01 / Reasoning and evaluation · Added 2026-10-26

Keep the Gaps

A small, finite reasoning instrument for asking what a declaration of evidence actually requires—and what remains possible when an assumption is removed.

Open the workbench ↗

The instrument

Three entirely synthetic ledgers each define finite state variables, constraints, optional premises, and claims. Changing a premise reselects a precomputed checker report. Each claim is entailed, refuted, unresolved, or the selected premises are inconsistent within the declared model.

Visitors can inspect compatible possibilities and counterexamples. The number of matching assignments is not a probability estimate or confidence score.

Recorded evidence

The archived standalone original is preserved without modifications from the packaged candidate.
Archived original file size: 1,466,549 bytes; SHA-256: 79eeac3bbd76c27916019d85867ddfe2c861205ad8f38210e7f1f383d6316901.
The package receipt for 0.1.0-review.2 records PASS_FINAL_PACKAGED_EXTRACTED_CANDIDATE_CHECKS, including 92 normal Python tests, 92 optimized Python tests, 13 fixture-fidelity tests, and 27 presentation tests. These are author-supplied checks, not independent validation.
A supplemental Chromium page.set_content smoke check rendered all three cases, toggled and reset premises, downloaded parseable report JSON, and raised no script errors. It did not navigate to a real deployed page.

Gallery navigation update · 2026-10-27: The public instrument adds only a scoped navigation bar and its styles. Scientific scripts, embedded data, and original instrument content are unchanged. The historical hashes identify the archived standalone original, not the navigation-adapted index.html.

Limits and unknowns

Green inside the model is not green outside it. The checker cannot authenticate sources, validate the natural-language encoding, discover omitted variables, infer causality, estimate real-world probability, enforce permission, or certify safety.

Ordinary live-page navigation and physical-device behavior are not verified by the supplemental browser check. The packaged receipt leaves native download verification in an ordinary browser, local file-protocol execution, mobile-device and assistive-technology checks, and broad browser compatibility unresolved.

Reproduction

Open the workbench and select all three examples.
Toggle one premise at a time and inspect changes in displayed claims and alternative assignments.
Reset the example and compare the state.
Download the checker report if supported, then compare its selected premise IDs to the interface.

Everything in the instrument is synthetic. The source HTML includes embedded checker reports and provenance. No remote API or runtime scientific verification service is claimed.

Status: finite packaged candidate, not independently adopted or validated.