Editorial standards

These are the rules every article on this site has to satisfy before it is published. They are kept in the site’s repository and this page is generated from that file, so what you read here is what the publishing pipeline actually enforces.

Avoltium Editorial Standards

The rules every article on avoltium.in must satisfy before it goes live.

Each rule exists because the failure it prevents actually reached production.

Where a rule is machine-checkable it is enforced by qc_check.py, which runs

as a gate inside generate_article.py and as a weekly site-wide audit. An

article that trips a BLOCKER is saved as a draft instead of published.


1. Images — depict the subject, not the words

**The rule: an image must show what the article is _about_. Never illustrate

the headline literally, word by word.**

This is the single most damaging failure mode this site has had. Real examples

that were live:

ArticleWrong imageWhy it happened
"BriHyNergy Ships 15MW PEM System"a cruise ship"ships" is a verb — it means delivers
"India's First Hydrogen-Powered Train"an LNG carrier vesselkeyword picked from the wrong noun
"Green Hydrogen Certification Portal"a blue jay birdno relationship at all
"BOP Testing for Well Control"a green office campusBOP is ambiguous (see §2)
"Ultrapure water — outsized risk"a passenger traintopic ignored entirely
"Gujarat Data Centre Policy"a politician beside a trainheadline keyword mismatch

How to choose an image

1. Read the article body — not the headline — and name the physical system

it actually discusses (an RO skid, a fuel cell stack, a compressor head).

2. Write the prompt describing that system.

3. Ask: "if a hydrogen engineer saw this picture next to this headline, would

it look researched or automated?" If automated, redo it.

Ambiguous words — never feed these straight into an image prompt

qc_check.py raises a warning when a headline contains any of these:

WordIn an energy headline it usually means
ships / shippingdelivers a product, not a marine vessel
BOPBalance of Plant or Blowout Preventer — disambiguate first
plantan industrial facility, not vegetation
trainrail transport (verify it is not "train a model")
cellan electrochemical cell
stackan electrolyzer stack
currentelectrical current

Image technical requirements

  • Minimum 900 px wide; target 1200 × 675 (16:9)
  • Aspect ratio must not exceed 2.4:1 — ultrawide banners crop badly in cards
  • Alt text is mandatory and must describe the image, not repeat the headline
  • Always pass a random seed to the image API — without it, identical prompts

return the identical cached picture for every article on that topic

Image credits

  • A third-party image (Wikimedia, CC-BY) keeps its credit block verbatim.

Attribution is a licence condition, not a courtesy.

  • When an image is replaced, the old credit must be removed in the same

edit. A credit describing an image that is no longer on the page is a false

statement about provenance.

  • AI-generated images are labelled where the reader sees them. The

caption directly under the image reads exactly AI-generated illustration.

A generated picture is never passed off as a photograph of real equipment,

a real site, or a real client installation.

Generated images — what must never reach the page

Each of these was live on the site before the rule existed:

DefectWhat the reader sawRule
Generator metadata as captionthe image service's raw request (its prompt, seed and model settings, in code form) printed under the picture on 8 articlesmetadata is stripped from every generated file before upload, and the caption is always set explicitly, never left for WordPress to fill from the file
Service watermarkan image-service logo in the corner of the featured imageno watermarks or third-party logos of any kind
Fake letteringinvented labels, gauges and part numbers on "engineering" imagesgenerated images contain no text at all; numbers belong in diagrams and tables
Literal readinga bedroom for "mixed-bed ion exchange"judged against the article, not the headline (see above)

Alt text describes what is visible ("centrifugal pump and piping on a

stainless steel skid"), never the headline repeated, and never the prompt's

style words ("photorealistic, high detail").


2. Title and content must agree

The title states what the body delivers. If a rewrite changes the subject,

the title changes with it.

Live failure: post 168 was titled *"BOP Testing and Maintenance: Best Practices

for Well Control"* — blowout preventers, oil & gas. A hydrogen-focused

rewrite prompt made the model read "BOP" as Balance of Plant, so the body

became a green-hydrogen article under an oil-well headline. Neither the title

nor the body was wrong on its own; together they misled the reader.

**Before rewriting anything, resolve every acronym in the title against the

existing body.** Retitle rather than let the mismatch stand.


3. Originality — the AdSense-critical rule

Never republish syndicated newswire copy. Duplicate content is the most

common reason AdSense applications are rejected.

When covering a news event:

1. The event is the hook — one opening paragraph, no more.

2. The rest is original engineering analysis: the constraints that decide

whether the thing works, balance-of-plant and water-treatment implications,

real efficiency and cost figures.

3. Never paraphrase a source sentence by sentence. Replace it with your own

technical argument.

Also prohibited:

  • Inventing quotes attributed to real, named people
  • Claiming Avoltium involvement in third-party projects
  • Publishing a post with an empty or stub body (thin content); minimum

2,500 characters, target 900+ words


4. House style for technical writing

Voice. Arun, Chief Engineer at Avoltium. Technical, specific, confident.

Written for practising engineers, not a general audience.

Substance. Include real numbers — efficiencies, current densities,

pressures, costs — wherever they support the argument. Specificity is what

separates this site from content farms.

Required structure.

1. Engineering Insight box — one-sentence takeaway, opens every article

2. <h2> / <h3> subheadings

3. At least one styled <blockquote>

4. Natural mentions of "electrolyzer" and "water treatment" where relevant

5. One internal link each to the electrolyzer and water-consumption

calculators — the site's most valuable original assets


5. Scientific notation — Unicode and HTML only

LaTeX must never reach the page. No backslash commands, no $…$ math.

It renders as literal \Delta T and \frac{...} garbage in the browser.

WriteNot
H<sub>2</sub>O, CO<sub>2</sub>H_2O, \text{CO}_2
A/cm<sup>2</sup>, Nm<sup>3</sup>A/cm^2
η, Δ, Ω, μ, °C, ±\eta, \Delta, \Omega
V<sub>cell</sub> = E<sub>rev</sub> + η<sub>anode</sub> + i·R<sub>ohmic</sub>$V_{cell} = ...$

Write equations as readable inline HTML. Division goes on one line with ÷ or

a bracketed numerator/denominator — never \frac.


6. HTML hygiene

  • Inline styles only. Never emit a <style> block into a post body; it

renders as visible CSS text in this theme.

  • CSS property names must be fully hyphenated. Generated copy has a

persistent habit of stripping hyphens — backgroundcolor, fontsize,

marginbottom — which silently kills the styling. sanitize_content()

repairs ~30 known cases and qc_check.py blocks any that survive.

  • No markdown fences ( ` ) in HTML output.
  • Hyphenated compounds stay hyphenated: hydrogen-powered, next-generation,

large-scale, state-of-the-art, zero-incident.

  • Internal links keep their hyphens: /electrolyzer-calculator/, not

/electrolyzercalculator/ — the unhyphenated form 404s.


7. Publishing controls

  • Cadence: Mon/Wed/Fri at 05:23 UTC. Daily posting at a round hour reads as

automated to search engines.

  • No duplicate topics. The publisher checks existing titles and only picks

uncovered subjects; when the pool is exhausted it uses dated review titles.

  • Categories are assigned per topic, never dumped into one bucket.
  • Comments and pingbacks are closed site-wide and by default. The site has

no moderation capacity and the only comments ever received were malware

links.

  • Every post needs an excerpt — without one, Google invents its own search

snippet from arbitrary body text.

  • Every third-party call retries with backoff. Pollinations and the Gemini

API both fail transiently, and a single unretried failure costs a whole

publishing slot. Two consecutive slots were lost this way: a Pollinations

HTTP 500 on 5 Aug parked a finished article as a draft, and on 7 Aug the

Gemini fallback chain burned through all three models in 0.6 s of 429/503

and produced nothing. Retry logic lives in resilient.py; the model chain

must retry within each model before falling through, so the chain outlasts

a short quota window.

  • Article text is generated before the image is uploaded. Text generation

is the step most likely to fail outright and the only one with nothing to

undo. Uploading first meant a dead Gemini chain stranded an orphaned image

in the media library (media 1117, 7 Aug).

  • A held draft is retried, not forgotten. recover_drafts.py runs daily,

regenerates missing images, re-runs the full QC suite, and publishes only

when every blocker clears. Drafts blocked for editorial reasons — thin

content, LaTeX, title mismatch — are reported and left for a human.

Advertising

  • At most one manual in-article unit per post. Auto Ads is enabled

site-wide; manual units stacked on top of it push a page over the ad-density

line. The unit is opt-in via the ADSENSE_SLOT_ID secret — unset, posts

carry no manual unit at all.

  • The unit anchors to an <h2> past the article midpoint, never mid-paragraph,

and is injected after the QC gate so its inline script text is not

counted toward the thin-content threshold.


8. Provenance — say where it came from, and never invent it

**The rule: an article states who stands behind it, when it was last updated,

and that AI assisted the drafting. A link is only ever presented as evidence

for a sentence if the article was written from it.**

Every published post carries a provenance block (editorial_provenance.py),

holding a named byline, the last-updated date, the AI-assistance disclosure,

and Further reading.

Further reading is not a citation list

The distinction is the whole rule. Further reading is authoritative

background on the subject. A source is the document a specific claim was

taken from, and it can only be recorded at the moment the claim is written.

When the first 49 posts were audited, none carried a source URL and not one

linked outside the site — they had been generated from topic prompts, not from

documents. A "Sources" list bolted on afterwards could only contain links

chosen because they looked like they might support the text. That is a

fabricated citation: an engineer who follows one, finds it does not say what

the sentence says, and stops trusting the site. It is the same failure as an

image file named "IISc hydrogen plant" that turns out to be a street with a

scooter on it — a label that promises evidence and does not carry it.

So: never add a citation retroactively. Fix it at generation time instead.

Checking a link before it ships

  • Fetch it. Only a 200 on the final URL may be published.
  • Read where it redirected to. Two energy.gov links returned 200 after

redirecting to unrelated offices, and the DOE "Hydrogen Shot" page 404s.

All three would have looked perfectly plausible pasted in unchecked.

  • A 403 is not permission to guess — IEA, IRENA and ISO block automated

fetches, so their pages cannot be verified this way and are not published

as verified.

The disclosure has to be true

The wording says *drafted with AI assistance and checked against these

standards* because that is exactly what the pipeline does: it drafts with a

model and gates on qc_check.py. It does not claim a human line-edited

the piece, because for most posts none has. Overstating review is worse than

disclosing none — it is the specific claim a reader can catch you in.

If a human does review an article, strengthen the wording for that article.


9. Enforcement

# Audit every published post; exit code 1 if any blocker is found
WP_USERNAME=... WP_APP_PASSWORD=... python qc_check.py
WhereWhenEffect
generate_article.pyevery publishblocker → saved as draft, workflow goes red
qc_audit.ymlweekly + manualsite-wide report, red run on any blocker
image_audit.ymlweekly (Mon) + manualvision review of every featured image; red run if any should be replaced
qc_check.py gateevery publishblocks generator JSON in captions/alt/body and forbidden names

Severities. BLOCKER — never reaches readers (thin content, broken CSS,

LaTeX, <style> blocks, missing image or alt text, duplicate title, broken

internal links). WARN — publishes but needs a human eye (ambiguous headline

term, small or ultrawide image, missing Engineering Insight box).

A warning on an ambiguous headline does not mean the image is wrong. It

means a human must confirm the picture reflects the article's meaning rather

than a literal reading of its words.


10. Accuracy — every claim must survive an engineer reading it

**The rule: a number, a standard, or a named organisation is either correct

and checkable, or it is not published.**

  • Numbers carry units and a basis. "55 kWh/kg" is incomplete; "≈55 kWh per

kg H₂ at system level, including balance of plant" is a claim that can be

checked. SI units, with the industry unit in brackets where readers expect it

(bar(g), Nm³/h).

  • Standards are cited by number and edition — ISO 22734:2019, ASME B31.12,

IEC 60079 — and only for what that standard actually covers. If the edition

cannot be confirmed, the sentence says so rather than guessing.

  • Physics before prose. Efficiencies stay below thermodynamic limits, mass

and energy balances close, and pressures, temperatures and concentrations are

in ranges real equipment runs at. A figure that would surprise a practising

engineer needs a source in the same sentence.

  • Superlatives need a source. "World's first", "largest", "100% efficient",

"zero maintenance" and "guaranteed" are flagged by the QC gate; each is

either attributed or softened.

  • Named companies and clients. Organisations are named only for what is

public. Former employers are not named on this site, and the projects or

clients of a former employer are described generically unless already public

and needed for the point. Experience is described by what was done

("a green hydrogen electrolyser start-up"), not by whom it was done for.

  • First-person experience matches the profile. "In our experience…" is used

only for work listed on the About page, never invented to add authority.

  • News is reported, not embellished. Figures in a news post match the

source announcement exactly, with the source named and linked.

  • Corrections. A factual error found after publishing is fixed in place

and the update date changes. It is never left because the post is old.


11. Image review — the picture is checked against the article

Rules in text cannot see a picture, so every featured image is also *looked

at*. image_audit.py sends each image together with the article's title and

opening to a vision model, which reports what is actually visible and scores

it:

CheckFails when
Relevance (0–10)3 or below: unrelated, or a literal misreading of a headline word
Watermark / logoany service mark, stock-site mark or unrelated brand
Fake textinvented labels, gauges, numbers or lettering
Technically wrongequipment an engineer would see as implausible or misleading

A failing image is replaced with the first candidate that itself scores 7 or

more with no flags: the article's diagram, then a correctly licensed

photograph, then a clean AI illustration labelled as such. If nothing passes,

the image is listed for a person to choose, and the old one stays rather than

being swapped for a second wrong picture. Weak alt text on images that pass is

replaced with a description of what the model saw.