How the platform keeps answers from claiming too much

What the platform refuses to claim

Every answer comes with a ceiling on it. The ceiling says how strong a claim that answer is allowed to support. If the evidence will not support what you asked for, the platform says so and tells you which evidence it did not have.

A refusal means the platform worked. It read the evidence it had, and the evidence did not support the step you asked for. An error is a different thing: an error means something broke.

Four rules produce those decisions. The mechanism comes from computer security, where a piece of data carries a label saying how sensitive it is, the label goes wherever the data goes, and certain operations are forbidden on labelled data. Here the label describes evidence instead of sensitivity, and it forbids claims the evidence cannot carry.

Say you measure how long it takes to drive across town at eight in the morning. You now know what happened. You do not know what would have happened if a lane were closed, and you do not know whether the traffic light caused the delay.

The platform keeps that difference attached to the result. Every result carries a label: observation, interpretation, reasoning, recommendation, revision. The label goes wherever the result goes. If someone later tries to use your traffic measurement as evidence about a cause, the label stops them.

Almost nobody reaches a wrong conclusion by lying. They reach it by taking a number measured under one set of conditions and using it to answer a different question. The label makes that swap visible the moment someone tries it.

5 grades. They run OBS to INT to RSN, and from RSN to either ACT or LRN. Those last two are not ranked against each other: neither counts as the stronger. Claim strength runs CS0 < CS1 < CS2 < CS3 < CS4. Each operation accepts only certain grades as input, and cannot produce a result above its own ceiling. The result carries that ceiling with it.

There are ten operations, and you can chain them: one operation's output becomes the next one's input. Measure something, compare the options, rank them. Ten operations means a hundred possible two-step chains.

Fifteen of those hundred get refused. Refused before anything runs, so no simulation starts and nothing is computed, the same way a spellchecker underlines a word before you have finished the sentence.

Nine of the fifteen are the same mistake. The first operation gives you an observation: you measured traffic, you know what happened. The second one needs more than that. The operation that explains something needs a cause. The one that picks between options needs grounds for a decision. A measurement gives you neither, so the chain is refused.

To explain a cause you need a simulation where you varied the thing you think caused it. To pick between options you need a comparison of those options. Those are the steps that raise an observation to something an explanation or a decision can stand on.

100 ordered pairs of the ten operations, of which 15 are rejected by static type checking before execution.

Every operation carries two sentences written into its definition: what it settles, and what it does not. Both were fixed when the operation was specified, before any result existed to attach them to.

Take the operation that measures variation across many runs. It tells you how far apart the runs came out. It does not give you a confidence interval, and it will not tell you how likely an outcome is in the world. A spread across runs of one model is evidence about that model, not about the world.

A frame is what matters to one person, written down in a form the platform can compute with. Someone might say: I care how long it takes my child to get to school, and I care about noise on my street.

Turning that into a computation can fail. Maybe the platform has no measure for one of the things they said. Maybe two of their concerns sit on scales that cannot sensibly be combined, the way you cannot add a temperature to a ranking. Maybe a community has said that a piece of information is not available for this use.

When that happens you do not get a number. You get a deliberation result: each thing the compiler could not turn into a computation, and why. Every compilation returns one or the other.

When a frame is applied to a finished result, every measure the person named gets one of six dispositions. Two of the six are refusals: the platform could not evaluate the measure, or it was blocked from doing so.

4 deliberation triggers: missing operationalization, prohibited claim, missing commensuration, sovereignty block. 6 dispositions when a frame reads a result: observed, observed as proxy, declared irrelevant, not observed, could not evaluate, sovereign blocked. 4 statuses a stated limit can take: binding, satisfied, violated, skipped.

Measured on 16 September 2026

Counted over the files tracked in the repository on 16 September 2026: Python under the platform package, Rust under the simulation core, Python under the MCP server, TypeScript under the browser client, and Python under the tests. These are physical lines, including blank lines and comments. Strip those out and the count would be lower.