The Epistemic Protocol Machine

A useful answer often takes more than one calculation. You want to explore an alternative, compare its consequences, understand a difference and decide what to try. The Epistemic Protocol Machine, or EPM, gives an AI assistant a structured way to pursue those questions while keeping its reasoning connected to evidence.

That matters when simulation becomes available to people who have never used it before. An AI assistant can explain a result fluently and still claim more than the calculation shows. A person should be able to ask an ambitious question and examine how the answer was reached. EPM decides what each result entitles the assistant to say. It does not make a model more accurate. It labels each result with the kind of work behind it and how strong a claim it supports, and it refuses steps the evidence cannot support.

Follow a human example · Read the technical reference

Take a question further

Imagine Rosa comparing three permitted spots for her pupusa truck. She explores when potential customers pass each one. Then she asks why one location looks promising and what changes on a day when more people work from home.

Those questions call for different work. Comparing activity around the three spots needs its own support. So do explaining why the pattern changes on a quieter office day and checking how much the comparison varies from one run to the next. EPM lets the assistant connect them while keeping what each answer establishes in view.

The person can follow the work, challenge an assumption and continue from an earlier result. The work has a life beyond a single answer in a chat.

Ten ways to take the question further

The protocol gives the assistant ten ways to take a question further. You can ask for the next useful step in ordinary language.

Each step has rules about the evidence it needs and what it can support. Some steps cannot follow others directly. A count of people passing a corner cannot, on its own, become a recommendation about where to park. The platform refuses that step before anything is calculated, and says why.

Each result says what kind of work produced it

Rosa's assistant does several kinds of work for her. Each result the platform returns with a claim is marked with the kind of work. EPM calls these marks grades.

The grades have an order: a measurement, then a simulation, then reasoning. A recommendation and a proposed change both reach beyond reasoning, but neither ranks above the other, so the platform will not combine them.

A busy corner is not proof of sales. A count of people passing says nothing about how many will stop for pupusas. Keeping the kinds of result apart keeps that difference visible.

Behind an answer to Rosa's question, the steps can run like this:

1. Look at the street today. The assistant reads what a source records about people passing each spot. Measured.
2. Describe each spot. It sets out each spot as a scenario, with its assumptions stated, such as how many office workers come in. Simulated.
3. Simulate. It runs each scenario. The result describes what the model produced, not what happened. Simulated.
4. Compare. It works out how the spots differ within the model. Reasoned.
5. Recommend. It ranks the spots against the goals and limits Rosa stated. Recommendation: Rosa decides.

A single simulation cannot go straight to a recommendation. The comparison has to come first.

How strongly a result can speak

Each result also comes with a word for how strong a claim it can make. EPM calls this claim strength. There are five words, from weakest to strongest: speculative, suggests, indicates, strongly indicates and demonstrates.

The assistant is given the word to use, and Rosa's browser shows the same word. A simulation result says "suggests": this model, run on these inputs, suggests the library corner is busier at lunch. Nothing on the platform says "demonstrates".

The words are an order, not a percentage or a probability. Each method the platform uses has its word set in advance, with a written reason, and the word does not change with how the numbers come out. If the inputs a method needs do not check out, there is no word at all: the platform refuses, or says it makes no claim.

Repeating, rephrasing or summarizing a result never strengthens its word. A later step, such as comparing two simulations, makes a different claim about the difference between them. It carries the word set for that way of comparing, and the earlier result keeps its own.

When the platform says no

Sometimes the evidence cannot support the step Rosa's assistant asks for. Then the platform refuses, with a short code and an explanation of what fell short and what would fix it. A refusal is a decision, not a breakdown. It records why the step cannot go ahead under the rules on evidence and access. An error is different: it means a calculation failed.

The explanation never names the records involved, so it cannot reveal data that Rosa is not allowed to see. A refusal is never used as evidence for a later answer.

A simulation can also finish without supporting a claim, for example when its authored structure has no scientific review for the intended use, or when the platform cannot confirm that the model ran as declared. Rosa still sees the numbers. The browser tells her that the platform makes no claim from them, and why.

The basis travels with the answer

A simulation result carries a record of the information, assumptions and calculations that produced it. Each saved result is fixed. If its basis later proves faulty, it can be withdrawn as evidence. New answers cannot build on a withdrawn result, while the original stays readable.

For Rosa, the comparison depends on where people go, when they are there and the assumptions about their workdays. Her next question can build on the comparison, with the further evidence it needs made clear.

This gives people something concrete to examine when an AI offers an explanation. They can return to the calculation, inspect its basis and ask for the work needed to explore another possibility.

Checking a model against what happened

Comparing a simulation with what actually happened is kept separate from the words above. Rosa approves such a check before she sees how it came out. So a disappointing result cannot be quietly dropped.

The check is recorded separately, and it never changes the words on a result, whichever way it comes out. A simulation result describes what the model shows, not what the street will do.

Explore the uncertainty

People need to know which assumptions matter and how a conclusion might change. Rosa can ask what happens if more people work from home, or if a nearby market moves to another day. Her assistant can compare those possibilities and explain where the evidence is stronger or weaker.

That can guide the next useful action: observe a corner on different days, plan a trial with her truck or explore a location that works across several plausible futures. Uncertainty becomes something she can respond to.

Evidence and purpose belong together

A well-supported comparison still leaves a human question: better for whom, and by which criteria? Perspectives make those points of view explicit. EPM governs what the evidence supports; Perspectives connect the analysis to what people care about, preserving disagreements and meaning that a calculation leaves out.

Together, they support a richer form of participation. People can examine the reasoning behind a proposal, question its definition of success, and develop alternatives whose evidence and priorities others can inspect.

What EPM does not do

EPM checks what the platform returns. It cannot check every sentence an assistant writes about it. The saved record lets Rosa compare the assistant's account with what the platform actually returned.

EPM also does not say how close a model is to the real street. Its words say how far a claim reaches beyond its method and inputs. The technical reference sets out the rules in full, and the technical overview explains how they fit with simulation, AI access and Perspectives.

The approach is protocol-like: it makes the rules of responsible analytical work part of the infrastructure through which an assistant acts.

See how the capabilities converge · Explore the public purpose

Keep a linked experiment accountable to its evidence

An experiment can connect physical calculations, network analysis and an authored model across several runs and time scales. Each result retains the evidence and conditions it actually used. A later step cannot turn numerical agreement, a repeated sample or a successful run into scientific standing that its sources do not supply.

You can ask what changed a conclusion, which uncertainty matters and what additional observation could help a decision. The record distinguishes measured inputs, modeled effects, value judgments and the assumptions connecting them. Scientific review applies to its stated structure and use; your approval determines what work may run.