Re: New issue: §15.4.5's "MUST generate any of the following" was written for one flat list and now governs two bullets; a careful reader took it to cover only the first

Hi Alexander,

Thanks for the response

On Wed, Sep 9, 2026, at 15:29, Leschinsky, Alexander wrote:
> Lucas,
> 
> real human here. Thanks. "no change" is a fair answer and I'll record it as yours, matter closed. Will also work on being less verbose in the future. 
> 
> One correction to the record: the copy that reached the list is truncated. It ends at the heading "Why we still think it is worth raising…" (https://www.google.com/url?q=https://lists.w3.org/Archives/Public/ietf-http-wg/2026JulSep/0279.html&source=gmail&ust=1789050315174000&sa=E). The two paragraphs that followed said that a model misreading is weak evidence, and that we claim no defect and no doubt about the six-field reading. Most likely a quoted-printable line-wrap artefact, not intent.

To clarify a little more, its fine to ask such questions to the list. And its fine to use tools to support gathering information and preparing text (whether AI or not).

What tends to be effective for list communication is concise problem statements, open questions or proposals placed front and center.

When referring to spec text, the history of how something came to be is generally not the most relevant thing to immediate discussion. Providing background or supporting evidence can be helpful, but is best provided after the key points so it doesnt distract passing readers that are typically time constrained. 

Cheers
Lucas

Alexander Leschinsky
G&L Geißendörfer & Leschinsky GmbH
> 
> On Wed, Sep 09, 2026 03:39 PM, Lucas Pardue <lucas@lucaspardue.com> wrote:
>> __
>> Hi,
>> 
>> Response in line
>> 
>> On Mon, Sep 7, 2026, at 08:48, Leschinsky, Alexander wrote:
>>> While building rule-ID-traceable conformance oracles for RFC 9110
>>> (
>>> https://www.google.com/url?q=http://gitlab.com/g6302/lintology&source=gmail&ust=1788848082913000&sa=E)
>>> we hit a reading question in §15.4.5 that cost us
>>> a review round, and we think the text can be made unambiguous cheaply.
>>> 
>>> 1. §15.4.5 states the 304 field duty as a verb phrase followed by two
>>> bulleted groups:
>>> 
>>> The server generating a 304 response MUST generate any of the
>>> following header fields that would have been sent in a 200 (OK)
>>> response to the same request:
>>> 
>>> * Content-Location, Date, ETag, and Vary
>>> 
>>> * Cache-Control and Expires (see [CACHING])
>>> 
>>> 2. That verb phrase is inherited verbatim from RFC 7232 §4.1, where it
>>> governs a single flat inline list of all six names:
>>> 
>>> > The server generating a 304 response MUST generate any of the
>>> > following header fields that would have been sent in a 200 (OK)
>>> > response to the same request: Cache-Control, Content-Location, Date,
>>> > ETag, Expires, and Vary.
>>> 
>>> So the wording was authored for one list and carried forward unchanged when
>>> the list became two groups.
>>> 
>>> 3. The grouping arose from a citation-tag question, not a normative one.
>>> httpwg/http-core issue #1025 (
>>> https://www.google.com/url?q=https://github.com/httpwg/http-core/issues/1025&source=gmail&ust=1788848082913000&sa=E
>>> ,
>>> "15.4.5 Linking", opened by reschke
>>> 2022-03-03, closed 2022-03-08) carries an [rfced] query asking whether
>>> Cache-Control and Expires should have cite tags given that the other
>>> fields did — and it quotes the then-current RFC 9110 text as a flat
>>> six-field list:
>>> 
>>> The server generating a 304 response MUST generate any of the
>>> following header fields that would have been sent in a 200 (OK)
>>> response to the same request: Cache-Control, Content-Location, Date,
>>> ETag, Expires, and Vary.
>>> 
>>> 4. The two-bullet layout was proposed the next day, in response to that
>>> citation question:
>>> 
>>> Maybe
>>> The server generating a 304 response MUST generate any of the
>>> following header fields that would have been sent in a 200 (OK)
>>> response to the same request:
>>> 
>>> * Content-Location, Date, ETag, and Vary
>>> * Cache-Control and Expires [CACHING]
>>> 
>>> The verb phrase is carried over character-for-character. Only the list
>>> is split. Whether "any of the following" still reads correctly across two
>>> groups is not raised.
>>> 
>>> 5. It was accepted in one line, and the issue closed four days later:
>>> 
>>> > That's good
>>> 
>>> Why it is worth a sentence of text. "Any of the following" reads naturally
>>> over a single list. Over two bullets, "the following" acquires a second
>>> possible
>>> referent — the groups rather than the field names — and the second group's
>>> parenthetical citation makes it look like a differently-sourced, possibly
>>> weaker
>>> class of obligation.
>>> 
>>> This is not hypothetical. In our own review of a cache-side rule, an LLM
>>> reviewer in our review chain (gpt-5.6-sol), working only from the published
>>> text, concluded that the MUST covered only the first bullet's four fields,
>>> with Cache-Control and Expires as optional cache-update metadata. Refuting
>>> that took three primary sources: RFC 7232 §4.1's
>>> flat six-field list, issue #1025 showing the grouping came from a citation
>>> question, and our own extracted text. The reviewer was wrong — and the text
>>> let
>>> a diligent reader, human or model, be wrong. We disclose that it was a
>>> model so
>>> the working group can weigh the evidence accordingly.
>>> 
>>> Suggested fix, any one of which resolves it:
>>> 
>>> - replace "any of the following header fields" with "each of the following
>>> header fields, where applicable" — this also removes the separate ambiguity
>>> of
>>> "any" reading as at least one rather than each that applies; or
>>> - state the six names in one flat list as RFC 7232 did, moving the [CACHING]
>>> citation to a following sentence; or
>>> - add one clause making the scope explicit: "…the following header fields
>>> (all
>>> six, across both bullets) that would have been sent…".
>>> 
>>> Why we think it survives the strongest counter-argument
>>> 
>>> The strongest defence is that the text is not actually ambiguous. "The
>>> following" introduces everything that follows the colon, bullets included;
>>> nothing in the layout narrows a MUST, and a parenthetical cross-reference is
>>> plainly a pointer rather than a normative qualifier. On that reading the
>>> reviewer who misread it simply misread it, and specifications are not
>>> obliged
>>> to be immune to careless reading.
>> The text to this point is an incredibly long-winded way of saying that the RFC "reader" - the model - lacks the fundamental comprehension capability to read a list of relevant items.
>> 
>> Misreaders gonna misread. Doing nothing is the right answer.
>>> 
>>> A second defence: the six-field duty has been in force since RFC 7232 and,
>>> so far as we know, no implementation has shipped only four fields because of
>>> this. Absent evidence of a real interoperability failure, an editorial
>>> clarification is churn against a stable document.
>>> 
>>> Why we still think it is worth raising, and what would settle it against us
>>> 
>> It seems like something was omitted here? Or perhaps there is no substantial justification for raising the topic. Since the use of a model to read an RFC was disclosed, one has to ask if a model was used to post this email. Did any human review the text before it was submitted? Will any human read this response?
>> 
>> Cheers 
>> Lucas
> 
> *G&L Geißendörfer & Leschinsky GmbH*
> Digital Media Engineering
> 
> https://www.gl-systemhaus.de
> Subscribe to our newsletter at https://www.gl-systemhaus.de/newsletter/
> 
> LinkedIn <https://www.linkedin.com/company/g%26l-geissendoerfer-%26-leschinsky-gmbh> - Twitter <https://twitter.com/glsystemhaus> - Facebook <https://www.facebook.com/glsystemhaus>
> Maarweg 149-161
> D - 50825 Köln
> Tel: +49 221 99809-0
> Fax: +49 221 99809-109
> 
> Geschäftsführer: Alexander Leschinsky
> Amtsgericht Köln, HRB 56772
> USt.-IdNr. DE814612929
> Steuernr. 217/5737/0885
> Datenschutzhinweise <https://www.gl-systemhaus.de/datenschutz/>

Received on Wednesday, 9 September 2026 14:41:01 UTC