Router CSI Lab
EN ZH

Corrections and data revisions: what changed and why

By Wu Jun · Updated 2026-11-02 · corrections, data revision, transparency, Wi-Fi CSI, evidence · 6 min read

Technical content becomes less trustworthy when errors disappear without a record. This page lists substantive corrections to methods, measurements, capability claims and public links. It does not list every spelling change or layout adjustment.

The operating rule is simple: preserve the failed version in the project record, explain the affected scope, correct both language editions, rerun the verification, and publish a new date or data version.

Correction log

Date What was wrong Current resolution
2026-11-02 Every page declared summary_large_image, but the HTML had no og:image or twitter:image, so social platforms could only render a text link. Post Inspector also mistook a firmware version for the publication date and reported no author. Added a 1200 × 630 JPEG generated from the real GL-MT6000 photo and the public 256-subcarrier heatmap. Every page now emits an absolute image URL, dimensions, MIME type and localized alternative text. Article pages also emit author, article:published_time and article:modified_time, preventing platforms from guessing dates from body text. A generator, JPEG boundary test and site-wide metadata gate are wired into deployment.
2026-11-02 The homepage, cross-page links and troubleshooting conclusion used wording such as “nobody else has written this”, “exclusive”, “root cause” and “proves”, exceeding the evidence and contradicting the article's own statement that firmware internals were not verified. The affected copy now uses bounded claims such as “based on this device's A/B test”, “the comparison supports an association” and “the firmware's internal mechanism remains unverified”. A new npm run test:claims gate scans content and page generators for regressions in unsupported exclusivity or causal-proof wording.
2026-11-01 The RSS feed advertised pages marked noindex, and English pages referenced a feed containing only Chinese items. Both feeds now exclude noindex pages. The Chinese feed and /en/rss.xml each contain 35 language-scoped items and declare their language. Four sitemap/RSS entries are listed in robots.txt; automated checks and Googlebot-UA requests confirm HTTP 200 application/xml.
2026-11-01 Organization, Article, AboutPage and founder were separate inline objects, so a search engine could not confirm that they described the same author and publisher. They now share stable @id values: /about/#person for the author and /#organization for the publisher. Organization publishing principles point to the methodology and correction pages, with an automated entity-graph regression check.
2026-10-31 The site kept only the GitHub profile after the old repository returned HTTP 404 to signed-out visitors, so it did not directly expose verifiable toolkit source. A new MIT toolkit repository and v0.1.2 Release are public. About, Tools, the footer and Organization sameAs now link directly to router-csi-toolkit, with anonymous access to CI results and the release checksum.
2026-10-29 The browser demo counted a final, unclosed CSI record as a complete 348th group. Grouping now closes only on chain_info bit 15. The public 60-second source contains 2,777 chain records, 347 complete groups and one tail fragment. The generator and page statistics were corrected and the source SHA-256 was rechecked.
2026-10-29 Some pages wrote 95.5 chain records/s as though it were a 95.5 Hz independent sample rate, and mixed closed groups with the later filtered analysis set. The 10-minute file is reported as 57,293 chain records, 7,609 bit-15 groups, about 12.68 closed groups/s, and 7,583 groups after a documented 25-group anomaly filter, about 12.64 analysed groups/s. The capture-quality audit now exposes the actual irregular intervals.
2026-10-29 A Home Assistant draft implied the router platform had an /api/csi/status endpoint and an MQTT publisher that it did not implement. The unsupported endpoint and topics were removed. The page now distinguishes the platform's REST/WebSocket surface from third-party ESPHome/MQTT routes and states that the router platform has no built-in MQTT publisher.
2026-10-29 A Raspberry Pi/Nexmon draft was too broad about current Pi 5 compatibility and rollback steps. The pages now state the modern Makefile.rpi and USE_VENDOR_CMD=1 path, appropriate kernel-page-size requirement, network-interface precautions, restore-wifi rollback and the need for a dedicated experiment card.
2026-10-29 An ESP32 page treated packet-by-packet variance as if it were calibrated human detection. The method now separates a motion proxy from event detection: capture-quality checks, per-subcarrier median/MAD baseline, time smoothing, hysteresis, independent session calibration and event-level false-alarm metrics. “No motion detected” is no longer presented as proof that a room is empty.
2026-10-29 The public site linked to a GitHub repository URL that returned HTTP 404 to signed-out visitors. The inaccessible URL was removed from About, the footer and Organization sameAs. The public GitHub profile is linked instead, while scripts and data remain available through the site's direct downloads and Zenodo record.
2026-10-29 The 404 template linked the Chinese language switch to a non-existent /zh/ path. The fallback now links to the Chinese root /. The automated link audit reports zero missing internal targets.
2026-10-29 The privacy and dataset pages said the public captures contained “no personal information” and treated a MAC-only rewrite as anonymisation. They now describe the files as MAC-de-identified or pseudonymised data, state that outer absolute timestamps, activity rhythms and within-session linkage remain, and link to a threat model, audit procedure and rules for a future time-shifted release.

Severity levels

Level Examples Required response
Critical leaked personal data, a non-existent citation used as proof, unsafe medical/safety conclusion remove or prominently flag first; notify affected readers where possible; correct the evidence chain
Major wrong sampling unit, claim about a hardware capability that was never implemented, incorrect reproducibility command correct both languages, rerun data checks, add the issue to this log
Minor broken link, 404 URL, formatting that obscures a command correct in the next maintenance batch; link failures are retained in the automated report

What the correction process covers

  1. Reproduce the issue. Reproduce the error or identify the authority for the correction.
  2. Find the affected scope. Search both languages and all pages that inherit the claim.
  3. Fix the source, not only the HTML. Correct the Markdown, generator, script or data package so a rebuild does not reintroduce the error.
  4. Recompute identifiers. If a published data file changes, update its SHA-256 and any affected counts. Do not silently replace a DOI under the same version.
  5. Run the technical checks. Build, content assertions, internal/external link audit and mobile rendering checks.
  6. Record the change. Update this page and the article's dateModified when the change is substantive.

The public 10-minute capture and toolkit are versioned separately [1] [2]. A future change to the raw data must be published as a new Zenodo version while the existing DOI remains resolvable [3].

Data revisions

The current 10-minute dataset keeps the original bytes and SHA-256. Corrections so far changed interpretation, documentation or derived demo data; they did not rewrite the original capture under the same filename.

Future data changes will be handled as follows:

AI assistance and responsibility

AI tools may help draft, translate, compare sources or generate audit scaffolding. They are not treated as a source and do not receive author responsibility. A claim is retained only when the author can connect it to a public source, a reproducible script, a raw measurement or an explicitly labelled limitation. The broader policy is in methodology.

Report a correction

Email okwujun@gmail.com with:

We normally reply within seven business days. A correction is not accepted merely because it is strongly worded: the claim and evidence are checked separately.

This log also has a limit: it proves that known issues were handled transparently, not that every future conclusion is infallible. The raw data, scripts, references and failure conditions remain the primary evidence.

References and data sources

Every external claim carries a source and an evidence grade (grading scheme). Grade E marks first-party, reproducible measurements from this project.

  1. Grade E · first-party measurement This project's 10-minute raw capture (57,293 records / 600.0 s, SHA-256 00373f83…95e2)
    Router CSI Lab · 2026
    okwujun.pw/en/datasets/ · verified: Reassembled chunks verified byte-identical (2026-10-11) · The local raw master has SHA-256 00373f83…95e2; the public de-identified derivative has SHA-256 527466ce…a440. They must not be confused.
  2. Grade E · first-party measurement Router CSI Toolkit — capture/analysis/watchdog scripts (MIT, public repository)
    Router CSI Lab · 2026
    github.com/jjjwu/router-csi-toolkit · verified: GitHub API verified the public repository, capture-quality preflight, successful CI and the v0.1.3 release on 2026-11-02; the archive SHA-256 is 59f144c0…adea7 · The public repository includes a dependency-free capture preflight; the browser checker uses the same container definition and evidence boundaries.
  3. Grade E · first-party measurement Router CSI Lab — GL-MT6000 (MT7986) Wi-Fi CSI dataset, 10 s sample + 10 min capture
    Router CSI Lab · 2026
    doi:10.5281/zenodo.23434789 · verified: CC BY 4.0; SHA-256 published; Zenodo DOI 10.5281/zenodo.23434789 · The dataset has a Zenodo DOI (10.5281/zenodo.23434789) and can be cited directly.