🔒 Fidelity check is locked

This audio↔story breakdown is insider-only. If you know the key, add it to the URL — if not, ask around. 😉

← back to the story

🎧 Audio ↔ 📄 Story · Fidelity Check

How close is the page to Rajesh's voice?

Comparing the 6:16 recording (auto-transcribed, ~847 words) against the published The Loud Fight (~1,509 words).  ·  ← back to the story

0/100faithfulness
Faithful core, expanded ending.

Every beat Rajesh actually recorded — the two tensions, the lunchtime shouting match, Sri stepping in, the word acceptable — survives intact and in his voice. But the recording stops where the V1 ships; the page keeps going, adding Venky's WhatsApp correction and a "two versions" P.S. that the audio never touches. The page is ~178% the length of the recording.

Where the score comes from

Story beats preservedDid every moment Rajesh told make it onto the page?78%
Facts & quotes accuracyNumbers, names, dates, and direct quotes84%
Tone & voice matchDoes it still sound like Rajesh telling it?88%
Verbatim wordingWord-for-word overlap — low by design for a written piece26%
Length retainedStory is 178% the length of the spoken version178%

▮ green = meaning preserved · ▮ amber = minor drift · ▮ grey = expected difference (rewrite, not loss)

✓ Kept faithfully

  • Stress-test by launching a project — after clarity on the art of the possible
  • The decade-old internal tool, in its third version
  • Dark mode — all in one room, two-three days, ship V1
  • Two tensions — the safety harness / demo-bias, and accuracy vs. acceptable
  • 2+2 as the binary-accuracy example
  • Lunchtime, ~12:30–1pm — the loud, stunning shouting match with Venky
  • Sri stepped in — cooled them off, gave space to reflect
  • High trust, low ego — and "credit to Sri" for the word acceptable
  • ~30 minutes → buy-in → V1 launched in a couple of days

~ Changed / dropped & added

  • ADDED: Venky's WhatsApp correction — who pushed / who paused, "I give up", the coffee, the sales sponsor — not in the recording at all
  • ADDED: the entire "two versions / Pondicherry, Voice IQ" P.S. section
  • ADDED: the four tidy lesson cards & the "dinner table" coda
  • "diverge-converge model" kept, but expanded into a fuller product-discovery riff
  • "Sri" on the page; recording also says it the same way
  • Recording ends at V1 launch; roughly the back half of the page has no audio source
Read it like this: the big number is faithfulness of meaning, not transcription accuracy. The recorded story is told faithfully — but the page deliberately grows past the recording, with a whole correction-and-P.S. arc Rajesh never spoke aloud. That's why this scores lower than a page that stays inside its audio: the part you can hear is honest, but you can't hear all of it.

Method. Audio transcribed locally with whisper.cpp (base.en). Verbatim (26%) and length (178%) are measured directly (difflib sequence ratio & word counts). Beats, facts, and tone are a beat-by-beat human/AI reading of transcript vs. page. Overall faithfulness is a weighted blend (beats 40% · facts 35% · tone 25%); verbatim is reported separately, not folded in.