The site needs a more reliable path from a published assertion to its exact evidence. The first work should prevent stale evidence from appearing under a different record URL, preserve keyboard position, and make a selected assertion or source shareable. The private workspace also needs reliable startup, readable policy tables and a direct path from a repeated-aircraft count to the records counted.
This plan covers the public investigation and the private research workspace as separate products. The public journey should continue to center on financial interests, corporate control, end users and policy decisions. Aircraft comparisons belong in the private follow-up tools unless a separately corroborated finding warrants publication. No aircraft timing result in this review establishes a passenger, meeting or policy intervention.
Review coverage and limits
Three Codex agents reviewed information architecture, accessibility and frontend/data behavior. Claude CLI and Kimi CLI completed separate source reviews. Root reproduced additional cases against the actual retained data and adjudicated the findings. The review inspected the public application, model, HTML, CSS, build and privacy checks, plus private navigation, evidence dialogs, source library, knowledge explorer and historical-comparison modules.
Executed local DOM checks reproduced the stale-source dialog, lost focus after loading more records, the skip link changing the active view, an optional-file failure blocking the private workspace, duplicate comparison loading that clears a typed filter, missing repeat-row evidence actions and the policy table becoming plain paragraphs. CSS contrast was calculated from declared colors. These are source and DOM observations, not browser-rendered results.
No browser, screen reader, viewport layout or real-user performance session was run. Mobile overflow, actual target geometry, accessibility-tree behavior and Core Web Vitals remain unverified. The plan includes those acceptance checks. The private source changed during the round to add ownership records and correct an outdated publication-state note; those small changes do not implement the navigation and accessibility work below.
What to preserve
- The public article remains readable with direct source links when JavaScript or the evidence library is unavailable.
- Public publication uses an explicit allowlist, separate from private captures, research logs, aircraft exports and account metadata.
- Public evidence text is escaped, URLs are constrained, diagrams use strict rendering, and maps offer text and source fallbacks.
- The public evidence modal has a name, a close control and working focus restoration in the tested DOM case.
- Roles, ownership, forecasts, receipts, source statements and inferences remain distinguishable. Keep these distinctions beside the relevant claim.
- Existing typography and color tokens should be reused. This review does not call for a framework replacement, a new brand or analytics.
Priority work
| ID | Priority | Observed problem or proposed improvement | Implementation direction | Acceptance check |
|---|---|---|---|---|
| U01 | P1 | A valid source dialog remains visible after navigation to a missing source ID. The URL and displayed evidence disagree. | Resolve the requested ID before display; replace stale content with an explicit unavailable-record state. Handle invalid claim/source combinations deliberately. | Navigate valid source to missing source, then valid claim to missing claim. No old record remains. Exercise direct entry and browser back/forward. |
| U02 | P1 | Opening evidence does not update the URL or expose a copyable citation. Search and selected-entity state are not shareable. | Give assertions, sources and entities stable links. Serialize relevant filters and selected records. Add Copy link and Copy citation controls with a clear success/failure announcement. | A copied link restores the same record and context in a new session. Closing and Back restore the prior list or article position. Invalid old-edition links explain the problem. |
| U03 | P1 | Public Show more replaces its focused button; DOM focus becomes BODY. Navigation does not set a destination focus. The skip fragment changes Evidence to Investigation. | Define one route/focus contract. Keep the active view for Skip to main. Append result batches or move focus deliberately. Preserve heading fragments and filter typing. | Keyboard-only use reaches new records without restarting at the header. Skip retains the current view and reaches main. Route changes announce/focus the destination; filtering does not steal focus. |
| U04 | P1 | An optional thesis-file transport failure blocks the otherwise valid private article and register. Late bootstrap reloads comparisons and erases a typed filter. | Isolate optional resources; load once through a shared initialization state; use per-section loading, retry and unavailable states. Discard stale route responses. | Reject each optional request independently. Article/register still work. Direct comparison entry makes one effective load and preserves typed filters when the core data settles. Rapid navigation cannot replace the destination with a late response. |
| U05 | P1 | The private policy comparison table becomes six pipe-separated paragraphs. | Reuse the reviewed Markdown/table build path or a deliberately limited shared renderer. Preserve links, headings and escaping. | The current policy comparison produces a semantic table with the expected header and four issue rows. Compare its text and links with the approved article. |
| U06 | P1 | The repeated-aircraft view has 22 aggregate rows and no action to inspect the events counted. | Add Inspect matching records, keyed by tail and hex. Show counted visit groups, distinct flight records, source pointers and exclusions. Keep current title and charter authority in separate dated fields. | Every displayed repeat count reconciles to its contributing records. Inspect works by keyboard. Multiple boundaries from one derived stop are not counted as separate visits. No meeting or occupant label is introduced. |
| U07 | P2 | The private dialog lacks an explicit accessible name. The declared public focus-ring color has only 2.533:1 contrast against the dark masthead. | Apply a named-dialog contract to every evidence trigger. Use a light or two-color focus indicator on dark surfaces, with deliberate focus return. | Inspect the browser accessibility tree and keyboard flow. Measure contrast against the actual adjacent colors. Check long dialogs and close controls at narrow widths and zoom. |
| U08 | P2 | The comparison table carries the wrapper class on the table itself, missing the descendant selector. Date controls lack visible labels and tables lack useful captions. | Use a named, keyboard-accessible scroll wrapper; visible UTC labels; captions and column/row headers. Provide a compact summary before detailed tables. | At 320 CSS pixels and 200% zoom, all cells and actions remain reachable within the table region. No accidental page-wide panning. Header associations are understandable with a screen reader. |
| U09 | P2 | The private source view inserts every card at once. The baseline snapshot rendered 976 cards and 15,559 descendant elements. Graph search repeatedly stringifies all node objects. | Page results, precompute a values-only search index and debounce input. Load route-specific datasets on demand. Keep full offline exports available. | Counts reconcile to the complete dataset; every page is reachable. A search for a JSON field name does not match every record merely because that key exists. Measure browser input responsiveness before setting a performance claim. |
| U10 | P2 | A journalist must know the vocabulary of the register to find the right evidence. Public text search has no date, source-type or status facets. | Add reviewed starting points for family control, compensation, counterparties/end users, policy chronology and missing communications. Add concise facets and Clear filters. | A reader can find the source for a named financial relationship, distinguish reporting from an original instrument, and recover from zero results without downloading raw JSON. Preserve the query in the URL. |
| U11 | P2 | A source dialog can combine many assertions without identifying the article passage that opened it. Some quotations are intentionally omitted. | Open the relevant assertion and locator first, with the broader source record secondary. Say selected excerpts rather than promise a quotation for every citation. Add edition and claim dates to copied citations. | Starting from an article paragraph, the first evidence shown supports that paragraph. A null excerpt displays its locator and source link, without an empty quotation or a promise of unavailable text. |
| U12 | P3 | Charts are easier to read as an overview than as a complete audit trail. Private graph results and neighbors are capped. | Add task-specific entry points, visible edge dates/types and an equivalent readable list. Make truncation explicit and provide pagination or a refined-search route. Preserve independent counts and source-family limits. | Follow one ownership or financing edge from chart to dated assertion and source, then return without losing selection. Every truncated list states its total and offers a way to inspect remaining records. |
| U13 | P3 | Divergent renderers, dense source lines and repeated helper logic make changes harder to review. Some fallback cases are latent schema risks rather than current failures. | Format source for review, share small audited formatting/URL/status utilities, and keep publication validation separate. Add explicit entity-ID joins and tests before replacing name-based display associations. | Existing output/schema/privacy tests still pass. A same-name fixture cannot attach an unrelated entity. No private dataset, account identifier or source map enters the public payload. |
P1 means evidence reliability or a central reporting task is impaired. P2 improves accessibility, resilience and efficient review. P3 improves exploration and maintenance after those contracts are stable. The severity labels describe this investigation's priorities; they are not security ratings or a declaration of accessibility noncompliance.
Proposed navigation and reading flow
Keep the public top level compact: Investigation, Evidence, Companies and control, Policy timeline, Connections, and Sources and methods. Add a short What the records establish / What remains missing panel near the article introduction. It should link to the relevant assertions rather than repeat a generic disclaimer. A Policy timeline can begin as a filtered evidence view; it does not require a separate application.
In the private workspace, group the primary reporting tools together: core thesis, corporate identities/control, financing and end users, policy/communications, and source search. Place aircraft, connector diagnostics and other background lanes in a secondary research-tools group. Keep their evidence accessible without giving them the same prominence as the central financial-policy question. The public navigation currently excludes aircraft tools; preserve that separation.
The main reporting sequence should be: question, dated assertion, exact passage or locator, original source, related records, remaining gap. An evidence detail view should preserve the originating paragraph or search, show whether the statement is documented, reported or inferred, and offer a stable citation. Display underlying ownership, legal title, management and service roles separately. Do not give a graph edge a stronger meaning than the supporting record.
Delivery sequence
- Evidence and state contracts: U01, U02, U03 and U04. Define URL, history, focus and loading behavior together. They affect every later screen. Add the reproduced cases as regression tests before changing the shared behavior.
- Reporting surfaces: U05, U06, U07, U08 and U11. Restore the policy table, make repeat counts inspectable, and finish modal/table accessibility. Reuse existing visual tokens and the public evidence patterns that already work.
- Finding and comparing evidence: U09, U10, U12 and U13. Add filters, paging, indexed search and chart navigation using the stable state contracts. Measure actual performance before deciding whether a worker or additional framework is needed.
A reasonable planning unit is a small change with its own before/after fixture, source-privacy check and deployment receipt. These are sequencing recommendations, not completed changes or calendar commitments. No purchase, new service, outreach or analytics is required.
Verification before calling the redesign complete
Retain the current build, application and publication-privacy checks. Add the reproduced failures to meaningful regression fixtures, including valid-to-invalid evidence links, late requests, optional-resource failures, filter restoration and repeated-aircraft count reconciliation.
Then perform a browser matrix at 320, 390, 768 and 1280 CSS pixels, 200% zoom, keyboard-only operation, reduced motion and a screen reader. Verify names and focus in the actual accessibility tree, dialog scrolling and return, horizontal table access, article print output and chart text alternatives. Record the tested browser and assistive-technology versions.
Use WCAG 2.2 AA as the proposed accessibility acceptance baseline, with documented exceptions and actual testing. Its minimum pointer-target criterion is 24 by 24 CSS pixels with exceptions; 44 by 44 is a useful stronger design goal, not an automatic WCAG 2.1 AA requirement. See W3C target-size guidance and the W3C modal-dialog pattern. Neither a DOM harness nor a contrast calculation establishes full conformance.
Keep the public edition's default-origin restrictions, lack of analytics, explicit export allowlist and identity separation. Technical diagnostics and this review belong in the private workspace. FOIA retrievals, requests and status work remain stopped.
Review adjudication
The CLI reviewers identified useful additional edge cases, but their proposed severities were not accepted wholesale. Missing/null URLs, empty quotations and unknown statuses are already constrained by the current public export schema or were not present in the retained edition; treat their remaining cases as robustness tests. The fixed count of forty entities is an intentional research-universe requirement. No framework migration, broad component rewrite or current claim of wrongful entity attribution is justified by those suggestions.
The source-only review covers the frontend implementation and its reporting workflows. A full rendered usability and accessibility pass remains a separate acceptance step. Native LanguageTool and Vale review is recorded for this plan; a permitted personal writing reference was not available within the investigation's source boundary, and the unavailable editorial-memory registry prevents a full editorial-gate claim.