The companion to Measuring What You Already Hear. That essay makes one argument and leaves out the inventory on purpose. This is the inventory: one line per thing, grouped, no narrative.
Getting material in
- Everything enters from a phone: photographs, screenshots, screen recordings, voice notes, links.
- Voice notes are transcribed locally; a hummed melody is treated as a source, not as a memo.
- Household financial records are photographed and filed into a private ledger — capability only; the records themselves never leave the box.
- Files are not left in the conversation. They are placed somewhere durable and referenced by link, so a document from three months ago is still reachable.
- Clips from social platforms are downloaded and read: audio pulled, transcribed, and measured.
- Video is no longer watched for information. An ingest line pulls the audio, transcribes it, and produces a digest.
- The digest is not read by hand either. It goes to an assistant that already holds the project context, with one question: what in here is useful to me. Two stages of compression, and the second one is against the reader’s own work rather than against the source.
Music: from hum to master
- Tempo and key derived from the raw hum, which carries no metadata.
- Style prompts written from measured production characteristics of a reference, never from its melody.
- Lyrics developed across several models with different failure modes, then chosen by the human.
- Generated masters taken apart into stems and reassembled: levels, low end, parts added that the generator never played.
- Speaker similarity measured against the artist’s own voice, calibrated on our own material, with a noise floor established from sibling takes.
- Tempo drift, kick-on-beat, stereo width, loudness and spectral balance measured on every candidate before a download slot is spent.
Video and publication
- Cover art generated locally at no per-image cost.
- Video assembled from stock footage, the artist’s own filming and generated shots.
- Edit grid derived from actual beat positions, not nominal tempo.
- Cuts placed a whole number of beats apart so they land on the bar; accents placed on measured kick onsets; visual events timed to drops.
- Subtitles aligned to the vocal by time, with their own motion and effects.
- Sound design layered underneath: atmospheres, transitions, radio chatter.
- Narration generated from script, transcribed back by a separate speech model, and compared word for word against the source — divergence raises an alert rather than waiting to be noticed by ear.
- Upload rail schedules the result on a channel, with metadata and thumbnail.
How the work is organised
- A command post holds the plan; working sessions own their areas and their own tools, credentials and permission boundaries.
- Sessions message each other directly: tasks out, results back.
- Peer review between sessions catches what a single session cannot, because the reviewer does not share the author’s assumptions.
- Long-running state is written to disk, not held in a conversation, so a restart costs nothing.
- Autonomous work driven from a ticket queue rather than from prompts.
Software work
- Pull requests opened, commented, reviewed and answered — including declining a review comment with a stated reason.
- Verification against the running thing rather than the diff: the browser opens the page, or a script drives the interface and reports what happened.
- Deployed version compared against the local one on a pull request; the difference is the review.
- Tests written to cover a feature immediately after building it, and run before the feature is called done.
- Measurement tools stamp their own method into every result, so “measured” and “extrapolated” never look alike.
Publishing the writing itself
- Essays and notes drafted, revised against the author’s corrections, and opened as pull requests rather than pasted into a CMS.
- Reviewed across sessions before merge: one reviewing session checked a published piece against its own numbers and found a claim stated more strongly than the data supported.
- Corrections shipped as their own pull request with the reasoning in the message, so the record shows what changed and why.
- Build run before every proposal; a page that does not build is not offered.
- Publication rules applied by the drafting session and checked on review: nothing personal, categories rather than instances, no paths, hosts or file names.
Hiring work
- Interview sandboxes built to order: a working environment with a task in it, so a conversation can be about what someone does rather than what they claim.
- Submitted CVs read for the things a ranking tool does not surface — where the gaps in a timeline are, where stated figures do not reconcile with each other, what could plausibly have gone wrong in a project described as smooth.
- Output is a question list rather than a score: what to ask this person to find out whether the experience behind a line is real.
- The same discipline as everywhere else in this list — the artefact is checked against itself before anyone is asked to trust it.
Not built yet
- Standup transcripts read against full project context — what was asked, what was reversed, who promised what. Blocked only by one platform’s refusal to export a transcript; other conferencing tools hand one over automatically.