Repository navigation
Engine 6 Preview
Updated 2 November 2026. The public product release is 5.5.0, with access to the separate experimental Engine 6.0 Preview. Engine 6.0.0-preview.1 is integrated into main through PR #37, merge commit d2525bf. The separate private Preview is online behind administrator authentication. Its original 47 hosted checks passed on 1 November; the exact artifact was restored and verified on 2 November after the public content promotion.
Published at 04:33 Europe/Rome on 2 November after staging verification. The exact same archive passed 49 production release checks, all 595 sitemap destinations and 30 source-file hash comparisons. The previous private Preview was restored at 04:38, passed 12 restoration checks, and preserved the full-row hashes of all 12 observed tables. No new live capture or daily window was counted.
The 5.5 release announcement links the private engine workspace, source-contract review, extraction comparison and JSON reports. Homepage, changelog, roadmap, both atlases, newsroom feeds and current EN/IT fact sheets identify product 5.5.0. Engine 6.0 is its experimental engine track, hosted separately with administrator authentication and private results.
The public artifact uses the established scanner. The Engine 6 source distribution on main stays restricted to private staging independently of the product number; neither a release label nor a selector enables public policy changes or notifications. Historical package versions and dated receipts below remain unchanged. Release validation and deployment receipt.
- Durable runs, tasks, attempts, leases and recovery after interruption.
- Frozen acquisition and derived evidence, with content hashes and source context.
- Source contracts covering document identity, scope, language, headings, anchors and terminal content.
- Structural extraction and clause comparison, exact old/new quotations and separate analysis/verification.
- Two-capture change confirmation; unresolved and high-impact cases remain held for review.
- Private Preview and same-capture extraction comparison, with persistent reports and JSON export. Reading a report does not repeat retrieval or inference.
- Authenticated hosted gateway, dedicated database and automatic Node worker with durable leadership, implementation-identity checks and recovery after actual lease expiry.
| Destination | Channel and result | Baselines and notifications |
|---|---|---|
| Main public site | Product 5.5.0 with the established scanner and evidence gates | Existing stable rules apply |
| Private staging Preview, online with administrator authentication | Preview: private observation, tasks and evidence | No official baseline/history updates; silent |
| Same separate staging environment | Compare extraction: two extractors applied to the same Preview capture | Private differences only; silent |
The hosted Preview accepts only Preview or Compare, in observe mode, with silent notifications and a maximum batch of 50 policies. Stable scans and scheduler entrypoints are unavailable there. The broader local channel selector was tested separately and must not be mistaken for a Stable scan option on the hosted trial.
Comparison is not two independent Internet retrievals and is not a comparison of AI models. Identical text does not establish complete source coverage. Pending, unavailable, not-requested and identical-but-unqualified outcomes remain distinct.
flowchart TD
A[Administrator] --> B{Choose environment}
B -->|Main site, online| C[Product 5.5 / established scanner]
C --> D[(Production database and existing evidence gates)]
B -->|Separate staging, online and private| E[Authenticated 6.0 Preview]
E --> F{Preview or Compare, at most 50 policies}
F --> G[(Dedicated private queue)]
H[Node instrumentation starts one leased worker] --> G
G --> I[Acquire and freeze source evidence]
I --> J[Source contract and extraction checks]
J --> K[Private outcome or review hold]
I --> L{Compare requested?}
L -->|Yes| M[Two extractors on the same capture]
M --> N[Private frozen report and JSON export]
The source is on main. The hosted operating boundary specifies the production-mode guard, allowed staging origin and dedicated database. Pages, APIs, feeds, RSC and prefetch responses require administrator authentication. The immutable preview identity retains private access even when the execution opt-in flag is missing.
Next.js Node instrumentation starts the hosted worker automatically after its guard passes; a separate legacy CLI worker must not be started for this environment. The worker is excluded during build, uses the same bundled implementation identity as HTTP, processes at most one hosted task at a time and does not deliver the notification outbox. Credentials remain private.
Candidate 3 passed 1,565 automated tests with 14 skipped and 22 checks against the actual generated standalone server. Candidate 2 separately passed 44 HTTP/worker checks, including real Apple retrieval, same-capture comparison and recovery after a process crash. The first Hostinger deployment failed integrity preflight before migrations because the provider wrapped next.config.ts. The reviewed correction recognizes only that exact transformation. The second deployment reached Completed / Current at 23:17 Europe/Rome, and 47 live checks subsequently passed on the hosted runtime.
The private administrator console is online. Actual hosted run 4e9eef90-0bbb-42ee-b945-922ac76ffffc completed in observation/silent mode. Apple returned HTTP 200 over direct retrieval and produced 28,124 Preview characters versus 2,089 from the stable extractor on the same response, with confirmation_pending and no AI analysis or publication. Three other selected sources remained held for source_contract_missing. Full-row digests showed no change to official policy, baseline/history/log, source-contract, subscriber or outbox state.
Administrator login, the hydrated channel selector and the saved completed report were separately verified in the browser. The report rendered one different result, three unavailable comparisons and 2,089 versus 28,124 Apple characters. The earlier browser JSON export was unverified after an observer timeout. On 2 November the actual downloaded file was found and parsed after a fresh browser click, matching the saved API report. Its download-event observer also timed out; that retained tooling result is separate from the verified completed file. An earlier initial GET timed out after 30 seconds and is retained; a subsequent GET returned the expected 307 in 576 ms. This was one successful 47-check suite, not two complete suites. Hosted deployment receipt.
At the earlier 03:24 Europe/Rome checkpoint on 2 November, the public homepage access banner and bilingual notebook were published through staging and an identical production archive, retaining runtime 5.2.0. Staging passed 11 required checks plus 22 content checks; production passed 22 content checks and nine internal runtime checks. The independent security test remains external. The same private Preview was restored at 03:28 Europe/Rome: 12 restoration checks passed and full-row digests across 12 observed tables stayed unchanged. The existing report was exported in the browser: its 22,649-byte JSON file equals the saved API response. No new capture or daily window was counted. Publication and restoration receipt. Future content-only staging promotions must likewise restore and verify the private Preview.
PR #39, candidate 050af19, corrects literal backslash escaping in extracted table cells. The original PR #37 email concerns three filesystem race instances already marked fixed and six intentional evidence-file writes dismissed with documented reasons. A subsequent main-branch analysis raised a separate incomplete-escaping alert, 91. The encoding defect was reproduced; an injection exploit was not established in the inspected text-only consumers.
The first CI execution caught an omitted whole-engine manifest regeneration and correctly disabled experimental execution; its 11 test failures and build rejection are preserved. The corrected candidate passed 1,569 local tests with 14 skipped, including four new table regressions, plus TypeScript, critical-test type checking, scoped lint and both identity checks. GitHub CI is evaluated separately on PR #39. An offline replay of the 12 saved 1 November captures preserved all 12 extracted-text hashes; it adds zero live dates and grants no completeness approval. The local comparison receipt has SHA-256 3921da7fd865938810af56dba3e02a926a9be2e549a0bce5a380c44b969af1db.
The candidate is not deployed: the private hosted Preview still uses its recorded r2 artifact. Because the extraction profile changes, its live qualification must start a new cohort. The existing 12-source, two-date cohort and its failures remain preserved in the original checkout. Detailed follow-up.
All six GitHub checks subsequently passed on 050af19, including CodeQL analysis 2073189355 with zero findings. PR #39 was integrated into main as 9443be0 at 22:48:06 UTC. This is source integration, not deployment. The new cohort engine-daily-cohort-20261101-table-escaping has manifest hash 395db5778b5fab8d15d2add676547547d27422fcedc5c991f26057beeacf0e38 and zero live windows at initialization. The existing daily automation now targets this candidate, with the next scheduled observation on 2 November at 21:00 Europe/Rome. Old captures are retained but cannot supply dates for the changed parser.
Post-merge CodeQL analysis 2073199042 completed against 9443be0, and GitHub now marks alert 91 fixed, not dismissed. The zero-result statement above refers to PR #39, not the entire repository's alert history: two earlier CodeQL alerts, 75 and 76 in scripts/submit-indexnow.mjs, remain open and were not reassessed in this Engine correction.
The dedicated database was bootstrapped from an explicit public-data allowlist: 18 companies, 113 configured policies, 110 prior public baseline texts and one bounded Codex-assisted source contract. Subscriber, session, private communication, access-log, notification and historical run data were not copied. A copied public baseline is not a newly verified Engine 6 baseline.
The Apple HTML retrieval override belongs only to the hosted smoke corpus. Both the separately recorded local capture and the new hosted capture produced 28,124 characters versus 2,089 from the stable extractor on each same response. Its outcome was confirmation_pending, without AI analysis or publication. This is extraction divergence, not a policy-change claim or universal completeness proof. Source review.
The frozen daily cohort remains unchanged: 12 sources, two accepted dates and a minimum actual span of 20.7712 hours. Source-by-source completeness, at least three dates and 48 elapsed hours, capacity for 500 distinct daily sources, managed-process survival and storage behaviour remain open qualifications.
See Validation and Release Status, the source-review runbook and the major-engine design.
Documentation checkpoint: 2 November 2026. Product 5.5.0 online; Engine 6 source on main; private Preview online with administrator authentication. Stable service · Engine 6 notebook · Validation and Release Status