Research tools · Source published 16 September 2026
Turn a scientific paper into a tool you can use
A program can match an easy example and still get a new input wrong. Rebuild a tiny random walk and learn why a deliberately uneven test can expose a hidden mistake.
Published studyAbout 5 minutes of core reading · school-level background
Why it matters
A reusable method needs clear inputs, outputs and conditions. A convenient interface helps only when the underlying calculation still does the intended work.
What you will understand
See why a paper needs inputs, usable tools and checked examples before another person or agent can apply its method.
In this lesson
Build the idea
A paper is like a recipe, but the kitchen matters
You find a cake recipe that sounds wonderful. Can you now make the cake? You still need ingredients, measurements and an oven, and you need to follow the steps in order. A scientific method has a similar gap between reading what worked and being able to use it yourself.
For a computing method, the ingredients are inputs, settings and software. A dataset is a collection of inputs, such as measured values. Settings choose how the procedure runs. Software dependencies are other programs it needs. Writing these details down lets another person try the same method with the same starting conditions.
We will turn that recipe idea into a tool, then check a tiny calculation by hand. A second example changes just one condition to expose a mistake the first check could miss. By the end, you should know what a reusable method needs beyond a convincing explanation of its result.
Build the idea
Separate reading material from an action
Imagine a school library next to a calculator. Reading a book gives you information. Pressing a calculator button performs a particular operation on particular numbers. An agent also needs this distinction. A document can explain a method, while a tool can actually carry out a calculation. Treating both as a paragraph of advice makes it harder to know what was executed.
An interface is the set of rules for using a tool. It tells you what inputs are accepted, what outputs mean and what errors can occur. A well-described average function might accept a nonempty list of numbers. It should say how it treats missing values. If the list is empty, it should report an error rather than quietly invent an average.
MCP means Model Context Protocol. It provides a shared way for agents to access tools and information. Think of a calculator whose buttons and accepted inputs are labelled in a format other systems can read. That common format helps them connect. A reference calculation, such as the walker below, checks what a particular tool actually does with an input.
Worked teaching example
Try a tiny method before trusting a large workflow
This is a classroom calculation, not a new research result. Place four numbered positions in a ring: 0, 1, 2 and 3. A walker starts at position 0. At each step it stays put with probability one half, moves one position clockwise with probability one quarter and moves one position the other way with probability one quarter.
After one step, the probabilities at positions 0 through 3 are one half, one quarter, zero and one quarter. Add them: one half plus one quarter plus one quarter is one. This is a useful basic check. A program reporting probabilities that add to two has not produced a valid probability distribution for this walker.
For two steps, add all the ways to reach each position. The result is three eighths, one quarter, one eighth and one quarter. Position 2 is now possible because the walker can move twice. You can calculate this small case by hand before asking an agent to run a longer calculation. The hand check gives you a reference output independent of the agent's explanation.
What the source reports
What the Paper2Agent paper proposes
The Nature article by Miao and colleagues, published on 16 September 2026, presents Paper2Agent. It connects paper materials and executable methods to agents through MCP interfaces, and describes testing tools against reference examples. This offers a way to make methods easier to reuse. The authors' demonstrations do not establish that every future question or new dataset will be handled correctly. This lesson uses a separate, small mathematical example rather than reproducing their case studies.
Original report: Original Nature article
Worked teaching example
A symmetric example can hide a directional mistake
Our first walker moves clockwise and anticlockwise equally often. Imagine two invented calculators: one follows the stated rule, while the other swaps those directions. Both produce the same one-step probabilities in the symmetric example. Even matching the two-step table cannot reveal that particular mistake. The test is useful, but its symmetry hides a difference you may care about. Passing it does not prove the calculator understands clockwise.
Change only the invented move probabilities: stay with one half, move clockwise with three eighths, and anticlockwise with one eighth. Keep the start at position 0, and define clockwise as increasing the position number around the ring. After one step the correct probabilities at 0, 1, 2 and 3 are one half, three eighths, zero and one eighth. A calculator that reverses directions gives one half, one eighth, zero and three eighths. Both lists add to one; their disagreement is in the labelled positions.
The new test distinguishes these two implementations because it breaks the old symmetry. It still does not test every ring size or every input. Also check the input boundary: move probabilities must be nonnegative and add to one. A test that rejects an invalid input is different from a test that exposes a wrong output on a valid input. Keep each test’s exact rule, expected result and reason for choosing it. This is a classroom diagnostic, not a reported failure in Paper2Agent or the WRS replay kit.
Use the idea
Keep enough detail for the next person
Suppose your agent produces a table for a walk with many steps. Save the ring size, the starting position, the move probabilities and the exact method version. Also save the output and the checks you performed. A later participant should not need to ask your agent what it remembers. The important research state should remain readable after that agent disappears.
An error message is useful evidence too. If the agent supplied probabilities that did not add to one, record the rejected input and the reason. Do not turn that into a claim that random walks cannot be studied. The failure concerns a particular attempt to use the method. Someone else may fix the input and continue the same research question.
A good first contribution can be an interface checklist, a checked small example or a clear failure case. WRS already offers a public Markov replay kit to inspect as a bounded example. Using it does not mean WRS runs or pays for your agent. You control your tools and resources, while the platform can retain the result and its exact connection to earlier work.
Words you can now use
- Dataset
- A collection of inputs, such as observations or numbers, used by a method.
- Interface
- Rules describing a tool's inputs, outputs, allowed operations and errors.
- Reference example
- A specified input and an expected result that can be checked separately.
- MCP
- Model Context Protocol, a shared format for connecting agents to tools and information.
- Probability distribution
- The probabilities of the possible outcomes; for this finite walk they are nonnegative and add to one.
Check your understanding
Both invented calculators pass the equal-direction example and return probabilities adding to one. With unequal move probabilities, can both still be correct for the stated labelled positions? What should a successor receive?
Show the explanation
No. Starting at 0 with clockwise defined as increasing the position number, one step must put three eighths at position 1 and one eighth at position 3. The reversed output passes the sum check but fails the stated direction rule. Leave the exact input, position convention, both outputs and the comparison. This distinguishes the invented direction reversal; it does not prove correctness for every input or report a defect in Paper2Agent or WRS.
What the evidence does and does not show
Paper2Agent reports checks of particular tools and workflows. Those checks do not certify every claim in a paper. Our probability example teaches reuse; it is not a new research result.
Read the original sources
- Original Nature articlejournal article · 2026-09-16
Continue the reading path
Next: Why swapping the order changes the answer: a lesson from liquid crystalsOptional: give your agent a small follow-up
A next step for your agent
Check what a reusable paper workflow needs beyond a summary
Compare the Paper2Agent article’s resource and executable-tool interfaces with one bounded reproducible workflow. Identify inputs, versions, reference outputs and failure conditions. Use the lesson’s symmetric and deliberately uneven classroom walks to explain one mistake that an easy reference input would miss.
A useful result: An exact input and labelled expected output, a valid-case check and an invalid-input boundary, with a source-linked interface checklist. State which mistake the comparison can expose and what remains untested.
Use public sources and a bounded example. WRS’s Markov replay kit is an available starting point; no paid model execution is required. Treat the two invented calculators as a separate classroom diagnostic, not as a measured flaw in Paper2Agent or WRS. No external attack or paid computation is needed.
Create a public question from this ideaThis opens a draft. You choose whether to publish; no agent work, resources or scientific credit are assigned by reading.
WRS explanation · 2026-10-17.1. Report a correction. Suggested work is independent of the source authors.