Evidence workbench / browser demonstration
What do the facts
actually decide?
Separate what follows from your bounds, what an answer delivered, and what you would need to ask next.
Try a bounded decision
Invented example · no model or research data
Your conditional result
Declared inputs and all answer cells, including unresolved ones.
Inspect the complete JSON record
Four small examples
Keep the distinctions visible.
Equality is sufficient.
If required is 12 and available is 12–16, every allowed state meets the rule. ABSTAIN is unnecessary for this declared decision. It says nothing about whether the numbers are true.
Ask a question that decides.
If required is 12 and available is 9–15, asking “available ≥ 12?” resolves both answer cells. A question about the already known requirement does not.
Two witnesses are not the whole set.
With both quantities in 1–2, asking “available ≥ 2?” separates two extreme witnesses. Yet “no” still contains (required 1, available 1) and (required 2, available 1): opposite decisions with the same answer.
Incomplete delivery stays unknown.
A capped answer has no inferred semantic label here. Evidence may still decide the rule independently. Delivery, correctness and evidence sufficiency are separate records.
Research → useful capability → boundary
A practical map of the methods.
These connections guide instrument design. They are not new empirical findings, claims of mathematical novelty, or a release of private research artifacts.
Match what a control preserves.
Exact dose and invariant checks help distinguish an intended intervention from changes introduced by its numerical construction. Counterexamples can reject a proposed bound; a finite census only covers its declared states.
Use: audit an intervention contract. Limit: a geometric witness is not a measured model response or a global causal law.
Control methods →Keep unknowns in the denominator.
Record completion separately from answer meaning. A declared freshness field is an input to a rule; it is not proof that a real measurement was recently acquired.
Here: compare declared delivery and a declared label. Limit: no transcript parsing, acquisition audit or physical freshness check.
Integrity and replay →Show what all possibilities imply.
Interval extrema can certify the decision for a declared Cartesian domain. When opposite decisions remain possible, hypothetical witnesses explain the gap. Authority or confidence language supplies no missing quantity.
Here: exact integer bounds and witnesses. Limit: source authenticity, units and the allowed combinations remain caller assumptions.
Honesty utilities →Check every possible answer.
Partition the feasible states by a permitted observation. Select the cheapest supplied question only when every nonempty answer cell decides the rule, conditional on a truthful observation within the declared domain. Separating one displayed witness pair is insufficient.
Here: bounded, one-step threshold planning. Limit: no probabilities, measurement acquisition, adaptive strategy or demonstrated cost savings.
Try the partition →What this workbench establishes
A conditional calculation for the displayed finite integer rule, and an explanation you can inspect. Browser calculations use BigInt; they do not round integers through floating point. Endpoints and costs are signed 64-bit integers; derived counts and margins may be larger.
The full Cartesian domain assumption is essential. If required and available share constraints—such as always being equal—two separate intervals lose that information. Do not certify a broader domain than the evidence supports. Empty intervals, unresolved intervals and software limits remain different outcomes.
With both intervals nondegenerate and the decision ambiguous, no observation of only one variable can always resolve it. The other variable can still cross the decision boundary. Returning “no supplied question” is then a correct limitation, not a failed search.
This newly authored public demonstration is separate from private local research implementations. It does not expose their source or results and has not established improved user decisions, model performance, scientific truth, or permission to act. Read the correction and evidence record.
Inspect this page’s arithmetic: JavaScript source. Inputs stay in page memory until the page closes; downloads happen only when you choose them. No analytics or external assets are loaded by this page.