Reports that Matter

Changelog

What has changed.

Improvements to the archive and how it reads

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.