# MEOK public consolidation and visitor-journey release 11 November 2026 • Existing public research Space • No new production model ## Delivered change The public model comparison could correctly reject an invalid input while leaving the previous successful table and receipt visible. The repaired presentation clears current output on edits and refusal, disables export, and ignores delayed receipt completions after input changes. The numerical core and all model/data bytes are unchanged. A new test exercises the actual `ui.mjs` with a minimal deterministic DOM, real frozen weights and a deliberately delayed cryptographic digest. Eighteen checks pass. The same harness fails on the exact previous public UI at the stale-result assertion. This is a discriminating regression control, not simply another all-pass counter. ## What was run again The original 37-check replay passes across 560 public synthetic transitions. The 38-check precision replay passes across 1,200 rows and six checkpoints (7,200 inference evaluations). The retained specialist presentation and join suites pass 18 and 87 checks respectively. These suites have different scopes and may share underlying examples; their counts must not be combined into a claim of independent capability measurements. An anonymous browser pass ran the original public model, 560-row replay, six-checkpoint replay and practice grading. It reproduced the old stale-success display after row 9999. The export button was clicked, but downloaded file contents were not verified in that pass. Later hosted acceptance is in a separate dated readback and is not inferred from source tests. ## Consolidated evidence, with limits The 980-parameter model is a synthetic process-state predictor, not Cosmos. Its original constrained test mean error is about 0.00288, with worst component error about 0.75. Exact rules perform better on this known synthetic task. The retained 68-parameter fitted-head experiment is held; lower average error is not automatic production promotion. The Memory Venturi and Academy remain with Council PR 4176. The eight draft Academy exercises reference all eleven SAFE layers, and 34 dated paths contain 170 stages. Those counts do not establish professional qualification, staffing, human learning gain or industry deployment. The existing five ORB banks remain separate. The categorical specialist receiver remains with PR 4180. Its byte matching, compatibility checks and presentation controls do not authenticate caller-supplied pins, establish model independence or authorise execution. Its retained evidence can be inspected without changing its algorithm. The private application now has a separately owned complete-source/saved-copy composition. Its source tests and dependency checks do not establish real bucket persistence or end-user identity workflows. This release neither modifies private records nor retries the earlier client-blocked credentialed browser job. ## Public journeys - Original process inference and refusal: `../index.html`. - Six-checkpoint replay and original practice pack: `../research-v03/index.html`. - Specialist receiver release and limits: `../integration-specialist-review/presentation-v1/index.html`. - Machine-readable release: `release.json`. - Executed test readback: `qualification.json`. - Dependency-ordered next gates: `next-steps.json`. ## Quality and publication policy Research results are dated and source-pinned. Prediction, observation, source integrity, authentication, authority and measurement admission remain distinct. A self-published hash is not an independent signature. A public demo is not customer acceptance. No guarantee of universal or 100-percent correctness is made. No large-model inference, new weight fitting, paid GPU, private-data publication, protected merge, authority change or new GSPC score is performed by this release. Original evidence remains available. Any genuinely new content or algorithmic result needs its own source, rights, test and promotion review. ## Mobile readback correction A real anonymous Chromium check found that a 390-pixel viewport widened to 511 pixels because the grid inherited the comparison table's intrinsic width. The bounded CSS correction gives grid tracks a zero minimum and keeps wide tables in their own horizontal-scroll container. The candidate produced a 390-pixel document width on the same browser page. Hosted acceptance after this publication remains a separately retained result. No numerical or model code changes. ## Completed post-publication acceptance The final public HTML/JavaScript/CSS journey was exercised in a fresh anonymous Chromium context using an already-installed browser, without login credentials or new dependencies. All 25 declared checks passed: valid inference, exact served UI/CSS hashes, actual downloaded receipt payload verification, refusal with cleared old output, recovery, empty-row rejection, 560-row and six-checkpoint replay, practice boundaries, eight internal navigation targets, and 390-pixel document layout. No page JavaScript errors were observed during that specific run. The download was saved and its sorted-JSON payload SHA256 independently recomputed. It explicitly remains unsigned, not persisted by the app and without authority. This establishes the tested export bytes, not the truth of predictions. The served readback verifies 15 pinned Hub files exactly. Twelve static responses match raw bytes. Two authored HTML files match after removing only the exact inspected host-variable/social-metadata additions. The root README is rendered to HTML and receives only a separate sanity check; full rendered-content equivalence is not claimed. Earlier attempts remain distinguishable: one browser driver read innerText from a collapsed details block and was corrected to inspect textContent; a later actual browser pass exposed the mobile overflow that was then repaired; a separate TinyFish follow-through was inconclusive. None is silently treated as a complete pass. See [POST-PUBLICATION.json](POST-PUBLICATION.json), [browser-acceptance.json](browser-acceptance.json) and [served-readback.json](served-readback.json). The tested application code is unchanged by this evidence-only update. Human editorial acceptance, private storage acceptance, actual authority-to-effect and capability promotion remain open with their existing owners. A concurrent specialist-lane update at 34773cafd575fb138dffdf1b506e9fbb8c1a952c added stronger source/test-count validation to its public reproduction helper and published its own acceptance record. The root UI, CSS and consolidated page were byte-identical to the tested version. This evidence-only publication preserves those 13 specialist changes and the separate owner-controlled [acceptance record](../integration-specialist-review/presentation-v1/acceptance-20261111/index.html). Our earlier 18-test run did execute all 18 tests; the new owner-controlled negative control addresses a different zero-test substitution weakness.