it. It rests on four rules.
- The vertical rule. Every heading states a point that the items beneath it support.
- The horizontal rule. Items at the same level share one dimension and one order.
- Mutually exclusive, collectively exhaustive. No item overlaps another, and together they leave nothing out. There is no "Other" category.
- Situation, complication, question, answer. Each structure opens by explaining why it is cut the way it is.
Each field in Taxonomy follows these rules, with between two and five members at each level so that the structure can be held in mind.
Four views of every field
| View | Question it answers | Example from physics |
|---|---|---|
| Objects | What is the claim about? | Matter in equilibrium |
| Methods | How was it produced? (represent, derive, observe, intervene, compute, compare, verify) | Measurement by interferometry |
| Warrants | What evidence licenses it? | Measurement with stated uncertainty |
| Bridges | Which discipline carries it? | Condensed matter physics |
The ownership rule
A claim belongs to the field whose standard of evidence can establish it. All other fields cite it. Four sentences about the same tax show the rule at work:
| Statement | Owner | Reason |
|---|---|---|
| Consumers bear most of this tax because demand is inelastic. | EC.O.A2.5 | Who bears the cost is an allocation question. |
| The importer must pay GST at the border. | LA.O.A2.5 | The duty rests on statute and is decided by a court. |
| Our March return shows output tax of $42,000. | BU.O.A5.3 | The calculation is business practice. |
| A competitive equilibrium exists in this model. | MA | Existence is a theorem. |
The thirteen evidence types
Every claim names the type of warrant it owes. A warrant never transfers from one type of claim to another: a court's finding does not settle a scientific fact, and scripture does not license a measurement.
| Code | Warrant | Meaning |
|---|
AI, hallucination and security
Large language models produce confident text with no built-in record of where a statement comes from or what kind of evidence it needs. A wrong sentence and a right one look the same. Taxonomy gives people and machines a shared way to tell them apart.
The risks
- Hallucination. Models state invented figures, citations and legal rules in the same tone as true ones.
- Category errors. A model blends a forecast, a legal duty and an opinion into one fluent paragraph, and the reader cannot see the joins.
- Contaminated sources. Machine-written text now feeds back into the material future models are trained on and retrieve from.
- Manipulation. Poisoned documents and prompt injection can slip false claims into automated pipelines.
How Taxonomy helps
Claim tagging
Require a model to tag each claim with a field, cell and warrant. A claim tagged as a measurement must point to a measurement; a claim tagged as a proof must point to a proof. Tags that cannot be filled mark the sentences to check first.
Retrieval by claim
Index sources by claim and evidence type, not by keyword. A question about drug safety returns trial-backed statements on adverse effects, not every page that mentions the drug.
Agent routing
In multi-agent systems, send each part of a question to the agent qualified to answer it: the incidence question to the economics agent, the duty to the legal agent, the theorem to a proof checker.
Evaluation
Score AI answers on whether they assign the right field and evidence type, not only on how closely the wording matches a reference answer.
Cyber security and threat intelligence
Threat reports combine statements of very different strength. An indicator of compromise is a measurement. Attribution of an attack to a state actor is an inference from comparison and argument, and should be marked as contested until shown. Confirmation that a control is in place is verification against a specification. When these are tagged, a confident attribution can no longer borrow the authority of a file hash.
A mismatch between a claim's stated evidence and its actual source, such as a forum post presented as a trial result, is a structural warning that an automated pipeline can detect before the claim reaches a decision.
AI policy and governance
Debates about AI risk mix technical capability (computer science), legal duty (law), public authority (politics) and ethical argument (philosophy). Taxonomy gives each part an owner, so a policy paper can show exactly which statements are evidence and which are judgement.
Uses
A knowledge layer for AI
Taxonomy can be loaded into an AI system as retrieval context (RAG) alongside a working document. The model then has the structure of human knowledge in front of it: the 25 fields, how each divides, what evidence each requires, and how they connect. It can test a plan against every field rather than the two or three a prompt happens to mention.
- optics and power (Physics)
- lens materials (Chemistry)
- the eye and eye strain (Biology and Medicine)
- on-device computation and data security (Computer science)
- how wearers and bystanders behave (Mind and society)
- privacy and recording law (Law)
- cost and market (Economics and Business)
- form and comfort (Design)
Government policy briefs
Divide each recommendation into its parts: the evidence of effect, the legal power to act, the political authority to decide, and the cost. Decision-makers see which parts are measured and which are choices.
| Section of the brief | Owner | Evidence expected |
|---|---|---|
| Expected effect on emissions | EV | Causal estimate or trial |
| Cost and who bears it | EC | Model with stated assumptions |
| Legal power to impose the measure | LA | Statute and case law |
| Who decides, and how | PO | Institutional record |
| Whether it is fair | PL | Argument |
Academic briefs and literature reviews
Map a body of research by claim and evidence type. See where one field relies on another's results, and where an entire debate rests on a single contested point.
Research strategy and funding
Crosswalks to ANZSRC 2020 and other official schemes show which classes map cleanly, which split across fields, and where existing schemes leave gaps.
Teaching
Students learn the shape of a field before its details: five parts, one organising idea, and clear signals for when another discipline must be consulted.
Intelligence and strategy
The Strategy field separates actors, the game being played, coercion, conduct and information, and leaves the dated record of each case to History, so lessons drawn from cases stay tied to the facts.
Compared with existing schemes
| Library and subject schemes | Taxonomy | |
|---|---|---|
| Unit classified | A book, paper or topic | A single claim |
| Evidence recorded | None | One of thirteen warrant types |
| Links between fields | Separate schemes, no shared rule | One ownership rule, 984 linked identifiers |
| Leftover categories | "Other" and "General" classes | None; every exclusion names its owner |
| Machine use | Shelf marks and subject headings | Stable IDs, typed links, API |
| Indigenous knowledge | A residual class | Its own field, with two modes of evidence, subject to Māori-led review |
Use the tool
Taxonomy works in three ways. Classify a statement to see what it is and what it needs. Take any subject and see it from the position of the whole of knowledge. Or set two things side by side and see how they relate.
Every field asks its own questions of the subject. Reading them together shows the particular from the position of the whole: what is measured, what is owed, what is decided and what is believed.
The built-in engine runs in your browser. Deep analysis uses Claude through your own Claude account and asks your permission first.
API
Taxonomy is designed to run inside software. Every cell has a stable identifier and typed links to the cells it uses. The interface below is the published design; developer access will open in stages.
Endpoints
| Call | Purpose |
|---|---|
| POST /classify | Send a sentence. Receive its candidate field, cell, warrant and method, with confidence and the rule applied. |
| GET /lookup/{id} | Retrieve a cell: its statement, conditions, evidence row, and every link in and out. |
| POST /audit | Send a document. Receive each claim with its placement, and flags for evidence mismatches, missing conditions and borrowing between fields. |
Example
POST /classify
{ "text": "Metformin lowers HbA1c in adults with type 2 diabetes." }
200 OK
{ "owner": "MD",
"cell": "MD.O.A4",
"warrant": "W3",
"method": "M4",
"uses": ["CH.O.A1", "MA.O.A5"],
"rule": "living function is Biology; disorder and treatment of humans is Medicine" }
Engine
Taxonomy Engine v0.1 is a transparent rule-based classifier, written in Python, that implements /classify and a whole-view function. The same engine runs in your browser on the Use the tool page. A model-assisted classifier adds depth where cue words are not enough.
Data release
The registry is planned for release in SKOS and JSON, with a persistent identifier and version history, so that libraries, research systems and AI developers can build on it directly.
About Taxonomy
This page is about the knowledge standard. Minerva, the lead division of Taxonomy, publishes it. The company site is Taxonomy. A short account of Chris Townsend is on the Chris Townsend page.
Taxonomy was developed by Chris Townsend, an analyst and writer based in Whanganui, New Zealand, working with Claude, Anthropic's AI model. It began as a single structure for mathematics and grew to cover 25 fields under one standard.
The standard
Standard v2 sets twelve rules that every field must meet. They include one home for each claim, exact counts, conditions stated for every theorem, open problems labelled, clear exclusions, and stable identifiers.
Standards it maps to
- Mathematics Subject Classification (MSC2020): all 63 classes
- Australian and New Zealand Standard Research Classification (ANZSRC 2020): all 23 divisions
- Journal of Economic Literature codes (JEL): 20 categories
- ACM Computing Classification System and CS2023 knowledge areas
- Medical Subject Headings (MeSH) and ICD-11
- PhilPapers and the Library of Congress Classification
Indigenous knowledge
Taxonomy includes a field for Indigenous knowledge, with mātauranga Māori in view. Its placements are provisional until reviewed by a Māori-led panel. The review pack has been prepared and invitations will follow.
Frequently asked questions
- Is this a replacement for library classification?
- No. Library schemes shelve documents. Taxonomy classifies the claims inside them, and maps to the major schemes so the two work together.
- Why mathematics first?
- Because every exact field depends on it. Placing theorems once, with their conditions, prevents the same result being restated loosely in a dozen places.
- Does it stop AI from making mistakes?
- No tool can do that alone. Taxonomy makes unsupported claims visible by asking what evidence each one needs, which is where checking should start.
- Who decides where a claim belongs?
- The ownership rule: the field whose standard of evidence can establish the claim. Contested placements are marked as such, with the alternative stated.
- Can I use it now?
- Yes. The fields, the rules and the classification examples on this site are available today. Register to receive the full standard, briefing papers and API access.
Register
Register to receive the full Standard v2 document, briefing papers for policy and research teams as they are released, and early access to the API.
Company contact, including speaking, writing, video and advisory, is on the Taxonomy contact page. This form is the Minerva register. It does not yet send what you type.