Reports that Matter

Changelog

What has changed.

Improvements to the archive and how it reads

2026-10-02 — Quotations and list items no longer split at page breaks (ingest v0.17.0)

A block quotation that ran over a page break used to arrive as two consecutive quotations, the second opening mid-sentence ("This group will be" / "known as the Tobacco Industry Research Committee"). The new quoteListRunOns pass joins them, and a word the break cut ("pro-" / "actively") is closed up. Ten reports declare it: Philip Morris loses 141 split quotations, Leveson 81, PSI 42, Lehman 19, Litvinenko 15, Saville 11, Chilcot 4, and 9/11, Challenger and Jack Smith one each, 316 in all. No paragraph link moves; PSI loses one word. Measured against the reference edition, Philip Morris's page-break join accuracy rises from 73.9% to 82.7%. The release also adds a layout oracle that measures the pipeline's headings, markers and paragraph boundaries against the PDF's own line layout.

2026-10-02 — Footnote markers open the right note (ingest v0.16.0)

In reports whose note numbers restart each chapter, one stray marker used to shift every later note onto the wrong chapter's text: 148 markers in the 9/11 Commission Report opened the wrong note, so "Haznawi in 6B." showed a note about Rohan Gunaratna instead of the airline record. Notes are now paired with their markers in reading order, so a stray marker disturbs nothing; the 9/11 chapter 1 date garble is corrected too. Moved: 9/11 (122 markers), Leveson (196), Deepwater Horizon (112), Litvinenko (94), Columbia (7, which gains a "Notes not linked in the text" section) and Lehman (1). Paragraph links are unchanged, except that 9/11's three-paragraph "Foresight—and Hindsight" now sits at the end of the chapter before it rather than on a page of its own.

2026-10-02 — The Lehman introduction now covers Repo 105

The Valukas Report's landing page now draws on Volume 3: its Repo 105 findings cite and quote the examiner's own Repo 105 analysis rather than the executive summary's précis of it — what the device was, the $38.6–$50.38 billion it took off the balance sheet at quarter-end, the absence of any business purpose, who was not told, and the claims against Lehman's officers and Ernst & Young — and the reading guide sends readers to the Repo 105 section. Three passages from it are added to the highlights.

2026-10-02 — Text fixes across twelve reports (ingest v0.15.0), and Repo 105 for Lehman

The Valukas Report now includes Volume 3, Repo 105 (the accounting-device analysis), as a section of the Lehman Brothers examiner's report, with its footnotes linked: 941 of its 1,094 notes, up from one. @rtm/ingest v0.15.0 carries a batch of fixes that moved most of the archive. Deepwater Horizon regains 279 paragraphs that had been fused into their neighbours, and Challenger 65. Litvinenko loses 241 running titles and tabs that had been set as paragraphs, and its lettered sub-items stop being cut into a quotation (prose run into a quotation: 251 down to 77). Leveson's page-break splits fall from 36 to 5 and its bogus "Chapter 7" section is gone. The Iraq Inquiry has all 88 of its contents-listed headings (68 before). 9/11 gets its contents as entries, 120 sub-headings and 371 endnotes linked that were not. Hillsborough, Saville, Columbia, Philip Morris, PSI and Jack Smith move by a handful of paragraphs each, mostly false footnote links removed. Paragraph ids moved in all of these, so any link into a moved paragraph on those pages may land at the top of its section.

2026-09-28 — An introduction for the Iraq Inquiry, and hero photographs for both

The Iraq Inquiry (Chilcot) Executive Summary now has a landing page: a standfirst, background, seven findings with quotations and a six-section reading guide. Both it and the Lehman Brothers examiner's report now carry a hero photograph — a Challenger 2 tank on patrol outside Basra in 2004, and the salvaged Lehman Brothers sign at Christie's in 2010 — sourced and credited in docs/design/2026-09-27-hero/sources.yaml.

2026-09-28 — Chilcot's numbered paragraphs are citable; a citable-density check

The Iraq Inquiry Executive Summary numbers its 892 paragraphs "13.", "20." straight through, at the left margin. Written to Markdown as a bare "20. …", each one was an ordered-list item to Markdown with no paragraph id: only 64 of 956 citable units on the page could be cited, quoted in the introduction, highlighted or linked — including every conclusion, such as paragraph 20 ("the diplomatic options had not at that stage been exhausted") and paragraph 432 (the legal basis was "far from satisfactory"). @rtm/ingest v0.14.1 adds escapeNumberedParagraphs(), an opt-in pass declared only by Chilcot; Litvinenko, Leveson, Hillsborough and Challenger also number "N." at the margin but are unaffected. Also new: a corpus-wide check that a report's paragraph ids per 1,000 words aren't implausibly low, the kind of check that would have caught this on day one (reportsthatmatter-4qw).

2026-09-28 — An introduction for the Lehman Brothers examiner report

The Valukas Report's landing page now opens with a standfirst, background, seven findings with quotations and a four-section reading guide. It covers Volume 1; Repo 105 is summarised from the Executive Summary, and the page says plainly that the full Repo 105 analysis (Volume 3) is coming.

2026-09-28 — A thirteenth report: the Iraq Inquiry (Chilcot); 9/11, PSI, Hillsborough, Challenger and Leveson text fixed

The Report of the Iraq Inquiry — Executive Summary (2016), Sir John Chilcot's account of the UK's decision to go to war: 69 of its 88 contents-listed section headings had no heading shape of their own — no capital, number or label — and previously ran straight into the paragraph after them; all are now recovered. Also in @rtm/ingest v0.14.0: the 9/11 Commission's printed Notes appendix now reads as linked endnotes under real chapter headings instead of running-together body text; PSI's footnotes that spill over a whole page no longer print as ordinary paragraphs (17 leaked down to 3); Hillsborough's Parts, Chapters and Appendices are recovered as real structure; Challenger drops its remaining caption and OCR-garble headings; and Leveson's headings ending "CHAPTER n"/"APPENDIX n" keep their number.

2026-09-28 — A twelfth report: the Lehman Brothers examiner (Volume 1)

Report of Anton R. Valukas, Examiner (2010), the court-appointed investigation into Lehman Brothers' collapse — scoped to Volume 1 of 9 (196 pages, 690 footnotes): the Introduction, Executive Summary and Procedural Background, and the Examiner's findings on Lehman's business and risk management. The official PDF opens with the whole nine-volume report's table of contents; it's excluded here as out of scope rather than ingested as loose bullet fragments (see the report's own README). The Examiner's "Repo 105" findings — the accounting device Lehman used to understate its leverage before each quarterly report — are in Volume 3, not yet on the site.

2026-09-28 — Page-break sentences, Deepwater's chapters, Philip Morris's findings, and Columbia's structure fixed

Four fixes shipped together in @rtm/ingest v0.13.0. Jack Smith: sentences split across a page break — including one a skewed scan had set as a block quotation — are rejoined; the three examples raised on GitHub, "false claims of election fraud", "monitored legal developments regarding" and "claims that they were making", now read as continuous prose. Deepwater Horizon: chapter, part and appendix headings are read from the contents instead of being folded into figure captions or the paragraph below them. United States v. Philip Morris: the opinion's 4,088 numbered Findings of Fact are citable paragraphs rather than an unlinked list, and its headings come from the contents rather than body or footnote text. Columbia: its section structure is read from the contents — fixing the two-column regression that had pinned it to an older pipeline release — and paragraph ids containing an apostrophe-like character (ʼ) resolve again across the archive.

2026-09-27 — A photograph of the event on every landing page

The other nine reports now open the way Jack Smith and Hillsborough do, with a photograph of what they are about in a band above the title: the towers burning from Brooklyn, the Challenger plume, Columbia breaking up over Texas, the Deepwater Horizon rig on fire, Lehman Brothers' headquarters on the night it went bankrupt, the seven tobacco chiefs swearing nicotine was not addictive, soldiers at a barricade in Derry on Bloody Sunday, Alexander Litvinenko in hospital in the photograph he asked to be released, and the last News of the World. Five of these replace the first research picks, which showed a place or an object rather than the event. Each carries a credit line linking to its source.

The Philip Morris report's landing page, with its photograph

All nine, desktop and phone.

2026-09-27 — Introductions for Challenger, Columbia, Deepwater Horizon and Philip Morris

The Challenger, Columbia, Deepwater Horizon and United States v. Philip Morris reports now open on the same landing page as Jack Smith and Wall Street: year, standfirst, background, what the report found (quoted, checked word for word), and where to start reading, all above the contents. Every report in the archive now has an introduction.

Deepwater Horizon's landing page

Before and after, in full.

2026-09-27 — A photograph of the event on two landing pages

The Jack Smith report and the Hillsborough Independent Panel's report now open with a photograph of what they are about, in a full-width band above the title: the gallows the crowd built in front of the Capitol on 6 January 2021, and the Leppings Lane end at Hillsborough on 15 April 1989. The colour is toned down so the photograph sits with the rest of the site, each carries a credit line linking to its source, and on those two pages it takes the place of the report's plate. Every other page is unchanged. This is a pilot: if it holds up, the other reports follow.

The Hillsborough report's landing page, with its photograph

Before and after, desktop and phone.

2026-09-27 — The 9/11 Commission Report's sections, and its chapter 11 notes, restored

Every numbered section of the 9/11 Commission Report is now its own section, titled as the report's contents titles it: "8.1 The Summer of Threat" rather than "THE SYSTEM WAS BLINKING RED" 8.1 THE SUMMER OF THREAT fused into one capitals heading. Six sections the text had lost into the paragraph below them are back, among them "9.2 September 11, 2001", "8.2 Late Leads—Mihdhar, Moussaoui, and KSM" and "11.2 Policy"; the contents now lists all 55. The report's notes are endnotes, but its notes pages were being read for footnotes, which produced five notes that nothing in the text pointed to; chapter 11's first thirteen notes had disappeared into one of them. They are back in the Notes under "11 Foresight—and Hindsight".

2026-09-27 — Wall Street's contents list only real sections

Quoted exhibits in Wall Street and the Financial Crisis — an awards-night script, an S&P memo header, a sales-goals slide — were being read as section headings in its contents. They no longer are, and the numbered findings that follow them, "Conflict Between Client Interests and Proprietary Trading in 2007", "Failing to Disclose Key Information to Investors", among them, read as their own paragraphs again instead of being swallowed into a caption's heading.

2026-09-26 — Fewer stray note links across seven reports; Wall Street's missing notes restored

In Wall Street and the Financial Crisis, a note whose text began with a number ("2009 OTS Annual Report…", "12 U.S.C. § 1820") was read as a new note with that number. That left 25 notes defined twice and threw the note count off, so on later pages the notes weren't found and printed as body text. The count now holds: no note is defined twice, and notes such as 65, 91–93 and 604–609 are back in the margin. Across the archive, numbers that aren't note markers are no longer linked as notes: case citations ("265 U.S. 182, 188"), years after a date ("May 3, 1991"), counts ("2008, 119 years to the day"), document numbers, and lists ("paragraphs 17, 19, 20 and 21"). That removed about 60 false links in Philip Morris, 70 in Deepwater Horizon and 35 each in Jack Smith and Leveson, and 21 Philip Morris notes that were printed as body text are notes again. About 17 Wall Street notes long enough to fill a whole page still print in the body; that is next.

2026-09-25 — Footnotes stay in the margin in the Jack Smith and Wall Street reports

In the Jack Smith report, a footnote that ran over a page break, or whose first line the PDF set above its own number, landed in the text as a block of raw citations. That happened twelve times, splitting five paragraphs mid-sentence, including one the landing page links to ("…the co-conspirators deliberately / [citations] / withheld from the elector nominees…"). Those citations are now back in their notes, and the paragraphs are whole. In Wall Street and the Financial Crisis, page numbers on the contents page and the count in "75 out of 75" were being read as note markers. Because each number's notes are matched in order, that gave later notes the wrong text: the Executive Summary's first note showed an unrelated Goldman Sachs note. The same contents-page fix corrects the Challenger and Philip Morris reports.

2026-09-25 — Reports open on a page about the report

The Jack Smith report and Wall Street and the Financial Crisis now open on a landing page instead of bare contents. It gives the year, a short account of what the report was about and why it matters, what it found with a few key passages quoted, and the sections to read first, all above the contents. Our introduction is marked as ours: only words in quotation marks are the report's, every one is checked word for word against the text, and each link opens it in context. The year a report was published now leads its row in the archive and sits in every report header.

The Jack Smith report's landing page

Before and after, in full.

2026-09-20 — Each report can say how its text was made

A report can now carry its own page, "About this edition", linked from its contents: where the text came from, what the pipeline had to do to read it, and the places where it still differs from the printed page. The Bloody Sunday Inquiry has the first one. It lists what was fixed (1,439 of 1,451 notes linked to their paragraphs, 157 subsection headings recovered) and what was not: about 50 paragraphs split at a page break, roughly 140 that end mid-sentence because of map labels, twelve notes whose marker is missing from the PDF. The other reports show no such link until they have one.

2026-09-18 — The Bloody Sunday Inquiry is published

Lord Saville's report into the killings in Derry on 30 January 1972 is now in the archive: Volume I, the Principal Conclusions and Overall Assessment and the background chapters, nine chapters and 1,451 notes. Its notes sit beneath each paragraph and restart at 1 every time, so the pipeline now reads them where they are printed and links each to its own paragraph. Before, 392 references pointed at a single note. The telegrams and orders it quotes in capitals stay quotations instead of breaking into headings, and each chapter's own contents list names its 157 subsections.

Paragraph 9.165 with its two notes beside it, under David's telegram of 10 January 1972

Screenshots, in full.

2026-09-18 — 18 Challenger scan errors leave the review queue

The first prose batch in the Challenger report's remaining OCR queue has been checked against the GPO scan. Twenty-five page-scoped corrections restore damaged footnote numbers and definitions, the SRM O-ring Task Force title, and a table heading; 18 suspect tokens are now resolved rather than left for another reviewer.

Footnote 25 on printed page 211, after correction

Before and after, in full.

2026-09-14 — A logo in the navbar, a press page, and two plates replaced

The pilcrow mark is back beside the site's name in the header — a plainer, unlettered version of the seal than the one dropped from there the day before, already proven legible at favicon size. A new /press page carries the logo, the logotype, and the two together, in the sizes the site itself ships.

Two of the ten reports' plates read poorly and are replaced. Columbia's showed an unrecognisable dark shape rather than anything identifiable as the report; it's now the shuttle stack in the Vehicle Assembly Building door, the same subject as a distinct candidate held from the earlier work. Challenger's report has no photographs anywhere in its own PDF — its plate was a hand-drawn diagram, same as the rejected direction for the mark beside it — so its plate now comes from NASA's own photograph of the smoke at the right booster's joint at liftoff, the evidence the Rogers Commission used to locate the failure.

Columbia's plate, before and after

Before and after, in full.

2026-09-13 — Every report gets a picture of its own evidence

Each of the ten reports now carries a plate: one image, drawn wherever possible from the report's own pages, and chosen to show the thing the inquiry turned on rather than the harm it investigated. The hole punched through Columbia's wing panel in the foam test. The polonium readings table by table in the Pine Bar. The stairwell deviation that failed the North Tower's evacuation. It sits beside each title in the archive, above the title on the report's own page, and on every share card, all in one grain so the ten read as a set.

The site's logo changes with it. The seal beside the name had 26 letters wrapped around a 30-pixel circle, so it only ever read as a smudge; the header is now the name alone. The browser-tab icon becomes a pilcrow (¶), the mark for a paragraph, set in the seal.

The archive, each report with its plate

Before and after, in full.

2026-09-12 — 166 wrong links stop pretending to be citations

Turning scanned text into a web page means guessing which fragments are web addresses. That guess was wrong far more than it looked: a dropped space at the end of a sentence, immediately before a short capitalised word, was routinely read as a link — "flight.At 8:46 it happened" became a live link to ".at", Austria's domain suffix, not a citation to anything. 155 of the 166 were exactly this, concentrated in the 9/11 Commission Report; the rest were short scanning errors that happened to end in a real country code.

Every one is now plain text again, and nothing that was a real link was touched — Leveson's own 94 citations to its own website, and every guardian.co.uk, FT.com, and similar bare citation elsewhere, still click through exactly as before.

2026-09-12 — 2,058 numbered paragraphs stop running into each other

A reader flagged it directly: in the Litvinenko Inquiry, paragraph 3.77 was printing as the tail end of 3.76's paragraph, not its own. The cause was the numbering itself — "3.77" prints flush against the margin, the same column an ordinary line of running text sits at, so nothing told the scan apart a genuine new paragraph from a line that just continues the one above. Wherever the page also happened to carry no blank line between paragraphs — common in the Litvinenko Inquiry, Leveson Inquiry, and Hillsborough Panel report, all three of which number this way — several numbered paragraphs in a row could run together into one.

The numbering is now read as what it is: Leveson gains 1,818 paragraphs back, Litvinenko 150, Hillsborough 90 — 2,058 in total, all previously fused into a neighbour. Nothing is reworded; every word is exactly where it was, just correctly split.

Read the corrected paragraph.

2026-09-11 — 210 of Challenger's Findings and Recommendations get their own paragraph back

The Committee's report structures every issue it investigated the same way — a question, a Finding, a Recommendation — but the label marking each one printed flush against the margin with nothing to set it apart from ordinary body text. The scan read it as part of whichever sentence came right before it: "...without certification and approval. Recommendations" looked like one run-on thought, and the numbered list that actually answered "Recommendations" followed as if unrelated. 214 of these labels were affected; 210 are now their own paragraph, exactly as printed.

Read the corrected report.

2026-09-11 — Footnote markers stop landing on dates and phone numbers

A handful of ordinary numbers in the prose were being turned into footnote links by mistake, because the linker only checked whether a number matched some footnote somewhere in the document — and a report with thousands of footnotes has almost every number 1–999 as some footnote's number. A date like "11, and 19 March 2003" had 11 and 19 turned into citations for unrelated footnotes; a repeated "Telephone orders: 020 7219 3890" block did the same to its digits. Both are now recognised and left alone. Six reports had at least one of these; Leveson and Hillsborough had the most.

Checking why Hillsborough's sidenotes stopped rendering after this fix turned up a separate, pre-existing gap: its citations are a numbered list, not superscript markers, and a handful of list fragments were being misread as footnote definitions that were never real notes. That's now tracked on its own rather than papered over.

2026-09-11 — Footnotes stop showing someone else's citation

When two different footnotes happened to share the same printed number — a report restarting its numbering per chapter, or a scan simply misreading a footnote's digits — the pipeline treated them as the same note. Every reference to that number showed whichever one of them won a silent merge, regardless of which note the reference actually pointed at. Fixed at the root: a reference is now resolved against its own note, in the order both were written, not through one shared lookup by number.

Seven of ten reports had this to some degree; two are worth naming. Deepwater Horizon was hit hardest — the same 1,226 sidenotes now each show only their own text instead of a merged blob, and the page dropped from 2.07 MB to 1.78 MB with nothing removed, just no longer duplicated. The Leveson Inquiry, which restarts numbering every chapter, had drifted ten pipeline releases behind everything else in the archive; fixing its footnotes meant catching it up, so its recall and heading detection also improved along the way.

Read the corrected Deepwater Horizon report.

2026-09-11 — Every page gets a real share-card image

Almost the whole site had no og:image — a link shared to Slack, iMessage, or X previewed as bare text everywhere except five curated quotes in one report. Every report now has its own default card (title, author, date), the site has one too, and a shared link always falls back to the most specific card that exists: a curated quote, then the report's own card, then the site's.

Reports that Matter — the site's default share card

Full before/after, including a couple of the new report cards: visual changelog.

2026-09-11 — Challenger's fidelity queue starts moving: 31 scanning slips fixed so far

Every report carries a review queue of places its scan might not read right — nearly 1,000 flagged spots across the archive, untouched until now. Challenger's is the largest queue, and it's the first to get a real pass: the Committee's own findings and recommendations, then its footnote citations, checked one by one against the source. 31 scanning errors corrected so far — "Recommendation" headings that had been printing as "Recomrnendation", footnote numbers reconstructed from OCR garble and confirmed against the scanned page images — and dozens more flagged spots checked and confirmed correct as printed. 310 of Challenger's flagged spots remain, mostly in the reproduced memos and equipment tables of its appendices.

Read the corrected report.

2026-09-05 — A report can be republished without redeploying the site

Until today, correcting a single line of one report meant rebuilding and redeploying the whole site, and every report's text travelled with the application code. Each report is now stored separately, under a fingerprint of its own contents, and the site is simply told which fingerprint to serve. All ten are live that way.

Two things follow from it. Correcting a report is now a change to that report, which is what it always should have been. And because a version is named by its contents, older versions remain exactly where they were — so a citation can point at the precise text it quoted, and putting a correction back is a single instruction rather than a rebuild.

Publishing goes through one gate that reads back everything it was given and refuses to point at a version that is missing anything, so a half-finished publish cannot reach a reader. Each report has its own key, which can change that report and nothing else.

2026-09-05 — Paragraph ids get a gate, and pages stop carrying their layout

The pre-rendered files also carried the site's own layout, which meant a styling change rewrote all six hundred of them and a corrected report could not be republished without redeploying the whole site. They now hold the report's text and nothing else, and the page is put together when it is asked for. All 578 pages the site serves are byte-for-byte what they were before — this is a change to where the work happens, not to what anyone reads.

2026-09-04 — The site stops carrying its own output

Report pages were pre-rendered into the repository and committed: 133 MB of generated HTML around 2 MB of application code, and 85% of all git history. They are now build output, regenerated in about three seconds and never committed.

Along the way, body.json is gone. It held every section of a report's HTML in one file — 55.5 MB across the archive, 19.0 MB for Leveson alone against a 25 MiB per-file ceiling — and a shared quote link parsed all of it to quote a single paragraph. A shared link now serves the same static page as the plain URL with only its <head> swapped, and takes the quoted words from the one section that holds them. Generated output drops from 133 MB to 77 MB, and the next report on the list, the Chilcot Inquiry at roughly twice Leveson's length, stops being blocked by a file-size limit.

All 31,918 indexed passages were compared against the previously published search index and are byte-for-byte identical, so nothing about what search finds has changed.

This also closes a latent hazard rather than a theoretical one: a documentation commit six days ago had deleted 413 generated files with no replacement, leaving main with no page artifacts at all for six of the ten reports. Deploys upload from disk after a manual pre-render, so the live site was never affected — but a deploy from a fresh clone would have dropped those six reports off the site. Build output committed to git can drift from what the build produces, and this one had.

Background and what comes next: content publishing.

2026-09-03 — Four more reports

The archive goes from six reports to ten:

Each report keeps its own repository under the reportsthatmatter organisation, holding the source PDF and the definition it is built from.

United States v. Philip Morris — the contents page of a 1,682-page opinion

Before/after for all four — including where Hillsborough's sectioning falls short — is in the visual changelog.

2026-09-01 — The Litvinenko report stops quoting itself

Most of the Litvinenko report's paragraphs were being shown as though they were quotations from somewhere else. The report numbers its paragraphs in the margin and indents the text beside them, and the pipeline read that indent as a quotation — so 865 of its 1,089 paragraphs were cut in half, the first line left as text and the rest set as a quote.

Paragraphs are whole again, and the passages that genuinely are quotations — witness evidence, expert statements — still read as quotations.

Chapter 2, with paragraphs whole and one real quotation

2026-08-29 — The Columbia report, and words that stopped breaking in half

The Columbia Accident Investigation Board report joins the archive: 248 pages on the loss of the shuttle and the decisions behind it. It had been ruled out once, because it is set in two columns and the text extractor read straight across them, welding an unrelated sentence into the middle of every paragraph. It is now read a column at a time.

Across every report, words the typesetter broke at the end of a line are joined back together — "Chal- lenger" was appearing as two words, and so were another two thousand in Columbia and six hundred in Leveson. Where a document never writes the word whole anywhere, the break is left visible rather than guessed at.

The Litvinenko report's footnotes also work properly now: around 230 of them appeared as bare numbers stuck to the end of a word, because the scan lost the space before the superscript.

The Columbia executive summary, reading in column order

Two contents pages also stopped inventing entries: the Jack Smith report had a court docket row listed as though it were a section of the report, and the Leveson contents page had been generating headings out of its own listings.

The Columbia text was published three times before it was right — the two columns kept finding new ways to run into each other, and each attempt looked correct until the page was actually read. The full account, with pictures, is in the visual changelog.

2026-08-28 — Thirty half-sentence headings gone, and citations that had quietly collided

The Wall Street report's contents page listed thirty headings that stopped mid-sentence — "Safeguard Against High Risk Products. Federal banking regulators should" — each one a numbered recommendation whose second half had been stranded in a separate block. They now read as the recommendations they always were, and five section pages that existed only because of them are gone.

Separately, a printed page number is not unique inside these reports: the Jack Smith report prints "2" on three different pages, the Challenger report collides sixteen times. Every occurrence carried the same link target, so a citation to the second page silently landed on the first. Each occurrence now has its own; links that already worked are unchanged.

Behind both: every report now records how it was built — which PDFs, in what order — and the pipeline checks its output against the original scans rather than, as it had been doing, against itself.

2026-08-28 — Leveson reads across the page as it does in print

Several passages in the Leveson Inquiry were broken at a page edge: running headers appeared in the text, and ordinary continuation lines were shown as quotations. The report now keeps paragraphs together across those page breaks and preserves only genuine indented quotations.

Operation Glade after the page-flow repair

2026-08-28 — The Litvinenko Inquiry gets its real contents page

The Litvinenko report's table of contents was listing sentence fragments — "On 23 November 2006, Alexander Litvinenko died at…", "G-BNWB)", "RESTRICTION ORDER" — because the ingest pipeline recognised A. and I. section markers but not a label that carries its own number, so every Part and Appendix heading was read as ordinary text while stray all-caps lines and numbered narrative sentences were promoted in their place. The pipeline now reads Part N / Chapter N / Appendix N headings (titles that wrap included), ignores a lone acronym or an aircraft registration on its own line, and keeps a numbered narrative sentence out of the heading structure. The contents page now shows Parts 1–10 with their chapters and all twelve appendices. Part 5's divider page has no text in the scan, so it is still absent and its chapters sit under Part 4.

2026-08-22 — A fifth report: Leveson on the press

Added: An Inquiry into the Culture, Practices and Ethics of the Press — Lord Justice Leveson's 2012 report into UK press conduct and regulation, prompted by the phone-hacking scandal. Four volumes bound into one continuous document: 2,022 pages, over a million words, the largest report in the archive by a wide margin.

The Wall Street and Challenger reports were also re-ingested under the current pipeline and republished, mostly a cleaner table of contents and a list-formatting fix, with nothing lost.

2026-08-22 — Report pages load without touching the source document

Every report page used to be rendered from its source markdown on the fly, which was slow enough to occasionally exceed the CPU a single request is allowed and return an error page instead. Report pages are now built once, ahead of time, and served as-is — pages load faster and hold up under a spike of traffic that would previously have knocked some of them offline. No visible change otherwise.

Search across every report. Find a name, a date, a bank — anywhere in the archive, not just the report you happen to be reading. A result is the exact matched passage, not a page to go hunting through: it shows the sentence, the report and section it's in, the printed page, and follows through to those exact words highlighted, the same as any shared link. Scope a search to one report from that report's own page, or search everything from the header.

Search results, term highlighted, across several reports

2026-08-21 — Quote the exact words, and keep what you find

Share part of a paragraph. Selecting half a sentence used to hand you a link to the whole paragraph — a broader claim than you meant. A shared link now names the words themselves, marks exactly those words when someone follows it, and previews that sentence rather than the paragraph around it. If the quoted words are no longer in the document, the link shows the paragraph they came from rather than marking whatever else matched.

Quoting longer passages, and lists that read in order. Selecting several sentences now produces a link to exactly those sentences rather than to the paragraph around them, and a quote can run across a paragraph break. Footnotes in the margin stay out of a selection made in the body. Separately, bulleted lists in the source documents are now rendered as lists: they used to be run together into a single paragraph, with a wrapped item's tail appearing after the item that followed it — a defect reported against the original site in 2016.

Highlights you can take with you. Save a passage and it is still marked when you come back. Everything you have kept is listed at /highlights and exports as Markdown — quote, source, printed page, and a link back to the exact words. It lives in your browser: no account, and none of it is sent to us.

The /highlights page, a saved quote with its source and an export

What other readers marked. A passage highlighted by another reader now shows the same way any highlight does, faintly at first and stronger as more readers mark it — no count printed, just the sense that this passage caught someone's attention too. Nothing is recorded until a reader shares or saves a passage, and nothing about a reader is kept, only a passage and how many people marked it.

A passage marked by six readers, washed stronger than one marked by a single reader

2026-08-09 — Contents pages read as a hierarchy

The contents page listed every section flat, with no way to tell a subsection from the part it belongs to. Sections now indent under their part:

Contents page, sections indented under their part

A long footnote also no longer drags the notes after it out of alignment or spills into the footer — clamped to about 8 lines with a "Show full note" toggle:

A clamped footnote, expanded on click

Full before/after: visual-changelog.

2026-08-02 — A third report, and better footnotes everywhere

Added: Investigation of the Challenger Accident — the House Science Committee's 1986 report into the loss of the Space Shuttle Challenger. 438 pages, 99.8% of the source text retained.

Footnotes. Nearly all of them are now found and placed beside the text: recall went from about three-quarters to around 97%. Where a footnote block had previously been missed, its citations were left stranded in the middle of the prose; that no longer happens.

Page numbers now work for reports that print them at the top of the page as well as the bottom. The Wall Street report puts them in a header and so had no page anchors at all — it now has 628.

Contents pages are cleaner. Numbered recommendations were being read as section headings, filling a contents list with half-sentences.

2026-08-02 — Reports split into sections

A report is no longer one enormous page. The Wall Street report was 2.9 MB — not a reading experience, a download. Each report now opens on its contents, and each section is a page of its own, following the document's own structure. The whole report is still available on one page for anyone who wants to search across it.

2026-08-02 — The archive is back

The previous version of this site is readable again at old.reportsthatmatter.org. Its old URLs redirect there.

2026-08-02 — The Wall Street report, properly converted

The PSI financial crisis report had been converted before the ingestion pipeline existed, and it showed. 8,178 blocks with a median length of 89 characters — prose paragraphs run to about 600 — because every line of the PDF had become its own paragraph. 1,641 stray page numbers left mid-text. Two headings in a 645-page document.

Re-ingested from the source PDF: 162 headings, 98.3% of the source text retained, no stranded page numbers, and paragraphs that are actually paragraphs.

Doing it exposed a real gap. Footnote numbers appear in two layouts — inline with their text, or alone on a line with the text beneath — and only the first was supported. This report had yielded 7 footnotes; it now yields 1,356, rendered as 2,725 sidenotes. The Jack Smith report gained from the same fix, 206 → 229.

The report moved to /reports/us-psi-financial-crisis, matching its source repository. The old URL redirects.

2026-08-02 — Sidenotes, and links that stay put

Sidenotes. Footnotes now sit in the margin beside the sentence they support rather than 70 KB away at the end of the document. A footnote you have to travel to is a footnote nobody reads, and in these reports the citation is the evidence. On narrow screens, where there is no margin, a note opens on tap.

Printed page anchors. The page numbers from the original document are now addressable — #page-46 — and shown in the margin. It is how these documents are actually cited, and it can be checked against the original PDF.

2026-08-01 — The site, rebuilt

A new design: off-white, mid-grey rather than black, a classical serif for the documents themselves and small monospaced type for everything structural. Fewer things, more space. Serious documents should look serious.

Behind it, a new ingestion pipeline — PDF in, clean Markdown out, deterministically — with automated fidelity checks that gate on what can actually be decided: that no text was invented, that none was lost, that no page furniture survived. Whether a scan was read correctly is a human judgement, so the pipeline surfaces the passages most likely to be wrong for review rather than hiding them behind a score.

First report through it: the Jack Smith Special Counsel report, Volume One — 169 pages, 99% of the source retained, every paragraph individually linkable.

The previous site remains readable in the archive; its URLs redirect there.