A view can remain ambiguous. A warning can change its own evidence. A convincing result can be selected from noise.
These three worked examples ask what makes a return informative. Each visual comes from an executable model. Use the controls to inspect the assumptions that make its result possible.
Exact model results and established methods are identified below. No human experiment or claim of a distinct advantage for the philosophy is presented.
The bounded program is complete: six models, ten exportable figures, and a final synthesis. The three sections below each lead to a deeper interactive experiment.
Galileo's successive views of Jupiter helped distinguish moons from fixed stars. Saturn's changing appearance posed another difficulty: understanding what the telescope could show. Our two-pixel model isolates that difference without pretending to reconstruct the historical telescope.
Scroll sideways to inspect the full figure at reading size.
A later observation helps here when the measurement changes what it can distinguish. Receiving a delayed record is a separate matter: its arrival changes what is available to the observer, without changing the record's content.
Scroll sideways to inspect the full figure at reading size.
There are only two static objects. Even source epochs measure total brightness; odd epochs measure left minus right. Measurements are noiseless. A fixed unknown wiring swap can preserve ambiguity across every view; a known left-hand reference through the difference channel resolves it. No amount of this reasoning establishes that real observations exhaust all possible objects or instrument faults.
Repeated noisy observations can improve precision. An old record can also yield a new implication. Neither possibility is denied by this noiseless example.
Continue with the noisy observation experiment: repeated measurements, a shared calibration budget, and a wiring-change counterexample.
Read the Galileo/Saturn argument · Evidence ledger · Model · Data
In this exact recurrence, prevention reduces defects from 80% to 20%. A controller uses the last observed rate to decide whether prevention is needed. When defects fall, it removes prevention. When they rise, it restores it. The response mechanism stays fixed throughout.
Scroll sideways to inspect the full figure at reading size.
| Policy | Defect probability | Total expected loss |
|---|
The conditional reference knows both action-specific risks. It is an oracle comparison, not a learner given the same information. Lower expected loss follows only under the displayed objective: defect probability + prevention cost.
The controller intervenes when its carried-over rate is at least 50%. For a baseline risk b, preventive effect e, and threshold τ, a two-cycle occurs when b − e < τ ≤ b. There is no sampling noise or mechanism drift. The equality convention matters.
An observed outcome under one action does not by itself identify the outcome under another. Action variation, causal knowledge, or another justified source is needed. A lower defect rate does not settle the cost or value of preventing defects.
Derivation and boundaries · Model · Data · Perdomo and colleagues
Continue with the learner that must acquire both action risks: a fixed horizon makes the cost of finding out part of the decision itself.
Every candidate in this example has true accuracy of 50%. Each is scored on 64 evaluation items. We select the candidate with the highest score, then compare what that selected score suggests with what an untouched evaluation would show.
| How the claim is evaluated | False-acceptance probability per search |
|---|
Scroll sideways to inspect the full figure at reading size.
A multiplicity correction accounts for the searched family. A fresh test evaluates the already selected candidate once, using evidence independent of that selection. These protect different stages of inquiry. Neither is a license to continue selecting after looking at the final result.
The calculation uses independent Binomial(64, 0.5) counts, exact inclusive upper-tail tests, and a fixed family of candidates. The nominal 5% single-test threshold is 40 correct answers; its attainable false-positive rate is about 2.997%, because counts are discrete. Bonferroni uses 0.05 divided by the number of candidates.
Each probability is per complete search under an all-null model. The fresh-test row evaluates one selected winner for every search; it is not restricted to searches that first passed the reused test. The fresh evaluation is independent and used once. Real models can have correlated errors, and this example does not reproduce a general adaptive analysis or the private reusable-holdout method.
Derivation and boundaries · Model · Data · Dwork and colleagues
Continue with shared evidence and one genuine signal: compare selecting the real advantage, accepting a lucky null, and purchasing fresh confirmation.
A loop can keep returning the answer it has made easiest to receive. The three extensions make different failures visible. This map compares their mechanisms; it is an interpretation of the models, not another measured result.
What can still be unresolved
The object and the instrument's wiring can change together while preserving the recorded evidence.
What the next inquiry needs
A constraint that distinguishes calibration histories. More of the same protocol cannot diagnose this reversal.
Inspect calibration and noise →What can still be unresolved
The outcome under prevention leaves the outcome without prevention unknown.
What the next inquiry needs
Action-specific evidence or justified causal knowledge, with the cost of learning included.
Inspect learning and action →What can still be unresolved
The record helped select the claim now being presented as confirmed.
What the next inquiry needs
Account for that selection or acquire independent evidence. A genuine signal can coexist with lucky nulls.
Inspect signal and selection →The framework holds that truth becomes actual as recognition through a living sensor and an instrument. These models contain no experiencers. They press on a narrower inference: a human's presence does not itself establish that the available record distinguishes the relevant alternatives.
This leaves two distinct questions: what became meaningful or transformative in the encounter, and what can distinguish its proposed explanation from the alternatives? The first can change which scientific question matters enough to pursue. It does not, by itself, settle the second.
The proposed CE practice is to connect a change in meaning or consequence to a revised question, a live alternative and a discriminating next step. It needs a comparison with an equally resourced human–instrument practice already using strong scientific methods. None of these models demonstrates an advantage for CE.
The full synthesis develops the argument, the causal connection to the “handshake,” and a remaining pressure on Doc III's treatment of beauty. Three questions remain for Alex's return:
The core text is unchanged by this synthesis. Synthesis evidence ledger · Completion checks
