The problem to solve

Zwikipedia is both a reference publication and a technical portfolio. It needs a process that preserves source boundaries and editorial history without making a relatively small creative website impossible to maintain. The development process is intended to do two things at once: support ambitious exploration and make finished work independently verifiable.

The project uses ZDP — Zwikipedia Development Protocol. The authoritative repository document defines the process for planning, implementing, verifying, releasing, documenting, and closing Zwikipedia work. It applies to human and AI contributors and scales safeguards and verification to the consequences of a change. ZDP does not itself choose the application architecture or grant authority to publish an article or deploy code. Those decisions and authorizations remain governed by the approved specifications, architecture records, and owner-controlled workflows.

A decision is not a deployment

Zwikipedia distinguishes a requested requirement, an approved design, a code change, a successfully executed test, an editorially approved article, and a visible production result. Each is a different kind of evidence. Confusing them can produce an impressive technical history without improving the page a visitor actually sees.

The project's prior handoff records identify that exact failure mode: substantial infrastructure progress had outpaced the visible design and article-completion work. The corrective principle is straightforward. Success must be demonstrated at the surface where the requirement matters. For layout, that means comparing the implemented page with the approved design at matching viewports. For publication, it means checking that the intended revision is the one readers can see.

AI as collaborator, not source of truth

AI assistants can help draft specifications, produce implementation candidates, review code, and develop article manuscripts. But an assistant may be working with outdated context, may infer an incorrect requirement, or may report its own work more generously than the resulting site warrants. The project therefore maintains explicit owner decisions, architectural records, and evidence of actual outcomes.

This is the serious operational response to the humorous Twelfth Principle of Destruction, under which the probability of two assistants sharing the same context decreases as their number increases.

Terrapy: confidence earned through verification

The project’s informal glossary distinguishes Terrafied—an expectation of impending trouble after an AI assistant makes a familiar-looking mistake—from Terrapy: confidence rebuilt through a sequence of small, independently verified successes. The terms are jokes with a practical lesson behind them. An agent’s apology, explanation, or restatement of a requirement is useful, but none is evidence that the underlying code or public page now complies.

One concrete example is the Crossroads immersion rule. When a prominent out-of-world notice interrupted an in-universe article, the relevant correction was not simply to move the notice. Out-of-world explanations had to remain behind deliberate reader interaction, closed by default, with the specifically permitted footer treatment as an explicit exception. The owner also identified missing preview text on Mary Flynn’s article. Restating either requirement did not establish completion. PR #71 subsequently delivered the correction with regression checks and independent production verification: the disclosures remained closed until deliberately opened, and the Crossroads previews retained their in-universe voice. This is evidence of that release, not a promise that later work cannot regress.

Proportionate engineering

A change should receive scrutiny appropriate to its risk. Adjusting a card heading's font size should not require the same process as modifying publication authorization or crossing a database security boundary. Tests, review, and rollback planning are useful; speculative infrastructure added solely because it is possible is not.

The process favors reuse of existing article models, editorial permissions, and publishing mechanisms. It also seeks to minimize repeated human copy-and-paste between AI agents when authorized tools can work directly on the source material.

Publication and owner authority

A manuscript is not a public page, and a staging preview is not publication approval. Owner approval of prose authorizes editorial progression but should not bypass the established controls on production publishing. Sensitive changes should retain the documented staging/canonical separation and owner-controlled release path.

An article about an external fictional world must also preserve its source boundary. Source quotation, owner interpretation, and a new Crossroads continuity are distinct kinds of information and must remain distinct after publication.

Evidence for completion

Useful evidence includes a live URL, the actual rendered content, visual comparisons against approved references, correct revision selection, relevant automated checks, and a clear statement of anything not yet verified. No single kind of evidence proves every requirement.

Related reading