Owner: Joschka Witte (admin@dadchecks.tech) — see the Impressum.
Last reviewed: 2026-11-16. Next review: 2027-05-15.
This page is a description of how editorial work on dadchecks.tech is actually
produced and checked. It is written to be verifiable, not reassuring. Where a
claim here would be nice to make but is not true of this site, it is not made.
1. How products are selected
Nothing is selected because a manufacturer sent something. Selection starts from
a question a reader actually has, and is filtered by three things:
- Fit for the site’s three categories. Smart Home, Home Lab, Dad Gaming.
A topic outside those categories is not covered, even if it is a good product. - Evidence availability. A product is only a candidate if its behaviour can
be established from evidence that survives review: manufacturer
documentation, datasheets, release notes and independently reproducible
sources. There is no first-hand testing on this site. Every product
that cannot be evidenced from published sources is written about as
research, or not at all. - Interest, not novelty. The criterion is “will this still be useful in a
year”, not “is this new this week”.
A product being new, popular, or already discounted is not a reason to cover it.
2. What “hands-on testing” means on this site
This site publishes research-based articles. It does not publish hands-on
tests. No article on this site claims that the owner tested a product, took
a measurement of the owner’s own, or made a purchase decision for the
reader’s benefit, and no title promises any of that. Measured on 2026-11-15
across all 22 of the site’s article bodies — the sixteen that were live on
that date and the six rewritten articles that were held — that claim is made
zero times, and 0 of the 22 titles promises any of it either.
Some of these articles describe installations the owner actually made: the
mini console and the Home Assistant migration are the owner’s own systems,
written up as a method for the reader rather than as a report of what was
found. Owning a system is not the same thing as testing a product, and the
site does not present it as one: those articles document the procedure and the
configuration from published documentation and vendor sources, and they do
not report a session, a measurement or a purchase as the owner’s own
experience.
“Hands-on” is therefore a term the site uses only to reject a claim: an
article that says it was tested, that a number was measured by the owner, or
that the owner’s own purchase supports the conclusion is not a hands-on
review, and it is reclassified into the editorial type it actually meets
before publication.
The bar a Research Guide has to clear is different, and it is the bar this
site actually works to:
- the conclusion is attributable to published sources — datasheets, manuals,
release notes, vendor documentation — and the source is named; - specifications and prices are current as of a stated capture date;
- arithmetic in the article is verifiable and correct;
- no first-person experience is asserted anywhere in the text.
If an article cannot meet that bar, it does not get published. It is
reclassified into the editorial type it actually meets, and its title is
written so that it does not claim what the article does not deliver.
This is enforced mechanically, not by good intentions. Every article is
classified into exactly one editorial type:
| Editorial type | Evidence bar |
|---|---|
| Hands-On Review | Owner confirmation, test notes, supporting evidence |
| Build Log | Build records, configurations, test evidence |
| Research Guide | Reliable citations and verified specifications |
| Comparison / Buying Guide | Verifiable data, transparent methodology |
| Opinion / Editorial | Clear separation of opinion from factual claims |
The Hands-On Review type is defined but currently unused: no published
article claims first-hand testing, and none of the sixteen articles that
were live on 2026-11-15 is classified as a Hands-On Review. The Build Log
type is in use — seven of those sixteen are classified as Build Logs
(re-measured 2026-11-15) — and what those seven carry is the procedure and
the configuration: the ordered steps, the settings and the cut-over plan.
They do not carry a first-hand test record. Both types are listed so that
the distinction is explicit, and so that a future genuinely hands-on article
would have a defined evidence bar to meet rather than an invented one.
A title that promises hands-on testing (“Tested for 30 days”, “What I bought”,
“My honest review”, “I built it”) while the body contains no first-person
testing evidence is treated as a critical finding, not a style note. It
blocks publication until the article is either substantiated or retitled.
3. How research-based reviews differ from personal tests
This site publishes research-based articles only, and does not blur that
into a personal test.
A research-based article reaches its conclusions from published
documentation, datasheets, release notes, and reproducible technical
information. Its conclusions are attributed to those sources. It does not claim
to have owned, opened, installed, or measured anything. It is a legitimate and
valuable form of content here — it is simply not a personal test, and it is
never presented as one.
A hands-on article would carry the owner’s first-hand experience. There
are none on this site at present, so every conclusion on this site is
attributed to a source rather than to personal experience. If a hands-on
article is ever published, its first-hand part is confined to what the owner
actually did, and everything else stays attributed.
The mechanical check for this distinction is the claim extractor: each factual
statement in an article is pulled out as a discrete claim and tagged by type —
first-person, technical specification, compatibility, recommendation, opinion,
and so on. First-person claims are then verified against owner confirmation,
not against prose that sounds confident.
4. How AI assists the editorial process
AI is used as a tool inside a checked editorial process, and the process is
described here accurately rather than defensively.
What AI actually does on this site:
- Research assistance. Gathering and summarising manufacturer documentation,
release notes, and datasheets, so that the source material itself is not replaced by a summary of it. - Draft assistance. Structuring a draft, tightening prose, and producing a
first pass at headings and summaries. - Mechanical checking. A validation pipeline runs over every article before
it can be published. It extracts claims, checks them against technical
feasibility, arithmetic, hardware and software compatibility, source quality,
currency, internal and cross-article consistency, and personal-experience
claims. It scores the article and enforces hard publication blocks. - Illustration. Header and thumbnail illustrations are AI-generated and are
labelled as illustrations. Product images in the shop are AI-generated
illustrations and are labelled as such. - Site maintenance. Routine upkeep such as checking that published links
still resolve.
What AI does not do:
- It does not decide whether a product is good. It drafts; the automated validation agents of the editorial pipeline check the draft and decide whether it may go live. The owner is accountable for what the site publishes, and is the provider named in the imprint.
- It does not confirm a first-person experience, because this site does not
publish first-person experience. Only the owner could, and there is none. - It is not a substitute for a source. A claim is not established because
several checks or several agents agree with it — agreement is not evidence. - It does not publish. The publication gate is separate from the AI tooling.
Its pipeline-level half is in force; its server-side WordPress half is not
yet deployed to production (see “How these rules are enforced” below).
Every article carries a visible notice that it was produced with AI assistance by the DadChecks Editorial desk and checked by the automated validation agents of the editorial pipeline. No individual person wrote or tested the article. See the AI Transparency Policy
for the full statement, including which legal transparency duties apply.
5. How affiliate links are handled
dadchecks.tech participates in affiliate programmes, principally Amazon
Associates and Awin. This is how the site is funded; it does not decide what
gets covered or what gets concluded.
The rules:
- The disclosure comes first. Every article containing affiliate links
carries a disclosure block at the top of the article body, before the
content. - Affiliate links are marked in the markup. Every affiliate link on the
site carriesrel="nofollow noopener sponsored", so it is identifiable as
advertising to search engines and to assistive technology, not only by eye.
The visible marking convention is an asterisk on the link text, explained by
the disclosure block. - The editorial conclusion is not for sale. A commission never influences
which product is recommended, how a product is described, or whether a
weakness is mentioned. Negative findings are published. - Price is stated with its date. Prices are captured with a verification
date. Prices move, so a price without a date is treated as a defect rather
than a fact. - A retailer listing is not a source. A product page on a retailer is the
weakest available source tier. A claim resting only on a retailer listing
does not meet the evidence bar on its own.
The German-language disclosure currently used on articles reads:
Affiliate-Links: Bei Links mit * erhalten wir eine Provision. Für dich ändert
sich der Preis nicht. Details in der Datenschutzerklärung.
No measured gap against the rules stated above is open (see
audit/site-defects.md), as of 2026-11-16. The two gaps this page previously
reported were closed by the article rewrite: re-measured on 2026-11-16, the
asterisk gap is gone — the marker is on every affiliate link in the site’s 22
article bodies, 173 of 173 (109 in the sixteen articles that were live on that
date, 64 in the six rewritten articles that were held), every one of them
carrying the full rel="nofollow noopener sponsored" — and the disclosure
block comes before the first affiliate link and before the article’s own first
paragraph in all 22 bodies, including the 8BitDo controller article, where it
used to come after the first affiliate link. The count is higher than the 142
of the same measurement on 2026-11-15 because the sixteen published articles
were re-rendered with computed cost tables, which carry more product links
than the hand-written bills they replaced; the marker was on every link at
both measurements.
The two gaps that were open in their place were measured on 2026-11-15, and both
were closed on 2026-11-16. In the sixteen articles that were live on that date,
ninety price entries carried no capture date, although the rule above treats a
price without a date as a defect rather than a fact: the six rewritten articles
stated a capture date for their bills, and the live sixteen did not state one at
all. That is corrected — every price cell in the sixteen published articles now
carries its capture date in the same cell as the figure, 128 of 128, from the
retailer listing capture of 2026-11-15. And in the six rewritten articles — held
and not publicly visible on that date — the lead paragraph sat above the disclosure
block, where the rule above puts the disclosure first; in the Zigbee stick article
that lead paragraph already named a product carrying the asterisk, so a reader
would meet the marker before the sentence that explains it. That is corrected
too — the disclosure block and the AI notice are now the first blocks in the body
of all six, above the article’s own first paragraph and above the first affiliate
link.
6. How factual corrections are made
Corrections are published, not silently applied.
When a factual error is found:
- It is logged. The affected claim is identified with the article, the
claim itself, the evidence that contradicts it, and the date. - It is routed. A claim that only the owner can settle waits for owner
evidence. Everything else is corrected by editing the article. The article
moves through a visible “Needs Update / Correction” state while this
happens. - The correction is made. The article is corrected and, where the error
was material to a reader’s decision, the corrected article says so
explicitly, stating what was wrong and when it was corrected. A correction
notice is not quietly deleted along with the error. - The shared fact registry is updated. Facts that recur across articles are
held in one registry, so a correction to a fact propagates to every article
that relies on it. A correction in one place does not leave contradictions
standing elsewhere. - Published content is not silently rewritten. Existing URLs are
preserved. Where a URL has to change, a redirect is added. Revisions and
originals are archived before a change, so a change is always auditable.
Anyone who finds an error can report it to admin@dadchecks.tech. Genuine
corrections are made and, where material, acknowledged.
7. Whether manufacturers can influence published conclusions
No. Manufacturers do not have a say in the conclusions on this site, and they
cannot buy a review.
Concretely:
- No product review is published as the result of a commercial arrangement.
Free review units, if they are ever accepted, are disclosed at the top of the
article and do not change the conclusion or the rating. Where a sample was
received, the article says so — including the possibility that the sample was
not kept. - No manufacturer or agency may review, approve, or pre-clear an article before
it is published. The validation pipeline flags commercial claims for
editorial review, and a commercial interest is not a route through the gate. - No publisher, agency, or advertiser can pay for a mention, a ranking
position, or the removal of a criticism. Commercial arrangements are
disclosed, and the editorial conclusion stands on its own. - A manufacturer disputing a factual statement does not get to overrule the
evidence. A documented error is corrected; a disagreement that is not backed
by evidence is not. - Commercial and editorial are kept structurally separate: a sponsor never
appears in the article body as an editorial voice, and no sponsored item is
presented as an independent recommendation.
To be transparent about a limit: this policy describes our process and our
commitment. It does not describe a third party’s behaviour. If a reader
believes a conclusion was influenced, the way to settle it is to ask —
editorial questions go to admin@dadchecks.tech, and they will be answered
directly, including with the reasoning behind a conclusion.
How these rules are enforced
The commitments on this page are backed by a validation pipeline that runs
before publication, not by a promise in prose.
Every article passes through a fixed workflow:
Idea → Researching → Draft → Verification → Needs Revision → Awaiting Owner
Evidence → Editorial Approval → Approved → Published → Needs Update /
Correction
A validation result determines the status an article may occupy:
- A material unsupported claim that only the owner can settle → Awaiting
Owner Evidence. A reviewer cannot edit an article into compliance with a
first-person claim; only the owner can. - Any mandatory override → Needs Revision.
- A score band of “revision required” or “hold” → Needs Revision.
- An unresolved critical finding → Needs Revision.
- Otherwise → Editorial Approval. Even a clean article is not published on the strength of its score alone; the automated validation agents must clear it first.
Publication is blocked regardless of score when any of these fire:
- fabricated personal testing
- fabricated purchases or measurements
- materially misleading product specifications
- unsafe technical instructions
- serious unresolved factual contradictions
- unverified or misleading commercial claims
- material legal or compliance issues
- missing required approval
Claims carry an evidence level:
| Level | Meaning |
|---|---|
| E3 — Direct evidence | Confirmed by the owner and supported where practical by measurements, screenshots, logs, photographs, invoices, or test records |
| E2 — Verified research | Supported by authoritative external documentation or independently reproducible technical information |
| E1 — Limited evidence | Plausible, but supported by weak, incomplete, outdated, or indirect sources |
| E0 — Unsupported | No reliable evidence, contradicted information, or an unverified first-hand claim |
Material E0 claims are blocked. Material E1 claims require qualification or
stronger research. Hands-on claims require genuine first-hand confirmation.
Sensitive, safety-related, or high-impact claims require stronger evidence than
ordinary low-risk observations.
Every claim is checked against nine verification categories. The nine
categories are: technical feasibility, real-world plausibility, mathematical
correctness, hardware/software compatibility, source quality, temporal
validity, internal consistency, cross-article consistency, and
personal-experience verification. Two further categories — commercial/link
transparency and editorial type honesty — sit alongside them.
One clarification on how this actually works, because the mechanics matter for
understanding what the check is worth: the nine categories are nine
independent checkers that each scan the whole article, not nine questions
answered per claim. Each checker decides which claims it applies to and
reports the ones it can speak to.
In the run over article 320 (the Raspberry Pi 5 article) against the text as it
stood before that article was rewritten — the baseline run, dated 2026-11-14, which
reproduced the figures first published here — the checkers produced 48 findings
across seven of the nine categories — technical feasibility and internal
consistency produced none on that article — with 40 of the article’s 60 claims
carrying at least one finding. Of those 60 claims, 36 were marked verified, 13
escalated and 11 left unresolved.
Re-measured the same day against the rewritten text, the same pipeline produced 30
findings across two of the nine categories — source quality and temporal validity —
with 30 of the article’s 46 claims carrying a finding: 16 verified, 30 escalated,
none unresolved, no critical finding, and a score of 80 out of 100. The gate still
refuses it, and the article is held rather than published: the report authorises no
publication, the status it assigns is “Needs Revision”, and missing owner approval
is one of the mandatory overrides.
A claim that no checker recognises is reported as unresolved and escalated to
the owner rather than passed silently — which is the safe direction to fail,
and which does mean the unsupported-claim counts run high on first-person
articles by construction.
A score alone never publishes anything. Articles are scored out of 100 on
technical correctness (25), factual verification and source strength (25),
internal and cross-article consistency (15), practical plausibility (15),
editorial quality (10), and commercial transparency and image authenticity (10).
The score is a diagnostic, not a licence.
The gate is enforced in the publishing tooling — and its server-side half is
not yet live. The same publication decision is computed by the validation
pipeline, and the pipeline is the only thing in this site that computes it. The
gate computes a decision; it never performs the publish. Publishing goes
through the workflow, and publishing requires the pipeline’s validation decision to come back clean.
Precisely, and without overclaiming: the pipeline-level gate is in force
today. A WordPress mu-plugin that enforces the same decision inside
WordPress — so it cannot be bypassed by calling the CMS API directly — has been
built, lint-verified on the host, and proven against a real WordPress instance
in a staging tree (ten publish-path attacks all held; seven controls passed,
proving it does not simply block everything).
It is not deployed to the production site. Verified on 2026-11-12: the
production wp-content/mu-plugins/ directory contains nine files — eight
mu-plugins and one stylesheet — and the publication gate is not among them;
the plugin exists on the host only as a staged file outside the production
tree. So today, the enforcement is a process rule and a tooling rule, not a
server-side rule.
That gap is a live risk and it is recorded here rather than smoothed over. It
is a deployment operation that requires owner approval, and it is filed as an
open item in audit/site-defects.md. Until it is closed, the honest position
is that no technical control prevents a publish from being issued directly
through the CMS API by an automated agent — only the workflow rule and the
owner-approval requirement stand in its way.
The gate is not a substitute for reading the article
The pipeline’s checkers are deterministic and rule-based. They catch arithmetic
errors, contradictions against known facts, missing dates, impossible
connections and missing disclosures. They do not judge whether a subjective
review is good. That remains an editorial judgement, and it is made by the owner when a conclusion needs one.
Related
- AI Transparency Policy — how AI is used, what is labelled,
and which legal transparency duties apply - Impressum — provider identification (§ 5 DDG)
- Datenschutzerklärung — privacy, cookies, and data rights
- Contact and collaboration — how to reach the editor