🕳️ you found the easter egg

The Singularity

Where sixteen years of school-board meetings collapse into a point you can query.

Behind this text, the corpus is falling into a black hole: dollar signs, section marks, stray digits from check registers — all of it spiraling past the photon ring while one very calm little computer keeps a safe orbit and takes notes. That is, roughly, an honest diagram of this project. Scroll on; the music is optional but recommended. ♪

🌀The compression cascade

Everything on this site is one long act of compression — gigabytes of civic paper trail squeezed into something a question can reach in milliseconds:

3.7 GBthe raw corpus — 419 meetings, ~2,800 documents, 2010–2026
344 MBthe queryable database — ~44,000 searchable chunks + three-tier AI summaries
~1 KBwhat a question actually retrieves — the paragraph that answers it

The newest layer does the same thing to video:

1.4 GBone board-meeting recording — 85 minutes of MP4
39 MBthe audio track alone
90 KBa speaker-attributed transcript — 13,767 words, searchable
15agenda chapters that seek the video — click one on any meeting page

📄How the archive works (the honest version)

Four steps: crawl every document off BoardDocs, extract the words, chunk them into ~800-word index cards, and summarize each document three ways. Each step sounds trivial. Three of the four bit us.

The dirty secret of extraction: a PDF does not contain a document — it contains drawing instructions. “Put a ‘B’ at (72, 400).” Closer to a painting than a text file. One library reads the words perfectly but scrambles their order; another keeps the order but splits $123,879,792 into 1 and 23,879,792. So the extractor runs both and merges them — one tool’s characters in the other tool’s order.

The summaries are the trick that makes search work. A query for “budget cuts” will never match a document that only ever says “expenditure reductions” — so a dense, verbose summary is indexed alongside every document, acting as a translation layer between how people ask and how bureaucracies write.

And because fluent output is the hardest kind to audit, every figure a summary asserts is machine-checked against the source — 6,151 numbers in the latest pass. This paranoia is earned: one innocent-looking line (TEXT_CAP = 6000) once fed the AI only the first two pages of everything, and 788 documents got fluent, plausible, nearly source-free summaries. A 100-page budget book, summarized from its table of contents. Nothing crashed. Everything read beautifully.

The dangerous failures produce confident, well-formed, wrong output. Crashes are a gift.

🎙️The archive taught the transcriber how to hear

Meeting recordings now get transcribed by a Whisper-class speech model (AssemblyAI’s Universal-3.5 Pro). Out of the box, a speech model has never heard of Boulan Park, mishears “Machesky,” and thinks TESPA is a typo. Names are exactly what a civic transcript cannot get wrong.

This is where the archive bootstraps itself. Those verbose summaries are dense with proper nouns — so a script mines them: every trustee and principal, every school building, every vendor on a bid tab, every acronym a board member has ever pronounced out loud. The result is a 361-term vocabulary that gets handed to the speech model as key terms before it listens to a single second of audio.

The paper trail teaches the ear. Documents → summaries → proper nouns → a transcriber that hears “Boulan Park Middle School security and paging” on the first try. Diarization splits the voices, a second pass pins real names on them, and the agenda items become chapters that jump the video to the moment a motion was moved. Cost of transcribing an entire board meeting this way: about forty cents.

It still wrote “Mr. Hauff” for Mr. Haupt six times — two different humans, one homophone apart, both real. Some entropy is irreducible.

The whole voice layer is open source: the transcription/ folder in the repo holds the pipeline scripts, the generated vocabulary, and a fully worked example meeting — with the methodology written up in docs/TRANSCRIPTION.md.

🧮A two-minute primer: compression is intelligence

In the 1940s Claude Shannon asked how far text can be compressed, and invented information theory to answer it. His definition: the information in a message is −log₂ p — “how many times must you halve the space of possibilities to land on this one?” Surprising messages carry many bits; predictable ones carry almost none.

Which means compression and prediction are the same job. If you can predict the next character well, you barely need to store it. Shannon proved it by using people as the predictor — show a human 100 letters of context, ask them to guess the next one — and measured English at roughly 1 bit per character, an 8× compression of plain ASCII, achievable only because a human brain is an excellent (if un-inspectable) model of language. Today we build those black boxes instead of interviewing them: a large language model is, quite literally, a compression engine you can talk to.

That is what this archive is. The verbose summaries are lossy compression tuned to future questions. The transcript compresses an afternoon of civic life into text you can grep. The index compresses “where is it?” from 3.7 GB to a millisecond. None of it reaches Shannon’s floor — but every useful system lives in the gap between ASCII and 1 bit per character.

Primer distilled from 3Blue1Brown’s “Reinventing Entropy — Compression is Intelligence, Part 1”, which you should absolutely watch instead of trusting my summary. (Fluent summaries, we established, deserve verification.)

🕳️The ultimate computer

Now the background art. Physicists have entertained a wonderful idea: the densest possible computer — the endpoint of Moore’s law, pursued without taste or restraint — is a black hole.

Jacob Bekenstein showed that the information a region of space can hold is finite, and scales with its surface area, not its volume — as if the universe stored everything on the boundary (the “holographic principle”). A black hole saturates that bound exactly: it is the most information per square inch that physics permits. Seth Lloyd ran the numbers on the “ultimate laptop” — one kilogram of matter computing as hard as physical law allows: ~10⁵⁰ operations per second on ~10³¹ bits, and its final form is, inconveniently, a small black hole. Maximum compression and maximum computation turn out to be the same object. Shannon would have smiled.

Hence the scene behind this page: the colorful matter in the accretion disk is the data on its way in — glowing hot precisely because it is being compressed — and the little computer in stable orbit is this website, reading the radiation and answering your questions from a safe distance.

This archive currently holds its sixteen years at a leisurely 344 MB, which leaves considerable headroom before the Bekenstein bound. We do not intend to test it.

Escape the event horizon — back to the archive
⚙️ Take it apart: github.com/akarpo/tsd-boarddocs — the crawler, the summarizer, the transcription pipeline, this very page. All of it.
♪ Background: Trent Reznor & Atticus Ross — Painted Sun in Abstract (The Social Network), playing via YouTube, muted until you say otherwise.
No pixels were harmed crossing the event horizon.

♪ Painted Sun in Abstract — Trent Reznor & Atticus Ross · The Social Network