The following paragraph in Section 3.0.3, IMHO has some unlukcy wordings:

]]
   One of the most common mistakes authors make that is easily repaired by user agents is to use a description, instead of a URL that links to a description. This means there is often plain text description in the content of an invalid longdesc attribute.  Converting such attributes to data URLs is a simple repair strategy that can help recover from cases where authors have made this mistake. 
[[

Problems - and suggested fixes:


Sentence 1: "One of the most common mistakes authors make".

             Question: How do you know? This is a spec for the future. Whether it is true that authors make that error *right now* or will make that error in the future, who knows? Please make that sentence past tence, about legacy content or something similar. As is, it sounds like we can expect this error to continue. (And if we can, then there is something seriously wrong ...)

              SUGGESTED FIX: "A common mistake in legacy content has been, "

Sentence 1: ", but easily repaired by user agents, "

             This phrase doeesn't add much (except that has the tone of  telling vendors that don't perform such repair, that they are, difficult people). Additionally, this new spec hopefully puts and end to this sort of error being common! Further more, I suggest that what this phrase says, is expressed in the last sentence (about converting to data URI).

              SUGGESTED FIX: Delete it.

Sentence 1: the phrase "… is to use a description"
 
               This prase doesn't sound like an error since the spec, after all, is about descriptions .... Also, while the content is *text*, it isn't necessarily a *independent description". Quite often, when there is text in the @longdesc attribute, that text is argually more like a "longer alt" than a independent description. I don't think the content _needs_ to be “blessed” with the quality word "description". It would increase the agreement around the spec, if the text used more neutral language.

              SUGGESTED FIX: "… is that authors typed a descriptive plain text string directly inside the attribute, instead of typing a URL to an independent description".

Sentence 2: The phrase "there is often plain text description". 

               Firstly, the phrase sounds like it should have been "there is often _a_ plain text description", so *a least*, add the " a ". However, the word "description" has the same problems as in sentence 1 - it is a positive word when it could have been a neutral word. Also, it seems like the condition expressed in the tail of the sentence could be emphasized more: "of an invalid longdesc attribute". Namley, it is when the content is invalid, that one often see the specific error that the content makes sence as plain text.

              SUGGESTED FIX:  "As a result, in legacy content, when the content of the longdesc attribute is invalid, one frequently finds that the content would make sense if interpreted as a plain text string."

The entire paragraph with all the changes above:

    ]] A common mistake in legacy content in legacy content has been that authors typed a descriptive plain text string directly inside the attribute, instead of typing a URL to an indepdendent description. As a result, in legacy content, when the content of the longdesc attribute is invalid, one frequently finds that the content would make sense if interpreted as a plain text string. Converting such attributes to data URLs is a simple repair strategy that can help recover from cases where authors have made this mistake. 
[[


If not in full, I hope for a partial acceptance.