Technical Manual · Local browser workbench

Research Alpha — Production Use Not Authorized

Open Web Steward Technical Manual

Installation, controlled inputs, candidate validation, inert preview, repair governance, evidence export, and developer replay for v0.1.0-alpha.2.

The ordinary runtime is one self-contained HTML workbench with inline CSS and JavaScript. The model is not bundled. The human chooses a model separately, copies a bounded capsule out, and pastes returned HTML back into the local review surface.

Requirements and exact download

Normal use

  • A desktop or laptop able to extract ZIP files.
  • A browser capable of opening the local workbench.
  • Optional access to any separately selected model or tool that can return HTML.
  • No account, activation, API key, server, model bundle, terminal, analytics, or UVLM network service.

Developer replay

  • Python compatible with the included tools and tests.
  • The dependencies listed in requirements-dev.txt.
  • Playwright and browser binaries only for browser evaluation.
  • Automated exact-release browser evidence is limited to Chromium 144.0.7559.96.

Verify, extract, and open

Keep the original ZIP unchanged as your identity anchor.

  1. Calculate SHA-256. Confirm that the source ZIP equals 84e1b46151cade62ad7fe0b34138fafd82929ab3ab9cd0acba910b0f810bff9f.
  2. Extract the ZIP. Use the operating system's normal extraction function. Do not work from inside the compressed archive.
  3. Open the workbench. Navigate to workbench/web_steward_workbench.html and open it in the browser.
  4. Begin with the included fixture. Press Load the included reference fixture before introducing your own product facts.
84e1b46151cade62ad7fe0b34138fafd82929ab3ab9cd0acba910b0f810bff9f  UVLM_Open_Web_Steward_Community_Edition_v0.1.0-alpha.2.zip

Review-evidence boundary. The exact package is designed for local opening, but the independent automated review environment blocked direct file:// navigation. That route remains NOT_EXECUTED in the published review rather than being relabeled PASS.

The three controlled input objects

The reference task is driven by machine-readable facts, not by free-form model improvisation.

Release facts

What may be claimed

Product identity, version, status, functions, artifacts, URLs, licenses, limitations, indexing posture, correction route, and authority boundaries.

Site brief

Who the page serves

Audience, user journey, primary action, supporting content, platform context, and editorial constraints.

Product theme

How the page may look

Declared design tokens, spacing, type scale, control dimensions, and bounded visual posture.

Strict JSON boundary. The validators reject duplicate keys, including Unicode-equivalent duplicate forms, instead of silently accepting ambiguous data.

Micro and Standard teaching capsules

The capsules are model-agnostic task contracts. They do not certify the model that receives them.

Micro capsule

For a small or limited-context model. It emphasizes the exact source of truth, one article plus one scoped style element, semantic essentials, visible status and limitations, 46-pixel controls, focus visibility, 320-pixel reflow, and prohibited active surfaces.

Standard capsule

For an ordinary local or hosted model. It adds a six-step method, complete page anatomy, platform profile, theme tokens, one primary and one secondary action, narrow-container behavior, forced-colors support, and explicit candidate-not-final language.

No live-model benchmark was run for the exact release. The comparative conditions for bare requests, monolithic prompts, Micro, Standard, and Standard plus repair remain NOT_EXECUTED. Do not claim universal model improvement.

Candidate identity and active-surface contract

Exact occurrence identity

The candidate SHA-256 is calculated over the exact UTF-8 bytes supplied, without Unicode normalization. Local occurrence IDs distinguish runs, requests, candidates, decisions, findings, receipts, and repairs. A stable content fingerprint separately binds equal candidate bytes across occurrences.

One bounded web surface

The candidate may contain exactly one top-level article.uvlm-ows-product and one top-level style. Every CSS selector must remain within the product root, and every URL must match a controlled destination.

Surfaces that fail closed

Scripts, forms, templates with active descendants, media, iframes, objects, embeds, base and link elements, event handlers, remote resources, dangerous URL schemes, inline style attributes, root-escaping and sibling selectors, generated counterfeit content, fixed overlays, masking and composites, transformations that move content off-screen, and pointer-event suppression.

Static analysis and inert preview are different lanes

Exact original

Source-fidelity and static checks evaluate the original candidate string. The host workbench does not insert that original directly into its own DOM.

Defensive transformation

Preview and geometry checks use an allowlist-based inert transformation inside a no-token sandboxed iframe with restrictive content-security policy. Inert about:blank links never replace controlled live links in a safe export.

Not a universal sanitizer. The preview reduces the active review surface for this exact contract. It is not permission to process arbitrary hostile HTML or deploy the candidate without independent security review.

Finding states and human review

StatusMeaning
PASSThe exact named check executed and passed.
WARNA bounded concern requires attention but is not represented as a deterministic failure.
FAILThe exact candidate violates the governed rule.
NOT_APPLICABLEThe rule does not apply to the candidate or platform condition.
NOT_EXECUTEDThe lane was not performed and remains open.

Human evidence lane. Keyboard-only task completion, screen-reader use, low vision and magnification, first-time cognitive usability, content accuracy, affected-person review, law, independent security, CMS behavior, and live publication remain human responsibilities.

Bounded repair governance

A repair is a controlled successor event, not a hidden overwrite.

  1. Observe a repairable FAIL or WARN. The finding must carry a bounded repair instruction.
  2. Choose REQUEST_REPAIR. Other human decisions do not create an executable repair packet.
  3. Bind the exact parent. The repair fingerprint is independent of rationale wording and identifies the exact parent candidate and named findings.
  4. Paste one direct successor. The packet can be consumed once.
  5. Revalidate and decide again. The successor does not inherit approval from the parent decision.

Evidence export

The workbench creates an uncompressed ZIP with sorted filenames and fixed ZIP metadata once a run state exists.

Preserved evidence

  • Exact controlled inputs and candidate bytes.
  • Native findings and finding set.
  • Human decisions and bounded repair packet.
  • Initial and successor lineage.
  • CoherenceLattice request, candidate, decision, repair, telemetry, and receipt wrappers.
  • A run manifest binding the exact UTF-8 bytes and lengths of every other export file.

Safe versus held export

A candidate without deterministic FAIL may export the exact validated article and scoped stylesheet. A held or rejected candidate exports an explicit held placeholder instead of promoted page markup.

Repeatability scope. Export bytes are repeatable from the same exact run state. Separate sessions intentionally receive new occurrence IDs and timestamps, so cross-session ZIP identity is not claimed.

CoherenceLattice integration boundary

Open Web Steward owns native web rules. CoherenceLattice wrappers standardize governed exchange anatomy.

  • Request
  • Candidate
  • Findings
  • Decision
  • Repair
  • Telemetry
  • Receipt

Native sovereignty. A generalized score cannot override a native web FAIL. The workbench emits no page-quality Ψ score and creates no PMR, Atlas, Sophia, memory, training, publication, deployment, or authority effect.

Developer replay

Run from the extracted release root after installing the declared development dependencies.

python tools/validate_release.py .
python -m pytest -q
python tools/run_browser_evaluation.py .

The browser command requires Playwright and browser engines. An unavailable engine or blocked route must remain NOT_EXECUTED.

Troubleshooting

The workbench opens but the model is not present

That is expected. Open Web Steward does not bundle or call a model. Generate a capsule, use your separately selected model, and paste only the returned HTML.

The browser blocks local-file behavior

Try another local folder or a user-approved local static server. Record the route used. Do not claim the blocked route passed.

A page looks good but validation fails

Read the first failed rule and exact evidence. Visual appeal cannot override source fidelity, active-surface, containment, privacy, or authority checks.

Validation passes but publication is still held

Complete the open human lanes: source accuracy, keyboard and assistive-technology review, security, legal, CMS, logged-out live retrieval, and final release authority.

Technical support

Include the version, exact source SHA-256, operating system, browser, reproduction steps, and redacted evidence. Do not email secrets or sensitive source material.