1 · "REGRESSIONS: NONE" was true as stated and much narrower than it sounded MOST IMPORTANT CORRECTED
Every pass from 4.28 to 4.51 reported "REGRESSIONS: NONE". That was measured against the 97-term list in the verification script, not against the 140-row master table. Only 52 of the 140 tracked keywords were in that list; 88 were never checked per-pass. Rebuilding the table found seventeen terms that fell, thirteen of them invisible to every per-pass check run all session.
The claims were literally true and too narrow to mean what they sounded like. The verification list should be regenerated FROM the master table, not maintained beside it — that is the fix, and it is logged as an open item. This report's keyword tables are the rebuilt measurement, not the per-pass one.
2 · The argument against building the eligibility section was circular CORRECTED
A reduced checklist was recommended instead of a full "how do I know if I have a case" section, partly on the grounds that procedural queries draw only four impressions on this URL.
That reasoning was circular and should not have been relied on. Search Console only reports queries for which the URL already surfaced. A page that does not cover a topic cannot earn impressions for that topic — so an absence of impressions is often evidence of the gap itself, not of absent demand. The same trap would have argued against the boating H3, which was built on exactly that reasoning going the other way. An external check run after the client pushed back: FindLaw publishes a dedicated "How To File a Wrongful Death Lawsuit" guide, and five competitor firms run dedicated pages of the same kind. The client's read was better than the data reading, and the reasoning error matters more than the outcome.
3 · A live typo, in bold, introduced by an earlier pass in this project FIXED
The Who Can File section read, in bold: "One definities by surprise." It should have read "One definition catches families by surprise." It was introduced by this project's own article pass and it was live.
Why the verification missed it: that pass checked whether the 768.18(2) content was present. It searched for the statutory substance — "minor children", "under 25" — found it, and passed. It never checked whether the sentence introducing that content was English. Presence testing is not proofreading.
4 · A metric that under-reported silently, twice in two passes FIXED
The "statute citations in the body" count used an allow-list pattern. Adding Chapter 766 and 733 broke it once; it was widened. The very next pass added Fla. Stat. 440.11, chapter 440 was not in the widened pattern either, and the measurement reported "23 → 23" — no increase after a new statute had just been added to the page.
Pattern now covers chapters 768, 766, 733, 440, 327 and 95. An allow-list metric under-reports every time the work moves outside the list, and it fails in the flattering direction — it never overstates, so it never looks wrong. It was caught only because a zero delta was implausible. Any future entry adding a new chapter must widen the pattern in the same commit.
5 · The attorney-vs-lawyer skew was overstated CORRECTED
Early findings described the page as using "lawyer" about 3.6× more than "attorney" (36 vs 10). Those were sums of the on-page column across every keyword row containing each word — and because longer phrases are counted inside shorter ones, that double-counts. Raw word counts: attorney family 17, lawyer family 28 — a 1.6× skew, not 3.6×. The skew is real and was still worth correcting, especially in the metas which contained zero "attorney", but it was never as severe as first stated. The underlying demand figures (157,290/mo attorney-form vs 190,670/mo lawyer-form) are unaffected.
6 · A keyword removed for good reasons, then recovered in grammatical English RECOVERED
"wrongful death attorney Miami" — 1,392 impressions, position 10.3, RANK 8 — was deleted along with the ungrammatical sentence hosting it, and the entry recording that deletion flagged it as the one item that could be wrong.
Recovered in the CTA as a reduced relative clause: "Every first conversation is with a wrongful death attorney Miami families can meet in person." That is ordinary English — an attorney whom Miami families can meet — and it is factual, because the firm has Miami and Hialeah offices. The verb was chosen carefully: "an attorney Miami families trust" or "choose first" would carry the same keyword and would be an unsubstantiated comparison under Rule 4-7.13(b)(2). "Can meet in person" states a fact the firm can prove.
The other one was declined, with reasons. "lawyer wrongful death case" — highest CPC in the table at $174.40, position 3.84 — cannot be written grammatically. Every construction either breaks the contiguity the count requires or breaks the English. It is a search query, not a phrase: it is how people type, not how anyone writes. And the page does not need it — four terms this URL ranks well for (wrongful death legal attorney at 2.7, attorney wrongful death at 5.7, accidental death lawyers at 3.9, wrongful death attorney miami fl at RANK 7) appear as exact strings zero times. Modern retrieval does not require contiguity, and the evidence for that is sitting in this page's own data.
Verification traps worth carrying forward
Stale renders and caches. A schema edit fetched at exactly the pre-edit byte size with the new markup absent, twice in a row — a seven-day output cache keyed on User-Agent. Double-fetching does not defeat it; both requests hit the same entry. Vary the User-Agent. Wrong environment. A CSS fix was applied to the production bundle while the user was on a dev server that never loads it. Check which asset list the environment actually uses before concluding a fix did not work. Blocked sources. bls.gov returns 403 to this environment on every path; myfwc.com returned 500. Neither is a dead source — but a figure that cannot be fetched should be marked as second-hand rather than presented like the ones that were parsed from the PDF.