Open Web Steward Community Edition:
Local-First Web Design with Human Authority Preserved.
Companion article · Public-interest web design
Open-source Alpha · Human Authority Preserved
Open Web Steward Community Edition
Local-first web design with human authority preserved
A model can produce fluent HTML in seconds. The difficult part is preserving source truth, containing active code, exposing skipped checks, authorizing only bounded repairs, and making sure a polished candidate never quietly becomes a published fact.
Generation was never the whole problem
Modern models can write headings, cards, buttons, and responsive CSS with remarkable speed. They can also invent a version, soften a limitation, substitute a plausible download link, claim a test was run when it was not, or produce a page whose active behavior is larger than the visible design suggests.
A human reviewer then faces two bad options: trust the page because it looks professional, or inspect every line as though no assistance occurred. Open Web Steward is designed for the space between those extremes.
Its central claim is modest: one product landing page can be generated and repaired more responsibly when the model receives exact controlled facts, the returned HTML is quarantined as a candidate, native checks remain visible, and the human decision is recorded as its own event.
A local workbench around a separately chosen model
The model is deliberately outside the trusted workbench.
- Controlled facts
- Micro or Standard capsule
- Separate model
- Untrusted candidate
- Inert preview
- Native findings
- Human decision
- Evidence export
The workbench does not require an account, activation, API key, model provider, server, analytics service, remote font, remote script, application network call, or persistent browser storage. It produces a task capsule that can be copied to a model selected by the user. The returned HTML then re-enters through a narrow candidate boundary.
Provider-free does not mean privacy-free. Copying a capsule or source facts into a hosted model is a separate external transfer. The selected model and interface determine what leaves the device. Open Web Steward makes that boundary explicit rather than pretending the local workbench controls it.
Controlled facts become the model's ceiling
The release-facts record identifies the product, version, status, licenses, functions, limitations, exact artifacts, correction route, and authority posture. A site brief describes the audience and user journey. A product theme supplies bounded visual tokens.
The capsule tells the model to use those objects as the source of truth and to return only one contained product article plus scoped CSS. It explicitly forbids invented versions, licenses, test results, downloads, capabilities, compliance, ranking, and publication authority.
Facts remain facts
Required status, limitations, licenses, and URLs must survive the transformation into page copy.
Design remains bounded
The model receives a controlled theme and structural contract rather than permission to modify the entire host site.
Claims remain reviewable
The system screens defined contradictions while acknowledging that general factual accuracy still needs human evidence.
The candidate is quarantined before it is admired
Open Web Steward hashes the exact UTF-8 candidate bytes without normalizing Unicode. The original candidate remains a string for static analysis and is not inserted directly into the host workbench DOM.
For visual and geometry checks, the system creates a defensive inert transformation in a sandboxed iframe with restrictive content-security policy. That preview can help a human see the proposed page and can support bounded contrast, focus, target-size, text-spacing, and overflow measurements.
Preview is not absolution. A defensive rendering is not proof that arbitrary HTML is safe. The Alpha's allowlist exists for one controlled product-page contract and does not replace independent security review.
A skipped test is allowed to stay skipped
Many software workflows collapse everything into green or red. That is dangerous when a browser engine, assistive-technology test, legal review, or live-site check was never performed.
Open Web Steward keeps PASS, WARN, FAIL, NOT_APPLICABLE, and NOT_EXECUTED separate. The distinction is not cosmetic. It prevents a static validator from claiming a screen reader was used, a browser from claiming legal review occurred, or a builder preview from pretending to be live-publication evidence.
Automated evidence
Can establish bounded behavior such as exact fact presence, active-surface rejection, computed indicators, responsive geometry in a tested engine, and export-lineage integrity.
Human evidence
Must still address content truth, affected people, keyboard use, assistive technology, low vision, cognitive load, law, security, CMS behavior, and live publication.
Human authority is represented, not implied
The workbench asks the human to choose a bounded posture. APPROVE means acceptable only within the declared local review scope. HOLD pauses the route. REJECT marks the candidate unsuitable. REQUEST_REPAIR is the only decision that authorizes one repair packet.
The repair packet names the observed findings, preserves the parent candidate, and can be consumed for only one direct successor. The successor must be revalidated and receives a new human decision. This prevents a vague instruction such as “fix everything” from becoming unlimited modification authority.
Repair without erasure. The original candidate, first decision, repair packet, successor, second findings, and final posture remain available in the exported lineage.
The export explains what happened
At the end of a run, the workbench can export the exact candidate, controlled inputs, findings, decisions, repair lineage, telemetry, CoherenceLattice wrapper objects, and a manifest binding every other artifact's exact bytes and length.
This evidence does not prove that the candidate is true or ready to publish. It makes the process inspectable. A future reviewer can see which inputs governed the model, which rules executed, which lanes remained open, who chose the repair posture, and which successor was actually reviewed.
Model-agnostic means interface freedom, not universal proof
The capsules are plain text, so a user can try a small local model, a larger hosted model, a non-model transformation tool, or a human author. Open Web Steward does not bundle or privilege one provider.
But interface compatibility is not evidence that every model performs equally well. The exact Alpha.2 release did not execute comparative live-model conditions. Any future improvement claim must identify the exact model, version, settings, source set, baseline, metric, uncertainty, and review burden.
The more honest claim is architectural: the workbench gives unlike candidate producers the same governed doorway and keeps native web findings and human decisions controlling.
Why release this as a public Alpha
Open web design should not require a subscription, proprietary prompt library, hidden validator, or remote service. The Community Edition publishes the workbench, schemas, rules, examples, tests, hostile fixtures, rights records, and exact review evidence so others can study, adapt, criticize, and improve the method.
Software and machine contracts are allocated under MPL-2.0. Educational prose, capsules, examples, and documentation are allocated under CC BY 4.0. The exact file-level ledger controls.
The Alpha label remains essential. Chromium automation and adversarial replay passed for the exact package, but Firefox, WebKit, assistive-technology review, independent security review, legal review, and live Squarespace behavior were not established by that evidence.
Try the bounded workflow
- Download and verify Alpha.2. Preserve the exact source ZIP and compare SHA-256.
- Open the local workbench. Load the reference fixture and generate the Standard capsule.
- Compare the included candidates. Validate the good and bad examples, inspect why one holds or rejects, and export the evidence.
- Only then use your own controlled product facts. Keep external-model privacy, affected-person review, and live publication as separate human decisions.
The steward is the process, not the model
A model may help write. A validator may help inspect. A hash may help identify. None of them should silently inherit the human's authority to decide what is accurate, appropriate, accessible, safe, and ready to enter the public record.
Controlled facts. Untrusted candidates. Visible evidence. Bounded repair. Human judgment preserved.
