EM-0022 execution plan

Source content

EM-0022 execution plan

Status: candidate validation complete; awaiting exact-head independent review on

codex/em-0022-case-entry-trust-agent-parity.

Objective

Make every primary human and agent entrance lead coherently into the accepted Case 001 evidence

experience. Replace the unlinked review badge with a sanitized, exact-head public receipt; turn

How We Know into a section index; label the existing repository corpus as Substrate; provide a

first-screen path into the unresolved lineage; and fail closed on agent-discovery or release-skew

regressions.

Change surface

1. Compile a public review-receipt projection from the already accepted, byte-bound independent

receipt without exposing reviewer temporary paths or unneeded raw review material.

2. Replace the duplicated How We Know homepage with a compact one-case section index that says

explicitly that no second case has yet been admitted.

3. Preserve /explore/ and its corpus, while naming that destination Substrate in navigation and

page copy so its role is legible.

4. Link the first-screen 4 + 1? evidence cell directly to the unresolved 2007 lineage.

5. Add deterministic cold-start agent discovery and whole-release identity-parity checks, then

exercise the same checks against the accepted Pages deployment.

Hard boundaries

• The accepted dossier, review receipt, feature manifest, research packet, counts, relations,

policy evaluations, and evidence conclusions remain byte-identical.

• No new case, lens, research claim, workflow, DNS, API/MCP deployment, package, or release work.
• The public review view is a sanitized projection of the accepted receipt, not a new review or a

claim that synthetic tests establish universal agent comprehension.

• Publication requires independent exact-head review and the normal protected squash path.

Verification

• targeted feature/interface/identity tests;
• cold-start discovery from generated llms.txt into stable Markdown and JSON twins;
• deliberate mixed-version mutation rejected by the release-identity verifier;
• no-script semantic HTML and native keyboard path into the unresolved lineage;
• desktop and mobile inspection;
• full make check with disclosure audit and deterministic rebuild;
• exact-head independent review, protected merge, Pages deployment, and live human/bot read-back.

Candidate verification

• the generated How We Know index contains one compact case row, explicitly says no second case

has been admitted, and no longer includes the homepage case hero;

• the primary navigation and /explore/ presentation use Substrate while preserving the route and

all self-describing corpus content;

• the sanitized review HTML, Markdown, and JSON expose the accepted reviewer, exact head/tree,

receipt digest, independence conditions, checked scope, counts, and limitations without the

receipt's temporary local paths;

• the first-screen 4 + 1? cell is a native link to #unresolved-lineage, and the target exists on

the case page with the existing visible focus treatment and no JavaScript;

• a cold-start fixture follows only llms.txt links into the stable Markdown and JSON twins and

recovers the exact question, verdict, 10 / 4 + 1? / 12 counts, and unresolved 2007 lineage;

• a whole-release verifier checks every generated HTML route and every JSON object that declares

release fields, recomputes the release-manifest identity, verifies every inventory path/digest/

byte binding and complete file closure, and rejects mixed-commit, forged-identity, and inventory

mutations;

• targeted feature/interface tests passed; the full suite now collects 75 tests;
• browser inspection at 1440 x 900 and 390 x 844 found one h1, zero scripts, no horizontal

overflow, responsive section and review layouts, exact reviewer/head display, and no private

temporary path disclosure;

• the browser driver did not dispatch Enter activation on the new native evidence link during the

local check; semantic link structure, destination, focus CSS, and click target are verified, and

exact keystroke activation remains an explicit post-deployment read-back;

• make check PYTHON=.venv/bin/python passed all 75 tests, disclosure audit returned no findings,

deterministic comparison build verified, and source state remained unchanged;

• immutable local receipt em0022-local-20260822T224248Z records the full check command.

External bot-user-agent, route, content-type, canonical, release-skew, and keyboard read-back remain

post-merge Pages acceptance checks. No hosted API or MCP runtime is asserted.