Every sentence on this wiki can be traced to a human being saying it out loud, on a specific date, into a microphone. That is the whole premise, and it is checkable rather than promised — so instead of describing the pipeline, this page takes one real sentence from a live article and follows it all the way down to the audio.

tracing this sentence — Analog to Digital Converter
The nominal bit count of a converter does not equal its usable resolution: a 24-bit part in a handheld power-measurement instrument returned roughly 18 effective bits at its highest sample rate and little more than 19 when slowed down.[218]
A hard technical claim. Watch the bundle passage swallow three speaker turns — the quote is Dave's question, the numbers are the guest's answer — and watch the claim get written so it never depends on which of them said what.
05 Article 1 sentence
Analog to Digital Converter

The nominal bit count of a converter does not equal its usable resolution: a 24-bit part in a handheld power-measurement instrument returned roughly 18 effective bits at its highest sample rate and little more than 19 when slowed down.[218]

The citation is a promise. Everything below is the receipt.
04 Packet — where judgment happens 1 of 96 claims
"kind": practice,
"claim_text": "The bit count on a converter's part number is not the number of usable bits. A 24-bit converter in a handheld power-measurement instrument returned roughly 18 effective bits at its highest sample rate and a little over 19 when slowed down.",
"quote_verbatim": "a 24 bit ADC doesn't actually give you 24 usable bits out",
"speaker": "Dave Jones",
"episode": 218
An extraction pass read all 242 passages below and wrote this. It is the only layer where a machine makes a judgment — so it is the layer that gets verified.
verify_packet.py byte-compare · packet quote vs. transcript
packet
source
03 Bundle — the evidence pack 1 of 242 passages
paragraph 455depth: explainsguest: Eric Van Wykspans 3 speaker turns — Chris Gammell + Dave Jones + Eric Van Wyk

Uh, because data. I could talk about that in future shows if we'd like. And of course you're, and of course you discovered that a 24 bit ADC doesn't actually give you 24 usable bits out. Sure. Of course, uh, we've been pulling, uh, in, in our highest speed settings, maybe 18 effective, um, and, and nudging a little bit over 19, uh, when we slow down and, and pay attention a little bit more.

Deterministic collection, no judgment: every passage in the corpus that touches this concept, gathered in one file for the extraction pass to read.
02 Census & canon — finding it at all 486 mentions of this concept
surface forms folded into one concept
adcadcsanalog-to-digital-convertercustom-analog-to-digital-converter
this concept across the corpus
211 episodes · 486 mentions · 239 explanatory
270,979 mentions were logged across the whole corpus, then folded down a 90,150-entry alias table so that “ADC”, “A to D” and “analog to digital” all land on one concept.
01 Transcript — raw ASR 16,283 words in this episode
Eric Van WykI don't see why they do that, but.
Chris GammellUh, because data. I could talk about that in future shows if we'd like.
Dave JonesAnd of course you're, and of course you discovered that a 24 bit ADC doesn't actually give you 24 usable bits out.
Eric Van WykSure. Of course, uh, we've been pulling, uh, in, in our highest speed settings, maybe 18 effective, um, and, and nudging a little bit over 19, uh, when we slow down and, and pay attention a little bit more.
Dave JonesAt the full, what? 8 K sample rate?
Machine transcription, errors and filler intact. Nothing here was corrected — correcting it would break the verification above. The rail marks how far the bundle passage above reached — 3 turns, more than one speaker. The packet's speaker field records who said the quote, not who supplied every detail in the claim, which is why a claim is written to stand on its evidence rather than on attribution.
00 The episode 1 of 719 episodes
218An Interview with Eric VanWyk - Meiotic Mountenance MooshimeterSeptember 29, 2014 · two engineers talking · listen →
Bottom of the stack. A person said this out loud, once, years ago, and now it is a citable sentence.

What the descent shows

The pipeline runs in the opposite direction to the way you just read it. Raw transcripts are swept for concept mentions; an alias table folds surface forms together and a speaker map repairs attribution; a graph of 4,016 concepts and 12,776 edges decides what deserves an article and which passages belong to it; an extraction pass turns passages into claims pinned to verbatim quotes; a verifier byte-compares every quote; a writer builds the article from the claims alone, never seeing the transcripts; and a linter rejects the result if any paragraph is uncited or any citation fails to resolve.

The layer that matters is the byte-compare. It is the one mechanism that makes the rest of it more than a plausible-sounding machine: a quote that was paraphrased, tidied, or invented does not survive it, and an article whose claims cannot be traced does not build.

What survives each stage

Words of raw transcript
11,000,000
719 episodes, machine-transcribed
Concept mentions logged
270,979
the census pass
Verified claims extracted
13,225
each pinned to a byte-checked quote
Episode citations published
6,625
across every live article
Articles live
153
of 412 planned

What this guarantees — and what it doesn’t

Guaranteed: every claim traces to a real, verbatim, byte-checked passage in a real episode you can go listen to. Nothing is sourced from a model’s general knowledge of electronics, from the web, or from vibes. 13,225 claims across the wiki carry a quote that has passed this check, and 6,625 citations resolve to a specific episode.

Not guaranteed: that the interpretation is right. The quotes are real, but an extraction can still miss sarcasm, lose context from five minutes earlier, or generalise an anecdote into a rule. ASR garble occasionally swallows a number. Speaker attribution is repaired but imperfect — where it stays doubtful, the claim is written so it doesn’t depend on who was talking, as the first trace above shows.

That residual risk is exactly what the Report button is for: highlight any sentence on any article and file an issue in one click. “The quote is real but the claim misreads it” is the failure mode we most want reported, because it is the one the machinery cannot catch by itself.

The numbers

Episodes transcribed719
Words of transcript~11 million
Concept mentions in the census270,979
Alias table entries90,150
Concept graph4,016 nodes, 12,776 edges
Verified claims extracted13,225
Articles live153 of 412 planned
Episode citations published6,625

Everything — transcripts, census, bundles, packets, tools, this site — is in the public repo. If you want to build a piece of it yourself, start at How to contribute.

The Amp Hour is Chris Gammell and Dave Jones, and it is theirs — this is an unaffiliated fan project built out of admiration for the show, quoting briefly and always with a link home. Go listen to the real thing.