One shared document replaces thousands of private agreements
A standard is a document that independent implementations agree to follow, so strangers can interoperate.
Your browser fetches a page from a server it has never met, built by people who never spoke to the people who built the browser. It works because both sides read the same document. Imagine there were no such document. Six vendors who want to work together must each strike a private agreement with every other: fifteen agreements. With a standard, each vendor implements one shared document. Six implementations, and any two of them can meet. The difference grows fast. For a hundred vendors, it is four thousand nine hundred and fifty agreements against one hundred. The bars are to scale. The price of the standard is a slow committee. The gain is that anyone can join without asking permission.
Why standards exist · Going deeper · 1 of 2
The cost of disagreeing
The cost of disagreement is interop failure: a page that renders in one browser only, a USB-C cable that delivers 100 W to one laptop and 5 W to another. Open standards buy permissionless interoperability; closed protocols such as iMessage cannot be implemented by a third party at all.
Why standards exist · Going deeper · 2 of 2
Who writes which rules
The IETF writes internet protocols (IP, TCP, TLS, HTTP, DNS). W3C and WHATWG split the web platform. IEEE covers Ethernet (802.3), Wi-Fi (802.11) and floating point (754). ISO covers C, C++ and ISO 27001. NIST publishes cryptographic baselines such as AES. ECMA’s TC39 owns JavaScript; the Unicode Consortium owns text.
How internet standards are made
Internet standards: rough consensus and running code
An RFC is reviewed in the open, published under a number, and never edited afterwards.
How does a document become the rule for the whole internet? It begins as an Internet-Draft that anyone can submit, and which expires after six months unless revised. It is argued over on a mailing list, then goes through a last call in its working group, a last call open to anyone, and a review board where every blocking objection must be resolved. That typically takes two to five years. Then it is published as an RFC with a number, and that text is never edited again. A revision is a new document that replaces the old one, as the RFC for TLS one point three replaced the one for one point two. Of more than nine and a half thousand RFCs, only about a hundred have reached the top status of Internet Standard. And the deciding principle is not a vote. It is rough consensus and running code.
How internet standards are made · Going deeper
The life of an RFC
Not every RFC is a standard: the status field distinguishes Standards Track, Best Current Practice, Informational, Experimental and Historic. Rough consensus means the chair judges that no unresolved technical objection remains, not that a majority voted. Errata can be filed, but the published text is immutable.
Ambiguity in specs
What a spec leaves open is where systems disagree
In a spec, MUST, SHOULD and MAY have exact meanings. Gaps in the text become differences in behaviour.
A specification is a contract between people who will never meet, so its wording is precise. Three capitalised words carry most of the weight. MUST: required, and an implementation that breaks it does not conform. SHOULD: recommended, deviate only with good reason. MAY: optional, each implementer decides. Now look at what happens where the text is silent. The specification for JSON does not say what it means for a key to appear twice. So one parser keeps the last value, another keeps the first, and a third reports an error. All three conform. Which means two systems that both support JSON can read the same message in two different ways. When you rely on a format, read what it does not promise.
Ambiguity in specs · Going deeper
Anatomy of a specification
Key words come from RFC 2119. Grammar is given in ABNF. A conformance section says what counts as conforming, and IANA Considerations says how new values, such as status codes, are registered. Skimmers read the abstract and security considerations; implementers read everything, then the errata and test suites.
Check yourself
A gateway and a backend both parse this message. The gateway keeps the first value of a repeated key, the backend the last. What happens?
The spec leaves repeated keys undefined, so both parsers conform while disagreeing. The check and the action see different data: a real class of security bug.
Semantic versioning
A version number is a promise about what can break
Semantic Versioning: MAJOR for breaking changes, MINOR for additions, PATCH for fixes.
Standards and libraries change. The question for everyone who depends on them is: can I take the new version safely? Semantic versioning answers with three numbers. The last, PATCH, rises for bug fixes and nothing else. The middle, MINOR, rises when features are added and old code still works. The first, MAJOR, rises when something that worked may now break. So from two point three point five, moving to two point three point six or two point four point zero should be safe. Moving to three point zero point zero means reading the release notes and expecting to change your code. For authors, a rule of thumb: if you can name three ways a change could break someone’s production, call it MAJOR. A cautious MAJOR costs a line in the notes. A wrong MINOR costs somebody an outage.
Hyrum’s law
Users depend on everything you do, not just what you promised
Hyrum’s Law: with enough users, every observable behaviour becomes part of the interface.
That promise has a weakness, and it has a name. This solid circle is what your documentation promises. The dashed circle is everything your callers can actually observe. It is always bigger. And with enough users, somebody depends on each thing in the gap. The order results happen to come back in. How long a call usually takes. The exact wording of an error message. A default value nobody documented. Change any of these and you have broken no promise, and you have broken somebody’s system. Three defences. Randomise what you do not promise: the Go language shuffles the order of its maps on purpose, so nobody can come to rely on it. Deprecate before you remove. And test the contract, not the quirks.
Hyrum’s law · Going deeper
Hyrum’s Law in practice
Observed by Hyrum Wright at Google, where refactors that were safe under the documented contract kept breaking callers. On the web, feature detection and polyfills are the compatibility tools: check whether a capability exists before using it, or supply it on older browsers.
Check yourself
The docs say “result order is not guaranteed”, but for five years results came back sorted. You release a faster version that returns them unsorted. What should you expect?
Hyrum’s Law: observable behaviour gets depended on whether or not it was promised. Treat the change as breaking, or announce it first.
Conformance testing
A standard is proven by two implementations passing one test suite
Conformance tests turn a document into something checkable. Independent implementations keep it honest.
A document cannot interoperate. Only implementations can. Start with the specification. Two teams implement it independently, each working from the text alone. Will their software talk to each other? Reading the same words is no guarantee. So the standard gets a shared suite of conformance tests, run against both. The web’s suite has over two million checks, run against every major browser. When two independent implementations pass the same tests, they can talk to each other, and the document was clear enough. This is why it matters that there are several browser engines. When one implementation dominates completely, whatever it does becomes the standard, and the document just describes it.
Conformance testing · Going deeper
Specification, implementation, tests
IETF hackathons put implementers in one room to connect their code; disagreements become errata. The annual Interop effort has browser vendors commit to passing agreed test areas. Passing the suite is necessary, not sufficient: production exposes middleboxes and timeouts the spec never anticipated. The healthiest state is one primary spec with two or more maintained implementations.
Check yourself
A protocol has a written spec, but everyone uses the same single implementation. That implementation changes a behaviour the spec left open. What is the standard now?
With no second implementation to compare against, callers code to what the software does. The document ends up describing the code.
Compliance
Compliance means proving, with evidence, that you followed the rules
A compliance framework is a set of controls plus an audit. The work is operating the controls, not passing the audit.
Some standards are voluntary. Others come with an auditor, a contract or a regulator, and that is compliance. Here is the most common one for software companies, month by month. Three months of readiness: write the policies, close the gaps. Then the observation window, typically six months, during which the controls must actually operate. And leave evidence. Every production login used a second factor. Every code change was reviewed. A backup was actually restored in a test. Then the audit. For each control the auditor picks about twenty-five items at random and checks every one. What comes out is a report your customers ask to see before they buy. The same pattern, different rules, applies to card payments, health data, personal data in Europe, and now AI systems.
Compliance · Going deeper
Frameworks by territory
SOC 2 is an attestation, an auditor’s opinion; ISO 27001 is a certification issued by an accredited registrar; FedRAMP is a government authorisation taking 12 to 18 months. PCI-DSS is contractual, through the card networks. GDPR requires breach notice within 72 hours, with fines up to 4% of global turnover or €20 million. The EU AI Act sorts systems into four risk tiers.
Accessibility standards
Accessibility is a written, testable standard, and often the law
WCAG defines accessibility as testable criteria in three levels. Level AA is the usual legal requirement.
One standard deserves its own page, because it turns something that sounds like goodwill into something you can test. The Web Content Accessibility Guidelines. Level A is the minimum, about thirty criteria. Every image has a text alternative. Everything works from the keyboard. Audio has captions. Level AA adds about twenty more: text contrast of at least four and a half to one, a visible focus indicator, text that can be enlarged to double size. Level triple-A adds twenty-eight more, and is aspirational: some content cannot meet it. Behind all of them are four questions. Can everyone perceive it, operate it, understand it, and can assistive technology parse it? And the law, in the United States and the European Union, points at level AA.
Accessibility standards · Going deeper
WCAG 2.2 and its neighbours
WCAG 2.2 was published in October 2023; version 3.0 is in development. WAI-ARIA defines the attributes that expose custom widgets to screen readers. EN 301 549 is the European harmonised standard. The next lesson covers how to build to these criteria.
Spec drift
Specs and reality drift apart in three ways
Code can lead the spec, ignore it, or fork it. Multiple implementations and public tests limit the damage.
A standard and the software that implements it are two different things, and they drift. There are three patterns. The specification follows the code. A browser ships a feature, it proves popular, and the document catches up a couple of years later. That is healthy, as long as they converge. The code never follows the specification. The document is finished and almost nobody ships it. The next version of the internet’s addressing, IPv6, is still mid-transition decades on. And the costly one: a vendor implements the standard, then extends it privately until users depend on the extensions and cannot leave. The defence against all three is the same: more than one independent implementation, and tests that anyone can run.
Spec drift · Going deeper
Three patterns of drift
Living standards (HTML, DOM, URL, Fetch) are updated continuously and have no version to cite; ISO and IEEE publish dated editions. “Embrace, extend, extinguish” was documented in the 1990s. Recent years show convergence too: coordinated browser interop work, and AI providers settling on compatible API shapes.
Summary
Between systems, correct means written, implemented twice, tested
Bodies, documents, implementations, tests, versions, audits. Next: interfaces, where systems meet people.
The whole lesson in one line. A body writes a document, in careful language. Independent teams implement it, and a shared test suite proves they agree. Over time, version numbers promise what an upgrade may break, and behaviour drifts regardless. And where the stakes are high, auditors and law make the standard mandatory. Everything earlier in this course, from the bytes of a text file to a secure connection, works between strangers because of this machinery. Next: interfaces, where all of it finally meets a person.
Summary · Going deeper
The standards landscape
A standard is a document. The bodies write it, the process gets them to done, and conformance is how reality decides whether the document was clear enough.
Lesson complete
You finished “Standards and Compatibility”
Answer the questions along the way to see your score here.