🔒 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 Venky's voice?

Comparing the 21:39 recording (auto-transcribed, ~3,420 words) against the published The Discovery Workshop (~1,540 words).  ·  ← back to the story

0/100faithfulness
Faithful — and tightened.

The page is a condensed cut of a longer recording — every story beat Venky tells survives, the names and dates line up, and the voice stays his. The spoken version is roughly twice the length; the page trims the architecture walk-through and the day-by-day Slack texture, not the story.

Where the score comes from

Story beats preservedDid every moment Venky told make it onto the page?93%
Facts & quotes accuracyNumbers, names, dates, and direct quotes86%
Tone & voice matchDoes it still sound like Venky telling it?92%
Verbatim wordingWord-for-word overlap — low by design for a written piece18%
Length retainedStory is 45% the length of the spoken version45%

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

✓ Kept faithfully

  • "Not a week in Bangalore with laptops" — three weeks of daily 4:30 p.m. calls (near-verbatim opener)
  • The room before the room — Slack channel opened ~21 July, nine engineers, "it was a Monday"
  • Sreenath's Pokémon paper — agentic problem = a sequence of decisions, not one prompt
  • The six-bucket prompt & the input-schema / output-scaffold "in the same breath" insight
  • Anish's ground-truth dataset — 30 confirmed proposals, "before evals was a word"
  • The soapbox — "future needs no designers/product owners," leaning harder into product ownership
  • Niraj's architecture + GraphRAG → Neo4j knowledge graph in production; the explainability layer
  • The fight & the coffee — Aug 4, James rattled, the 12:30 coffee, "we agree on the coffee"
  • TC Copilot naming, the Veo film, co-creation, the birthday booking & "aila loves madness"

~ Changed / dropped

  • Recording is ~2× longer — the page is a condensed cut, not a 1:1 retelling
  • Booking date: recording says 20 August (Anish's birthday); page narrates "20th of August" but the day-5 section reads as if it's the same week
  • The blockquotes ("The first Gen-AI proposal just converted…", the coffee reflection) are shaped paraphrases, not transcribed quotes
  • Sreenath's open-questions list (MongoDB vector search, HubSpot/Artemis mapping) — dropped
  • The full AWS pipeline walk-through (Lambda → knowledge base → Bedrock → storage) compressed to one line
  • Anish's hand-drawn "option 1 / option 2" output mock — dropped
  • Andy / "Ian" named in the recording as a TC; page keeps Andy, the third consultant role trimmed
  • Recording aside: "I'll boot you out of this channel" framing softened on the page
Read it like this: the big number is faithfulness of meaning, not transcription accuracy. A score around 90 means "trust this page as Venky's story." The 18% verbatim and 45% length bars are expected — the recording is long and discursive; the page is the same journey, tightened for reading.

Method. Audio transcribed locally with whisper.cpp (base.en). Verbatim (18%) and length (45%) 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.