🔒 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 T's voice?

Comparing the 8:53 recording (auto-transcribed, ~1,220 words) against the published I Was the Sceptic (~850 words).  ·  ← back to the story

0/100faithfulness
Faithful retelling.

Every story beat in the recording survives on the page — the heated argument, the night he messaged back, the Chrome plugin, the moment the AI reasoned about the UI. It's tightened and re-worded for reading, and the recording covers the whole arc end to end. Nothing on the page is invented.

Where the score comes from

Story beats preservedDid every moment Rajesh told make it onto the page?92%
Facts & quotes accuracyNumbers, names, dates, and direct quotes88%
Tone & voice matchDoes it still sound like Rajesh telling it?90%
Verbatim wordingWord-for-word overlap — low by design for a written piece32%
Length retainedStory is 70% the length of the spoken version70%

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

✓ Kept faithfully

  • The conversation Sree started — "how can we start using AI? How can we start leveraging it?"
  • The product-person reflex — heard it as "AI is there, what do we do now?"
  • "Too premature… it might just be hype" & the other-levers argument
  • The hype list — internet boom, big data, mobile, crypto
  • The heated exchange & going home to reflect
  • "Just: wonder" — the reframe to the art of the possible, and the same-night message
  • Word-guessing machine → convert arc, intact
  • The Chrome plugin — CRO programme, CrUX data, API key, NPM scaffolding, Windsurf, 6–7 lines, 15 min + 5 min debug
  • The magical moment — AI reasoned the benchmark was supporting info, kept it subtle
  • "way before the models became powerful" & "Sree had latched onto this earlier"
  • The teaser — next chapter is the full web app + converting sceptical friends

~ Changed / dropped

  • Travelopia office not named — page says only "in the office"
  • The tool name ("Deepak Bear" in the transcript) → genericized to "a tool"
  • "crux score" spelled out as CrUX — Chrome User Experience data on the page
  • "WinServe" (mis-heard) → correctly written as Windsurf
  • Pronoun cleanup — transcript wavers "she/he" for Sree; page settles on a consistent "him"
  • The UI praise — recording also notes the plugin's first UI was already nicely designed; page folds this into the benchmark moment
  • "art impossible" (transcription slip) → rendered as "art of the possible"
Read it like this: the big number is faithfulness of meaning, not transcription accuracy. A score of 90 means "trust this page as Rajesh's story." The recording runs longer than the page and covers the whole arc, so the 70% length bar is the page tightening the telling — not the page reaching past the audio. The 32% verbatim bar is healthy — if it were near 100, the page would just be a transcript, not a voice.

Method. Audio transcribed locally with whisper.cpp (base.en). Verbatim (32%) and length (70%) 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.